Deep learning-based belt conveyor fault detection method and apparatus

Abstract

This application discloses a deep learning-based fault detection method and apparatus for belt conveyors, relating to the field of fault detection technology. The method includes: maintaining a global fault detection model in a cloud computing center and deploying a lightweight fault detection model on each edge computing gateway; using the lightweight fault detection model of the edge computing gateway to perform hierarchical fault detection on the DAS monitoring dataset, extracting abnormal data subgraphs, and uploading them to the cloud computing center; using an improved optimization algorithm to optimize adaptive dynamic threshold parameters for the next hierarchical fault detection, and performing periodic baseline reset; using the global fault detection model in the cloud computing center to perform deep fault detection on the abnormal data subgraphs, generating intelligent maintenance decisions, and sending them to the edge computing gateway. This solves the problems of high false alarm and false negative rates, poor model performance and adaptability, and insufficient stability in existing technologies.

Description

Technical Field

[0001] This invention relates to the field of fault detection technology, and in particular to a method and apparatus for fault detection of belt conveyors based on deep learning. Background Technology

[0002] Belt conveyors, as common continuous conveying equipment in industrial systems, are widely used in coal mines, power plants, ports, and tunnel construction. Their operational stability and safety are crucial to the entire production system. In practical applications, belt conveyors may experience various malfunctions such as tearing, misalignment, foreign object jamming, and overheating. Therefore, real-time monitoring and intelligent alarm systems are particularly important.

[0003] Currently, traditional belt conveyor fault detection relies heavily on point-based sensing devices (such as belt misalignment switches, temperature sensors, and pull-cord switches) to collect localized data. While these methods are simple to deploy and low in cost, they suffer from limitations in deployment points, insufficient coverage, and susceptibility to interference. They also make it difficult to continuously monitor the entire conveying system, especially in long-distance conveying lines.

[0004] In recent years, with the development of deep learning technology, autoencoders, as an unsupervised learning method, have been widely used in the field of equipment fault detection. For example, existing technology discloses a reactor fault detection method based on autoencoders. This method trains the autoencoder to learn the feature representation of the reactor's normal operating state and uses the reconstruction error to make fault judgments, showing good performance in fault detection of static equipment such as reactors.

[0005] However, applying existing self-encoder fault detection technology directly to belt conveyor monitoring scenarios presents significant technical challenges:

[0006] 1) Spatial Distributed Multi-Point Heterogeneity Problem: Belt conveyors may involve hundreds to thousands of spatially distributed monitoring points along their route. The signal characteristics of each point vary significantly due to factors such as geographical location, equipment status, and environmental conditions. Existing unified models or unified threshold strategies cannot adapt to this large-scale spatial heterogeneity, easily leading to some points being overly sensitive and generating high false alarm rates, or not sensitive enough and resulting in high false alarm rates.

[0007] 2) Threshold drift problem under time-varying operating conditions: When a belt conveyor operates under different loads, temperatures, humidity, and other conditions, the signal characteristics of its normal state will drift over time. Existing technologies mostly use static thresholds determined during the training period, which can easily lead to a mismatch between the threshold and the actual normal state distribution during long-term operation, resulting in an increased misjudgment rate.

[0008] 3) Real-time processing challenges of high-frequency, large-scale data: DAS monitoring networks generate high-frequency, high-density data streams, placing higher demands on model inference speed and computational resource consumption. Existing complex model structures are unsuitable for large-scale parallel deployment and real-time processing in industrial settings.

[0009] 4) Stability issues during long-term unattended operation: Belt conveyor monitoring systems need to operate unattended for extended periods in harsh industrial environments, requiring the detection algorithm to possess adaptive capabilities and long-term stability. Existing technologies lack effective mechanisms to address issues such as threshold drift and model degradation during long-term operation.

[0010] Therefore, there is an urgent need for a fault detection method specifically designed for belt conveyor monitoring scenarios. This method needs to systematically address unique technical challenges such as spatially distributed heterogeneity, adaptability to time-varying operating conditions, high-frequency large-scale real-time processing, and long-term stable operation. Summary of the Invention

[0011] This invention provides a method and device for fault detection of belt conveyors based on deep learning, which solves the problems of high false alarm rate and false negative rate, poor model performance and adaptability, and insufficient stability in the existing technology.

[0012] In a first aspect, embodiments of the present invention provide a fault detection method for belt conveyors based on deep learning, the method comprising:

[0013] Based on federated learning and knowledge distillation mechanisms, a global fault detection model is maintained in the cloud computing center, and a lightweight fault detection model is deployed in each edge computing gateway.

[0014] Based on incremental update and dynamic threshold determination mechanism, a lightweight fault detection model of edge computing gateway is used to perform hierarchical fault detection on DAS monitoring dataset of belt conveyor, extract abnormal data sub-graphs, and upload them to cloud computing center.

[0015] An improved optimization algorithm is used to optimize the adaptive dynamic threshold parameters of each edge computing gateway. The optimized adaptive dynamic threshold parameters are then used for the next hierarchical fault detection and periodic baseline reset is performed.

[0016] Using the global fault detection model of the cloud computing center, deep fault detection is performed on the abnormal data subgraph. Based on the obtained deep fault detection results, intelligent maintenance decisions are generated and sent to the corresponding edge computing gateway.

[0017] The technical solution provided in this application has at least the following beneficial effects:

[0018] By explicitly modeling the physical transport relationships between points using a graph structure, the fault detection model can learn global graph structure information and temporal information, fundamentally solving the spatial heterogeneity problem and achieving high consistency in detection accuracy. The hierarchical incremental update mechanism concentrates computational resources in suspicious areas, avoiding the huge overhead of high-frequency computation across the entire graph, and can meet the real-time processing requirements of high-frequency data while ensuring high detection accuracy. The combination of adaptive dynamic threshold and periodic baseline reset can adapt to long-term operating condition drift and has self-recovery capability, ensuring reliability for long-term unattended operation in harsh industrial environments. Compared with training a model independently for each point, this invention only requires a lightweight model and a global model, simplifying model management and reducing deployment and maintenance costs, thus achieving system-level optimization.

