Research article Special Issues

Involvement of smart technologies in an advanced supply chain management to solve unreliability under distribution robust approach

  • The proposed study described the application of innovative technology to solve the issues in a supply chain model due to the players' unreliability. The unreliable manufacturer delivers a percentage of the ordered quantity to the retailer, which causes shortages. At the same time, the retailer provides wrong information regarding the amount of the sales of the product. Besides intelligent technology, a single setup multiple unequal increasing delivery transportation policy is applied in this study to reduce the holding cost of the retailer. A consumed fuel and electricity-dependent carbon emission cost are used for environmental sustainability. Since the industries face problems with smooth functioning in each of its steps for unreliable players, the study is proposed to solve the unpredictable player problem in the supply chain. The robust distribution approach is utilized to overcome the situation of unknown lead time demand. Two metaheuristic optimization techniques, genetic algorithm (GA) and particle swarm optimization (PSO) are used to optimize the total cost. From the numerical section, it is clear the PSO is 0.32 % more beneficial than GA to obtain the minimum total cost of the supply chain. The discussed case studies show that the applied single-setup-multi-unequal-increasing delivery policy is 0.62 % beneficial compared to the single-setup-single-delivery policy and 0.35 % beneficial compared to the single-setup-multi-delivery policy. The sensitivity analysis with graphical representation is provided to explain the result clearly.

    Citation: Soumya Kanti Hota, Santanu Kumar Ghosh, Biswajit Sarkar. Involvement of smart technologies in an advanced supply chain management to solve unreliability under distribution robust approach[J]. AIMS Environmental Science, 2022, 9(4): 461-492. doi: 10.3934/environsci.2022028

    Related Papers:

    [1] Biljana Ilić, Dragica Stojanovic, Gordana Djukic . Green economy: mobilization of international capital for financing projects of renewable energy sources. Green Finance, 2019, 1(2): 94-109. doi: 10.3934/GF.2019.2.94
    [2] Mukul Bhatnagar, Sanjay Taneja, Ercan Özen . A wave of green start-ups in India—The study of green finance as a support system for sustainable entrepreneurship. Green Finance, 2022, 4(2): 253-273. doi: 10.3934/GF.2022012
    [3] Nabil Haque, Sungida Rashid . Host country characteristics attracting climate projects through public-private partnerships. Green Finance, 2019, 1(4): 405-428. doi: 10.3934/GF.2019.4.405
    [4] Shahinur Rahman, Iqbal Hossain Moral, Mehedi Hassan, Gazi Shakhawat Hossain, Rumana Perveen . A systematic review of green finance in the banking industry: perspectives from a developing country. Green Finance, 2022, 4(3): 347-363. doi: 10.3934/GF.2022017
    [5] Laura Grumann, Mara Madaleno, Elisabete Vieira . The green finance dilemma: No impact without risk – a multiple case study on renewable energy investments. Green Finance, 2024, 6(3): 457-483. doi: 10.3934/GF.2024018
    [6] Goshu Desalegn . Insuring a greener future: How green insurance drives investment in sustainable projects in developing countries?. Green Finance, 2023, 5(2): 195-210. doi: 10.3934/GF.2023008
    [7] Wen-Tien Tsai . Green finance for mitigating greenhouse gases and promoting renewable energy development: Case study in Taiwan. Green Finance, 2024, 6(2): 249-264. doi: 10.3934/GF.2024010
    [8] María Cantero-Saiz, Sergio Sanfilippo-Azofra, Begoña Torre-Olmo, Violeta Bringas-Fernández . ESG and bank profitability: the moderating role of country sustainability in developing and developed economies. Green Finance, 2025, 7(2): 288-331. doi: 10.3934/GF.2025011
    [9] Yafei Wang, Jing Liu, Xiaoran Yang, Ming Shi, Rong Ran . The mechanism of green finance's impact on enterprises' sustainable green innovation. Green Finance, 2023, 5(3): 452-478. doi: 10.3934/GF.2023018
    [10] Raja Elyn Maryam Raja Ezuma, Nitanan Koshy Matthew . The perspectives of stakeholders on the effectiveness of green financing schemes in Malaysia. Green Finance, 2022, 4(4): 450-473. doi: 10.3934/GF.2022022
  • The proposed study described the application of innovative technology to solve the issues in a supply chain model due to the players' unreliability. The unreliable manufacturer delivers a percentage of the ordered quantity to the retailer, which causes shortages. At the same time, the retailer provides wrong information regarding the amount of the sales of the product. Besides intelligent technology, a single setup multiple unequal increasing delivery transportation policy is applied in this study to reduce the holding cost of the retailer. A consumed fuel and electricity-dependent carbon emission cost are used for environmental sustainability. Since the industries face problems with smooth functioning in each of its steps for unreliable players, the study is proposed to solve the unpredictable player problem in the supply chain. The robust distribution approach is utilized to overcome the situation of unknown lead time demand. Two metaheuristic optimization techniques, genetic algorithm (GA) and particle swarm optimization (PSO) are used to optimize the total cost. From the numerical section, it is clear the PSO is 0.32 % more beneficial than GA to obtain the minimum total cost of the supply chain. The discussed case studies show that the applied single-setup-multi-unequal-increasing delivery policy is 0.62 % beneficial compared to the single-setup-single-delivery policy and 0.35 % beneficial compared to the single-setup-multi-delivery policy. The sensitivity analysis with graphical representation is provided to explain the result clearly.



    Machinery has become an integral part of human life, especially considering the technological advancements related to Industry 4.0. Failures occurring on such crucial machinery lead to unplanned downtime, ultimately resulting in loss in economic aspects [1]. This is catastrophic in industry and public transport since these failures stop production, causing hassles to the public. Hence, machine diagnostics is of high importance in such instances. Fault prediction in an early stage dramatically improves the machine's lifetime, reducing costs and preventing downtimes.

    Air and oil leaks are two of the predominant operational failures in public transport modalities, especially in metro trains, which are our prime objective. Air leaks are prone to occur in the dryer component, whereas oil leaks are prone to occur in the compressor component [2]. Various sensors, like pressure transducers, pneumatic sensors, motor current, etc., are used to analyze and diagnose the faults in metro trains [3,4,5]. An abnormal change can be observed in these sensors in the event of a fault in that component. Continuous monitoring of the vulnerable components with the sensors mentioned above can help identify the occurrence of a fault in that component [6].

    Predictive maintenance has been an emerging technology in machine diagnostics, aiming to predict faults early and perform maintenance to prevent catastrophic events [7]. Also, anomaly detection based on sensor data on an early scale will reduce maintenance expenses and avoid downtime. Data curation, data pre-processing, diagnosis, and decision-making are the critical aspects of predictive maintenance [8]. This data-driven approach has been proven effective due to the vast availability of data and intelligent algorithms for automated analysis [9].

    Artificial intelligence plays a promising role in fault prediction and predictive maintenance. Several machines and deep learning algorithms are trained on continuous data from sensors attached to the target machinery [10]. The proposed work uses machine and deep learning algorithms for anomaly detection (air and oil failure) on the air production unit of metro trains and real-time dashboard development, which is the first of this work as per our knowledge. The following section explains the existing works on machine usage and deep learning algorithms for predictive maintenance.

