An intelligent fault diagnosis method and system for explosion-proof motors in the natural gas industry

Abstract

The present invention discloses an intelligent fault diagnosis method and system for explosion-proof motors in the natural gas industry, the method comprising the following steps: S10: acquiring multi-source heterogeneous data of explosion-proof motors; S20: performing enhancement processing on the multi-source heterogeneous data of explosion-proof motors to obtain an enhanced multimodal data set of explosion-proof motors; S30: performing feature extraction on the enhanced multimodal data set of explosion-proof motors using a deep learning model to obtain a high-dimensional feature vector; S40: based on the high-dimensional feature vector, a multi-task learning network is used to perform fault classification and fault severity evaluation to obtain a fault type probability distribution and a fault severity index; S50: generating an intelligent diagnosis report of the explosion-proof motor including maintenance suggestions according to the fault type probability distribution and the fault severity index. The method can make full use of multi-source heterogeneous data and realize an explosion-proof motor fault diagnosis method of intelligent diagnosis, so as to improve the diagnostic efficiency and accuracy and realize preventive maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment maintenance in the natural gas industry, and in particular to an intelligent fault diagnosis method and system for explosion-proof motors in the natural gas industry. Background Art

[0002] As the core equipment of the natural gas industry, the safe and reliable operation of explosion-proof motors is directly related to the stability and safety of the entire production system. However, the existing fault diagnosis methods for explosion-proof motors mainly rely on manual experience judgment and regular maintenance, which has many shortcomings.

[0003] First, manual diagnosis is inefficient and it is difficult to detect potential faults in a timely manner, resulting in an increased risk of equipment failure. Second, the accuracy of diagnosis is limited by personnel experience and subjective factors, which can easily lead to misjudgments or missed judgments, affecting the accuracy of maintenance decisions. Third, traditional methods fail to make full use of multi-source heterogeneous data, such as vibration, temperature, current and other multi-dimensional information, resulting in one-sided diagnostic results that cannot fully reflect the status of the equipment. In addition, existing technologies lack intelligent means and cannot achieve automatic classification and severity assessment of faults, making it difficult to provide timely and targeted maintenance recommendations. Finally, due to the lack of a comprehensive grasp of the operating status of the equipment, preventive maintenance strategies are difficult to formulate and implement, often leading to over-maintenance or under-maintenance of equipment, increasing maintenance costs and downtime risks.

[0004] The limitations of traditional fault diagnosis methods are becoming increasingly prominent. There is an urgent need for an explosion-proof motor fault diagnosis method that can make full use of multi-source heterogeneous data and realize intelligent diagnosis, so as to improve diagnostic efficiency and accuracy, achieve preventive maintenance, and thus ensure safe and efficient production in the natural gas industry. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides an intelligent fault diagnosis method and system for explosion-proof motors in the natural gas industry to solve the above technical problems.

[0006] To achieve the above objectives, in a first aspect, a method for intelligent fault diagnosis of explosion-proof motors in the natural gas industry is provided, which comprises the following steps:

[0007] Acquire multi-source heterogeneous data of explosion-proof motors;

[0008] Performing enhancement processing on the multi-source heterogeneous data of the explosion-proof motor to obtain an enhanced multimodal data set of the explosion-proof motor;

[0009] Using a deep learning model to extract features from the enhanced multimodal data set of the explosion-proof motor to obtain a high-dimensional feature vector;

[0010] Based on the high-dimensional feature vector, a multi-task learning network is used to perform fault classification and fault severity assessment to obtain a fault type probability distribution and a fault severity index;

[0011] An intelligent diagnosis report including maintenance suggestions for the explosion-proof motor is generated according to the fault type probability distribution and the fault severity index.

[0012] In a second aspect, an intelligent fault diagnosis system for explosion-proof motors in the natural gas industry is provided, which includes:

[0013] Multiple sensors are installed on explosion-proof motors to collect the operating data of explosion-proof motors in real time;

[0014] A data acquisition device, connected to the plurality of sensors, for receiving and preprocessing the operation data collected by the plurality of sensors to obtain preprocessed multi-source heterogeneous data;

[0015] An edge computing device, connected to the data acquisition device, is used to perform preliminary analysis and screening on the preprocessed multi-source heterogeneous data to obtain the preliminarily analyzed multi-source heterogeneous data;

[0016] A remote server, connected to the edge computing device via a network, for receiving multi-source heterogeneous data that has been preliminarily analyzed and performing fault diagnosis;

[0017] The remote server comprises:

[0018] A data acquisition module, used to acquire multi-source heterogeneous data transmitted from edge computing devices and subjected to preliminary analysis;

[0019] A data enhancement module, used for performing enhancement processing on the multi-source heterogeneous data after preliminary analysis to obtain an enhanced multi-modal data set of the explosion-proof motor;

[0020] A feature extraction module, used for extracting features from the enhanced multimodal data set of the explosion-proof motor using a deep learning model to obtain a high-dimensional feature vector;

[0021] A multi-task learning module, used to perform fault classification and fault severity assessment using a multi-task learning network based on the high-dimensional feature vector to obtain a fault type probability distribution and a fault severity index;

[0022] The diagnostic report generating module is used to generate an intelligent diagnostic report of the explosion-proof motor including maintenance suggestions according to the fault type probability distribution and the fault severity index.

[0023] Furthermore, it also includes a terminal device connected to the remote server via a network, and used for receiving and displaying the intelligent diagnosis report.

[0024] The above technical solution has the following beneficial technical effects:

[0025] First, the method realizes the complementarity and synergy between data through the fusion and enhanced processing of multi-source heterogeneous data, breaks through the limitations of the traditional single data source, and makes the diagnosis results more comprehensive and reliable. Secondly, the deep learning model adopted by the present invention can adaptively learn and extract high-dimensional features without artificially designed features, which not only greatly improves the efficiency and accuracy of feature extraction, but also is a capability that traditional machine learning methods do not have. Furthermore, the present invention introduces multi-task learning into the field of fault diagnosis, and realizes fault classification and severity assessment at the same time, which not only improves the generalization ability of the model, but also realizes the collaborative optimization of diagnostic tasks. In addition, the method of the present invention has the ability of self-learning and continuous optimization. As the amount of data increases, the diagnostic performance will continue to improve. This characteristic of becoming more intelligent with use is not possessed by traditional static diagnostic methods. Finally, the intelligent diagnostic report generated by the present invention not only provides fault information, but also includes personalized maintenance suggestions based on the historical data and current status of the equipment. This ability of active prediction and precise maintenance far exceeds the scope of traditional passive response maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings are used to better understand the present invention and do not constitute an improper limitation of the present invention.

[0027] Figure 1 It is a flow chart of an intelligent fault diagnosis method for explosion-proof motors in the natural gas industry according to an embodiment of the present invention;

[0028] Figure 2 is a flowchart of step S20 of an embodiment of the present invention;

[0029] Figure 3 is a flowchart of step S30 of an embodiment of the present invention;

[0030] Figure 4 is a flowchart of step S40 of an embodiment of the present invention;

[0031] Figure 5 is a flowchart of step S50 of an embodiment of the present invention;

[0032] Figure 6 is a flowchart of step S60 of an embodiment of the present invention;

[0033] Figure 7 is a flowchart of step S70 of an embodiment of the present invention;

[0034] Figure 8 It is a functional block diagram of an intelligent fault diagnosis system for explosion-proof motors in the natural gas industry according to an embodiment of the present invention;

[0035] Fig. 9 It is a specific functional block diagram of an intelligent fault diagnosis system for explosion-proof motors in the natural gas industry according to an embodiment of the present invention;

[0036] Fig.10 It is a schematic diagram of a simplified fault evolution tree of an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0038] The embodiment of the present invention provides an intelligent fault diagnosis method and system for explosion-proof motors in the natural gas industry. By acquiring and enhancing multi-source heterogeneous data, using a deep learning model for feature extraction, and using a multi-task learning network for fault classification and severity assessment, an intelligent diagnosis report containing maintenance recommendations is finally generated. This method is conducive to improving diagnostic efficiency, improving diagnostic accuracy, improving data utilization, realizing intelligent diagnosis, and strengthening preventive maintenance capabilities, thereby comprehensively improving the operation and maintenance level of explosion-proof motors in the natural gas industry. Embodiment 1

[0039] like Figure 1 As shown, this embodiment provides an intelligent fault diagnosis method for explosion-proof motors in the natural gas industry, which is executed by a remote server and includes the following steps:

[0040] S10: Acquire multi-source heterogeneous data of explosion-proof motors.

[0041] Multi-source heterogeneous data refers to various data types from different sensors in different formats. In this embodiment, these data include but are not limited to:

[0042] Vibration data is obtained by using piezoelectric accelerometers. Piezoelectric accelerometers are installed on the bearing seat and housing of the explosion-proof motor. Three piezoelectric accelerometers are installed on each bearing seat to measure the radial vibration (X-axis and Y-axis) and axial vibration (Z-axis) of the bearing. The piezoelectric accelerometer is directly connected to the data acquisition device through a shielded cable.

[0043] The current data is obtained by using Hall effect current sensor. The Hall effect current sensor is installed on each phase line of the explosion-proof motor input end and is directly connected to the data acquisition device through an analog signal line.

[0044] Temperature data is obtained by using thermocouples and thermal resistors. Thermocouples and thermal resistors are installed on the stator windings, bearings and housing surfaces of explosion-proof motors respectively. These sensors are directly connected to the data acquisition equipment through compensation wires.

[0045] The sound data is obtained by using an explosion-proof electret microphone array. The explosion-proof electret microphone array is arranged around the explosion-proof motor and is directly connected to the data acquisition device through an audio cable.

[0046] The magnetic flux data is obtained by using a Hall effect magnetic flux sensor. The Hall effect magnetic flux sensor is installed at the stator core slot of the explosion-proof motor. The Hall effect magnetic flux sensor is directly connected to the data acquisition device through a shielded twisted pair cable.

[0047] The oil analysis data is obtained by using an online oil analyzer. The online oil analyzer is installed in the explosion-proof motor bearing lubricating oil circulation system and is directly connected to the data acquisition device through a data cable.

[0048] Infrared thermal imaging data is obtained by using an infrared thermal imager, such as a fixed infrared thermal imager. The fixed infrared thermal imager is installed at a suitable location around the explosion-proof motor and is directly connected to the data acquisition device via an Ethernet cable.

[0049] Gas concentration data is obtained by multi-gas detectors. The multi-gas detectors are installed in the confined space around the explosion-proof motor and are directly connected to the data acquisition device via a serial bus.

[0050] Load torque data is obtained by using a torque sensor. The torque sensor is installed between the output shaft of the explosion-proof motor and the load, and is directly connected to the data acquisition device through a shielded twisted pair or a multi-core shielded cable.

[0051] Combined with reference Figure 8 , all these sensors are directly connected to the central data acquisition device to form a unified data acquisition system. The data acquisition device is responsible for synchronously collecting data from all sensors and transmitting this data to the edge computing device for preliminary processing and analysis. As the next-level processing unit of the data acquisition device, the edge computing device is responsible for preliminary processing and analysis of the collected raw data in preparation for subsequent in-depth analysis.

[0052] Distributed sensor network refers to a network system formed by arranging the above-mentioned sensors in key parts of explosion-proof motors in a decentralized manner and interconnecting them through fieldbus (such as Profibus-DP or Modbus) or industrial Ethernet (such as EtherCAT). This network structure can realize multi-point synchronous acquisition, improve the comprehensiveness and reliability of data, reduce wiring complexity, and improve the flexibility and scalability of the system.

[0053] Edge computing devices refer to small computing units deployed near the data source, such as industrial embedded computers or programmable logic controllers (PLCs). For example, an explosion-proof industrial computer equipped with an Intel Core i5 processor, 8GB RAM, and a 256GB SSD can be used as an edge computing device. These devices perform preliminary processing and compression on the collected multi-source heterogeneous data, including the following steps: First, data filtering is performed to remove obvious noise and outliers to ensure the quality of the data. Then, according to the characteristics of different data types, the data is downsampled and the sampling rate is appropriately reduced to reduce data redundancy. Next, the basic statistical features of the data in the time domain and frequency domain are extracted for further analysis and application. Finally, the data is compressed using a lossless or low-loss compression algorithm to reduce the data volume and ensure the efficiency of data transmission and storage. The purpose of this step is to reduce the amount of data transmitted to the remote server and reduce the network bandwidth requirements while retaining the key information of the data. After processing, the data packets that have been preliminarily cleaned, compressed, and feature extracted are obtained. These data packets contain the main features of the original data, but the volume is significantly reduced.

[0054] Finally, the processed multi-source heterogeneous data is transmitted to a remote server via a wide area network (such as a 4G / 5G mobile network or a dedicated optical fiber network) for further in-depth analysis and diagnosis. This type of data collection can effectively reduce data transmission and storage costs while ensuring data quality.

[0055] S20: performing enhancement processing on the multi-source heterogeneous data of the explosion-proof motor to obtain an enhanced multimodal data set of the explosion-proof motor.

[0056] Enhancement processing includes a series of data processing and enhancement techniques that help eliminate noise, standardize data, synchronize different data sources, and increase data diversity. This process involves signal noise reduction, data standardization, time synchronization, and data enhancement. Through these processes, the original multi-source heterogeneous data is converted into a high-quality, unified format enhanced multimodal dataset.

[0057] S30: Using a deep learning model to perform feature extraction on the enhanced multimodal data set of the explosion-proof motor to obtain a high-dimensional feature vector.

[0058] Deep learning models refer to complex neural network structures that can automatically learn data features, including but not limited to convolutional neural networks, recurrent neural networks, etc. These models can process multimodal data and extract temporal and spatial features. High-dimensional feature vectors are compact and information-rich data representations of model outputs, which capture key features and patterns in the original data and provide strong support for subsequent fault diagnosis.

[0059] S40: Based on the high-dimensional feature vector, a multi-task learning network is used to perform fault classification and fault severity assessment to obtain a fault type probability distribution and a fault severity index.

[0060] A multi-task learning network is a neural network structure that can perform multiple related tasks simultaneously. In this case, it mainly performs two tasks: fault classification and severity assessment. The fault type probability distribution represents the probability of various possible fault types, and the fault severity index is a continuous value that quantifies the severity of the fault. This multi-task learning method can make full use of the correlation between tasks and improve the accuracy and comprehensiveness of diagnosis.

[0061] S50: Generate an intelligent diagnosis report including maintenance suggestions for the explosion-proof motor according to the fault type probability distribution and the fault severity index.

[0062] The smart diagnosis report is a comprehensive document that not only contains the fault diagnosis results, but also provides targeted maintenance recommendations. The content of the smart diagnosis report can include the identified fault type and its probability, fault severity assessment, fault description and potential impact, and specific maintenance recommendations. These recommendations are generated based on the fault diagnosis results and preset maintenance rules, providing maintenance personnel with a clear and practical guide to action. Embodiment 2

[0063] like Figure 2 As shown, in this embodiment, step S20 includes the following sub-steps:

[0064] In step S21, an adaptive threshold denoising algorithm based on wavelet packet transform is used to denoise the multi-source heterogeneous data of explosion-proof motors to obtain denoised multi-source heterogeneous data. The algorithm first applies wavelet packet transform to each type of data to decompose the signal into sub-bands of different frequencies and time scales, and then applies an adaptive threshold to each sub-band for denoising. The threshold calculation formula is: , where T is the threshold, σ is the estimated value of the subband noise standard deviation, and N is the signal length. Finally, the signal is reconstructed through soft threshold processing and inverse wavelet packet transform to obtain the denoised data. The specific processing process is as follows: First, soft threshold processing is performed on each subband coefficient after wavelet packet decomposition. The soft threshold function is defined as: , where x is the subband coefficient, T is the threshold, () +Indicates taking the positive part. This process can effectively suppress noise while retaining the main features of the signal. After processing, coefficients less than the threshold are set to zero, and coefficients greater than the threshold are reduced. Then, the processed coefficients are inversely transformed by wavelet packet transform, the information of each sub-band is recombined, and the time domain signal is reconstructed. The reconstructed signal obtained in this way is the denoised data, which retains the main features of the original signal while effectively removing the noise component.

[0065] In step S22, the denoised multi-source heterogeneous data is uniformly standardized using the standard score method to obtain a standardized data matrix. For each type of data x, its mean μ and standard deviation σ are calculated, and then standardized using the formula z = (x - μ) / σ. After this processing, different types of data are converted to the same scale with a mean of 0 and a standard deviation of 1, forming a standardized data matrix.

[0066] In step S23, the dynamic time warping (DTW) algorithm is used to synchronize the different data sources in the standardized data matrix to obtain a time-aligned multimodal data sequence. The DTW algorithm achieves sequence alignment of different lengths or sampling rates by constructing a distance matrix and a cumulative distance matrix and backtracking the optimal path. The distance matrix is ​​a two-dimensional matrix, in which each element represents the distance between corresponding points in two sequences (using Euclidean distance). The cumulative distance matrix is ​​constructed based on the distance matrix, and each of its elements represents the minimum cumulative distance from the starting point to the current point. The construction process starts from the lower left corner and gradually fills the matrix. The value of each element is the current distance plus the minimum value of the three adjacent elements on the left, below, and below the left. Backtracking the optimal path means starting from the upper right corner of the cumulative distance matrix and moving toward the lower left corner in the direction of the minimum cumulative distance to find the best correspondence between the two sequences. This path defines how to align the two sequences. DTW alignment is performed on all data sources, and finally a time-aligned multimodal data sequence is obtained.

[0067] In step S24, based on the time-aligned multimodal data sequence, data enhancement is implemented through a conditional generative adversarial network (CGAN) to generate simulated data samples. CGAN consists of a generator G and a discriminator D. The task of the generator G is to generate false data. It receives random noise and conditional information (such as equipment working conditions) as input and outputs generated data samples. The task of the discriminator D is to distinguish between real data and generated false data. It receives data samples and conditional information as input and outputs a scalar indicating the degree of authenticity. By alternately training the generator G and the discriminator D, the entire network is gradually optimized. Specifically, when training the discriminator D, real data and generated false data are used as input, and the goal is to maximize the predicted value of real data and minimize the predicted value of false data. When training the generator G, the goal is to maximize the predicted value of the discriminator D for the generated data, that is, to deceive the discriminator D into judging the generated false data as real data. This process is repeated until the generator G can generate simulated data samples that are similar to the real data in statistical characteristics. After training is completed, the generator G is used to generate a large number of simulated data samples to achieve data enhancement.

[0068] In step S25, the simulated data samples generated by CGAN are merged with the original time-aligned multimodal data sequence to finally obtain an enhanced multimodal dataset of the explosion-proof motor. This dataset contains not only the original real data, but also a large amount of simulated data, which greatly expands the number and diversity of training samples, provides a solid foundation for subsequent feature extraction and fault diagnosis model training, and helps to improve the generalization ability and diagnostic accuracy of the model.

[0069] By adopting the adaptive threshold denoising algorithm based on wavelet packet transform, the noise in multi-source heterogeneous data can be effectively removed while retaining the important features of the signal. Compared with the traditional fixed threshold denoising algorithm, this method can better adapt to the noise characteristics of different types of data, improve the signal-to-noise ratio of the data, and lay a good foundation for subsequent analysis.

[0070] The standard score method is used to perform unified standardization and convert data of different types and dimensions to the same scale. This not only solves the problem of inconsistent scales of multi-source heterogeneous data, but also eliminates the dimensional differences between different data sources, making subsequent analysis and modeling more accurate and reliable.

[0071] Using the dynamic time warping (DTW) algorithm for time synchronization can effectively handle the time offset and sampling rate differences between different data sources, and achieve accurate alignment of multimodal data. Compared with simple linear interpolation or resampling methods, DTW can better preserve the timing characteristics of the data and improve the accuracy of multimodal data fusion.

[0072] This embodiment achieves data enhancement through conditional generative adversarial networks (CGANs), which can not only generate a large number of simulated data samples, but also ensure the similarity of the generated data with the original data in statistical characteristics. This method significantly expands the number and diversity of training samples, especially for situations where fault samples are scarce, and can effectively improve the generalization ability and diagnostic accuracy of the model.

[0073] This multi-stage processing approach not only solves complex problems that are difficult to deal with with a single technology, but also produces synergistic effects. For example, the denoising process improves the effect of subsequent normalization and time alignment, while time alignment provides a basis for CGAN to generate more accurate simulation samples. Embodiment 3

[0074] like Figure 3 As shown, in this embodiment, step S30 specifically includes the following steps: S31: using a time convolution network enhanced by an attention mechanism to process the time series data in the enhanced multimodal data set of the explosion-proof motor and extract time series features; S32: using a graph convolutional neural network to process the data reflecting the spatial relationship between sensors in the enhanced multimodal data set of the explosion-proof motor and extract spatial features; S33: fusing the time series features and the spatial features to obtain fused features; S34: using a variational autoencoder to perform nonlinear dimensionality reduction on the fused features to obtain the final high-dimensional feature vector. The specific implementation process of each step is described in detail below and given examples:

[0075] S31: Use the Attention-enhanced Temporal Convolutional Network (ATCN) to process time series data and extract time series features.

[0076] ATCN consists of multiple layers of temporal convolutional blocks and attention layers. Each temporal convolutional block contains a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation function. Dilated convolution is used to expand the receptive field and capture long-term dependencies. The attention layer uses a self-attention mechanism to calculate the correlation between different time steps.

[0077] In an example, suppose an explosion-proof motor is equipped with 10 sensors, each of which collects data for 1000 time steps. The input shape of ATCN (Attention Temporal Convolutional Network) is (10, 1000, 1). After 3 layers of temporal convolution blocks and 1 layer of attention layer processing, the output shape becomes (10, 1000, 64), where 64 is the number of channels in the last layer.

[0078] Specifically, the settings of the 3-layer temporal convolutional blocks are as follows: the convolution kernel size of the first layer of convolutional blocks is 3, the dilation rate is 1, the input channel is 1, and the output channel is 32. The convolution kernel size of the second layer of convolutional blocks is 3, the dilation rate is increased to 2, the input channel is 32, and the output channel is 48. The convolution kernel size of the third layer of convolutional blocks is still 3, the dilation rate is increased to 4, the input channel is 48, and the output channel is 64. Each convolutional block is followed by a batch normalization layer and a ReLU activation function.

[0079] After these three convolution blocks, the network includes an attention layer, which processes the output of the third convolution block, keeping the time step and number of channels unchanged. By introducing the attention mechanism, the network can focus on important time step features, which helps capture the key patterns of time series data.

[0080] This network design allows for capturing features at different time scales by increasing the dilation rate layer by layer: the first layer has a dilation rate of 1, capturing local, short-term patterns; the second layer has a dilation rate of 2, capturing medium-range patterns; the third layer has a dilation rate of 4, capturing longer-range patterns. By increasing the dilation rate layer by layer, the network can effectively expand the receptive field and capture multi-scale time series features.

[0081] The advantage of this design is that it can expand the receptive field of the network without significantly increasing the number of parameters and computational complexity, thereby effectively learning the characteristics of different time scales and better understanding and predicting time series data.

[0082] S32: Use Graph Convolutional Network (GCN) to process the spatial relationship data between sensors and extract spatial features.

[0083] This step first constructs the adjacency matrix A of the sensor network, which represents the connection relationship between sensors. Then, multi-layer graph convolution operations are used, and the calculation formula for each layer is: ,in l Indicates the index of the current layer ( l =0,1,2,…,L−1,L is the total number of layers), D is the degree matrix, A is the adjacency matrix, It is l The node characteristics of the layer, is a learnable weight matrix, and σ is an activation function. This formula performs the following operations: normalizes the adjacency matrix A, and adds the normalized adjacency information to the node features of the current layer. Multiply, then apply the weight matrix Perform a linear transformation and finally introduce nonlinearity through the activation function σ.

[0084] The adjacency matrix A is a matrix that represents the connection relationship between nodes in a graph. In sensor networks, it can be constructed based on the physical location, correlation, or other predefined relationships between sensors. For a graph containing n nodes, the dimension of the adjacency matrix A is n×n. If there is a connection between node i and node j, then A[i][j]=1, otherwise A[i][j]=0.

[0085] The degree matrix D is a diagonal matrix, where the i-th diagonal element represents the degree of the i-th node in the graph (i.e., the number of edges connected to it). The degree matrix is ​​obtained as: D[i][i]=∑A[i], which is the sum of the i-th row in the adjacency matrix A. In order to balance the influence of different nodes, the degree matrix is ​​normalized and calculated. .

[0086] Weight Matrix It is a learnable parameter of each layer of the graph convolutional network GCN, which is randomly assigned during initialization and optimized through the back-propagation algorithm. Its function is to convert the node features of the previous layer into new feature representations, learn the feature relationships in the graph structure, and continuously adjust as the network training progresses.

[0087] The activation function σ is a function used to introduce nonlinear transformations, allowing the model to learn complex nonlinear relationships. Activation functions include ReLU, sigmoid, and tanh. The choice of activation function is predetermined when the model is built and does not need to be learned during training. It improves the expressiveness of the model in graph convolutional networks and improves its ability to capture complex data patterns.

[0088] Node feature matrix represents the feature vector of each node in the lth layer, where is the input feature matrix. As the number of layers increases, the It is obtained from the node features of the previous layer through the graph convolution calculation formula, and the feature representation of the node is constructed layer by layer.

[0089] In an example, assume that 10 sensors form a graph structure and the adjacency matrix A is 10x10. The input feature matrix shape is (10, 64), and after 2 layers of graph convolution (each layer has 32 output channels), the final output shape is (10, 32).

[0090] S33: Fuse the temporal features and the spatial features to obtain fused features.

[0091] In this step, the temporal features output by ATCN are globally averaged pooled in the time dimension to obtain the temporal feature vectors of each sensor. These temporal feature vectors are then concatenated with the spatial feature vectors output by GCN in the channel dimension. In one example, the ATCN output (10, 1000, 64) becomes (10, 64) after global average pooling, and is concatenated with the GCN output (10,32) to obtain the fused feature (10, 96).

[0092] S34: Use Variational Autoencoder (VAE) to perform nonlinear dimensionality reduction to obtain the final high-dimensional feature vector.

[0093] VAE consists of an encoder and a decoder. The encoder maps the input features to the latent space and outputs the mean μ and variance σ. 2 . The latent variable z = μ + ε * σ is sampled, where ε is sampled from a standard normal distribution. The decoder attempts to reconstruct the original features from z. The training goal is to minimize the reconstruction error and the KL (Kullback-Leibler) divergence, which is used to evaluate the similarity between a distribution and a target distribution.

[0094] In one example, the input fusion feature (10, 96) is used, and the encoder consists of two fully connected layers to map the features to a 32-dimensional latent space. The decoder also consists of two fully connected layers to try to reconstruct the 96-dimensional features. Finally, a 32-dimensional high-dimensional feature vector is obtained as the input for fault diagnosis.

[0095] The beneficial technical effects of this method include the following aspects: First, the attention mechanism-enhanced temporal convolutional network can not only capture the long-term and short-term dependencies of time series data, but also adaptively focus on important time steps and features, significantly improving the effect of time series feature extraction. Second, the introduction of graph convolutional neural network enables the model to make full use of the spatial relationship information between sensors, which is often ignored in traditional methods. By modeling the sensor network as a graph structure, GCN can capture complex spatial correlations and provide rich spatial features. Third, the fusion of temporal features and spatial features realizes the effective integration of multimodal information, so that subsequent fault diagnosis can consider information in both time and space dimensions at the same time. Finally, the use of variational autoencoders not only realizes nonlinear dimensionality reduction, but also learns the probability distribution of data, which provides a more robust and interpretable feature representation for subsequent anomaly detection and fault diagnosis. This multi-level and multi-angle feature extraction and fusion method greatly improves the expressiveness and discriminability of features, and provides more reliable and accurate input for fault diagnosis of explosion-proof motors. Embodiment 4

[0096] like Figure 4As shown, in this embodiment, step S40 specifically includes:

[0097] S41: constructing a multi-task learning network, wherein the multi-task learning network includes a shared underlying feature extraction layer, and a fault classification subnetwork and a fault severity assessment subnetwork connected to the shared underlying feature extraction layer;

[0098] In the fault classification subnetwork, the output of the shared underlying feature extraction layer is connected to a first fully connected layer, and the first fully connected layer is used to extract fault classification related features; the output of the first fully connected layer is processed by a soft maximum activation function to obtain a probability distribution of each fault type;

[0099] In the fault severity assessment subnetwork, the output of the shared underlying feature extraction layer is connected to a second fully connected layer, and the second fully connected layer is used to extract fault severity related features; the output of the second fully connected layer is processed by a regression activation function to obtain a fault severity index;

[0100] In specific implementation, the shared underlying feature extraction layer can adopt a multi-layer convolutional neural network structure, for example, including 3 convolutional layers and 3 pooling layers. Each convolutional layer uses a 3x3 convolution kernel, and the activation function selects ReLU. The pooling layer uses a 2x2 maximum pooling.

[0101] The first fully connected layer in the fault classification subnetwork may contain 128 neurons and use a ReLU activation function. A softmax activation function is used to convert the output of the fully connected layer into a probability distribution of each fault type.

[0102] The second fully connected layer in the fault severity assessment subnetwork can also contain 128 neurons and use the ReLU activation function. The regression activation function selects a linear activation function and directly outputs the fault severity index.

[0103] S42: Adopting a gradient descent optimization algorithm, based on a multi-task loss function, the network parameters of the shared underlying feature extraction layer, the fault classification subnetwork and the fault severity assessment subnetwork are optimized to achieve end-to-end training of fault classification and fault severity assessment.

[0104] In this step, the Adam (Adaptive Moment Estimation) optimizer is used as the specific implementation of the gradient descent algorithm. The learning rate is initially set to 0.001, and the learning rate decay strategy is used to reduce the learning rate to 0.9 times the original value every 10 cycles (epochs). The batch size is set to 64.

[0105] S43: After the training is completed, the high-dimensional feature vector is input into the trained multi-task learning network to obtain the probability distribution of the fault type and the corresponding fault severity index.

[0106] After training, the network can be used for fault diagnosis. The high-dimensional feature vector to be diagnosed is input into the network, and the network will output the probability distribution of the fault type and the corresponding fault severity index. The fault type with the highest probability is selected as the diagnosis result, and the fault severity index output is directly used to evaluate the severity of the fault.

[0107] In a further embodiment, step S42 may specifically include:

[0108] S421: Design a multi-task loss function, which includes: a cross entropy loss between the probability distribution output by the fault classification subnetwork and a pre-labeled true fault type label; a mean square error loss between the fault severity index output by the fault severity assessment subnetwork and a pre-labeled true fault severity label; wherein the cross entropy loss and the mean square error loss are weightedly combined based on their respective corresponding weight coefficients to obtain the multi-task loss function, which is used for optimizing the training of network parameters of the multi-task learning network.

[0109] The specific design of the multi-task loss function is as follows:

[0110] L_total = α * L_classification + β * L_severity;

[0111] Among them, L_classification is the cross entropy loss of fault classification, L_severity is the mean square error loss of fault severity, and α and β are the corresponding weight coefficients. Initially, α=β=0.5 is set, indicating that the two tasks have the same importance.

[0112] S422: During the training process, according to the learning difficulty and convergence speed of each task, the weight coefficients of the fault classification task loss and the fault severity assessment task loss in the multi-task loss function are adjusted, and the weight coefficients are used for the weighted combination of the two parts of the loss in step S421.

[0113] In order to dynamically adjust the weight coefficients, a gradient-based adaptive weight method is used. Specifically, the gradient norm of each task loss relative to the shared layer is calculated, and the weight is adjusted according to the ratio of the gradient norm. For example:

[0114] α= ||∇L_classification|| / (||∇L_classification|| + ||∇L_severity||);

[0115] β= ||∇L_severity|| / (||∇L_classification|| + ||∇L_severity||).

[0116] Among them, ||∇L_classification|| and ||∇L_severity|| represent the gradient norms of classification loss and severity loss with respect to the shared layer, respectively. This approach can automatically balance the learning process of the two tasks.

[0117] S423: Using a gradient descent optimization algorithm, based on the multi-task loss function, iteratively update the network parameters of the shared underlying feature extraction layer, the fault classification subnetwork, and the fault severity assessment subnetwork until a preset convergence condition or a maximum number of iterations is reached.

[0118] In this step, the model uses the Adam optimizer mentioned above to update parameters. The training process first needs to initialize all network parameters to ensure that the model starts from a good starting point.

[0119] For each training batch, the model goes through the following steps: First, forward propagation is performed to calculate the probability distribution of fault types and the fault severity index, which is the key process for the model to make predictions based on the input data. Next, based on the model's predicted output, the value of the multi-task loss function is calculated to measure the accuracy of the model's predictions. After the loss function value is calculated, the model will perform backpropagation to calculate the gradient of the loss function relative to each parameter. Subsequently, the Adam optimizer is used to update the network parameters based on these gradients to make the model more accurate after each batch. After each cycle (epoch), the model will evaluate the performance on the validation set to check the generalization ability of the model, ensure that there is no overfitting during the training process, and that the performance continues to improve. Finally, the above training and validation steps are repeated until the preset convergence conditions are met (for example, the performance of the validation set no longer improves) or the maximum number of iterations (for example, 100 epochs) is reached, thus completing the entire model training process.

[0120] Through the above steps, an end-to-end multi-task learning network is implemented, which can simultaneously perform fault classification and fault severity assessment, providing comprehensive decision support for equipment maintenance and management.

[0121] The advantages of the fault diagnosis method based on multi-task learning proposed in the embodiment of the present invention are:

[0122] First, by constructing a multi-task learning network that includes a shared underlying feature extraction layer, a fault classification subnetwork, and a fault severity assessment subnetwork, unified modeling of fault classification and fault severity assessment is achieved. This structure can not only make full use of the correlation between different tasks and improve the efficiency of feature extraction, but also significantly reduce the number of model parameters, computational complexity, and storage requirements while ensuring diagnostic accuracy.

[0123] Secondly, the multi-task loss function design adopted by this method cleverly combines the cross entropy loss of fault classification and the mean square error loss of fault severity assessment by weighted combination. This design enables the model to optimize two related but different task objectives at the same time, achieving mutual promotion and restriction between tasks during the training process. In particular, by dynamically adjusting the task weight coefficients, the model can adaptively balance the learning difficulty and convergence speed of different tasks, effectively solving the task conflict problem in multi-task learning, and significantly improving the overall performance and generalization ability of the model.

[0124] Finally, this method realizes end-to-end training and reasoning of fault diagnosis. By directly inputting high-dimensional feature vectors into the trained multi-task learning network, the probability distribution of fault types and the corresponding fault severity index are obtained at one time, which greatly simplifies the diagnosis process. This not only improves the diagnosis efficiency and reduces the errors that may be introduced in the intermediate links, but also provides a more comprehensive and reliable basis for equipment maintenance decisions. Embodiment 5

[0125] like Figure 5 As shown, in this embodiment, step S50 may specifically include:

[0126] S51: Determine a fault type from the fault type probability distribution based on a preset fault type probability threshold.

[0127] Specifically, in this step, the fault type probability threshold is set to 0.6. For the fault type probability distribution output by the multi-task learning network, the fault type with a probability value greater than 0.6 is selected as the diagnosis result. For example, if the fault type probability distribution output by the network is: {bearing fault: 0.7, stator winding fault: 0.2, rotor fault: 0.1}, the determined fault type is a bearing fault.

[0128] S52: Mapping the fault severity index to a predefined severity level to obtain a fault severity rating.

[0129] Specifically, this step predefines five severity levels: minor, medium, severe, dangerous, and urgent. The fault severity index ranges from 0 to 1, using the following mapping rule: 0 ≤ index < 0.2: minor, 0.2 ≤ index < 0.4: medium, 0.4 ≤ index < 0.6: severe, 0.6 ≤ index < 0.8: dangerous, 0.8 ≤ index ≤ 1: urgent. For example, if the fault severity index is 0.55, the corresponding fault severity rating is "serious".

[0130] S53: According to the fault type and the fault severity rating, the corresponding fault description, fault cause and potential impact on the performance and safety of the explosion-proof motor are retrieved from a pre-established fault diagnosis knowledge base.

[0131] Specifically, this step pre-establishes a fault diagnosis knowledge base containing various fault types, severity levels, and corresponding descriptions, causes, and impacts. For example, for a "bearing fault" with a severity of "serious", the following information can be retrieved: Fault description: The bearing has obvious vibration and noise, and the temperature rises significantly; Fault cause: It may be caused by insufficient lubrication, severe bearing wear, or damage to the internal parts of the bearing; Potential impact: It may cause accelerated bearing wear, increase energy consumption, and reduce motor efficiency. In severe cases, it may cause the bearing to get stuck, causing the motor to shut down.

[0132] S54: Generate targeted maintenance suggestions based on the fault severity rating and preset maintenance rules.

[0133] Specifically, this step formulates corresponding maintenance rules according to the fault severity rating. For example, for a bearing fault of the "severe" level, the following maintenance recommendations are generated:

[0134] The recommended maintenance time is to immediately arrange shutdown and maintenance, no more than 48 hours; specific maintenance operation steps: shut down and cut off the power, disassemble the motor housing and bearing cover, check the bearing status, replace the bearing if necessary, check and add lubricating oil, reassemble the motor, and conduct trial operation and vibration test; the required spare parts list includes new bearings, seals, and lubricating oil.

[0135] S55: Generate an intelligent diagnosis report, which includes the following contents: the fault type identified from the fault type probability distribution and its corresponding probability value; the fault severity rating and its corresponding fault severity index; the fault description, fault cause and potential impact retrieved and integrated from the fault diagnosis knowledge base; targeted maintenance suggestions, including recommended maintenance time, specific maintenance operation steps and required spare parts list; if there are multiple fault types, they are sorted according to the comprehensive score obtained by the weighted combination of the probability value of the fault type and the fault severity index.

[0136] Specifically, in this step, all the information obtained in the previous steps is integrated into a structured intelligent diagnosis report. For multiple fault types, the comprehensive score is calculated using the following formula: Comprehensive score = 0.6 * fault type probability + 0.4 * fault severity index, and then sorted from high to low according to the comprehensive score, with fault types with higher scores being displayed first.

[0137] S56: Push the intelligent diagnosis report to the terminal device of the maintenance personnel.

[0138] Finally, the generated intelligent diagnosis report is pushed to the terminal device of the maintenance personnel through the preset communication channel (mobile application, email system or internal enterprise communication platform). This ensures that the maintenance personnel can obtain fault diagnosis results and maintenance suggestions in a timely manner, so as to respond quickly and take appropriate maintenance measures. Embodiment 6

[0139] In some embodiments, the method further includes the following steps: S60: constructing a digital twin model of the explosion-proof motor; S70: using the digital twin model to perform fault simulation and predictive maintenance.

[0140] In some embodiments, step S60 specifically includes: S61: obtaining design parameters, material properties and geometric structure data of the explosion-proof motor; S62: establishing a three-dimensional geometric model of the explosion-proof motor based on the geometric structure data; S63: based on the three-dimensional geometric model, and in combination with the acquired design parameters and material properties of the explosion-proof motor, constructing a multi-physical field coupling model of the explosion-proof motor, which includes electromagnetic field, thermal field and mechanical field; S64: based on the historical operation data and experimental data of the explosion-proof motor, calibrating and verifying the parameters of the multi-physical field coupling model to obtain a calibrated and verified digital twin model.

[0141] like Figure 6 As shown, specifically, in step S60, a digital twin model of the explosion-proof motor is constructed. This process includes four sub-steps:

[0142] S61: First, obtain the design parameters (such as rated power, rated voltage, rated speed, etc.), material properties (such as magnetic material properties of stator and rotor, thermal conductivity of insulation materials, etc.) and geometric structure data (such as stator outer diameter, air gap length, number of slots, etc.) of the explosion-proof motor. These data can be obtained from motor design documents, material databases and CAD drawings.

[0143] S62: Based on the acquired geometric structure data, use 3D modeling software (such as SolidWorks or CATIA) to build an accurate 3D geometric model of the explosion-proof motor. This model includes all key components such as stator, rotor, bearing, housing, etc.

[0144] S63: Based on the 3D geometric model, combined with the design parameters and material properties, a multi-physics coupling model of the explosion-proof motor is constructed. This model includes electromagnetic fields (used to simulate the electromagnetic performance of the motor), thermal fields (used to simulate heat distribution and transfer), and mechanical fields (used to simulate stress and deformation). Multi-physics simulation software (such as COMSOL Multiphysics) can be used to implement this step.

[0145] S64: Finally, the multi-physics coupling model is calibrated and verified using the historical operating data of the explosion-proof motor (such as time series data of temperature, vibration, current, etc.) and experimental data (such as performance test data under different working conditions). Through repeated iterations and optimization, a digital twin model that can accurately reflect the actual behavior of the explosion-proof motor is obtained.

[0146] In some embodiments, step S70 includes: S71: using the Monte Carlo method to perform fault scenario simulation on the digital twin model of the explosion-proof motor to obtain simulation results under various possible fault scenarios, including response data and performance change data of the explosion-proof motor; S72: based on the simulation results, identifying potential fault modes and constructing a fault evolution tree; S73: using the fault evolution tree to analyze the development paths and potential impacts of different fault modes; S74: dynamically updating the remaining service life prediction results of the explosion-proof motor based on the real-time monitored explosion-proof motor operation data; S75: based on the development path of the fault mode and the potential impact, as well as the updated remaining service life prediction results of the explosion-proof motor, determining an optimized predictive maintenance strategy, which includes maintenance schedules and specific maintenance measures.

[0147] like Figure 7 As shown, specifically, in step S70, the constructed digital twin model is used to perform fault simulation and predictive maintenance. This process includes five sub-steps:

[0148] S71: Use the Monte Carlo method to simulate various possible fault scenarios on the digital twin model. For example, parameters such as bearing friction coefficient and winding insulation performance can be randomly changed to simulate different degrees of bearing failure or insulation aging. Through a large number of simulations, the response data (such as vibration characteristics, temperature distribution, etc.) and performance change data (such as efficiency reduction, power factor change, etc.) of the explosion-proof motor under various fault scenarios are obtained.

[0149] Other examples of using the Monte Carlo method to simulate possible failure scenarios of explosion-proof motors are as follows: For rotor imbalance, the rotor mass distribution parameters can be randomly changed to simulate different degrees of rotor imbalance. For stator core looseness, the contact stiffness between the stator core and the base is randomly adjusted to simulate different degrees of stator core looseness. For air gap eccentricity, the relative position of the rotor center and the stator center is randomly changed to simulate various combinations of static eccentricity and dynamic eccentricity. For winding turn-to-turn short circuit, several turns in the stator winding are randomly selected, their resistance values ​​are changed, and turn-to-turn short circuits of different degrees and positions are simulated. For reduced efficiency of the cooling system, various situations of reduced efficiency of the cooling system are simulated by randomly reducing the thermal conductivity or flow rate of the cooling system. For poor bearing lubrication, the viscosity and oil film thickness of the bearing lubricant are randomly adjusted to simulate different degrees of poor lubrication. For partial discharge of the stator winding, the dielectric strength of the insulating material is randomly changed to simulate partial discharge phenomena of different positions and intensities. For rotor bar breakage, several rotor bars are randomly selected, their conductivity is reduced, and rotor bar breakage of different degrees and positions is simulated. For shaft bending, the bending deformation of the shaft is randomly added to simulate the shaft bending caused by external force or material fatigue. For end cover deformation, the geometric parameters of the end cover are randomly changed to simulate the end cover deformation caused by thermal stress or mechanical stress. Through these diverse fault scenario simulations, the performance and response of explosion-proof motors under various possible fault conditions can be fully evaluated, so as to formulate a more comprehensive and effective predictive maintenance strategy.

[0150] S72: Based on the simulation results, potential fault modes (such as bearing failure, stator winding short circuit, rotor bar breakage, etc.) are identified and a fault evolution tree is constructed. This fault evolution tree shows the correlation and possible development paths between different fault modes.

[0151] Specifically, fault mode identification is the first step in building a fault evolution tree. Based on the Monte Carlo simulation results in S71, a variety of potential fault modes can be identified, including bearing failure, stator winding short circuit, rotor bar breakage, air gap eccentricity, stator core loosening, shaft bending, reduced cooling system efficiency, and partial discharge caused by insulation aging. These fault modes constitute the basic nodes of the fault evolution tree.

[0152] Fault feature extraction is to determine the unique signal characteristics for each fault mode. For example, a bearing fault may be manifested as an enhanced vibration signal of a specific frequency; a stator winding short circuit may cause phase current imbalance and increased harmonics; a broken rotor bar may cause periodic fluctuations in torque and speed; and an air gap eccentricity may cause magnetic field fluctuations of a specific frequency. These characteristic signals provide an important basis for subsequent fault diagnosis and prediction.

[0153] Fault severity quantification is to establish quantitative evaluation indicators for each fault mode. For example, the severity of bearing faults can be quantified by vibration amplitude or spectrum energy, the degree of winding short circuit can be indicated by phase current imbalance, and the severity of rotor broken bars can be measured by torque ripple amplitude.

[0154] Fault correlation analysis aims to explore the interrelationships and causal relationships between different fault modes. For example, a bearing failure may cause shaft bending, reduced cooling system efficiency may accelerate insulation aging, and air gap eccentricity may increase bearing loads and accelerate bearing wear. This analysis helps understand the complex interactions between faults and lays the foundation for building a more accurate fault evolution tree.

[0155] Constructing a fault evolution tree is to integrate the analysis results of the previous steps into a structured model. This tree structure includes nodes representing various fault modes, edges representing the evolution relationship between faults, and weights representing the probability or speed of fault evolution. For example, Fig.10 As shown, a simplified fault evolution tree may start from bearing failure, branch to shaft bending, vibration enhancement and temperature increase, and then further develop to air gap eccentricity, stator core looseness and insulation aging.

[0156] Adding a time dimension to the fault evolution tree can indicate the time required for a fault to develop from one state to another. This is estimated based on historical data and expert experience. At the same time, a probability value is assigned to each branch in the fault evolution tree, indicating the likelihood of the fault developing along that path. These probabilities can be obtained by statistically analyzing historical data and simulation results, making fault prediction more accurate.

[0157] Dynamic updating of the fault evolution tree is key to ensuring its continued accuracy. As new operating data and simulation results are obtained, the fault evolution tree needs to be continuously updated and optimized. This can include adding new fault modes, adjusting evolution paths, updating probability values, etc. In addition, considering the impact of environmental factors (temperature, humidity, dust, etc.) on fault evolution and using these factors as additional dimensions of the fault evolution tree can further improve the accuracy of the model.

[0158] Finally, using graphical tools (such as D3.js) to visualize the fault evolution tree can make complex fault relationships more intuitive and easy to understand. Through this detailed fault evolution tree, we can better understand the complex relationship between various failure modes of explosion-proof motors, predict the development trend of faults, and provide an important basis for formulating predictive maintenance strategies. This method not only considers the development of a single fault, but also the interaction and cascading effects between multiple faults, thereby achieving more comprehensive and accurate fault prediction and prevention.

[0159] S73: Use the constructed fault evolution tree to analyze the development paths and potential impacts of different fault modes. For example, you can evaluate how a bearing fault gradually develops and eventually causes the motor to shut down, and the impact on the motor performance and life in the process.

[0160] Specifically, using the fault evolution tree to analyze the development paths of different fault modes first requires determining the starting fault point. Taking bearing faults as an example, the analysis can start from the appearance of tiny cracks on the inner ring of the bearing. In the fault evolution tree, this initial fault may be represented as a root node. Starting from this node, multiple possible development paths can be traced. For example, one path may be: tiny cracks → crack expansion → increased vibration → increased noise → increased temperature → lubrication failure → severe wear → bearing jamming → motor shutdown. Another path may involve other problems caused by bearing failures: tiny cracks → increased vibration → shaft misalignment → air gap eccentricity → increased stator winding stress → accelerated insulation aging. In this way, multiple possible development paths of the fault can be comprehensively mapped.

[0161] Analyzing the potential impact of the fault development path requires defining specific performance indicators and thresholds for each fault state. Continuing with the bearing fault example, the following indicators and thresholds can be defined: vibration amplitude (normal: <2mm / s, mild: 2-5mm / s, moderate: 5-10mm / s, severe: >10mm / s), temperature rise (normal: <10°C, mild: 10-20°C, moderate: 20-40°C, severe: >40°C), noise level (normal: <70dB, mild: 70-80dB, moderate: 80-90dB, severe: >90dB). By monitoring the changes in these indicators, the severity and development speed of the fault can be evaluated. For example, if the vibration amplitude rises from 3mm / s to 7mm / s within a week, it can be judged that the fault is rapidly deteriorating and requires timely intervention.

[0162] In order to evaluate the impact of faults on the life of the motor, it is necessary to establish a relationship model between the severity of the fault and the expected remaining life. This can be constructed based on historical data and expert experience. For example, for bearing faults, a simplified model can be established: when the vibration amplitude reaches 5mm / s, the expected remaining life is 70% of the original design life; when it reaches 8mm / s, the remaining life is 50%; when it reaches 12mm / s, the remaining life is only 20%. This model needs to consider multiple factors, such as the severity of the initial fault, the speed of fault development, and the operating environment conditions. Through this analysis, the remaining service life of the equipment can be predicted more accurately, providing a basis for maintenance decisions.

[0163] An important application of fault evolution trees is the analysis of the cascading effects of faults. It is necessary to identify the mutual influence and acceleration between different faults. For example, the vibration caused by a bearing fault may cause shaft misalignment, which in turn causes air gap eccentricity. This cascading effect can be quantified by defining an influence coefficient. Assume that every 1mm / s increase in bearing vibration will cause an increase in shaft misalignment of 0.01mm, and every 0.1mm of shaft misalignment will increase the air gap eccentricity by 0.05mm. In this way, it is possible to track how an initial fault gradually affects the entire system and quantify the degree of impact at each stage.

[0164] Considering the impact of external factors on the development of faults is key to a complete analysis. It is necessary to model the relationship between environmental conditions, operating conditions, and maintenance practices and the rate of fault development. For example, for bearing faults, a temperature influence factor can be defined: for every 10°C increase in ambient temperature, the fault development rate increases by 20%. Similarly, a load influence factor can be defined: for every 10% increase in load, the fault development rate increases by 15%. In this way, the development trend of faults under different conditions can be more accurately predicted.

[0165] Based on the above analysis, corresponding response strategies can be formulated for different fault scenarios. This includes determining the best maintenance time point and selecting appropriate maintenance methods. For example, for bearing failures, the following strategy can be formulated: when the vibration amplitude reaches 4mm / s, increase the monitoring frequency to once a week; when it reaches 6mm / s, plan to replace the bearing at the next shutdown; when it reaches 9mm / s, immediately arrange shutdown for replacement. At the same time, the cascading effect of the fault also needs to be considered, and related issues such as shaft alignment and stator winding status may need to be checked and handled at the same time.

[0166] Finally, the fault evolution tree model needs to be continuously updated and optimized. As new data is accumulated and new knowledge is acquired, model parameters such as fault development path, influence coefficient, threshold, etc. should be reviewed and adjusted regularly. For example, if a new bearing material is found to significantly delay fault development, the time scale and influence factor in the fault evolution tree need to be adjusted accordingly. Through this continuous improvement process, it can be ensured that the fault analysis model remains accurate and effective at all times, providing strong support for the reliable operation of explosion-proof motors.

[0167] S74: Based on the real-time monitored explosion-proof motor operation data (such as vibration, temperature, current, etc.), combined with the prediction results of the digital twin model, dynamically update the remaining service life prediction of the explosion-proof motor. This process can be achieved using machine learning algorithms (such as recursive neural networks or long short-term memory networks).

[0168] Specifically, real-time monitoring of explosion-proof motor operation data is the basis for dynamically updating the remaining service life prediction. This process involves the deployment of multiple sensors and the establishment of a data acquisition system. Monitoring parameters include vibration, temperature, current, voltage, speed, etc. of the motor. For example, accelerometers can be installed at key locations of the motor (bearing seat, stator housing) to monitor vibration; thermocouples or infrared sensors can be used to monitor temperature; current transformers and voltage transformers can be used to monitor electrical parameters. These sensors continuously collect data through a data acquisition system (PLC or dedicated data acquisition card), and the sampling frequency may range from a few times per second to thousands of times per second, depending on the characteristics of the monitored parameters. After preliminary processing (filtering, averaging), the collected raw data is transmitted to the central processing system for further analysis.

[0169] Data preprocessing and feature extraction are key steps in dynamically updating life prediction. Raw data usually contains noise and redundant information and needs to be cleaned and converted. For vibration data, spectrum analysis may be required to extract characteristic frequencies and amplitudes; for temperature data, the average, maximum, and rate of change may need to be calculated; for current data, harmonic components and imbalance may need to be analyzed. For example, for bearing vibration data, fast Fourier transform (FFT) can be used to extract characteristic frequencies, such as inner race fault frequency, outer race fault frequency, etc.

[0170] Establishing a dynamic life prediction model is the core of achieving real-time updates. This involves the application of machine learning or statistical models. Available methods include regression analysis, neural networks, support vector machines, etc. Taking neural networks as an example, a multi-layer perceptron model can be constructed, where the input layer includes processed feature data (vibration amplitude, temperature, current imbalance, etc.), the hidden layer performs nonlinear transformation, and the output layer gives the predicted remaining life. The initial training of the model can be based on historical data and expert knowledge, but the key is to design a mechanism that enables the model to continuously self-adjust and optimize based on new monitoring data. For example, an online learning algorithm can be used to update the model parameters slightly whenever a new data point is received, so that the prediction results can reflect the changes in the motor state in a timely manner.

[0171] The integration of real-time data and model updates are key links in dynamic prediction. Whenever new monitoring data arrives, the system needs to quickly process the data and integrate it into the existing prediction model. This process involves multiple steps: first, the new data is preprocessed and feature extracted; then, these new features are input into the prediction model to obtain a preliminary estimate of the remaining life; then, the system compares this new estimate with the previous prediction results. If the difference is significant, it may be necessary to trigger model recalibration or parameter adjustment. For example, if the new vibration data shows a sudden increase in amplitude that exceeds the previously predicted range, the system needs to adjust the vibration-related weight parameters in the model to reflect the impact of this sudden change on life.

[0172] Considering the impact of changes in the operating environment and operating conditions on life prediction is also an important aspect of the dynamic update process. The operating environment of explosion-proof motors may change, such as temperature fluctuations, humidity changes, changes in dust concentration, etc. At the same time, the operating conditions may also change, such as load fluctuations, changes in start-stop frequency, etc. These factors may significantly affect the life of the motor. Therefore, it is necessary to incorporate these environmental and operating condition parameters into the prediction model. For example, a temperature compensation factor can be introduced to appropriately reduce the predicted remaining life when the ambient temperature rises; or a load factor can be introduced to accelerate the calculation of life consumption when operating at high load for a long time. This dynamic adjustment ensures that the prediction results can reflect changes in the actual operating conditions of the motor.

[0173] Visualization and early warning of prediction results are the final link of the dynamic update system. The processed prediction results need to be presented to users in an intuitive and easy-to-understand way. This may include real-time updated remaining life curves, health status dashboards, failure probability distribution diagrams, etc. At the same time, the system should be able to automatically generate early warning information based on the prediction results. For example, when the predicted remaining life is lower than a certain threshold (20% of the design life), the system can automatically issue maintenance recommendations; when a sharp drop in remaining life is detected, an emergency inspection alarm may need to be triggered. These visualization and early warning functions not only help operators intuitively understand the health status of the motor, but also provide timely and reliable basis for maintenance decisions.

[0174] Finally, continuous model evaluation and improvement are key to ensuring the long-term effectiveness of the dynamic prediction system. It is necessary to regularly compare the prediction results with the actual observed motor performance degradation to evaluate the accuracy and reliability of the prediction. If it is found that the prediction results deviate significantly from the actual situation, it is necessary to analyze the reasons and adjust the model accordingly. This involves retraining the machine learning model, adjusting the feature extraction method, updating the environmental impact factor, etc. For example, if it is found that the model always underestimates the life consumption in a high temperature environment, the temperature compensation factor may need to be recalibrated. Through this continuous evaluation and improvement process, it can be ensured that the dynamic life prediction system can adapt to changing operating conditions and emerging failure modes, and provide long-term and effective support for the reliable operation of explosion-proof motors.

[0175] S75: Finally, based on the development path of the failure mode, the potential impact, and the updated remaining service life prediction results, an optimized predictive maintenance strategy is determined. This strategy includes the best maintenance schedule (such as overhaul before the next planned shutdown) and specific maintenance measures (such as replacing bearings, cleaning windings, etc.). This digital twin-based predictive maintenance strategy can maximize the service life of explosion-proof motors while minimizing maintenance costs and downtime.

[0176] Specifically, determining the optimized predictive maintenance strategy first requires a comprehensive analysis of the development path of the failure mode, potential impacts, and updated remaining service life prediction results. Taking bearing failure as an example, it may have been known through fault evolution tree analysis that its development path is: small cracks → increased vibration → increased temperature → lubrication failure → severe wear. At the same time, through the dynamic life prediction model, it may be obtained that the current expected remaining life of the bearing is 40% of the original design life. Based on this information, it is necessary to determine the key intervention point, that is, maintenance before the failure develops to an irreversible stage. For example, it may be determined that when the vibration amplitude reaches 6mm / s or the temperature rises by more than 25°C, maintenance should be scheduled, and this time point may correspond to the expected remaining life falling to about 35%.

[0177] Scheduling maintenance requires balancing multiple factors, including the health of the equipment, production schedule, spare parts availability, and maintenance resources. Based on the above bearing failure example, assume that the vibration amplitude is predicted to reach a critical value of 6mm / s in 3 months. Maintenance needs to be scheduled before this time point. Taking into account the production schedule, maintenance may be chosen on a weekend after 2.5 months to minimize the impact on production. At the same time, it is necessary to ensure that the necessary spare parts (new bearings) and maintenance personnel are ready before this time point. In addition, if there are multiple devices that need maintenance, it may be necessary to develop a priority list to arrange the maintenance order based on the health status and importance of each device.

[0178] The formulation of specific maintenance measures needs to be based on the characteristics and development stage of the failure mode. Continuing with the example of bearing failure, if the problem is discovered in the early stages (when vibration just begins to increase), maintenance measures may include relubrication, adjusting shaft alignment, or replacing bearings. Specifically, the following maintenance plan may be formulated: first, shut down and remove the bearing cover; then, check the actual condition of the bearing, and if it is just insufficient lubrication, clean and relubricate it; if minor wear is found, the bearing may need to be replaced; at the same time, check the alignment of the shaft and adjust it if necessary; finally, reinstall and conduct a test run to ensure that vibration and temperature return to normal levels. In this process, detailed maintenance data also needs to be recorded for subsequent analysis and model optimization.

[0179] Dynamic adjustment of maintenance strategies is key to ensuring that predictive maintenance continues to be effective. As new data accumulates and predictive models are updated, maintenance strategies need to be regularly evaluated and adjusted. For example, if it is found that the life of a certain type of bearing under specific operating conditions is shorter than expected, it may be necessary to advance the maintenance time point or increase the frequency of inspections. On the other hand, if certain preventive measures (regular lubrication) significantly extend the life of the equipment, the strategy may be adjusted to increase the frequency of these preventive maintenance. This dynamic adjustment process can be achieved by setting up a feedback loop: after each maintenance, compare the actual problems found with the predicted results, and use this information to fine-tune the predictive model and maintenance strategy.

[0180] Considering maintenance costs and benefits is an important aspect of optimizing predictive maintenance strategies. Maintenance costs (including direct costs such as spare parts and labor, and indirect costs such as downtime losses) need to be weighed against potential benefits (avoided failure losses and extended equipment life). For example, for a critical explosion-proof motor, even if the maintenance cost is higher, more frequent inspections and a more conservative replacement strategy may be chosen because once a serious failure occurs, it may lead to huge safety risks and economic losses. On the contrary, for non-critical equipment, a more flexible strategy may be adopted, allowing the equipment to operate to a state closer to the failure point. Through this cost-benefit analysis, the optimal maintenance strategy can be formulated for each type of equipment.

[0181] Integrating the maintenance needs of multiple subsystems is an important consideration in developing a comprehensive predictive maintenance strategy. Explosion-proof motors typically contain multiple subsystems, such as bearing systems, insulation systems, cooling systems, etc. Each subsystem may have different failure modes and life characteristics. An optimized maintenance strategy needs to consider the mutual impact and possible collaborative maintenance opportunities between these subsystems. For example, if the forecast shows that the bearing needs to be replaced within 3 months, and the insulation system is expected to need inspection within 4 months, it may be chosen to perform both tasks at the same time in 3 months to reduce overall downtime. This integrated approach not only improves maintenance efficiency, but also helps to discover and solve potential system-level problems.

[0182] Finally, the execution and feedback of maintenance strategies are key links in closed-loop optimization. After each maintenance activity is performed, it is necessary to record in detail the actual problems found, the measures taken and their effects. This information should be used to verify and improve the predictive model and maintenance strategy. For example, if the actual wear of the bearing is found to be more serious than predicted during the scheduled maintenance, this indicates that it is necessary to adjust certain parameters in the predictive model or consider adding certain influencing factors. At the same time, the feedback from maintenance personnel is also a valuable source of information, which may reveal factors not considered in the predictive model or new failure modes. Through this continuous data collection, analysis and strategy adjustment, the predictive maintenance strategy can be continuously optimized, its accuracy and effectiveness can be improved, and ultimately the best balance between the reliability and economy of the explosion-proof motor operation can be achieved.

[0183] This technical solution has the following beneficial technical effects:

[0184] First, by constructing a high-precision digital twin model, this method achieves a comprehensive and accurate simulation of explosion-proof motors. This includes not only the geometric structure, but also the coupling effects of multiple physical fields such as electromagnetic, thermal and mechanical. This multi-dimensional, high-fidelity digital representation goes beyond the traditional single physical field simulation, enabling a more comprehensive understanding and prediction of the behavior of explosion-proof motors under various complex operating conditions. By calibrating and validating the model using actual operation and experimental data, the accuracy and reliability of the digital twin model are greatly improved, a level of accuracy that is difficult to achieve with traditional simulation methods.

[0185] Secondly, using the Monte Carlo method to simulate fault scenarios can simulate a large number of possible fault scenarios, including extreme situations that are difficult to observe or reproduce in actual operation. This not only expands the scope of knowledge of potential fault modes, but also reveals some minor but potentially serious faults that are easily overlooked in traditional methods. Through this comprehensive simulation, we can better understand the causes, development and impacts of faults, so as to develop more effective prevention and response strategies.

[0186] Furthermore, constructing a fault evolution tree provides a novel fault analysis method. This method not only identifies various potential failure modes, but also reveals the complex correlations and evolution paths between them. This systematic fault analysis method enables a better understanding of the root causes and development trends of faults, thereby achieving more accurate fault prediction and prevention. The innovation of this method lies in its ability to capture the interactions and cascading effects between faults, which is difficult to achieve with traditional fault analysis methods.

[0187] Finally, by dynamically updating the remaining service life prediction with real-time data and formulating an optimized predictive maintenance strategy based on it, this method achieves truly intelligent and personalized maintenance. This method not only takes into account the current status of the equipment, but also combines historical data and future predictions to achieve dynamic optimization of maintenance strategies. This forward-looking and adaptive maintenance method greatly improves the accuracy and efficiency of maintenance, maximizes the life of the equipment, and minimizes maintenance costs and downtime. This intelligent maintenance strategy based on data-driven and model prediction far exceeds the traditional fixed cycle or simple threshold-triggered maintenance methods. Embodiment 7

[0188] like Figure 8 As shown, this embodiment provides an intelligent fault diagnosis system for explosion-proof motors in the natural gas industry, which includes:

[0189] Multiple sensors are installed on explosion-proof motors to collect the operating data of explosion-proof motors in real time;

[0190] A data acquisition device, connected to the plurality of sensors, for receiving and preprocessing the operation data collected by the plurality of sensors to obtain preprocessed multi-source heterogeneous data;

[0191] An edge computing device, connected to the data acquisition device, is used to perform preliminary analysis and screening on the preprocessed multi-source heterogeneous data to obtain the preliminarily analyzed multi-source heterogeneous data;

[0192] A remote server, connected to the edge computing device via a network, for receiving multi-source heterogeneous data that has been preliminarily analyzed and performing fault diagnosis;

[0193] The remote server comprises:

[0194] A data acquisition module, used to acquire multi-source heterogeneous data transmitted from edge computing devices and subjected to preliminary analysis;

[0195] A data enhancement module, used for performing enhancement processing on the multi-source heterogeneous data after preliminary analysis to obtain an enhanced multi-modal data set of the explosion-proof motor;

[0196] A feature extraction module, used for extracting features from the enhanced multimodal data set of the explosion-proof motor using a deep learning model to obtain a high-dimensional feature vector;

[0197] A multi-task learning module, used to perform fault classification and fault severity assessment using a multi-task learning network based on the high-dimensional feature vector to obtain a fault type probability distribution and a fault severity index;

[0198] The diagnostic report generating module is used to generate an intelligent diagnostic report of the explosion-proof motor including maintenance suggestions according to the fault type probability distribution and the fault severity index.

[0199] Furthermore, it also includes a terminal device connected to the remote server via a network, and used for receiving and displaying the intelligent diagnosis report.

[0200] like Fig. 9 As shown, in some embodiments, the data enhancement module includes:

[0201] A noise reduction unit, used for performing noise reduction processing on the multi-source heterogeneous data of the explosion-proof motor by using an adaptive threshold noise reduction algorithm based on wavelet packet transform to obtain the multi-source heterogeneous data after noise reduction;

[0202] A standardization unit, used for performing a unified standardization process on the denoised multi-source heterogeneous data using a standard score method to obtain a standardized data matrix;

[0203] A time synchronization unit, used for performing time synchronization on different data sources in the standardized data matrix using a dynamic time warping algorithm to obtain a time-aligned multimodal data sequence;

[0204] A data generation unit, configured to generate simulated data samples by implementing data enhancement through a conditional generative adversarial network based on the time-aligned multimodal data sequence;

[0205] A data merging unit is used to merge the simulation data samples with the time-aligned multimodal data sequence to finally obtain an enhanced multimodal data set of the explosion-proof motor.

[0206] In some embodiments, the feature extraction module includes:

[0207] A time series feature extraction unit, used to process the time series data in the enhanced multimodal data set of the explosion-proof motor using a time convolution network enhanced by an attention mechanism, and extract time series features;

[0208] A spatial feature extraction unit, used for processing the data reflecting the spatial relationship between sensors in the enhanced multimodal data set of the explosion-proof motor using a graph convolutional neural network to extract spatial features;

[0209] A feature fusion unit, used for fusing the temporal feature and the spatial feature to obtain a fused feature;

[0210] The dimension reduction unit is used to perform nonlinear dimension reduction on the fused features by using a variational autoencoder to obtain a final high-dimensional feature vector.

[0211] In some embodiments, the multi-task learning module includes:

[0212] A network construction unit, used to construct a multi-task learning network, wherein the multi-task learning network includes a shared underlying feature extraction layer, and a fault classification subnetwork and a fault severity assessment subnetwork connected to the shared underlying feature extraction layer;

[0213] A network training unit, configured to optimize the network parameters of the shared underlying feature extraction layer, the fault classification subnetwork, and the fault severity assessment subnetwork by using a gradient descent optimization algorithm based on a multi-task loss function, so as to achieve end-to-end training of fault classification and fault severity assessment;

[0214] The result output unit is used to input the high-dimensional feature vector into the trained multi-task learning network after the training is completed, so as to obtain the probability distribution of the fault type and the corresponding fault severity index.

[0215] In some embodiments, the network training unit includes:

[0216] A loss function design subunit, used to design a multi-task loss function, wherein the multi-task loss function includes a cross entropy loss between a probability distribution output by a fault classification subnetwork and a pre-labeled true fault type label, and a mean square error loss between a fault severity index output by a fault severity assessment subnetwork and a pre-labeled true fault severity label;

[0217] A weight adjustment subunit, used to adjust the weight coefficients of the fault classification task loss and the fault severity assessment task loss in the multi-task loss function according to the learning difficulty and convergence speed of each task during the training process;

[0218] A parameter optimization subunit is used to iteratively update the network parameters of the shared underlying feature extraction layer, the fault classification subnetwork and the fault severity assessment subnetwork based on the multi-task loss function using a gradient descent optimization algorithm until a preset convergence condition or a maximum number of iterations is reached.

[0219] In some embodiments, the diagnostic report generation module includes:

[0220] a fault type determination unit, configured to determine a fault type from the fault type probability distribution based on a preset fault type probability threshold;

[0221] a severity rating unit, used for mapping the fault severity index to a predefined severity level to obtain a fault severity rating;

[0222] A knowledge retrieval unit, for retrieving corresponding fault descriptions, fault causes and potential impacts on the performance and safety of the explosion-proof motor from a pre-established fault diagnosis knowledge base according to the fault type and the fault severity rating;

[0223] A maintenance suggestion generating unit, configured to generate targeted maintenance suggestions based on the fault severity rating and preset maintenance rules;

[0224] A report generation unit, used to generate an intelligent diagnosis report, which includes the fault type and probability value, fault severity rating and index, fault description, fault cause, potential impact and maintenance suggestions;

[0225] The report pushing unit is used to push the intelligent diagnosis report to the terminal device of the maintenance personnel.

[0226] In some embodiments, the system further comprises:

[0227] Digital twin model building module, used to build the digital twin model of explosion-proof motors;

[0228] The fault simulation and predictive maintenance module is used to perform fault simulation and predictive maintenance using digital twin models.

[0229] In some embodiments, the digital twin model building module includes:

[0230] A data acquisition unit, used to acquire design parameters, material properties and geometric structure data of the explosion-proof motor;

[0231] A geometric modeling unit, used to establish a three-dimensional geometric model of the explosion-proof motor according to the geometric structure data;

[0232] A multi-physics field modeling unit, used to construct a multi-physics field coupling model of the explosion-proof motor based on the three-dimensional geometric model and in combination with the acquired design parameters and material properties of the explosion-proof motor, which includes electromagnetic field, thermal field and mechanical field;

[0233] The model calibration unit is used to calibrate and verify the parameters of the multi-physics field coupling model based on the historical operation data and experimental data of the explosion-proof motor to obtain a calibrated and verified digital twin model.

[0234] In some embodiments, the fault simulation and predictive maintenance module includes:

[0235] A fault scenario simulation unit, used to perform fault scenario simulation on the digital twin model of the explosion-proof motor using the Monte Carlo method, to obtain simulation results under various possible fault scenarios, including response data and performance change data of the explosion-proof motor;

[0236] A failure mode analysis unit, used to identify potential failure modes and construct a failure evolution tree based on the simulation results;

[0237] A fault impact analysis unit, used to analyze the development paths and potential impacts of different fault modes using the fault evolution tree;

[0238] A life prediction unit is used to dynamically update the remaining service life prediction result of the explosion-proof motor according to the real-time monitored operation data of the explosion-proof motor;

[0239] The maintenance strategy optimization unit is used to determine an optimized predictive maintenance strategy including a maintenance schedule and specific maintenance measures based on the development path of the failure mode and the potential impact, as well as the updated remaining service life prediction result of the explosion-proof motor.

[0240] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0241] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, any one of the above methods is implemented.

[0242] The present invention also provides a computer device. The computer device of the embodiment of the present invention includes: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided by the present invention.

[0243] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present invention. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the above-mentioned module, program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0244] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis method for explosion-proof motors in the natural gas industry, characterized in that: The following steps are involved: S10: Acquire multi-source heterogeneous data of the explosion-proof motor from multiple sensors installed on or around the explosion-proof motor; S20: performing enhancement processing on the multi-source heterogeneous data of the explosion-proof motor to obtain an enhanced multimodal data set of the explosion-proof motor; S30: extracting features from the enhanced multimodal data set of the explosion-proof motor using a deep learning model to obtain a high-dimensional feature vector; S40: Based on the high-dimensional feature vector, a multi-task learning network is used to perform fault classification and fault severity assessment to obtain a fault type probability distribution and a fault severity index; S50: generating an intelligent diagnosis report including maintenance suggestions for the explosion-proof motor according to the fault type probability distribution and the fault severity index; Wherein, step S10 specifically includes: Acquire sound data from an explosion-proof electret microphone array arranged around the explosion-proof motor; The magnetic flux data is obtained from the Hall effect magnetic flux sensor installed at the stator core slot of the explosion-proof motor; The oil analysis data is obtained from the online oil analyzer installed in the lubricating oil circulation system of the explosion-proof motor bearing; Obtain infrared thermal imaging data from infrared thermal imagers installed around explosion-proof motors; The load torque data is obtained by detecting a torque sensor installed between the output shaft of the explosion-proof motor and the load; Wherein, step S30 specifically includes: S31: using a temporal convolutional network enhanced by an attention mechanism to process the time series data in the enhanced multimodal data set of the explosion-proof motor to extract time series features; S32: using a graph convolutional neural network to process data reflecting the spatial relationship between sensors in the enhanced multimodal data set of the explosion-proof motor, and extracting spatial features; S33: fusing the temporal feature and the spatial feature to obtain a fused feature; S34: using a variational autoencoder to perform nonlinear dimensionality reduction on the fused features to obtain a final high-dimensional feature vector; Wherein, step S40 specifically includes: S41: constructing a multi-task learning network, wherein the multi-task learning network includes a shared underlying feature extraction layer, and a fault classification subnetwork and a fault severity assessment subnetwork connected to the shared underlying feature extraction layer; In the fault classification subnetwork, the output of the shared underlying feature extraction layer is connected to a first fully connected layer, and the first fully connected layer is used to extract fault classification related features; the output of the first fully connected layer is processed by a soft maximum activation function to obtain a probability distribution of each fault type; In the fault severity assessment subnetwork, the output of the shared underlying feature extraction layer is connected to a second fully connected layer, and the second fully connected layer is used to extract fault severity related features; the output of the second fully connected layer is processed by a regression activation function to obtain a fault severity index; the regression activation function selects a linear activation function; S42: Adopting a gradient descent optimization algorithm, based on a multi-task loss function, optimizing network parameters of the shared underlying feature extraction layer, the fault classification subnetwork, and the fault severity assessment subnetwork, to achieve end-to-end training of fault classification and fault severity assessment; S43: After the training is completed, the high-dimensional feature vector is input into the trained multi-task learning network to obtain the probability distribution of the fault type and the corresponding fault severity index.

2. The method according to claim 1, characterized in that Step S20 specifically includes: S21: using an adaptive threshold noise reduction algorithm based on wavelet packet transform to perform noise reduction processing on the multi-source heterogeneous data of the explosion-proof motor to obtain noise-reduced multi-source heterogeneous data; S22: performing unified standardization processing on the denoised multi-source heterogeneous data using a standard score method to obtain a standardized data matrix; S23: Using a dynamic time warping algorithm to time synchronize different data sources in the standardized data matrix to obtain a time-aligned multimodal data sequence; S24: Based on the time-aligned multimodal data sequence, data enhancement is implemented through a conditional generative adversarial network to generate simulated data samples; S25: merging the simulation data samples with the time-aligned multimodal data sequence to finally obtain an enhanced multimodal data set of the explosion-proof motor.

3. The method according to claim 1, characterized in that Step S42 specifically includes: S421: Design a multi-task loss function, where the multi-task loss function includes: The cross entropy loss between the probability distribution output by the fault classification subnetwork and the pre-labeled true fault type labels; The mean square error loss between the fault severity index output by the fault severity assessment subnetwork and the pre-labeled true fault severity label; The cross entropy loss and the mean square error loss are weightedly combined based on their respective corresponding weight coefficients to obtain the multi-task loss function, which is used for optimizing the network parameters of the multi-task learning network. S422: During the training process, according to the learning difficulty and convergence speed of each task, the weight coefficients of the fault classification task loss and the fault severity assessment task loss in the multi-task loss function are adjusted, and the weight coefficients are used for the weighted combination of the two parts of the loss in step S421; S423: Using a gradient descent optimization algorithm, based on the multi-task loss function, iteratively update the network parameters of the shared underlying feature extraction layer, the fault classification subnetwork, and the fault severity assessment subnetwork until a preset convergence condition or a maximum number of iterations is reached.

4. The method according to claim 1, characterized in that: Step S50 specifically includes: S51: determining a fault type from the fault type probability distribution based on a preset fault type probability threshold; S52: Mapping the fault severity index to a predefined severity level to obtain a fault severity rating; S53: According to the fault type and the fault severity rating, retrieve corresponding fault descriptions, fault causes, and potential impacts on the performance and safety of the explosion-proof motor from a pre-established fault diagnosis knowledge base; S54: generating targeted maintenance suggestions based on the fault severity rating and preset maintenance rules; S55: Generates an intelligent diagnosis report, which includes the following contents: The fault types identified from the fault type probability distribution and their corresponding probability values; Fault severity ratings and their corresponding fault severity indexes; Fault descriptions, fault causes and potential impacts retrieved and integrated from the fault diagnosis knowledge base; Targeted maintenance recommendations, including recommended maintenance time, specific maintenance steps and a list of required spare parts; S56: Push the intelligent diagnosis report to the terminal device of the maintenance personnel.

5. The method according to claim 1, characterized in that The following steps are also included: S60: Building a digital twin model of explosion-proof motors; S70: Utilize the digital twin model to perform fault simulation and predictive maintenance.

6. The method according to claim 5, characterized in that Step S60 specifically includes: S61: Obtain design parameters, material properties and geometric structure data of the explosion-proof motor; S62: Establishing a three-dimensional geometric model of the explosion-proof motor according to the geometric structure data; S63: Based on the three-dimensional geometric model and in combination with the acquired design parameters and material properties of the explosion-proof motor, construct a multi-physics field coupling model of the explosion-proof motor, which includes an electromagnetic field, a thermal field, and a mechanical field; S64: Based on the historical operating data and experimental data of the explosion-proof motor, the multi-physics field coupling model is calibrated and verified to obtain a calibrated and verified digital twin model.

7. The method according to claim 5, characterized in that Step S70 includes: S71: Performing fault scenario simulation on the digital twin model of the explosion-proof motor using the Monte Carlo method to obtain simulation results under various possible fault scenarios, including response data and performance change data of the explosion-proof motor; S72: Based on the simulation results, identifying potential failure modes and constructing a failure evolution tree; S73: Analyze the development paths and potential impacts of different failure modes using the fault evolution tree; S74: dynamically updating the remaining service life prediction result of the explosion-proof motor according to the real-time monitored operation data of the explosion-proof motor; S75: Based on the development path of the failure mode and the potential impact, as well as the updated remaining service life prediction result of the explosion-proof motor, determine an optimized predictive maintenance strategy, which includes a maintenance schedule and specific maintenance measures.

8. An intelligent fault diagnosis system for explosion-proof motors in the natural gas industry, characterized in that: include: Multiple sensors are installed on or around explosion-proof motors to collect the operating data of explosion-proof motors in real time; The operation data of the explosion-proof motor includes: sound data, which is detected by an explosion-proof electret microphone array arranged around the explosion-proof motor; magnetic flux data, which is detected by a Hall effect magnetic flux sensor installed at the notch of the stator core of the explosion-proof motor; oil analysis data, which is detected by an online oil analyzer installed in the lubricating oil circulation system of the explosion-proof motor bearing; infrared thermal imaging data, which is detected by an infrared thermal imager installed around the explosion-proof motor; load torque data, which is detected by a torque sensor installed between the output shaft of the explosion-proof motor and the load; A data acquisition device, connected to the plurality of sensors, for receiving and preprocessing the operation data collected by the plurality of sensors to obtain preprocessed multi-source heterogeneous data; An edge computing device, connected to the data acquisition device, is used to perform preliminary analysis and screening on the preprocessed multi-source heterogeneous data to obtain the preliminarily analyzed multi-source heterogeneous data; A remote server, connected to the edge computing device via a network, for receiving multi-source heterogeneous data that has been preliminarily analyzed and performing fault diagnosis; The remote server comprises: A data acquisition module, used to acquire multi-source heterogeneous data transmitted from edge computing devices and subjected to preliminary analysis; A data enhancement module, used for performing enhancement processing on the multi-source heterogeneous data after preliminary analysis to obtain an enhanced multi-modal data set of the explosion-proof motor; A feature extraction module, used for extracting features from the enhanced multimodal data set of the explosion-proof motor using a deep learning model to obtain a high-dimensional feature vector; A multi-task learning module, used to perform fault classification and fault severity assessment using a multi-task learning network based on the high-dimensional feature vector to obtain a fault type probability distribution and a fault severity index; A diagnostic report generating module, configured to generate an intelligent diagnostic report including maintenance suggestions for the explosion-proof motor according to the fault type probability distribution and the fault severity index; Wherein, the feature extraction module includes: A time series feature extraction unit, used to process the time series data in the enhanced multimodal data set of the explosion-proof motor using a time convolution network enhanced by an attention mechanism, and extract time series features; A spatial feature extraction unit, used for processing the data reflecting the spatial relationship between sensors in the enhanced multimodal data set of the explosion-proof motor using a graph convolutional neural network to extract spatial features; A feature fusion unit, used to fuse the temporal feature and the spatial feature to obtain a fused feature; A dimensionality reduction unit, used for performing nonlinear dimensionality reduction on the fused features by using a variational autoencoder to obtain a final high-dimensional feature vector; Wherein, the multi-task learning module includes: A network construction unit is used to construct a multi-task learning network, wherein the multi-task learning network includes a shared underlying feature extraction layer, and a fault classification subnetwork and a fault severity assessment subnetwork connected to the shared underlying feature extraction layer; in the fault classification subnetwork, the output of the shared underlying feature extraction layer is connected to a first fully connected layer, and the first fully connected layer is used to extract fault classification related features; the output of the first fully connected layer is processed by a soft maximum activation function to obtain a probability distribution of each fault type; in the fault severity assessment subnetwork, the output of the shared underlying feature extraction layer is connected to a second fully connected layer, and the second fully connected layer is used to extract fault severity related features; the output of the second fully connected layer is processed by a regression activation function to obtain a fault severity index; the regression activation function selects a linear activation function; A network training unit, configured to optimize the network parameters of the shared underlying feature extraction layer, the fault classification subnetwork, and the fault severity assessment subnetwork by using a gradient descent optimization algorithm based on a multi-task loss function, so as to achieve end-to-end training of fault classification and fault severity assessment; The result output unit is used to input the high-dimensional feature vector into the trained multi-task learning network after the training is completed, so as to obtain the probability distribution of the fault type and the corresponding fault severity index.

Images