[0019] In one alternative implementation, the fault detection model is constructed based on the GNN-LSTM-AE algorithm, and the fault detection model includes a graph structure feature extraction module based on the GNN algorithm, a temporal feature extraction module based on the LSTM algorithm, and a fault detection module based on the AE algorithm. The graph structure feature extraction module and the temporal feature extraction module are both connected to the fault detection module.

[0020] In one alternative implementation, based on federated learning and knowledge distillation mechanisms, a global fault detection model is maintained at the cloud computing center, and a lightweight fault detection model is deployed at each edge computing gateway, including:

[0021] The initial teacher model is built in the cloud computing center using the GNN-LSTM-AE algorithm, and a lightweight initial student model is built at each edge computing gateway using the GNN-LSTM-AE algorithm.

[0022] Based on the private sample set of each edge computing gateway, the initial student model is trained locally, the model update amount of the obtained local student model is extracted, and it is encrypted and uploaded to the cloud computing center.

[0023] Based on the federated learning mechanism, the initial teacher model is updated according to the model update amount of all edge computing gateways to obtain the updated teacher model.

[0024] Based on the cloud sample set in the cloud computing center, the updated teacher model is globally trained to obtain a global teacher model, and the corresponding teacher model parameters and soft targets are extracted and sent to all edge computing gateways.

[0025] Based on the knowledge distillation mechanism, according to the teacher model parameters and soft objectives, at each edge computing gateway, the local student model is distilled and trained to obtain an updated student model, and the teacher model update process is re-executed.

[0026] Repeat the above steps until the performance of the updated student model converges. Use the final student model of each edge computing gateway as a lightweight fault detection model and the final teacher model of the cloud computing center as the global fault detection model.

[0027] In one optional implementation, based on incremental updates and dynamic threshold determination mechanisms, a lightweight fault detection model of the edge computing gateway is used to perform hierarchical fault detection on the DAS monitoring dataset of the belt conveyor, extract abnormal data sub-graphs, and upload them to the cloud computing center, including:

[0028] Using the DAS monitoring network, the initial DAS monitoring dataset of the belt conveyor is collected and uploaded to the edge computing gateway within the communication range;

[0029] Using an edge computing gateway, the initial DAS monitoring dataset is preprocessed to obtain a standard DAS monitoring dataset, which is then transformed into a DAS monitoring data graph.

[0030] Using a lightweight fault detection model, the reconstruction error of each node in the DAS monitoring data graph is extracted, and an adaptive dynamic threshold for the edge computing gateway is generated based on the reconstruction error of all nodes according to the dynamic threshold determination mechanism.

[0031] If the reconstruction error of any node exceeds the adaptive dynamic threshold, then the abnormal data subgraph, including several of its neighboring nodes, is extracted centered on the abnormal node that exceeds the threshold, and the abnormal data subgraph is uploaded to the cloud computing center.

[0032] In one alternative implementation, an edge computing gateway is used to preprocess the initial DAS monitoring dataset to obtain a standard DAS monitoring dataset, which is then transformed into a DAS monitoring data graph, including:

[0033] Using an edge computing gateway, each initial DAS monitoring data in the initial DAS monitoring dataset is cleaned, filtered, normalized, and segmented by a sliding window to obtain several standard DAS monitoring datasets for each time window.

[0034] Each monitoring point in the DAS monitoring network is taken as a node, and the corresponding standard DAS monitoring data and the spatial location information of the monitoring point are combined into node features.

[0035] Based on the physical connection relationship of the belt conveyor, define the edges between different monitoring points, and initialize the edge weights according to the physical distance and / or signal correlation.

[0036] Based on all nodes, node features, and the edges and edge weights between nodes, a graph transformation is performed on the standard DAS monitoring dataset to obtain the DAS monitoring data graph.

[0037] In one alternative implementation, the improved optimization algorithm is the ECSA algorithm.

[0038] In one alternative implementation, an improved optimization algorithm is used to optimize the adaptive dynamic threshold parameters of each edge computing gateway. The optimized adaptive dynamic threshold parameters are then used for the next hierarchical fault detection, and periodic baseline resets are performed, including:

[0039] The adaptive dynamic threshold parameters of the edge computing gateway are encoded into individual vectors of the ECSA algorithm, and the ECSA population parameters and maximum number of iterations are set.

[0040] Based on the ECSA population parameters, the initial ECSA population is obtained by initializing using the Tent chaotic mapping sequence; each ECSA individual in the ECSA population corresponds to an alternative adaptive dynamic threshold parameter.

[0041] Based on the private sample set, the fitness function is used to obtain the fitness value of each initial ECSA individual, and the initial ECSA individual with the best fitness value is taken as the optimal solution.

[0042] Based on the Levy flight strategy and convergence factor, the nest positions of all initial ECSA individuals except the optimal solution are updated to obtain the updated first ECSA population.

[0043] Based on the reconstruction probability, several first ECSA individuals are randomly selected from the updated first ECSA population to rebuild nests, resulting in several updated second ECSA individuals.

[0044] According to the dynamic reverse mechanism, several first ECSA individuals are randomly selected from the updated first ECSA population with reverse probability to reverse, resulting in several updated third ECSA individuals.

[0045] Based on the private sample set, the fitness function is used to obtain the fitness value of each updated first, second or third ECSA individual, and the updated ECSA individual with the best fitness value is updated as the optimal solution;

[0046] When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, the iterative update of the ECSA population is terminated, and the optimal solution of the current iteration is output.

[0047] The individual vector of the ECSA individual corresponding to the optimal solution is decoded to obtain the optimized adaptive dynamic threshold parameter, and the optimized adaptive dynamic threshold parameter is used for the next hierarchical fault detection.

[0048] Perform periodic baseline resets for each edge computing gateway.