    The following are the contributions of the proposed work:

    1. To develop a deep learning algorithm for the simultaneous identification of the type and location of the fault, along with GPS quality monitoring from sensor data in metro trains.

    2. To integrate an explainable AI technique into the model's prediction and highlight the key sensors contributing to the fault.

    3. To develop a dashboard integrating the sensor data analytics as visual graphs, the deep learning model, and the explainable AI results for analysis by engineers.

    This work will be of great aid to maintenance engineers for fault analysis in sensor values and assessment of GPS signals in real time. This predictive maintenance application can aid in reducing the downtime and service costs of machine parts, if found damaged.

    The following is the outline of the research paper. Section 2 briefs about the existing work relevant to the field of interest. Section 3 explains the material description, the pre-processing techniques used, the different machine learning algorithms, the training parameters, and the methodologies for dashboard development. Section 4 represents the proposed methods' results, graphs, and supporting diagrams. Section 5 compares the proposed work with existing works and draws significant conclusions and future scope.

    The detection of air leakage from the pneumatic door of a train is an attempt to reduce train downtime. Deep learning algorithms were applied for automatic feature extraction from extensive data obtained from continuous monitoring by sensors for the task of fault detection [11]. The OSR (open set recognition) concept was used for multi-task classification to predict the known class and detect unknown samples. A lightweight convolutional neural network (CNN) model streamlined with the OSR technique was trained to predict the air leakage. An 8-layer neural network consisting of 6 convolutions and two dense layers was used for air leakage. This model was trained using an SGD (stochastic gradient descent) optimizer with a learning rate of 0.001 for a batch size of 64.

    The server air leakage in the breaking pipe results in breaking issues and decreases the train reliability [12]. Due to the visual constraints for air leak detection, the paper proposes a framework for the simultaneous prediction of the type and severity of air leakage using anomaly detection methods based on the on-and-off logs of the compressor. Around 632,683 data points were collected from May 2016 to October 2016 from 178 VIRM trains, of which 6957 are labeled as "Air Leakage" and 625,726 are labeled as "Normal". They have used a logistic classifier model for two different classes of compressor behavior for each separate train. One defines the boundary by separating two classes under everyday situations, and the other models the distribution of the compressor idle time and run time separately using logistic functions. It also further detects the context of compressor idle time erroneously classified as a compressor run time, aiding in anomaly detection. A density-based unsupervised clustering approach is adopted for anomaly detection before four weeks and can pre-filter anomalies to prevent false alarms.

    The challenges encountered by traditional manufacturing companies during their transition to intelligent factories, notably the scarcity of historical data for training machine learning models, were addressed by Mohan Rajashekarappa et al. [13]. A novel approach of artificially inducing anomalies for data labeling was introduced, and it underscored the importance of proactive readiness for potential future disruptions in newly installed systems. Through two experiments focused on air leakage detection, the proposed methodology demonstrates exceptional performance with RUS-Boosted bagged trees, yielding 98.73% accuracy, 99.40% precision, recall of 99.21%, and an F1 score of 99.30% on the test data.

    The critical issue of energy efficiency and fault detection in air conditioning systems emphasizes their intricate nature and substantial energy consumption impact [14]. The study comprises two essential components: First, it investigates the ramifications of various faults within the air conditioning system on its coefficient of performance (COP), shedding light on the potential energy wastage associated with these faults. The research convincingly demonstrates that different faults lead to varying degradation levels in the COP. Second, the paper evaluates the effectiveness of three supervised learning classifier models in classifying these faults: deep learning, support vector machine (SVM), and multi-layer perceptron (MLP). The research assesses the performance of these classifiers across six distinct fault classes, revealing that different faults indeed exert varying negative impacts on the COP.

    Predicting air failure of the air production unit (APU) in metro trains. The dataset used for this task was MetroPT, a 6-month analysis of metro trains in Portugal comprising analog, digital, and GPS sensors [15]. The GPS information was excluded from the dataset, and the timestamp was encoded using the label encoding technique. A random forest classifier algorithm was used for the multi-class classification of air failure prediction. The data was undersampled and then split into training and testing sets. A feature importance visualization technique was employed to identify the root cause of the air failure. The random forest classifier produced 85% and 97% accuracies for the binary and multi-class classification tasks, respectively.

    A deep learning neural network for anomaly detection in metro trains was developed by Davari et al. [16]. The algorithms used for this task were the sparse autoencoder and variational autoencoder. This work is an unsupervised learning approach for anomaly detection of air failures in trains. The MetroPT dataset was used for this work with sensors placed in the air production unit. The two versions of the autoencoder were used for sensor data reconstruction, and a low-pass filter was used to perform anomaly detection and detect faults. The autoencoder algorithms using the digital data produced precision, recall, and F1 scores of 44%, 13%, and 32% better than that of the algorithms trained on the analog data.

    An expert system for the multi-objective optimization of equipment was developed for highway optimization by Ali et al. [17]. The particle swarm optimization was used to simultaneously optimize the time, cost, and quality of the equipment for construction. This method reduced the time and cost by 35.4% and 39.1%, respectively. The application of predictive maintenance in concrete manufacturing was done by Alshboul et al. [18]. Seven different classification algorithms were used, out of which the cat boost classifier produced an F1-score of 0.985, an accuracy of 0.984, a recall of 0.983, and a ROC curve area of 0.984. A comparative analysis of machine learning algorithms for concrete strength estimation was performed by Alshboul et al. [19]. Three machine learning algorithms, namely, XGBoost, LighGBM, and genetic programming, were used, out of which the LightGBM and XGBoost algorithms surpassed other studied algorithms with a coefficient of determination of 95.74% and 93.27%, respectively.

    The proposed work aims to develop a decision process for failure prediction and identification of the failure type and location using machine learning and multi-task models using deep learning algorithms.

    The proposed work adopts the following workflow consisting of different blocks: data acquisition, data pre-processing, feature pre-processing, visualization, model development, validation, and deployment. Figure 1 represents the proposed workflow visually.

    Figure 1.  Pictorial representation of the proposed workflow.

    The dataset used for this work is named MetroPT [17,18], comprising of sensor information related to urban metro trains in Portugal collected during the year 2022. The dataset comprises different analogue, digital, and GPS sensors continuously capturing data from the metro trains for six months. The MetroPT dataset has been curated to develop AI algorithms for automated fault prediction and predictive maintenance of metro trains based on sensor data. Table 1 represents the different sensors, their description, and the unit of measurement used for acquiring the MetroPT dataset.

    Table 1.  Name, description, type, and units of the different sensors used in the metroPT dataset acquisition.
    Name Description Type of sensor Unit
    TP2 Compressor pressure Analog Bar
    TP3 Pneumatic panel pressure Analog Bar
    H1 Pressure of the valve that is activated when the pressure exceeds 10.2 bar Analog Bar
    DV_Pressure Pressure drop due to water discharge by air dryers Analog Bar
    Reservoirs Air tank pressure Analog Bar
    Oil temperature Temperature of oil in compressor Analog Celsius
    Flowmeter Airflow Analog m3/h
    Motor current Current flowing in the motor Analog Ampere
    Comp Electric signal of the compressor based on the air intake Digital -
    DV Electric Electric signal of compressor outlet Digital -
    Towers Specifies the two towers based on the action of air drying Digital -
    MPG Trigger to start the compressor when the pressure is less than 8.2 bar Digital -
    LPS Trigger when the pressure is less than 7 bar Digital -
    The pressure switch Trigger when pressure is detected in the pilot control valve Digital -
    Oil level Trigger when the oil level is less than the threshold Digital -
    Caudal Impulses Trigger for the air flowmeter Digital -
    GPS Longitude Longitude position Analog °
    GPS Latitude Latitude position Analog °
    GPS Speed Speed Analog Km/h

     | Show Table
    DownLoad: CSV

    The dataset comprises around 15 million sensor data records from Jan 1, 2022 to Jun 30, 2022. The dataset obtained from the source is not directly labeled. The dataset owners have provided information about the failure, like the start time, end time, type, and location of the failure. Based on the start and end times, the appropriate timestamps were found and the values between those timestamps were coded according to the fault type. Table 2 shows the statistical description of the dataset for the parameters, namely, TP2, TP3, H1, DV_pressure (DVP), Reservoirs, Oil_temperature (OT), Flowmeter, Motor_current (MoC), COMP, DV Electric (DVE), Towers, MPG, LPS, Pressure_switch (PrS), Oil_level (OL), Caudal_impulses (CaI), GPS Longitude (GPSLong), GPS Latitude (GPSLat), GPS Speed, GPS Quality, month, day, hour, minute, second. The statistical values of mean, standard deviation (std), minimum value, 25%, 50%, 75%, and maximum values for a dataset are evaluated here. Table 3 represents the label code for the different types of faults occurring in the metro train.

    Table 2.  Statistical decription of the dataset.
    TP2 TP3 H1 DVP Reservoirs OT Flowmeter MoC
    mean 0.947 8.989 8.038 -0.019 1.63 65.843 20.128 2.040
    std 2.836 0.667 2.846 0.185 0.064 5.931 3.578 2.198
    min -0.03 0.937 -0.033 -0.036 1.349 18.575 18.8347 -0.009
    25% -0.009 8.492 8.332 -0.0279 1.608 61.825 18.9748 0.0024
    50% -0.007 8.996 8.876 -0.025 1.635 66.475 19.03 0.007
    75% -0.006 9.506 9.438 -0.025 1.667 70.575 19.040 3.837
    max 10.806 10.38 10.368 8.11 1.791 80.174 37.008 9.3375
    COMP DVE Towers MPG LPS PrS OL CaI
    mean 0.892 0.107 0.946 0.892 0.004 0.0 0.0 0.002
    std 0.309 0.309 0.225 0.309 0.068 0.0 0.0 0.047
    min 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
    25% 1.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0
    50% 1.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0
    75% 1.0 0.0 1.0 1.0 0.0 0.0 0.0 0.0
    max 1.0 1.0 1.0 1.0 1.0 0.0 0.0 1.0
    GPSLong GPSLat GPSSpeed GPSQuality
    mean -7.880 37.578 8.592 0.912
    std 2.443 11.649 14.096 0.282
    min -8.69 0.0 0.0 0.0
    25% -8.66106 41.1696 0.0 1.0
    50% -8.658 41.1858 0.0 1.0
    75% -8.583 41.212 16.0 1.0
    max 0.0 41.240 286.0 1.0
    month day hour minute second
    mean 1.0 16.046 13.139 29.507 29.499
    std 0.0 8.919 6.444 17.318 17.318
    min 1.0 1.0 0.0 0.0 0.0
    25% 1.0 8.0 9.0 15.0 14.0
    50% 1.0 16.0 14.0 30.0 29.0
    75% 1.0 24.0 19.0 45.0 44.0
    max 1.0 31.0 23.0 59.0 59.0

     | Show Table
    DownLoad: CSV
    Table 3.  Faults occurring in the metro train along with their corresponding label code.
    Label Code Corresponding fault
    0 Air leak in air dryer
    1 Air leak in client chamber
    2 Oil leak in compressor
    3 No fault

     | Show Table
    DownLoad: CSV

    Timestamp data cannot be processed by machine learning and deep learning algorithms. Hence, it needs to be processed. This issue was tackled by extracting the month, week, day, hour, minute, and second information from the timestamp using the Pandas functionalities. Finally, the timestamp feature is removed from the dataset. Two new columns, one for the type of failure and another for the location of the failure, were created from the labels. These columns are created as these are the labels for the multi-task model. Table 4 represents the different failure types and location codes.

    Table 4.  Label code for the type and location of the failure.
    Label Code Type of failure Location of failure
    0 Air leak Air dryer
    1 Oil leak Client chamber
    2 No failure Compressor
    3 Not reported No location

     | Show Table
    DownLoad: CSV

    The final step in the data pre-processing was the operation of undersampling, especially in the dataset of stage 1. The stage-1 dataset comprises 2, 12,104 samples under the fault condition, about 1.63% of the entire data. Around 98.4% of the data belongs to the regular class, proving the dataset is highly imbalanced. Hence, the samples from the regular class were undersampled. Around 28, 00,000 randomly selected samples were taken from the entire dataset, which was used as the data for the first stage. The final dataset comprises about 25 input features and three target vectors. Table 5 represents the description of all the targets in the final dataset.

    Table 5.  Description and classes for the target vectors in the final processed dataset.
    Name of target vector Description Classes
    Type Code for the type of failure 0- Air failure
    1- Oil failure
    2- Normal
    Location Code for the location of the failure 0- Air dryer
    1- Client
    2- Compressor
    3- No location
    GPS quality Quality of the GPS sensor 0- Good
    1- Bad

     | Show Table
    DownLoad: CSV

    The dataset has no null or duplicate values since the dataset is obtained from sensors that continuously monitor the trains. The feature scaling technique, normalization, was adopted to bring all the features to the same scale (0-1). Appropriate data visualization techniques were used for the dataset's univariate, bivariate, and multivariate analysis. The dataset comprises continuous (analogue sensors) and categorical features (digital sensors); we have split them for data visualization. A histogram with a kernel density estimator function is used to visualize the continuous features in the dataset. A pie chart is used to visualize the categorical features. Finally, the information was visualized using a map that represents the train's route along with the train's speed. Figures 2, 4, 3, and 5 represent the visualization plots obtained from the continuous, categorical, GPS, and entire dataset, respectively. After visualization, a stratified split of ratio 80:20 was made for the training and testing sets, respectively.

    Figure 2.  Bar plot for the digital sensors in the dataset.
    Figure 3.  Map visualization of the GPS information of the metro train. The intensity of the red dots represents the train's speed at those points.
    Figure 4.  Histogram plot with a kernel density estimator for the analog sensors in the dataset.
    Figure 5.  Heatmap visualization for the entire dataset along with the target variables.