[0049] In one optional implementation, a global fault detection model from the cloud computing center is used to perform deep fault detection on the abnormal data subgraph. Based on the obtained deep fault detection results, an intelligent maintenance decision is generated and sent to the corresponding edge computing gateway, including:

[0050] Input the abnormal data subgraph into the global fault detection model of the cloud computing center;

[0051] The graph structure feature extraction module of the global fault detection model is used to extract the graph structure features of the abnormal data subgraphs.

[0052] The temporal feature extraction module of the global fault detection model is used to extract the temporal features of the node features in the abnormal data subgraph.

[0053] Based on the graph structure features and temporal features, the fault detection module of the global fault detection model is used to call the softmax function to perform deep fault detection and obtain the deep fault detection results.

[0054] Extract the belt conveyor operating status corresponding to the abnormal data subgraph, and input the belt conveyor operating status and deep fault detection results into the pre-built intelligent maintenance decision model;

[0055] Based on the operating status of the belt conveyor and the results of deep fault detection, an intelligent maintenance decision model is used to generate intelligent maintenance decisions, which are then sent to the corresponding edge computing gateway.

[0056] In one alternative implementation, the intelligent maintenance decision model is built based on the PPO algorithm, and the intelligent maintenance decision model is equipped with an agent and an experience playback pool.

[0057] Secondly, embodiments of the present invention provide a deep learning-based belt conveyor fault detection device for implementing a belt conveyor fault detection method. The device includes:

[0058] The model building unit is used to maintain a global fault detection model in the cloud computing center based on federated learning and knowledge distillation mechanisms, and to deploy a lightweight fault detection model on each edge computing gateway.

[0059] The hierarchical fault detection unit is used to perform hierarchical fault detection on the DAS monitoring dataset of the belt conveyor based on incremental update and dynamic threshold determination mechanism, using a lightweight fault detection model of edge computing gateway, extracting abnormal data sub-graphs, and uploading them to the cloud computing center.

[0060] The threshold parameter optimization unit is used to optimize the adaptive dynamic threshold parameters of each edge computing gateway using an improved optimization algorithm, and uses the optimized adaptive dynamic threshold parameters for the next hierarchical fault detection and periodic baseline reset.

[0061] The deep fault detection unit uses the global fault detection model of the cloud computing center to perform deep fault detection on abnormal data subgraphs. Based on the obtained deep fault detection results, it generates intelligent maintenance decisions and sends them to the corresponding edge computing gateway.

[0062] A third aspect of this invention provides an electronic device, which includes:

[0063] At least one processor; and a memory communicatively connected to the at least one processor; wherein,

[0064] The memory stores instructions that can be executed by at least one processor, such that the at least one processor can perform the method proposed in the first aspect of the present invention.

[0065] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in the first aspect of the present invention. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention;

[0067] Figure 2 This is a flowchart of the steps of a belt conveyor fault detection method based on deep learning provided in an embodiment of the present invention;

[0068] Figure 3 This is a functional unit diagram of a belt conveyor fault detection device based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0069] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0070] The present invention will be further described below with reference to the accompanying drawings.

[0071] Reference Figure 1 , Figure 1This is a schematic diagram of the electronic device structure of the hardware operating environment involved in the embodiments of the present invention.

[0072] like Figure 1 As shown, the electronic device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk drive. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0073] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on the electronic device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0074] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating device, a data storage module, a network communication module, a user interface module, and electronic programs.

[0075] exist Figure 1 In the electronic device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the electronic device of the present invention can be set in the electronic device. The electronic device calls the deep learning-based belt conveyor fault detection device stored in the memory 1005 through the processor 1001 and executes the deep learning-based belt conveyor fault detection method provided in the embodiment of the present invention.

[0076] Reference Figure 2 The present invention provides a deep learning-based fault detection method for belt conveyors, the method comprising:

[0077] S201: Based on federated learning and knowledge distillation mechanisms, a global fault detection model is maintained in the cloud computing center, and a lightweight fault detection model is deployed in each edge computing gateway.

[0078] S202: Based on incremental update and dynamic threshold determination mechanism, a lightweight fault detection model of edge computing gateway is used to perform hierarchical fault detection on the distributed acoustic sensing (DAS) monitoring dataset of belt conveyor, extract abnormal data sub-graphs, and upload them to the cloud computing center.

[0079] S203: Using an improved optimization algorithm, optimize the adaptive dynamic threshold parameters of each edge computing gateway, use the optimized adaptive dynamic threshold parameters for the next hierarchical fault detection, and perform periodic baseline reset.

[0080] S204: Using the global fault detection model of the cloud computing center, perform deep fault detection on the abnormal data subgraph, generate intelligent maintenance decisions based on the obtained deep fault detection results, and send them to the corresponding edge computing gateway.

[0081] The technical solution provided in this application has at least the following beneficial effects:

[0082] By explicitly modeling the physical transport relationships between points using a graph structure, the fault detection model can learn global graph structure information and temporal information, fundamentally solving the spatial heterogeneity problem and achieving high consistency in detection accuracy. The hierarchical incremental update mechanism concentrates computational resources in suspicious areas, avoiding the huge overhead of high-frequency computation across the entire graph, and can meet the real-time processing requirements of high-frequency data while ensuring high detection accuracy. The combination of adaptive dynamic threshold and periodic baseline reset can adapt to long-term operating condition drift and has self-recovery capability, ensuring reliability for long-term unattended operation in harsh industrial environments. Compared with training a model independently for each point, this invention only requires a lightweight model and a global model, simplifying model management and reducing deployment and maintenance costs, thus achieving system-level optimization.

[0083] In one alternative implementation, the fault detection model is constructed based on the Graph Neural Network (GNN)-Long Short-Term Memory (LSTM)-Autoencoder (AE) algorithm, and the fault detection model includes a graph structure feature extraction module based on the GNN algorithm, a temporal feature extraction module based on the LSTM algorithm, and a fault detection module based on the AE algorithm. Both the graph structure feature extraction module and the temporal feature extraction module are connected to the fault detection module.