    Multi-task learning is the ability of a neural network to simultaneously obtain multiple outputs from a single input. In this work, based on the sensor information, the multi-task neural network is designed to simultaneously predict the type and location of the failure in the metro train. The multi-task neural network has common pre-processing layers, followed by branches, each corresponding to a particular task. Hence, the multi-task neural network uses parallel processing to simultaneously identify the type and location of the fault and assess the GPS quality.

    The multi-task neural network comprises shared layers and task-specific layers. The input layer, a single hidden layer with four neurons, and regularization layers like dropout and batch normalization are common for both tasks. In contrast, there are individual output layers for each task. The output layer comprises three neurons for the task of failure identification, one neuron for the task of location identification, and one neuron for the task of GPS quality identification. Figure 6 represents the architecture diagram of the proposed multi-task model.

    Figure 6.  Architecture diagram for the proposed multi-task model.

    The abovementioned multi-task neural networks were trained using the Adam optimizer and a combination of categorical and binary cross-entropy loss functions. The batch size was set to 5000, and the models were trained for 20 epochs. A hybrid loss function was used to train the multi-task model since two tasks (fault type and location identification) were multi-class. In contrast, the task of GPS quality identification is binary. Equation 3.1 represents the hybrid loss function used to train the multi-task model.

    LF=2i=11NnJMk=1ykjlog(p(ykj)i1NNL2p=1ylplogPlp (3.1)

    Evaluation is an essential component of the proposed workflow. Model evaluation is done to identify the model's performance on the testing set. Evaluation of the testing set is essential to identify if the trained model has overfit or underfit. The following performance metrics are derived from the confusion matrix.

    The confusion matrix comprises four values: true positive, true negative, false positive, and false negative. The diagonal elements of the matrix indicate the correctly classified samples (true positive and true negative), and the non-diagonal elements indicate the misclassified samples (false positive and false negative). The following are the different metrics used to evaluate the trained models.

    Accuracy is defined as the ratio of the correct classifications to that of the total classifications. Accuracy is considered the gold standard metric for the evaluation of classification algorithms. The formula for accuracy is mentioned in Equation 3.2.

    Accuracy=TP+TNTP+TN+FP+FN (3.2)

    Precision is defined as the ratio of true positives to that of the total positives. Precision is one of the metrics used to analyze a model's performance in class imbalance conditions. The formula for precision is mentioned in Equation 3.3.

    Precision=TPTP+FP (3.3)

    Recall is defined as the ratio of true positives to that of total samples. Recall is another metric that is used to analyze multi-class classification under the condition of class imbalance. Equation 3.4 represents the mathematical formula for recall.

    Recall=TPTP+FN (3.4)

    AUC-ROC is expanded as the area under the regional operating characteristic curve. The ROC curve is the plot between the false positive and true positive values. The area under that curve is termed an AUC score. An AUC value of less than 0.5 is considered a terrible score, a score of 0.5 is considered a random guess, and a score of more than 0.8 is considered a good score.

    Model interpretability has been a focus and requirement for which the demand has risen in recent years. Many machine learning and deep learning algorithms are considered black boxes, providing output for input data without any logical interpretations. This is needed in areas like healthcare, where life-concerning critical decisions are made. In such instances, providing interpretability to the model by providing explanations of the predictions can greatly aid clinicians.

    In this work, the local interpretable model agnostic explanations (LIME) [19] is used to derive the interpretations of the complex ensemble classifier. As the name suggests, LIME works locally, meaning it works on individual data samples. Also, this method is model agnostic, meaning it works for all models. LIME works by mapping a simple interpretable model (like linear regression) on a complex model [20]. The local region of the data space is considered, where synthetic samples are generated based on original samples. These synthetic samples are labeled based on the prediction of the complex model. Then, a simple interpretable linear regression is trained on the synthetic labeled data. The coefficients of the trained linear regression model represent the interpretations of the complex model on the local space. The LIME tabular function from the LIME library is used to get the interpretations for individual data samples.

    The LIME function explains the type of fault, location of the fault, and quality of the GPS sensor, respectively. The tabular explainer from LIME was used to explain the predicted instances. A LIME explainer was used for each task: failure type, failure location, and GPS quality. The LIME explainer generates a figure that shows the input that contributed to that particular class (positive) and the features that contributed to the counter-class (negative). Hence, we can understand the features that positively and negatively contribute to a particular class from the plot.

    The ultimate aim of the work is to develop a dashboard for real-time data analytics and predictions. A website was developed using a Python-based web development tool. The website will request the data recorded from the sensors as an Excel sheet of CSV (comma-separated values). The data visualization and prediction tasks are done simultaneously upon receiving the data. On the data visualization task, a stacked area chart of the continuous features, a stacked bar chart for the categorical features, and a map depicting the speed and route of the train are made. For the data prediction task, the latest values are processed and sent to the trained multi-task model for predictions on the failure type, failure location, and quality of the GPS, which are displayed on the website. Finally, the LIME explanations for the model on the provided sample are given for all three tasks: failure type identification, failure location identification, and GPS quality assessment.

    The developed multi-task model was trained on the training set with the above-mentioned epochs and batch size. Table 6 represents the performance metrics of the model on the training and testing datasets for the fault type and location identification, respectively.

    Table 6.  Performance comparison of the trained multi-task model.
    Set Time consumption per epoch (ms) Metric Type identification Location identification GPS quality identification
    Training 43 Loss 0.0031 0.0020 0.0029
    Accuracy 99.94 99.998 99.99
    Precision 99.99 100 100
    Recall 100 100 100
    AUC 1 1 1
    Testing 1 Loss 0.0035 0.0026 0.0033
    Accuracy 98.89 99.12 99.24
    Precision 99.56 99.67 99.84
    Recall 99.92 99.93 99.93
    AUC 1 1 1

     | Show Table
    DownLoad: CSV

    The trained multi-task ANN model produced 98.89%, 99.12% and 99.24% accuracy for failure type identification, failure location identification and GPS quality assessment. Also, the trained model's precision, recall and AUC values are high, indicating that the model overcomes class imbalance issues.

    The performance plots representing the values of the performance metrics for each epoch in the training phase are shown in Figure 7. The loss gradually decreases in each task's epoch for fault, location, and GPS quality, as shown in Figure 7(a). In contrast, the recall, AUC, and precision are increasing for each epoch, as shown in Figure 7(b), Figure 7(c), and Figure 7(d), respectively. This shows no fluctuations in the training phase and no signs of varying gradients.

    Figure 7.  Performance plots of the proposed multi-task model on the training set: (a) loss rate, (b) precision, (c) recall, (d) AUC.

    The performance plots represent the values of the performance metrics for each epoch in the testing phase. The loss gradually decreases in each task's epoch, as shown in Figure 8(a). In contrast, the recall, AUC, and precision are increasing for each epoch for fault, location, and GPS quality as shown in Figure 8(b), Figure 8(c), and Figure 8(d). The precision, recall, and AUC values for GPS quality reached the maximum in the initial epochs and remained the same for the rest. Also, fluctuations in the precision values are observed for the tasks of fault type and fault location identification.

    Figure 8.  Performance plots of the proposed multi-task model on the testing set: (a) loss rate, (b) precision, (c) recall, (d) AUC.

    The confusion matrices on the testing set represent that the trained model has produced high true negatives and true positives. In contrast, it has produced very few false positives and false negatives. The confusion matrix for the fault type is shown in Figure 9(a), The fault location confusion matrix is shown in Figure 9(b) and the confusion matrix for GPS quality is shown in Figure 9(c).

    Figure 9.  Confusion matrices for the trained multi-task model on the testing sets of all tasks: (a) fault type, (b) fault location, (c) GPS quality.

    The classification report for fault type classification indicates high precision, recall, and F1 scores for all three classes, showing no sign of class imbalance. The classification report for the trained multi-task model for fault type identification on the testing set is shown in Table 7. The classification report for the trained multi-task model for fault location identification on the testing set is shown in Table 8. The classification report for the trained multi-task model for GPS quality assessment on the testing set is shown in Table 9.

    Table 7.  Classification report for the trained multi-task model for fault type identification on the testing set.
    Classes Precision Recall F1-Score
    0 0.98 1.00 0.99
    1 1.00 1.00 1.00
    2 1.00 1.00 1.00

     | Show Table
    DownLoad: CSV
    Table 8.  Classification report for the trained multi-task model for fault location identification on the testing set.
    Classes Precision Recall F1-Score
    0 1 1 1
    1 1 1 1
    2 1 1 1
    3 1 1 1

     | Show Table
    DownLoad: CSV
    Table 9.  Classification report for the trained multi-task model for GPS quality assessment on the testing set.
    Classes Precision Recall F1-Score
    0 1 1 1
    1 1 1 1

     | Show Table
    DownLoad: CSV

    A website was developed using Streamlit and hosted online using the Streamlit share, see [21] for the website URL. Figures 10 and 11 represent the snips of the developed website about data visualization and predictive modeling. Figures 12, 13, and 14 represent the LIME explanations for the tasks of failure type identification, failure location identification, and GPS quality assessment, respectively. Figure 10 represents the area chart for the continuous features and the bar chart for the categorical features. From this bar, anomalies in the sensor values can be visually identified. Figure 11 represents the map plot for the GPS data; the dots represent the map's route based on the latitude and longitude values, whereas the dot's intensity depicts the speed. This graph identifies the train's route and the crucial locations at which the train went fast/slow. Also, the multi-task neural network predictions are mentioned on the website for each dataset instance.

    Figure 10.  The data visualization part of the developed website shows the plot for continuous and categorical features of the testing data.
    Figure 11.  Map representing the speed and route of the train along with the predictions made by the multi-task model.
    Figure 12.  LIME explanations for the model prediction on random samples for failure type identification.
    Figure 13.  LIME explanations for the model prediction on random samples for failure location identification.
    Figure 14.  LIME explanations for the model prediction on random samples for the task of GPS quality assessment.

    Table 10 compares the performance of the proposed multi-task neural network to that of the existing works related to the application of predictive maintenance in train fault analysis using sensor data.

    Table 10.  Comparative analysis of the results of the proposed work to that of the existing works.
    Author Algorithm Result
    Rajashekarappa et al. [13] RUS Boosted Classifier 98.73% accuracy
    Najjar et al. [15] Random Forest Classifier 97% accuracy
    Davari et al. [16] Autoencoder 44% improvement in precision compared to baseline
    Proposed work Multi-task Artificial Neural Network 99% average accuracy

     | Show Table
    DownLoad: CSV

    Based on our knowledge, only two works [15,16] use artificial intelligence for the predictive maintenance of urban metro trains. Najjar et al. [15] worked on predicting air failure of the air production unit (APU) in metro trains. The dataset used for this task was MetroPT, a 6-month analysis of metro trains in Portugal comprising analog, digital, and GPS sensors. The GPS information was excluded from the dataset, and the timestamp was encoded using the label encoding technique. A random forest classifier algorithm was used for the multi-class classification of air failure prediction. The data was undersampled and then split into training and testing sets. A feature importance visualization technique was employed to identify the root cause of the air failure. The proposed work produced testing accuracies of 84% and 87% on the binary and multi-class classification tasks and F1 scores in the ranges of 0.83–0.5 and 0.73–0.97 for the binary and multi-class classification tasks. The proposed model produced better results than the work and has also included oil failures in addition to air failures. Davari et al. [16] developed a deep learning neural network for anomaly detection in metro trains. The algorithms used for this task were the sparse autoencoder and variational autoencoder. This work is an unsupervised learning approach for anomaly detection of air failures in trains. The sparse autoencoder trained on the digital data produced 42% more than those trained on analog data. Also, the variational autoencoder performed better than the sparse autoencoder by 37%. The proposed work considered both analog, digital, and GPS sensors worked on both air and oil failures, and produced state-of-the-art results.

    The multi-task model has produced excellent results on the training and testing datasets. The performance plots prove that the model has been trained perfectly and shows no signs of overfitting or underfitting. Also, the confusion matrices and classification report suggest that the trained model is generalized and does not exhibit any signs of class imbalance.

    The proposed work has some advantages in comparison to the existing works. The proposed work addresses all issues faced in the metro trains (air and oil failures). Another advantage is the excellent results obtained by the trained models. The third advantage is agility; the proposed multi-task model has taken less time to train and predict batch data. Finally, the explainable AI technique, namely LIME, has been implemented to provide interpretations to the outputs given by the multi-task model. This provides belief in the model prediction and can also be useful for engineers to deeply analyze the issue.

    Figure 12 represents the LIME explanation for a local instance related to the task of fault type prediction. As observed from the plot, the predicted instance is air failure with 100% confidence, and the features positively and negatively contribute to the prediction.

    Figure 13 represents the LIME explanation for a local instance related to the task of fault location prediction. As observed from the plot, the predicted instance is a client with 100% confidence, and the features positively and negatively contribute to the prediction. Motor current, DV pressure, and the day positively contributed to the prediction, followed by GPS speed and reservoir. The GPS longitude, minute, and flowmeter have negatively contributed to the prediction.

    Figure 14 represents the LIME explanation for a local instance related to the task of GPS quality prediction. As observed from the plot, the predicted instance is air failure with 100% confidence, and the features positively and negatively contribute to the prediction. Oil temperature has majorly contributed positively, whereas H1, day, and minute have majorly contributed negatively. For all plots, the range or condition of the input features is given, which might be of great use for the fault analysis.

    One reason for achieving good results is the split of the output labels into fault type and location. This reduces the number of interdependent classes in each stage, which might have improved the performance of the algorithms. Also, we observed a performance rise of 6% when the features were standardized. Considering the deep learning aspects, we developed a multi-task model which splits the classes into individual tasks, allowing for more attention, resulting in better results and quicker periods.