[0084] In this embodiment, the graph structure feature extraction module is used to process spatial structure data. In the belt conveyor scenario, GNN is responsible for understanding the physical connection relationships (edges) between various monitoring points ("nodes" in the graph). It can learn how a fault at one point affects its neighboring points, thereby capturing the spatial propagation and correlation features of the fault. The temporal feature extraction module is used to process time series data. The vibration signals collected by DAS are typical time series data. LSTM is good at learning long-term dependencies in data and can remember vibration patterns over a period of time, thereby identifying abnormal states that evolve over time, such as the development process of a fault from minor to severe. The fault detection module is used for unsupervised anomaly detection. An autoencoder is an unsupervised learning model that works by learning how to properly "compress" and "reconstruct" data. When normal features are input, it can reconstruct well; when abnormal features are input, the reconstruction effect will be poor, resulting in a large "reconstruction error". By monitoring the magnitude of this reconstruction error, it is possible to determine whether a fault has occurred.

[0085] In one alternative implementation, based on federated learning and knowledge distillation mechanisms, a global fault detection model is maintained at the cloud computing center, and a lightweight fault detection model is deployed at each edge computing gateway, including:

[0086] S2011: Using the GNN-LSTM-AE algorithm, build an initial teacher model in the cloud computing center (such as using the Graph Sampling and Aggregation (GraphSAGE) algorithm or the Graph Attention Network (GAT) structure), and use the GNN-LSTM-AE algorithm to build a lightweight initial student model (such as fewer GNN layers and fewer LSTM hidden units) at each edge computing gateway.

[0087] S2012: Based on the private sample set of each edge computing gateway, the initial student model is trained locally, the model update quantity of the obtained local student model is extracted, and it is encrypted and uploaded to the cloud computing center.

[0088] S2013: Based on the federated learning mechanism, the initial teacher model is updated according to the model update amount of all edge computing gateways to obtain the updated teacher model;

[0089] In this embodiment, large enterprises typically have multiple factories or multiple conveyor lines. Due to business competition or safety regulations, the monitoring data from each location is difficult to share centrally, forming "data silos." This limits the possibility of using global data to train high-performance models, and directly centralizing data poses a risk of privacy leakage.

[0090] To address the data silo problem, a federated learning framework is adopted. Edge computing nodes in each factory area or conveyor line independently train fault detection models using private sample sets. Only encrypted or abstracted model updates (rather than the original data) are uploaded to the cloud computing center for aggregation, generating a global model that incorporates global knowledge. This global model is then distributed to each edge computing gateway, achieving "the model moves while the data remains still." This significantly improves the model's generalization ability and robustness while protecting data privacy.

[0091] The formula is:

[0092]

[0093] In the formula, Number of iterations t+ 1, t Global model parameters; Edge computing gateway The amount of model parameter updates; t This represents the current iteration number; To the number of iterations t Edge computing gateway The importance of weight; For edge computing gateway indication; K Calculate the number of gateways at the edge;

[0094]

[0095] In the formula, An edge computing gateway generated based on the amount of sample data in a private sample set. The importance of weight; Edge computing gateway ,j The amount of sample data in the private sample set; j For edge computing gateway indication;

[0096] S2014: Based on the cloud sample set of the cloud computing center, perform global training on the updated teacher model to obtain the global teacher model, and extract the corresponding teacher model parameters and soft targets, and send them to all edge computing gateways.

[0097] In this embodiment, the cloud computing center uses a cloud sample set to perform multiple rounds of global training on the updated teacher model to further improve its performance. Then, the cloud computing center uses the teacher model to perform inference on a batch of public data (or simulated data) to obtain soft targets (such as the probability distribution of reconstruction error or intermediate layer features). The cloud computing center then distributes the teacher model parameters and soft targets to all edge gateways.

[0098] S2015: Based on the knowledge distillation mechanism, according to the teacher model parameters and soft objectives, at each edge computing gateway, the local student model is distilled and trained to obtain an updated student model, and the teacher model update process is re-executed.

[0099] In this embodiment, the global teacher model is reconstructed at each edge computing gateway based on the teacher model parameters, and together with the local student model, a teacher-student knowledge distillation architecture is formed, using its private sample set for distillation training.

[0100] The formula is:

[0101]

[0102] In the formula, This represents the total loss during distillation training; The standard reconstruction loss for private sample sets in a teacher-student knowledge distillation architecture; Output the KL divergence loss of the soft objective between the student model and the teacher model; This is the loss coefficient;

[0103] S2016: Repeat the above steps until the performance of the updated student model converges. Use the final student model of each edge computing gateway as a lightweight fault detection model and the final teacher model of the cloud computing center as the global fault detection model.

[0104] In one optional implementation, based on incremental updates and dynamic threshold determination mechanisms, a lightweight fault detection model of the edge computing gateway is used to perform hierarchical fault detection on the DAS monitoring dataset of the belt conveyor, extract abnormal data sub-graphs, and upload them to the cloud computing center, including:

[0105] S2021: Use the DAS monitoring network to collect the initial DAS monitoring dataset of the belt conveyor and upload it to the edge computing gateway within the communication range;

[0106] In this embodiment, DAS sensing optical cables are laid along the belt conveyor, the sampling frequency is set to 20kHz, the spatial resolution reaches 5 meters, a single conveyor line can monitor 1000 points, and a DAS monitoring network is constructed. The DAS monitoring data includes vibration signals of the monitoring points, etc.

[0107] S2022: Using an edge computing gateway, the initial DAS monitoring dataset is preprocessed to obtain a standard DAS monitoring dataset, which is then transformed into a DAS monitoring data graph.

[0108] S2023: Using a lightweight fault detection model, the reconstruction error of each node in the DAS monitoring data graph is extracted, and based on the dynamic threshold determination mechanism, an adaptive dynamic threshold for the edge computing gateway is generated according to the reconstruction error of all nodes.