    However, the study has some limitations. First, the dataset was undersampled to 30, 00,000 data points, roughly 20% of the entire dataset. The second one was the expansion of the target vectors into new columns, which introduced more computations and the need to train more models. While this approach was considered to improve the holistic performance of the models, this resulted in the creation of multiple datasets and, ultimately, multiple ML models for training, leading to computational costs. The third one was the limited selection of machine learning algorithms. Many good machine learning algorithms like support vector machine and ensemble learning techniques were not implemented due to the computational constraints and long training durations (the SVM algorithm did not train even after 30 minutes!).

    Hence in this work, a multi-task model was developed for the identification of failures simultaneously. The proposed method has produced 98.89%, 99.12%, and 99.24% accuracies in the testing set for failure type, failure location, and GPS quality predictions, respectively, exceeding the state-of-the-art methods. The model produced 99.56%, 99.67%, and 99.84% precision in the testing set for failure type, failure location, and GPS quality predictions, respectively. The high accuracy and precision values indicate the good performance of the model and no signs of class imbalance. The deep learning model took 43 seconds for training and 1 second for inferencing for test data, showing fast predictions, needed for predictive maintenance applications. Moreover, a real-time interactive dashboard was developed, performing dynamic data visualization and predictions. Finally, using the LIME explainable AI technique provides explanations for the predictions, adding belief and better analysis for engineers. The developed system would be advantageous for engineers to perform fault analysis and predictive maintenance effectively. Future work will use a database to store the streaming data and deploy the system in real time. Also, we will develop deep learning algorithms on the entire dataset and employ online learning strategies to update the learned model in real time.

    Pratik Vinayak Jadhav: Data Curation and Analysis, Research Design and Methodology, writing draft. Sairam V. A: Data Curation and Analysis, Research Design and Methodology, writing draft. Siddharth Sonkavade: Data Curation and Analysis, Research Design and Methodology, writing draft. Shivali Amit Wagle : Conceptualization, Supervision, Project Development, Writing, Review, and Editing. Preksha Pareek: Conceptualization, Supervision, Project Development, Writing, and Review. Ketan Kotecha: Writing, Review, and Editing, Funding and Resources. Tanupriya Choudhury: Data Analysis, Writing, Review, and Editing.

    The authors declare no conflict of interest.