[0109] The formula is:

[0110]

[0111] In the formula, Number of iterations t Adaptive dynamic threshold; Number of iterations t The mean of the reconstruction error of the standard DAS monitoring dataset within the sliding window; This is the sensitivity adjustment coefficient, and it is recommended to take a value between 2.0 and 2.5. It is one of the adaptive dynamic threshold parameters that need to be optimized later. Number of iterations t The standard deviation of the reconstruction error of the standard DAS monitoring dataset within the sliding window;

[0112] S2024: If the reconstruction error of any node exceeds the adaptive dynamic threshold, then take the abnormal node that exceeds the threshold as the center, extract the abnormal data subgraph including several of its neighboring nodes, and upload the abnormal data subgraph to the cloud computing center.

[0113] In one alternative implementation, an edge computing gateway is used to preprocess the initial DAS monitoring dataset to obtain a standard DAS monitoring dataset, which is then transformed into a DAS monitoring data graph, including:

[0114] S20221: Using an edge computing gateway, perform data cleaning, filtering, normalization, and sliding window segmentation on each initial DAS monitoring data in the initial DAS monitoring dataset to obtain several standard DAS monitoring datasets for each time window.

[0115] S20222: Treat each monitoring point in the DAS monitoring network as a node, and merge the corresponding standard DAS monitoring data and the spatial location information of the monitoring point into node features;

[0116] S20223: Define the edges between different monitoring points based on the physical connection relationship of the belt conveyor, and initialize the edge weights based on physical distance and / or signal correlation.

[0117] S20224: Based on all nodes, node features, and the edges and edge weights between nodes, perform graph transformation on the standard DAS monitoring dataset to obtain the DAS monitoring data graph.

[0118] In one alternative implementation, the improved optimization algorithm is the Enhanced Cuckoo Search Algorithm (ECSA).

[0119] In one alternative implementation, an improved optimization algorithm is used to optimize the adaptive dynamic threshold parameters of each edge computing gateway. The optimized adaptive dynamic threshold parameters are then used for the next hierarchical fault detection, and periodic baseline resets are performed, including:

[0120] S2031: Encode the adaptive dynamic threshold parameters of the edge computing gateway into individual vectors of the ECSA algorithm, and set the ECSA population parameters and the maximum number of iterations;

[0121] In this embodiment, the adaptive dynamic threshold parameters include sensitivity adjustment coefficient, adaptive dynamic threshold upper and lower limit parameters, smoothing coefficient, etc.

[0122] S2032: Based on the ECSA population parameters, the initial ECSA population is obtained by initializing using the Tent chaotic mapping sequence; each ECSA individual in the ECSA population corresponds to an alternative adaptive dynamic threshold parameter.

[0123] The formula is:

[0124]

[0125] In the formula, For the initial ECSA population, the first i The initial ECSA individual; For the first i One chaotic variable; These are the upper and lower bounds of the search space; i For ECSA individual indicators;

[0126]

[0127] In the formula, For the first i- One chaotic variable; compared with random initialization, chaotic initialization can ensure that the population is evenly distributed in the solution space, thus enhancing diversity.

[0128] S2033: Based on the private sample set, use the fitness function to obtain the fitness value of each initial ECSA individual, and take the initial ECSA individual with the best fitness value as the optimal solution;

[0129] In this embodiment, a dynamic threshold determination mechanism is set for the candidate adaptive dynamic threshold parameters of each initial ECSA individual, and a lightweight fault detection model is used to perform hierarchical fault detection based on a private sample set, and a fitness function is used to obtain the fitness value.

[0130] The formula is:

[0131]

[0132] In the formula, For ECSA individuals X fitness value; For ECSA individuals X The false alarm rate calculated on the private sample set; For ECSA individuals X The false negative rate calculated on the private sample set; The fitness coefficient;

[0133] S2034: Based on the Levy flight strategy and convergence factor, update the nest positions of all initial ECSA individuals except the optimal solution to obtain the updated first ECSA population;

[0134] The formula is:

[0135]

[0136] In the formula, Number of iterations t+ 1 of i The first updated ECSA individual (Bird's Nest); This is the convergence factor, which controls the strength of the approach to the optimal solution; Step size factor; for Levy Distributed random numbers; b for Levy Step length, and b ∈[1,2]; Number of iterations t The i One ECSA individual, which is the initial ECSA individual in the initial iteration;

[0137]

[0138] In the formula, These are the maximum and minimum values ​​of the convergence factor; This represents the maximum number of iterations. t This represents the current iteration number; , To adjust the parameters; It is the hyperbolic tangent function;

[0139] S2035: Randomly select several first ECSA individuals from the updated first ECSA population with reconstruction probability to rebuild nests, and obtain several updated second ECSA individuals;

[0140] In this embodiment, the reconstruction probability is used. Randomly select a subset of bird nests, and for each nest, generate a random number between [0,1]. ,if > If the nest is found, the nest is "discovered". For a discovered nest, a new nest is randomly generated in the search space. This simulates the behavior of the host bird abandoning the old nest and rebuilding a new nest after discovering foreign eggs.

[0141] The formula is:

[0142]

[0143] In the formula, Number of iterations t+ 1 of i A new second ECSA individual (Bird's Nest); A random number between [0, 1]; This is the disturbance range control factor;

[0144] S2036: According to the dynamic reverse mechanism, several first ECSA individuals are randomly selected from the updated first ECSA population with reverse probability to reverse, resulting in several updated third ECSA individuals.

[0145] In this embodiment, the reverse probability is used. A portion of bird nests are randomly selected for reverse learning to generate a random number between [0, 1]. ,if < If so, then reverse learning will be performed;

[0146] The formula is:

[0147]

[0148] In the formula, Number of iterations t+ The dynamic inverse of 1 yields the first i A newer third ECSA individual; Number of iterations t The i The reverse solution for each updated ECSA individual; The fitness function;

[0149]

[0150] In the formula, Number of iterations t The reverse center point determined by the dynamic boundary of the current search space; It is a dynamic inverse factor that increases with iteration (intensified inverse exploration in later stages).

[0151] S2037: Based on the private sample set, use the fitness function to obtain the fitness value of each updated first, second or third ECSA individual, and update the updated ECSA individual with the best fitness value as the optimal solution;

[0152] S2038: When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, terminate the iterative update of the ECSA population and output the optimal solution of the current iteration;

[0153] S2039: Decode the individual vector of the ECSA individual corresponding to the optimal solution to obtain the optimized adaptive dynamic threshold parameter, and use the optimized adaptive dynamic threshold parameter for the next hierarchical fault detection.

[0154] In this embodiment, before the next hierarchical fault detection is performed, the optimized adaptive dynamic threshold parameters are applied to the dynamic threshold determination mechanism to update the adaptive dynamic threshold.

[0155] By combining the historical steady-state sample mixing mechanism, 10% of the normal error samples from the historical stable operation period are selected and mixed into the current sliding window statistics of the private sample set. These historical samples come from the model training stage or the stable operation period. This mixing mechanism maintains the representativeness of the statistical distribution and prevents the random fluctuations of the current sample from affecting the threshold calculation.

[0156] Based on the optimized adaptive dynamic threshold parameters, the upper and lower limits of the adaptive dynamic threshold are set according to the dynamic protection mechanism of the upper and lower thresholds:

[0157] The formula is:

[0158]

[0159] In the formula, For adaptive dynamic threshold upper and lower limits; The mean and standard deviation during the training phase; This is the upper limit parameter for adaptive dynamic thresholds, which is one of the optimized adaptive dynamic threshold parameters, and generally does not exceed 3.5; It is an adaptive dynamic threshold lower limit parameter, which is one of the optimized adaptive dynamic threshold parameters, and generally does not exceed 1.5;

[0160] Standard deviation lower limit protection setting equal to 0.1 times When the error is too concentrated within the sliding window, causing the standard deviation to approach zero, this lower limit is used to prevent oversensitivity caused by an excessively small threshold.

[0161] The adaptive dynamic threshold is updated based on the exponential moving average smoothing mechanism, and the formula is as follows:

[0162]

[0163] In the formula, Number of iterations t+ 1. Smoothed and updated adaptive dynamic threshold; Number of iterations t+ 1. Updated adaptive dynamic threshold; This is the smoothing coefficient, one of the optimized adaptive dynamic threshold parameters, with a recommended value of 0.2. 0.3;

[0164] This mechanism effectively suppresses high-frequency fluctuations in the threshold and ensures the stability of threshold changes. Through the synergistic effect of these multiple steady-state control strategies, new technical effects are achieved. The threshold can adaptively adjust to changes in normal operating conditions within a range of ±20%. The coefficient of variation of threshold stability is controlled within 10% during long-term operation, and the sensitivity retention rate of anomaly detection exceeds 95%, effectively solving the drift problem of traditional dynamic threshold methods.

[0165] S20310: Perform periodic baseline reset for each edge computing gateway.

[0166] In this embodiment, a periodic baseline reset is performed every 24 hours, automatically using the standard DAS monitoring dataset from the most recent stable period to recalculate the baseline statistics in the dynamic threshold module. The system employs a 5-step incremental adjustment strategy for updates, with each step adjusting 20% ​​of the total adjustment. Parameter updates are completed gradually over 5 time steps to avoid the impact of threshold mutations on the stability of the edge computing gateway. Simultaneously, the performance of the lightweight fault detection model is monitored (e.g., by reconstructing the KL divergence of the error distribution). When performance degradation is detected, the lightweight fault detection model is automatically incrementally learned using new normal data to adapt to the slow changes in device status and environment.

[0167] In one optional implementation, a global fault detection model from the cloud computing center is used to perform deep fault detection on the abnormal data subgraph. Based on the obtained deep fault detection results, an intelligent maintenance decision is generated and sent to the corresponding edge computing gateway, including:

[0168] S2041: Input the abnormal data subgraph into the global fault detection model of the cloud computing center;

[0169] In this embodiment, a global fault detection model in the cloud with more abundant computing resources is invoked to analyze the abnormal data subgraph and output more accurate fault types (such as idler roller faults, belt misalignment), locations, severity, and confidence levels.

[0170] S2042: Use the graph structure feature extraction module of the global fault detection model to extract the graph structure features of the abnormal data subgraph;

[0171] S2043: Use the temporal feature extraction module of the global fault detection model to extract the temporal features of node features in the abnormal data subgraph;

[0172] S2044: Based on the graph structure features and temporal features, the fault detection module of the global fault detection model is used to call the softmax function to perform deep fault detection and obtain the deep fault detection results.

[0173] S2045: Extract the operating status of the belt conveyor corresponding to the abnormal data subgraph, and input the operating status of the belt conveyor and the deep fault detection results into the pre-built intelligent maintenance decision model;

[0174] S2046: Based on the operating status of the belt conveyor and the results of deep fault detection, use the intelligent maintenance decision model to generate intelligent maintenance decisions, obtain intelligent maintenance decisions, and send them to the corresponding edge computing gateway.

[0175] In one alternative implementation, the intelligent maintenance decision model is built based on the Proximal Policy Optimization (PPO) algorithm, and the intelligent maintenance decision model is equipped with an agent and an experience replay pool.

[0176] In one optional implementation, based on the operating status of the belt conveyor and the results of deep fault detection, an intelligent maintenance decision model is used to generate an intelligent maintenance decision, which is then sent to the corresponding edge computing gateway, including:

[0177] S20461: Input the deep fault detection results (type, location, severity, and confidence level) and the current operating status of the belt conveyor (load, speed, etc.) into the state space of the intelligent agent of the intelligent maintenance decision model;

[0178] Status space: includes fault type (e.g., idler failure), location (e.g., at 2350 meters), severity (e.g., medium), confidence level (e.g., 80%), belt conveyor operating status (load, speed, etc.), spare parts inventory (available / unavailable), location and status of the nearest maintenance personnel, current production plan (e.g., whether it is during peak period), etc.

[0179] S20462: Randomly extract several historical experiences from the experience replay pool, and write the action set of several historical experiences into the action space of the intelligent agent of the intelligent maintenance decision model.

[0180] Action space: including "immediately dispatch the nearest maintenance personnel to replace it", "wait for the next planned downtime window to handle it together", "order spare parts and then repair it at an opportune time", "only record the monitoring and do not handle it for the time being", etc.

[0181] S20463: The intelligent agent selects an action in the action space based on the current state of the state space and its policy network to obtain an intelligent maintenance decision, which is then sent to the corresponding edge computing gateway.

[0182] In this embodiment, the formula for the reward function of the intelligent maintenance decision-making model is:

[0183]

[0184] In the formula, Number of iterations t The reward value; For maintenance costs, downtime losses, and production losses; Weights for the reward function;

[0185] In this embodiment, the model can learn the optimal maintenance strategy that maximizes long-term returns (taking into account maintenance costs, downtime, production losses, etc.), thus achieving a leap from "passive response" to "proactive prediction and intelligent decision-making".

[0186] S20464: Collect real-time experience generated by intelligent maintenance decisions, and update the intelligent maintenance decision model based on real-time experience and some historical experience to obtain an updated intelligent maintenance decision model.

[0187] This invention also provides a deep learning-based fault detection device for belt conveyors, referring to... Figure 3 The diagram shows a functional unit diagram of a deep learning-based belt conveyor fault detection device 300 according to the present invention. The device may include the following units:

[0188] Model building unit 301 is used to maintain a global fault detection model in the cloud computing center based on federated learning and knowledge distillation mechanisms, and to deploy a lightweight fault detection model in each edge computing gateway.

[0189] The hierarchical fault detection unit 302 is used to perform hierarchical fault detection on the DAS monitoring dataset of the belt conveyor based on the incremental update and dynamic threshold determination mechanism and the lightweight fault detection model of the edge computing gateway, extract abnormal data sub-graphs, and upload them to the cloud computing center.

[0190] The threshold parameter optimization unit 303 is used to optimize the adaptive dynamic threshold parameters of each edge computing gateway using an improved optimization algorithm, and to use the optimized adaptive dynamic threshold parameters for the next hierarchical fault detection and periodic baseline reset.

[0191] The deep fault detection unit 304 is used to perform deep fault detection on the abnormal data subgraph using the global fault detection model of the cloud computing center, generate intelligent maintenance decisions based on the obtained deep fault detection results, and send them to the corresponding edge computing gateway.

[0192] Based on the same inventive concept, another embodiment of the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0193] Memory, used to store computer programs;

[0194] The processor, when executing the program stored in the memory, implements the deep learning-based belt conveyor fault detection method of the present invention.

[0195] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus. The communication interface is used for communication between the aforementioned terminal and other devices. The memory can include Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0196] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0197] Furthermore, to achieve the above objectives, embodiments of the present invention also propose a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the deep learning-based belt conveyor fault detection method of the present invention.

[0198] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable hardware devices (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0199] The embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0200] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0201] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0202] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. "And / or" indicates that either one or both can be chosen. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes the element.

[0203] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A fault detection method for belt conveyors based on deep learning, characterized in that, The method includes: Based on federated learning and knowledge distillation mechanisms, a global fault detection model is maintained in the cloud computing center, and a lightweight fault detection model is deployed in each edge computing gateway. The fault detection model is constructed based on the GNN-LSTM-AE algorithm, and includes a graph structure feature extraction module constructed based on the GNN algorithm, a temporal feature extraction module constructed based on the LSTM algorithm, and a fault detection module constructed based on the AE algorithm. The graph structure feature extraction module and the temporal feature extraction module are both connected to the fault detection module. include: The initial teacher model is built in the cloud computing center using the GNN-LSTM-AE algorithm, and a lightweight initial student model is built at each edge computing gateway using the GNN-LSTM-AE algorithm. Based on the private sample set of each edge computing gateway, the initial student model is trained locally, the model update amount of the obtained local student model is extracted, and it is encrypted and uploaded to the cloud computing center. Based on the federated learning mechanism, the initial teacher model is updated according to the model update amount of all edge computing gateways to obtain the updated teacher model. Based on the cloud sample set in the cloud computing center, the updated teacher model is globally trained to obtain a global teacher model, and the corresponding teacher model parameters and soft targets are extracted and sent to all edge computing gateways. Based on the knowledge distillation mechanism, according to the teacher model parameters and soft objectives, at each edge computing gateway, the local student model is distilled and trained to obtain an updated student model, and the teacher model update process is re-executed. Repeat the above steps until the performance of the updated student model converges. Use the final student model of each edge computing gateway as a lightweight fault detection model and the final teacher model of the cloud computing center as a global fault detection model. Based on incremental update and dynamic threshold determination mechanism, a lightweight fault detection model of edge computing gateway is used to perform hierarchical fault detection on DAS monitoring dataset of belt conveyor, extract abnormal data sub-graphs, and upload them to cloud computing center. An improved optimization algorithm is used to optimize the adaptive dynamic threshold parameters of each edge computing gateway. The optimized adaptive dynamic threshold parameters are then used for the next hierarchical fault detection and periodic baseline reset is performed. Using the global fault detection model of the cloud computing center, deep fault detection is performed on the abnormal data subgraph. Based on the obtained deep fault detection results, intelligent maintenance decisions are generated and sent to the corresponding edge computing gateway.

2. The deep learning-based fault detection method for belt conveyors according to claim 1, characterized in that, Based on incremental updates and dynamic threshold determination mechanisms, a lightweight fault detection model using an edge computing gateway is employed to perform hierarchical fault detection on the DAS monitoring dataset of a belt conveyor, extracting abnormal data sub-graphs and uploading them to the cloud computing center, including: Using the DAS monitoring network, the initial DAS monitoring dataset of the belt conveyor is collected and uploaded to the edge computing gateway within the communication range; Using an edge computing gateway, the initial DAS monitoring dataset is preprocessed to obtain a standard DAS monitoring dataset, which is then transformed into a DAS monitoring data graph. Using a lightweight fault detection model, the reconstruction error of each node in the DAS monitoring data graph is extracted, and an adaptive dynamic threshold for the edge computing gateway is generated based on the reconstruction error of all nodes according to the dynamic threshold determination mechanism. If the reconstruction error of any node exceeds the adaptive dynamic threshold, then an abnormal data subgraph, including several of its neighboring nodes, is extracted centered on the abnormal node that exceeds the threshold, and the abnormal data subgraph is uploaded to the cloud computing center.

3. The deep learning-based fault detection method for belt conveyors according to claim 2, characterized in that, Using an edge computing gateway, the initial DAS monitoring dataset is preprocessed to obtain a standard DAS monitoring dataset, which is then transformed into a DAS monitoring data graph, including: Using an edge computing gateway, each initial DAS monitoring data in the initial DAS monitoring dataset is cleaned, filtered, normalized, and segmented by a sliding window to obtain several standard DAS monitoring datasets for each time window. Each monitoring point in the DAS monitoring network is taken as a node, and the corresponding standard DAS monitoring data and the spatial location information of the monitoring point are combined into node features. Based on the physical connection relationship of the belt conveyor, define the edges between different monitoring points, and initialize the edge weights according to the physical distance and / or signal correlation. Based on all nodes, node features, and the edges and edge weights between nodes, a graph transformation is performed on the standard DAS monitoring dataset to obtain the DAS monitoring data graph.

4. The deep learning-based fault detection method for belt conveyors according to claim 3, characterized in that, The improved and optimized algorithm is the ECSA algorithm.

5. The deep learning-based fault detection method for belt conveyors according to claim 4, characterized in that, An improved optimization algorithm is used to optimize the adaptive dynamic threshold parameters of each edge computing gateway. The optimized adaptive dynamic threshold parameters are then used for the next hierarchical fault detection, and periodic baseline resets are performed, including: The adaptive dynamic threshold parameters of the edge computing gateway are encoded into individual vectors of the ECSA algorithm, and the ECSA population parameters and maximum number of iterations are set. Based on the ECSA population parameters, the initial ECSA population is obtained by initialization using the Tent chaotic mapping sequence; each ECSA individual in the ECSA population corresponds to an alternative adaptive dynamic threshold parameter. Based on the private sample set, the fitness function is used to obtain the fitness value of each initial ECSA individual, and the initial ECSA individual with the best fitness value is taken as the optimal solution. Based on the Levy flight strategy and convergence factor, the nest positions of all initial ECSA individuals except the optimal solution are updated to obtain the updated first ECSA population. Based on the reconstruction probability, several first ECSA individuals are randomly selected from the updated first ECSA population to rebuild nests, resulting in several updated second ECSA individuals. According to the dynamic reverse mechanism, several first ECSA individuals are randomly selected from the updated first ECSA population with reverse probability to reverse, resulting in several updated third ECSA individuals. Based on the private sample set, the fitness function is used to obtain the fitness value of each updated first, second or third ECSA individual, and the updated ECSA individual with the best fitness value is updated as the optimal solution; When the number of iterations reaches the maximum number of iterations or the fitness value of the optimal solution meets the requirements, the iterative update of the ECSA population is terminated, and the optimal solution of the current iteration is output. The individual vector of the ECSA individual corresponding to the optimal solution is decoded to obtain the optimized adaptive dynamic threshold parameter, and the optimized adaptive dynamic threshold parameter is used for the next hierarchical fault detection. Perform periodic baseline resets for each edge computing gateway.

6. The deep learning-based fault detection method for belt conveyors according to claim 5, characterized in that, Using a global fault detection model from the cloud computing center, deep fault detection is performed on the abnormal data subgraph. Based on the obtained deep fault detection results, intelligent maintenance decisions are generated and sent to the corresponding edge computing gateway, including: Input the abnormal data subgraph into the global fault detection model of the cloud computing center; The graph structure feature extraction module of the global fault detection model is used to extract the graph structure features of the abnormal data subgraphs. The temporal feature extraction module of the global fault detection model is used to extract the temporal features of the node features in the abnormal data subgraph. Based on the graph structure features and temporal features, the fault detection module of the global fault detection model is used to call the softmax function to perform deep fault detection and obtain the deep fault detection results. Extract the belt conveyor operating status corresponding to the abnormal data subgraph, and input the belt conveyor operating status and deep fault detection results into the pre-built intelligent maintenance decision model; Based on the operating status of the belt conveyor and the results of deep fault detection, an intelligent maintenance decision model is used to generate intelligent maintenance decisions, which are then sent to the corresponding edge computing gateway.

7. The deep learning-based fault detection method for belt conveyors according to claim 6, characterized in that, The intelligent maintenance decision-making model is built based on the PPO algorithm, and the intelligent maintenance decision-making model is equipped with an intelligent agent and an experience playback pool.

8. A deep learning-based belt conveyor fault detection device, used to implement the belt conveyor fault detection method as described in any one of claims 1-7, characterized in that, The device includes: The model building unit is used to maintain a global fault detection model in the cloud computing center based on federated learning and knowledge distillation mechanisms, and to deploy a lightweight fault detection model on each edge computing gateway. The hierarchical fault detection unit is used to perform hierarchical fault detection on the DAS monitoring dataset of the belt conveyor based on incremental update and dynamic threshold determination mechanism, using a lightweight fault detection model of edge computing gateway, extracting abnormal data sub-graphs, and uploading them to the cloud computing center. The threshold parameter optimization unit is used to optimize the adaptive dynamic threshold parameters of each edge computing gateway using an improved optimization algorithm, and uses the optimized adaptive dynamic threshold parameters for the next hierarchical fault detection and periodic baseline reset. The deep fault detection unit uses the global fault detection model of the cloud computing center to perform deep fault detection on abnormal data subgraphs. Based on the obtained deep fault detection results, it generates intelligent maintenance decisions and sends them to the corresponding edge computing gateway.

Images