    [1] Frank AG, Dalenogare LS, Ayala NF (2019) Industry 4.0 technologies: Implementation patterns in manufacturing companies. International Journal of Production Economics 210: 15–26. https://doi.org/10.1016/j.ijpe.2019.01.004 doi: 10.1016/j.ijpe.2019.01.004
    [2] Bueno AF, Godinho Filho M, Frank AG (2020) Smart production planning and control in the Industry 4.0 context: A systematic literature review. Computers & Industrial Engineering 149: 106774. https://doi.org/10.1016/j.cie.2020.106774 doi: 10.1016/j.cie.2020.106774
    [3] Sardar SK, Sarkar B, Kim B (2021) Integrating Machine Learning, Radio Frequency Identification, and Consignment Policy for Reducing Unreliability in Smart Supply Chain Management Processes 9: 247. https://doi.org/10.3390/pr9020247 doi: 10.3390/pr9020247
    [4] Tseng ML, Tran TPT, Ha HM, et al. (2021) Sustainable industrial and operation engineering trends and challenges Toward Industry 4.0: A data driven analysis. Journal of Industrial and Production Engineering 38: 581–598. https://doi.org/10.1080/21681015.2021.1950227 doi: 10.1080/21681015.2021.1950227
    [5] Sarkar M, Pan L, Dey BK, et al. (2020) Does the Autonomation Policy Really Help in a Smart Production System for Controlling Defective Production? Mathematics 8: 1142. https://doi.org/10.3390/math8071142 doi: 10.3390/math8071142
    [6] Sarkar M, Chung BD (2021) Effect of Renewable Energy to Reduce Carbon Emissions under a Flexible Production System: A Step Toward Sustainability. Energies 14: 215. https://doi.org/10.3390/en14010215 doi: 10.3390/en14010215
    [7] Jwo JS, Lin CS, Lee CH (2021) Smart technology–driven aspects for human-in-the-loop smart manufacturing. The International Journal of Advanced Manufacturing Technology 114: 1741–1752. https://doi.org/10.1007/s00170-021-06977-9 doi: 10.1007/s00170-021-06977-9
    [8] El Cadi AA, Gharbi A, Dhouib K, et al. (2021) Joint production and preventive maintenance controls for unreliable and imperfect manufacturing systems. Journal of Manufacturing Systems 58: 263–279. https://doi.org/10.1016/j.jmsy.2020.12.003 doi: 10.1016/j.jmsy.2020.12.003
    [9] Sarkar B, Guchhait R, Sarkar M, et al. (2019) How does an industry manage the optimum cash flow within a smart production system with the carbon footprint and carbon emission under logistics framework? International Journal of Production Economics 213: 243–257. https://doi.org/10.1016/j.ijpe.2019.03.012 doi: 10.1016/j.ijpe.2019.03.012
    [10] Roy MD, Sana SS (2021) Production rate and lot-size dependent lead time reduction strategies in a supply chain model with stochastic demand, controllable setup cost and trade-credit financing. RAIRO: Recherche Opérationnelle 55: 1469. https://doi.org/10.1051/ro/2020112 doi: 10.1051/ro/2020112
    [11] Jeong H, Karim RA, Sieverding HL, et al. (2021) An Application of GIS-Linked Biofuel Supply Chain Optimization Model for Various Transportation Network Scenarios in Northern Great Plains (NGP), USA. BioEnergy Research 14: 612–622. https://doi.org/10.1007/s12155-020-10223-7 doi: 10.1007/s12155-020-10223-7
    [12] Sedehzadeh S, Seifbarghy M (2021) Redesigning a fast-moving consumer goods supply chain considering social responsibility and logistical restrictions: case study in an Iranian food company. Environmental Science and Pollution Research 2021: 1–16. https://doi.org/10.1007/s11356-021-14760-2 doi: 10.1007/s11356-021-14760-2
    [13] Guchhait R, Pareek S, Sarkar B (2019) How Does a Radio Frequency Identification Optimize the Profit in an Unreliable Supply Chain Management? Mathematics 7: 490. https://doi.org/10.3390/math7060490 doi: 10.3390/math7060490
    [14] Hoque M (2020) A manufacturer-buyers integrated inventory model with various distributions of lead times of delivering equal-sized batches of a lot. Computers & Industrial Engineering 145: 106516. https://doi.org/10.1016/j.cie.2020.106516 doi: 10.1016/j.cie.2020.106516
    [15] Sarkar B, Saren S, Sinha D, et al. (2015) Effect of unequal lot sizes, variable setup cost, and carbon emission cost in a supply chain model. Mathematical Problems in Engineering 2015. https://doi.org/10.1155/2015/469486 doi: 10.1155/2015/469486
    [16] Hota SK, Sarkar B, Ghosh SK (2020) Effects of unequal lot size and variable transportation in unreliable supply chain management. Mathematics 8: 357. https://doi.org/10.3390/math8030357 doi: 10.3390/math8030357
    [17] Ai Y, Xu Y (2021) Strategic sourcing in forward and spot markets with reliable and unreliable suppliers. International Journal of Production Research 59: 926–941. https://doi.org/10.1080/00207543.2020.1711987 doi: 10.1080/00207543.2020.1711987
    [18] Giri BC, Majhi JK, Chaudhuri K (2021) Coordination mechanisms of a three-layer supply chain under demand and supply risk uncertainties. RAIRO-Operations Research 55: S2593–S2617. https://doi.org/10.1051/ro/2020101 doi: 10.1051/ro/2020101
    [19] Hlioui R, Gharbi A, Hajji A (2017) Joint supplier selection, production and replenishment of an unreliable manufacturing-oriented supply chain. International Journal of Production Economics. 187: 53–67. https://doi.org/10.1016/j.ijpe.2017.02.004 doi: 10.1016/j.ijpe.2017.02.004
    [20] Scarf H (1958) A min-max solution of an inventory problem. Studies in the mathematical theory of inventory and production
    [21] Pal B, Adhikari S (2022) Optimal strategies for members in a two-echelon supply chain over a safe period under random machine hazards with backlogging. Journal of Industrial and Production Engineering 2022: 1–18. https://doi.org/10.1080/21681015.2021.2001767 doi: 10.1080/21681015.2021.2001767
    [22] Elfarouk O, Wong KY, Wong WP (2022) Multi-objective optimization for multi-echelon, multi-product, stochastic sustainable closed-loop supply chain. Journal of Industrial and Production Engineering 39: 109–127. https://doi.org/10.1080/21681015.2021.1963338 doi: 10.1080/21681015.2021.1963338
    [23] Chen Y, Feng Q, Senior Member I, et al. (2021) Modeling and analyzing RFID Generation-2 under unreliable channels. Journal of Network and Computer Applications 178: 102937. https://doi.org/10.1016/j.jnca.2020.102937 doi: 10.1016/j.jnca.2020.102937
    [24] Entezaminia A, Gharbi A, Ouhimmou M (2021) A joint production and carbon trading policy for unreliable manufacturing systems under cap-and-trade regulation. Journal of Cleaner Production 293: 125973. https://doi.org/10.1016/j.jclepro.2021.125973 doi: 10.1016/j.jclepro.2021.125973
    [25] Romero-Silva R, Shaaban S, Marsillac E, et al. (2021) The Impact of Unequal Processing Time Variability on Reliable and Unreliable Merging Line Performance. International Journal of Production Economics 2021: 108108. https://doi.org/10.1016/j.ijpe.2021.108108 doi: 10.1016/j.ijpe.2021.108108
    [26] Sardar SK, Sarkar B (2020) How Does Advanced Technology Solve Unreliability Under Supply Chain Management Using Game Policy? Mathematics 8 :1191. https://doi.org/10.3390/math8071191 doi: 10.3390/math8071191
    [27] Sarkar B, Omair M, Kim N (2020) A cooperative advertising collaboration policy in supply chain management under uncertain conditions. Applied Soft Computing 88 :105948. https://doi.org/10.1016/j.asoc.2019.105948 doi: 10.1016/j.asoc.2019.105948
    [28] Park K, Lee K (2016) Distribution-robust single-period inventory control problem with multiple unreliable suppliers. OR spectrum 38 :949–966. https://doi.org/10.1007/s00291-016-0440-4 doi: 10.1007/s00291-016-0440-4
    [29] Ullah M, Sarkar B (2020) Recovery-channel selection in a hybrid manufacturing-remanufacturing production model with RFID and product quality. International Journal of Production Economics 219: 360–374. https://doi.org/10.1016/j.ijpe.2019.07.017 doi: 10.1016/j.ijpe.2019.07.017
    [30] Dalenogare LS, Benitez GB, Ayala NF, et al. (2018) The expected contribution of Industry 4.0 technologies for industrial performance. International Journal of Production Economics 204: 383–394. https://doi.org/10.1016/j.ijpe.2018.08.019 doi: 10.1016/j.ijpe.2018.08.019
    [31] Wang S, Wan J, Zhang D, et al. (2016) Towards smart factory for industry 4.0: a self-organized multi-agent system with big data based feedback and coordination. Computer Networks 101: 158–168. https://doi.org/10.1016/j.comnet.2015.12.017 doi: 10.1016/j.comnet.2015.12.017
    [32] Chen B, Wan J, Shu L, et al. (2017) Smart factory of industry 4.0: Key technologies, application case, and challenges. Ieee Access 6: 6505–6519. https://doi.org/10.1109/ACCESS.2017.2783682 doi: 10.1109/ACCESS.2017.2783682
    [33] Dey BK, Pareek S, Tayyab M, et al. (2021) Autonomation policy to control work-in-process inventory in a smart production system. International Journal of Production Research 59: 1258–1280. https://doi.org/10.1080/00207543.2020.1722325 doi: 10.1080/00207543.2020.1722325
    [34] Bhuniya S, Pareek S, Sarkar B, et al. (2021) A Smart Production Process for the Optimum Energy Consumption with Maintenance Policy under a Supply Chain Management. Processes 9: 19. https://doi.org/10.3390/pr9010019 doi: 10.3390/pr9010019
    [35] Sarkar M, Sarkar B (2019) Optimization of safety stock under controllable production rate and energy consumption in an automated smart production management. Energies 12: 2059. https://doi.org/10.3390/en12112059 doi: 10.3390/en12112059
    [36] Goyal SK (1988) "A joint economic-lot-size model for purchaser and vendor": A comment. Decision sciences 19:236–241. doi: 10.1111/j.1540-5915.1988.tb00264.x
    [37] Hoque M (2013) A manufacturer–buyer integrated inventory model with stochastic lead times for delivering equal-and/or unequal-sized batches of a lot. Computers & operations research 40: 2740–2751. https://doi.org/10.1016/j.cor.2013.05.008 doi: 10.1016/j.cor.2013.05.008
    [38] Hariga M, Gumus M, Daghfous A (2014) Storage constrained vendor managed inventory models with unequal shipment frequencies. Omega 48: 94–106. https://doi.org/10.1016/j.omega.2013.11.003 doi: 10.1016/j.omega.2013.11.003
    [39] Garai A, Sarkar B (2022) Economically independent reverse logistics of customer-centric closed-loop supply chain for herbal medicines and biofuel. Journal of Cleaner Production 334: 129977. https://doi.org/10.1016/j.jclepro.2021.129977 doi: 10.1016/j.jclepro.2021.129977
    [40] Sarkar B, Debnath A, Chiu AS, et al. (2022) Circular economy-driven two-stage supply chain management for nullifying waste. Journal of Cleaner Production 339: 130513. https://doi.org/10.1016/j.jclepro.2022.130513 doi: 10.1016/j.jclepro.2022.130513
    [41] Sarkar B, Bhuniya S (2022) A sustainable flexible manufacturing–remanufacturing model with improved service and green investment under variable demand. Expert Systems with Applications 202: 117154. https://doi.org/10.1016/j.eswa.2022.117154 doi: 10.1016/j.eswa.2022.117154
    [42] Choi SB, Dey BK, Kim SJ, et al. (2022) Intelligent servicing strategy for an online-to-offline (O2O) supply chain under demand variability and controllable lead time. RAIRO-Operations Research 2022 https://doi.org/10.1051/ro/2022026 doi: 10.1051/ro/2022026
    [43] Sarkar B, Ullah M, Sarkar M (2022) Environmental and economic sustainability through innovative green products by remanufacturing. Journal of Cleaner Production 332: 129813. https://doi.org/10.1016/j.jclepro.2021.129813 doi: 10.1016/j.jclepro.2021.129813
    [44] Majumder A, Jaggi CK, Sarkar B (2018) A multi-retailer supply chain model with backorder and variable production cost. RAIRO-Operations Research 52: 943–954. https://doi.org/10.1051/ro/2017013 doi: 10.1051/ro/2017013
    [45] Tang S, Wang W, Cho S, et al. (2018) Reducing emissions in transportation and inventory management:(R, Q) Policy with considerations of carbon reduction European Journal of Operational Research 269: 327–340. https://doi.org/10.1016/j.ejor.2017.10.010 doi: 10.1016/j.ejor.2017.10.010
    [46] Bhuniya S, Sarkar B, Pareek S (2019) Multi-product production system with the reduced failure rate and the optimum energy consumption under variable demand Mathematics 7: 465. https://doi.org/10.3390/math7050465 doi: 10.3390/math7050465
    [47] Mishra M, Hota SK, Ghosh SK, et al. (2020) Controlling Waste and Carbon Emission for a Sustainable Closed-Loop Supply Chain Management under a Cap-and-Trade Strategy. Mathematics 8: 466. https://doi.org/10.3390/math8040466 doi: 10.3390/math8040466
    [48] Manna AK, Mondal R, Akbar Shaikh A, et al. (2021) Single-manufacturer and multi-retailer supply chain model with pre-payment based partial free transportation. RAIRO-Operations Research 55: 1063–1076. https://doi.org/10.1051/ro/2021053 doi: 10.1051/ro/2021053
    [49] Lin YJ (2008) Minimax distribution free procedure with backorder price discount. International Journal of Production Economics 111: 118–128. https://doi.org/10.1016/j.ijpe.2006.11.016 doi: 10.1016/j.ijpe.2006.11.016
    [50] Moon I, Choi S (1998) TECHNICAL NOTEA note on lead time and distributional assumptions in continuous review inventory models. Computers & Operations Research 25: 1007–1012. https://doi.org/10.1016/S0305-0548(97)00103-2 doi: 10.1016/S0305-0548(97)00103-2
    [51] Centobelli P, Cerchione R, Del Vecchio P, et al. (2021) Blockchain technology for bridging trust, traceability and transparency in circular supply chain. Information & Management 2021: 103508 https://doi.org/10.1016/j.im.2021.103508 doi: 10.1016/j.im.2021.103508
    [52] Kshetri N (2021) Blockchain and sustainable supply chain management in developing countries. International Journal of Information Management 60: 102376. https://doi.org/10.1016/j.ijinfomgt.2021.102376 doi: 10.1016/j.ijinfomgt.2021.102376
    [53] Shen C, Pena-Mora F (2018) Blockchain for cities—a systematic literature review. Ieee Access 6: 76787–76819. https://doi.org/10.1109/ACCESS.2018.2880744 doi: 10.1109/ACCESS.2018.2880744
    [54] Centobelli P, Cerchione R, Del Vecchio P, et al. (2021) Blockchain technology design in accounting: Game changer to tackle fraud or technological fairy tale? Accounting, Auditing & Accountability Journal 2021. https://doi.org/10.1108/AAAJ-10-2020-4994 doi: 10.1108/AAAJ-10-2020-4994
    [55] Zhang H, Hou JC (2005) Maintaining sensing coverage and connectivity in large sensor networks. Ad Hoc Sens Wirel Networks 1: 89–124.
    [56] Hefeeda M, Ahmadi H (2007) A probabilistic coverage protocol for wireless sensor networks. 2007 IEEE International Conference on Network Protocols 2007: 41–50. https://doi.org/10.1109/ICNP.2007.4375835 doi: 10.1109/ICNP.2007.4375835
    [57] Hota SK, Ghosh SK, Sarkar B (2022) A solution to the transportation hazard problem in a supply chain with an unreliable manufacturer AIMS Environmental Science 9: 354–380. https://doi.org/10.3934/environsci.2022023 doi: 10.3934/environsci.2022023
  • This article has been cited by:

    1. Piotr F. Borowski, Nexus between water, energy, food and climate change as challenges facing the modern global, European and Polish economy, 2020, 6, 2471-2132, 397, 10.3934/geosci.2020022
    2. Christopher Schulz, Udisha Saklani, The future of hydropower development in Nepal: Views from the private sector, 2021, 179, 09601481, 1578, 10.1016/j.renene.2021.07.138
    3. Fanqi Zou, Tinghui Li, Feite Zhou, Does the Level of Financial Cognition Affect the Income of Rural Households? Based on the Moderating Effect of the Digital Financial Inclusion Index, 2021, 11, 2073-4395, 1813, 10.3390/agronomy11091813
    4. Divya Narain, Hoong Chen Teo, Alex Mark Lechner, James E.M. Watson, Martine Maron, Biodiversity risks and safeguards of China’s hydropower financing in Belt and Road Initiative (BRI) countries, 2022, 5, 25903322, 1019, 10.1016/j.oneear.2022.08.012
    5. Edoardo Borgomeo, Bill Kingdom, Judith Plummer-Braeckman, Winston Yu, Water infrastructure in Asia: financing and policy options, 2022, 0790-0627, 1, 10.1080/07900627.2022.2062707
    6. Judith Plummer Braeckman, Sanna Markkanen, Perceptions of Risk in Relation to Large Hydropower Projects: A Finance Perspective, 2021, 1556-5068, 10.2139/ssrn.4011266
    7. Emmanuel Yamoah Tenkorang, Francis Enu-Kwesi, Franklin Bendu, Pon Souvannaseng, Evolving Lending Regimes and the Political Economy of Dam Financing in Ghana, 2022, 1556-5068, 10.2139/ssrn.4013584
    8. Yan Yang, Bo Yang, Zijun Xin, Green finance development, environmental attention and investment in hydroelectric power: From the perspective of environmental protection law, 2024, 69, 15446123, 106167, 10.1016/j.frl.2024.106167
    9. Pon Souvannaseng, Fast Finance and the Political Economy of Catastrophic Dam Collapse in Lao PDR: The Case of Xe Pian-Xe Namnoy, 2024, 97, 0030-851X, 261, 10.5509/2024972-art7
  • Reader Comments
  • © 2022 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(2148) PDF downloads(138) Cited by(8)

Figures and Tables

Figures(8)  /  Tables(3)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog