A system-level composite diagnosis method and system for complex equipment under multiple working conditions

Abstract

The application belongs to the field of fault diagnosis and big data technology, and particularly relates to a system-level composite diagnosis method and system for complex equipment under multiple working conditions, a graph structure model is constructed based on the physical connection relationship of the equipment, key components are modeled as nodes and connection relationships are modeled as edges; subsequently, node features are extracted and global graph-level features are aggregated through a physical structure fusion graph attention network (PSI-GAT) to accurately capture the fault signal propagation and coupling effect between components; then, relying on a multi-task collaborative diagnosis model with a hierarchical structure, information fusion between tasks is realized through hard parameter sharing and cross-attention mechanism, and multi-component fault diagnosis results are synchronously output; finally, a gradual transfer learning strategy with staged unfreezing parameters is adopted to enable the model to quickly adapt to new working conditions. The application significantly improves the composite fault diagnosis accuracy and multi-task collaboration efficiency, and provides reliable technical support for the safe and stable operation of complex equipment throughout its life cycle.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of fault diagnosis and big data, and particularly relates to a system-level composite diagnosis method and system for complex equipment under multiple working conditions. BACKGROUND

[0002] As a core support carrier in modern industrial production, transportation and other fields, the system-level fault diagnosis of complex equipment is a key link to ensure the safe and stable operation of the equipment throughout its life cycle, and is directly related to the operation safety, production continuity and comprehensive maintenance cost. With the increasing integration and intelligence of complex equipment, the physical connection and functional coupling between components are becoming increasingly close, and the fault signals under multiple working conditions show the characteristics of propagation, coupling and uncertainty, and the challenges faced by traditional fault diagnosis techniques are becoming increasingly prominent.

[0003] The main solution relies on traditional data-driven methods and independent diagnosis schemes based on single components. Traditional data-driven methods take machine learning and ordinary deep learning models as the core, extract time domain and frequency domain features of fault signals, build classification models to realize fault identification, and some schemes introduce graph models to model component relationships, but do not deeply integrate device physical structure information. Diagnosis methods based on single components model and analyze each component independently, and judge the fault state by combining artificial experience or fixed rules. In addition, when facing new working conditions, traditional transfer learning methods mostly use the strategy of pre-training with a fixed working condition set and then directly fine-tuning to try to improve model adaptability.

[0004] The existing solution is a multi-task fault diagnosis scheme based on ordinary graph attention network (GAT). This kind of scheme preliminarily models the connection relationship between components through a graph model, realizes multi-component fault collaborative diagnosis by adopting a basic parameter sharing mode, and adapts to different working conditions by means of transfer learning. However, the existing technology still has the following defects: low composite fault recognition accuracy; insufficient multi-task collaborative diagnosis capability; weak cross-working-condition adaptability, difficult to meet the dynamic adaptation demand under multiple working conditions; limited intelligence level, lag response to dynamic changes under complex scenarios, and partial dependence on artificial experience intervention. This is the deficiency of the prior art.

[0005] Therefore, it is necessary to provide a system-level composite diagnosis method and system for complex equipment under multiple working conditions to solve the defects in the prior art. SUMMARY

[0006] The present application aims at the defects of the prior art, such as low accuracy of composite fault recognition, insufficient multi-task collaborative diagnosis capability, weak cross-condition adaptability, difficulty in meeting the demand of dynamic adaptation in multiple conditions, limited intelligent level, lag response to dynamic changes in complex scenarios, and partial dependence on manual experience intervention, and provides a system-level composite diagnosis method and system for complex equipment in multiple conditions to solve the above technical problems.

[0007] To achieve the above object, the present application provides the following technical scheme:

[0008] In a first aspect, the present application provides a system-level composite diagnosis method for complex equipment in multiple conditions, which specifically comprises the following steps:

[0009] Step S1, system structure modeling step, in which:

[0010] Based on the physical connection relationship of the complex equipment, a graph structure model is constructed; the key components are modeled as graph nodes of the graph structure model, the physical connection relationship between the key components is modeled as directed edges of the graph structure model, and the connection topological relationship of the graph structure model is represented in the form of an edge index matrix; and a structured input is provided for subsequent feature extraction;

[0011] Step S2, graph feature extraction step, in which:

[0012] Based on the graph structure model obtained by system structure modeling, a three-layer stacked physical structure fusion graph attention network (PSI-GAT) is used to extract node features and optimize training stability layer by layer through a multi-head attention mechanism, inter-layer batch normalization and residual connection, and then generate global graph-level features through node feature enhancement and attention pooling, thereby providing unified basic feature support for subsequent multi-task collaborative diagnosis;

[0013] Step S3, multi-task collaborative diagnosis step, in which:

[0014] Based on the global graph-level features obtained by graph feature extraction, a multi-task collaborative diagnosis model with a hierarchical structure is constructed, the lower layer provides unified basic feature representation through hard parameter sharing, the upper layer performs task inter soft parameter sharing and complementary information fusion through cross-attention mechanism, and each key component fault diagnosis result is output through feature enhancement and exclusive classification head operation, and all tasks are optimized through a joint loss function, thereby improving the accuracy and efficiency of composite fault diagnosis;

[0015] Step S4, progressive transfer learning step, in which:

[0016] For new operating conditions of complex equipment under multiple working conditions (such as new lines, load modes or component configurations), based on the pre-trained model of the source working condition, a phased parameter unfreezing strategy is adopted, combined with a regularization objective function and a low learning rate fine-tuning, to guide the model to gradually adapt to the target domain data from the classification layer to the feature extraction layer, effectively alleviate the problem of catastrophic forgetting and negative transfer, realize the efficient self-adaptation of the diagnosis model across working conditions, and ensure the accuracy of fault diagnosis under new working conditions.

[0017] Further, the method further comprises the steps of data acquisition and preprocessing, in which steps:

[0018] A plurality of types of sensors, including vibration sensors, current sensors and acoustic sensors, are deployed, including vibration sensors, current sensors and acoustic sensors, and are installed at specified positions of key components of the complex equipment to collect three-axis acceleration, three-phase current, speed and sound intensity and other multi-channel original operating signals; the synchronous acquisition of all sensor signals is realized through a multi-channel data acquisition system to ensure that the data is aligned in the time dimension.

[0019] A fixed-length sliding window is used to segment the collected continuous time series original signals, and each window is used as an independent analysis sample.

[0020] Time domain and frequency domain statistical features are extracted for each signal segment sample, including mean, standard deviation, peak value, root mean square, kurtosis, skewness, waveform factor, pulse factor, etc. Frequency domain features include spectral energy, spectral entropy, main frequency amplitude and energy proportion of key frequency bands, etc.

[0021] The Pearson correlation coefficient between each extracted feature and the fault label of each key component of the equipment is calculated, sorted by correlation score, and the features with strong correlation with faults are selected to form the final feature set of each key component to reduce the data dimension and remove redundant information.

[0022] Z-score standardization processing is performed on the filtered features to eliminate the dimensional differences between different features; the standardization parameters are calculated based on the training set data only and are applied to the validation set and test set simultaneously to avoid data leakage and ensure the stability of subsequent model training. After the above processing, each key component corresponds to a standardized feature vector, which is used as the feature basis for the graph node in the system structure modeling stage.

[0023] Further, the step S1 further comprises:

[0024] Based on the physical composition and functional core of the complex equipment under multiple working conditions, the key components of the complex equipment are selected and determined, and each key component is modeled as a node in the graph structure, and each node is associated with its corresponding physical attributes and operating feature dimensions.

[0025] The complex equipment takes the bogie system as a typical application example, and the key component nodes thereof specifically include a traction motor, a gear box, a left axle box, and a right axle box. The feature basis of each node is a feature vector obtained by preprocessing multi-dimensional physical signals collected in the running process of the complex equipment.

[0026] According to the actual physical connection relationship and signal propagation path among the key components of the complex equipment under multiple working conditions, the rigid connection and power transmission correlation among the components are modeled as edges in a graph structure. Taking the bogie system as an example, the physical connection relationship specifically includes a power transmission physical connection between the traction motor and the gear box, a rigid support physical connection between the gear box and the left axle box, and a rigid support physical connection between the gear box and the right axle box. Each physical connection corresponds to two opposite directed edges in the graph structure to represent the bidirectional propagation characteristics of fault signals among the components of the complex equipment.

[0027] The directed edge connection relationship among the nodes is formally defined by an edge index matrix. The first row elements of the edge index matrix represent the starting node index of each directed edge, the second row elements represent the ending node index of each directed edge, and each column of the matrix corresponds to an independent directed edge. The edge index matrix is specifically as follows:

[0028]

[0029] Wherein, 0 represents the traction motor; 1 represents the gear box, 2 represents the left axle box, and 3 represents the right axle box.

[0030] The edge index matrix completely represents the physical connection topology and fault signal propagation path among the key components of the complex equipment under multiple working conditions, and provides a structured input basis for subsequent feature extraction of the graph neural network.

[0031] Further, the step S2 specifically further includes:

[0032] Step S201, a step of defining the PSI-GAT network structure, in which:

[0033] The PSI-GAT is a three-layer stacked graph attention network. To construct the PSI-GAT, the number of attention heads of each layer and the inter-layer optimization mechanism need to be determined. The first layer adopts an 8-head attention mechanism to capture diversified interaction modes of adjacent nodes. The second layer adopts a 4-head attention mechanism to aggregate associated information of a wider neighborhood. The third layer adopts a single-head attention mechanism for final feature aggregation and refinement. Batch normalization and residual connection are introduced between layers to enhance the stability of model training. The output features of the nodes of the third layer The calculation rule is:

[0034] ​

[0035] wherein, is the intermediate representation of node i, is the residual connection matrix of the layer, is the exponential linear unit activation function, is the batch normalization operation.

[0036] Step S202, a step of generating a node intermediate representation, in which:

[0037] For each node i in the graph structure model and its features , in the kth attention head, the attention score of the node i and the neighbor node j is calculated , and the calculation formula is:

[0038]

[0039] wherein, and are learnable parameters, || represents vector splicing, and LeakyReLU is an activation function.

[0040] The attention score is normalized to obtain the attention weight , and the calculation formula is:

[0041]

[0042] wherein, represents the neighbor set of the node i, is an exponential function;

[0043] Based on the normalized attention weight , the features of the neighbor node j are weighted and summed to obtain the output representation of the node i in the kth attention head , and the calculation formula is:

[0044]

[0045] wherein, is an activation function, is a learnable parameter;

[0046] The outputs of the K attention heads are then combined, transformed by the learnable parameter , and activated by σ to obtain the intermediate representation of the node i , and the calculation formula is:

[0047]

[0048] wherein, is a learnable parameter, is an activation function;

[0049] Step S203, a step of global graph-level feature aggregation, in which:

[0050] Based on 8-head attention mechanism, the attention score calculation, weight normalization and feature aggregation are completed, the intermediate representation is obtained by merging the outputs of the 8 attention heads, and the PSI-GAT first layer node feature is obtained by processing according to the output feature rule of l=1. ;

[0051] Based on 4-head attention mechanism, the above score calculation, normalization and aggregation process are repeated, the intermediate representation is obtained by merging the outputs of the 4 attention heads, and the PSI-GAT second layer node feature is obtained by adding a residual connection according to the output feature rule of l. ;

[0052] Based on single-head attention mechanism, the final feature refinement is completed, the intermediate representation is obtained by aggregation, and the PSI-GAT third layer node feature is obtained by adding a residual connection according to the output feature rule of l. .

[0053] The node feature of the PSI-GAT third layer output is enhanced to obtain the enhanced node feature , and the calculation formula is:

[0054]

[0055] wherein, is a learnable parameter, is a bias parameter, and ReLU is an activation function, is the node feature of the PSI-GAT third layer output, is a normalization function;

[0056] The enhanced node feature is processed through an attention-specific multi-layer perception (MLP_attn) to obtain node attention weight , and the calculation formula is:

[0057]

[0058] wherein, σ is a σ activation function, is a multi-layer perception, is the enhanced node feature;

[0059] Based on the node attention weight, element-wise multiplication and weighted average operation are performed on the enhanced features of all nodes to obtain the global graph-level feature , and the calculation formula is:

[0060]

[0061] where N is the total number of nodes, is the node attention weight, is the enhanced node feature.

[0062] Further, the step S3 specifically further comprises:

[0063] A hierarchical multi-task collaborative diagnosis model is constructed, the lower layer of which adopts a hard parameter sharing manner, and all diagnosis tasks share global graph-level features as unified basic feature representations; the upper layer adopts a soft parameter sharing manner, and information interaction and fusion between different diagnosis tasks are realized through a cross-attention mechanism. The diagnosis tasks aim at fault diagnosis of key components of complex equipment under multiple working conditions, and in a typical application example, correspond to fault diagnosis tasks of four components of a bogie system, i.e., a traction motor, a gear box, a left axle box and a right axle box.

[0064] Based on the global graph-level features, a task-specific linear transformation is performed on each diagnosis task to generate an intermediate feature representation of the task, and the transformation formula is:

[0065]

[0066] wherein, and are learnable parameters of task t, is the intermediate feature representation of task t, is the global graph-level feature.

[0067] For each diagnosis task t, the intermediate feature representation of the task is taken as a query , and the intermediate feature representations of all other diagnosis tasks are taken as keys and values respectively, and the relevance weight between task t and other tasks s is calculated through a cross-attention mechanism, and the calculation formula is:

[0068]

[0069] wherein, is the dimension of the key , and a Softmax function is used to normalize the attention score.

[0070] The output results corresponding to the relevance weights of all other tasks s and the current task t are averaged and aggregated, and after a dimension projection operation (Projection), they are superimposed with the intermediate feature representation of the current task t to obtain an enhanced task feature , and the calculation formula is:

[0071]

[0072] wherein N is the total number of diagnosis tasks, for mapping the fused features to a consistent dimension, wherein the consistent dimension is the intermediate feature representation of the current task t, wherein the intermediate feature representation of the current task t is used to query . .

[0073] wherein the enhanced task features are input into a classification head specific to the task t, wherein the classification head is a neural network, wherein the classification head is activated by an activation function to obtain the final component fault classification result corresponding to the task t, and the calculation formula is:

[0074]

[0075] wherein and are learnable parameters of the classification head specific to the task t, is an activation function, is the component fault classification result corresponding to the task t.

[0076] wherein the classification results of all diagnosis tasks are spliced to obtain a spliced prediction vector, and based on the prediction vector and the corresponding real label vector, a joint loss function is calculated, all learnable parameters of the multi-task collaborative diagnosis model are updated through back propagation, the collaborative optimization of all diagnosis tasks is realized, and the overall diagnosis performance of the model is improved.

[0077] Further, the step S4 further comprises:

[0078] defining the boundaries of source domain and target domain data, wherein the source domain data is complex equipment operation data under multiple working conditions in the pre-training stage, and the pre-training model parameters correspond thereto; the target domain data is:

[0079]

[0080] wherein is the total number of target domain samples, is the input feature of the target domain, is the corresponding real fault label;

[0081] setting a loss function for quantifying the difference between the prediction result and the real label, a balance coefficient λ for regulating the weight of the regularization term, and learning rates at each stage.

[0082] The progressive transfer learning includes three stages.

[0083] The first stage only unfreezes the classification head parameters, specifically including:

[0084] Freezing all level parameters of the physical structure fusion graph attention network (PSI-GAT) in the feature extraction step, while freezing the cross-attention layer parameters in the multi-task collaborative diagnosis step, only unfreezing the learnable parameters of the exclusive classification head corresponding to each diagnosis task in the multi-task collaborative diagnosis step, including the classification head weight parameters and bias parameters ;

[0085] Based on the target domain labeled data Training to quickly adapt the classification head to the fault label distribution under the new working condition as the training target. The optimization objective function used in the training process is:

[0086]

[0087] Where, is the current parameter to be optimized, is the source domain pre-training parameter, is a regularization term used to preserve the effective features learned from the source domain; is the loss function, and λ is the balance coefficient;

[0088] The second stage unfreezes the feature layer and cross-attention layer parameters, specifically including:

[0089] Unfreezing the last two layers of parameters of PSI-GAT in the feature extraction step (i.e. all learnable parameters of the second layer 4-head attention mechanism layer and the third layer single-head attention mechanism layer), while unfreezing the cross-attention layer parameters in the multi-task collaborative diagnosis step, keeping the learnable parameters of the first layer of PSI-GAT in a frozen state;

[0090] Continue to use the target domain labeled data Training, using the target function, update the unfreezed parameters through backpropagation; to adapt the feature extraction module and the cross-task information fusion module to the signal features (such as high-frequency vibration, fault features under special load) of the new working condition.

[0091] The third stage is full parameter end-to-end fine-tuning, specifically including:

[0092] Unfreeze all learnable parameters of the model, including the first layer parameters of PSI-GAT, all parameters of the cross-attention layer, the exclusive classification head parameters of each diagnosis task, and all learnable matrix and vector parameters in the model, and use a low learning rate for training;

[0093] Based on the target domain labeled data End-to-end fine-tuning is performed to continuously optimize all parameters through the target function to ensure that global parameters are adapted to new working conditions in coordination, thereby further improving the diagnostic accuracy and robustness of the model under new working conditions.

[0094] After training, the final diagnostic model adapted to the new working condition is saved, which reuses the effective feature extraction capability of the source domain pre-training while adapting to the fault feature distribution of the target domain new working condition, and can be directly used for real-time fault diagnosis of complex equipment under new working conditions in multiple working conditions.

[0095] Further, the method further comprises a step S5 of outputting the diagnosis result and applying a closed loop, in which:

[0096] The complete diagnostic model optimized by the progressive transfer learning is deployed to the target application scene, and real-time running data of the complex equipment under multiple working conditions is inputted;

[0097] Through the deployed model, the input data is processed, the global graph-level features are extracted by the PSI-GAT, the cross-task information fusion and classification head operation of the multi-task cooperative diagnosis module are performed, and the fault diagnosis results of each key component are outputted. The results are the fault probabilities and corresponding classification labels outputted by the sigma activation function, which directly reflect the health status of each key component;

[0098] Based on the attention weight in the PSI-GAT and the cross-attention mechanism calculation result of the multi-task cooperative diagnosis, the propagation strength and interaction relationship of the fault signals between each key component are represented, the fault source and the related affected components are determined, and the explainability of the diagnosis result is improved.

[0099] The fault diagnosis results and explainability representation information of each key component are outputted synchronously to provide clear component health status reference for equipment maintenance personnel, support maintenance decision making, and realize accurate positioning and efficient disposal of faults.

[0100] In a second aspect, the application provides a system-level composite diagnosis system for complex equipment under multiple working conditions, comprising a system structure modeling module, a graph feature extraction module, a multi-task cooperative diagnosis module and a progressive transfer learning module.

[0101] The system structure modeling module, in which:

[0102] Based on the physical connection relationship of the complex equipment, a graph structure model is constructed, the key components are modeled as graph nodes of the graph structure model, the physical connection relationship between the key components is modeled as directed edges of the graph structure model, and the connection topological relationship of the graph structure model is represented in the form of edge index matrix. The structured input is provided for subsequent feature extraction.

[0103] The graph feature extraction module, in which:

[0104] Based on the graph structure model obtained by system structure modeling, through a three-layer stacked physical structure fusion graph attention network (PSI-GAT), a multi-head attention mechanism, inter-layer batch normalization and residual connection are adopted to extract node features layer by layer and optimize the training stability, and then the node feature enhancement and attention pooling are performed to aggregate and generate global graph-level features, providing unified basic feature support for subsequent multi-task collaborative diagnosis.

[0105] The multi-task collaborative diagnosis module, in which:

[0106] Based on the global graph-level features obtained by graph feature extraction, a multi-task collaborative diagnosis model with hierarchical structure is constructed, the lower layer provides unified basic feature representation through hard parameter sharing, the upper layer performs soft parameter sharing and complementary information fusion between tasks through cross-attention mechanism, and the fault diagnosis results of each key component are output through feature enhancement and exclusive classification head operation, and all tasks are optimized through joint loss function, improving the accuracy and efficiency of composite fault diagnosis.

[0107] The progressive transfer learning module, in which:

[0108] For new operating conditions of complex equipment under multiple operating conditions (such as new lines, load modes or component configurations), based on the pre-trained model of the source operating condition, a phased parameter unfreezing strategy is adopted, combined with the regularization objective function and low learning rate fine-tuning, to guide the model to gradually adapt to the target domain data from the classification layer to the feature extraction layer, effectively alleviate the catastrophic forgetting and negative transfer problems, realize efficient self-adaptation of the diagnosis model across operating conditions, and ensure the accuracy of fault diagnosis under new operating conditions.

[0109] Further, the system further comprises a data acquisition and preprocessing module, in which:

[0110] A variety of sensors are deployed, including vibration sensors, current sensors and acoustic sensors, which are respectively installed at specified positions of key components of complex equipment (such as the drive end and fan end of the traction motor in the bogie system, the input and output shafts of the gear box, the shaft box end cover, etc.), to collect multi-channel raw operating signals such as three-axis acceleration, three-phase current, speed and sound intensity; through a multi-channel data acquisition system, the synchronous acquisition of all sensor signals is realized, ensuring that the data is aligned in the time dimension.

[0111] A fixed-length sliding window is used to segment the collected continuous time series raw signals, and each window is used as an independent analysis sample.

[0112] Time and frequency domain statistical features are extracted for each signal segment sample, including mean, standard deviation, peak value, root mean square, kurtosis, skewness, waveform factor, pulse factor, etc. for time domain features, and spectral energy, spectral entropy, main frequency amplitude and energy proportion of key frequency band, etc. for frequency domain features.

[0113] The Pearson correlation coefficient between each extracted feature and the failure label of each key component of the equipment is calculated, the correlation scores are sorted, and the features with strong correlation with failure are selected to form the final feature set of each key component, so as to reduce the data dimension and remove redundant information.

[0114] The Z-score standardization processing is performed on the screened features to eliminate the dimensional differences between different features; the standardization parameters are calculated based on the training set data only and are applied to the validation set and the test set simultaneously to avoid data leakage and ensure the stability of subsequent model training. After the processing, each key component corresponds to a standardized feature vector, which is used as the feature basis of the graph node in the system structure modeling stage.

[0115] Further, the system structure modeling module specifically further includes:

[0116] Based on the physical composition and functional core of the complex equipment under multiple working conditions, the key components of the complex equipment are screened and determined, and each key component is modeled as a node in the graph structure, and each node is associated with its corresponding physical attribute and operating feature dimension.

[0117] The complex equipment takes a bogie system as a typical application example, and the key component nodes specifically include a traction motor, a gear box, a left axle box and a right axle box, and the feature basis of each node is a feature vector of a multi-dimensional physical signal collected during the operation of the complex equipment after preprocessing.

[0118] According to the actual physical connection relationship and signal propagation path between the key components of the complex equipment under multiple working conditions, the rigid connection and power transmission association between the components are modeled as edges in the graph structure. Taking the bogie system as an example, the physical connection relationship specifically includes: power transmission physical connection between the traction motor and the gear box, rigid support physical connection between the gear box and the left axle box, and rigid support physical connection between the gear box and the right axle box; each physical connection corresponds to two opposite directed edges in the graph structure to represent the bidirectional propagation characteristics of the fault signal between the components of the complex equipment.

[0119] The directed edge connection relationship between the nodes is formally defined by an edge index matrix, the first row elements of the edge index matrix represent the starting node index of each directed edge, the second row elements represent the ending node index of each directed edge, and each column of the matrix corresponds to an independent directed edge. The edge index matrix is specifically:

[0120]

[0121] Wherein, 0 represents the traction motor; 1 represents the gear box, 2 represents the left axle box, and 3 represents the right axle box.

[0122] Through the edge index matrix, the physical connection topologies and fault signal propagation paths among key components of complex equipment under multiple working conditions are completely represented, thereby providing a structured input basis for subsequent feature extraction of a graph neural network.

[0123] Further, the graph feature extraction module specifically further includes:

[0124] A three-layer stacked graph attention network (PSI-GAT) is constructed, and the number of attention heads of each layer and the interlayer optimization mechanism are specified; the first layer adopts an 8-head attention mechanism to capture diversified interaction modes of adjacent nodes; the second layer adopts a 4-head attention mechanism to aggregate associated information of a wider neighborhood; and the third layer adopts a single-head attention mechanism to finally aggregate and refine features; batch normalization and residual connection are introduced between layers to enhance the stability of model training. The output features of the nodes of the first layer The calculation rule is:

[0125]

[0126] wherein, is the intermediate representation of node i, is the residual connection matrix of the lth layer, is an exponential linear unit activation function, is a batch normalization operation.

[0127] For each node i and its feature hi in the graph structure model, in the kth attention head, the attention score of node i and its neighbor node j is calculated as The calculation formula is:

[0128]

[0129] wherein, and are learnable parameters, || represents vector splicing, and LeakyReLU is an activation function.

[0130] The attention score is normalized to obtain the attention weight The calculation formula is:

[0131]

[0132] wherein, Ni represents the neighbor set of node i, and exp is an exponential function.

[0133] Based on the normalized attention weight , the feature of the neighbor node j is weighted and summed to obtain the output representation of node i in the kth attention head​ , the calculation formula is:

[0134]

[0135] wherein, is an activation function, is a learnable parameter;

[0136] The outputs of the K attention heads are then combined, and the intermediate representation of node i is obtained by transforming through the learnable parameter and activating through σ, and the intermediate representation of node i is obtained , the calculation formula is:

[0137]

[0138] wherein, is a learnable parameter, is an activation function;

[0139] Based on the 8-head attention mechanism, the attention score calculation, weight normalization and feature aggregation are completed, and the intermediate representation is obtained by combining the outputs of the 8 attention heads. Then, according to the output feature rule of l=1, the first layer node feature of PSI-GAT is obtained .

[0140] Based on the 4-head attention mechanism, the score calculation, normalization and aggregation process are repeated, and the intermediate representation is obtained by combining the outputs of the 4 attention heads. Then, according to the output feature rule of l attention, the residual connection is added to obtain the second layer node feature of PSI-GAT .

[0141] Based on the single-head attention mechanism, the final feature refinement is completed, and the intermediate representation is obtained by aggregation. Then, according to the output feature rule of l single, the residual connection is added to obtain the third layer node feature of PSI-GAT .

[0142] The node feature output by the third layer of PSI-GAT is enhanced, first performing linear transformation through a learnable parameter and a bias, then activating through ReLU, and then normalizing through LayerNorm to obtain the enhanced node feature , the calculation formula is:

[0143]

[0144] wherein, is a learnable parameter, is a bias, ReLU is an activation function, is the node feature output by the third layer of PSI-GAT;

[0145] The enhanced node features are processed by an attention dedicated multi-layer perception (MLP_attn) to obtain node attention weights through a sigma activation function , and the calculation formula is

[0146]

[0147] wherein, sigma is a sigma activation function, is a multi-layer perception, is an enhanced node feature;

[0148] Based on the node attention weight, element-wise multiplication and weighted average operation are performed on the enhanced features of all nodes to obtain global graph-level features , and the calculation formula is

[0149]

[0150] wherein, N is the total number of nodes, is the node attention weight, is the enhanced node feature.

[0151] Further, the multi-task collaborative diagnosis module specifically further comprises:

[0152] A multi-task collaborative diagnosis model with a hierarchical structure is constructed, the lower layer of the model adopts a hard parameter sharing manner, and all diagnosis tasks share global graph-level features as unified basic feature representation; the upper layer adopts a soft parameter sharing manner, and information interaction and fusion between different diagnosis tasks are realized through a cross-attention mechanism. The diagnosis tasks aim at fault diagnosis of key components of complex equipment under multiple working conditions, and in a typical application example, correspond to fault diagnosis tasks of four components of a traction motor, a gear box, a left axle box and a right axle box of a bogie system, and the total number of tasks N is 4.

[0153] Based on the global graph-level features, a task-specific linear transformation is performed on each diagnosis task to generate an intermediate feature representation of the task, and the transformation formula is

[0154]

[0155] wherein, and are learnable parameters of task t, is an intermediate feature representation of task t, is a global graph-level feature.

[0156] For each diagnosis task t, the intermediate feature representation of the task is taken as a query , and the intermediate feature representations of all other diagnosis tasks are taken as keys and values respectively The relevance weights between task t and other tasks s are calculated using a cross-attention mechanism, and the calculation formula is as follows:

[0157]

[0158] in, For key The Softmax function is used to normalize the attention score in terms of dimensions.

[0159] The outputs corresponding to the relevance weights of all other tasks s with the current task t are averaged and aggregated. After dimensionality projection, these aggregated outputs are superimposed with the intermediate feature representation of the current task t to obtain the enhanced task features. The calculation formula is:

[0160]

[0161] Where N is the total number of diagnostic tasks. Used to map the fused features to... Consistent dimensions Let be the intermediate feature representation of the current task t, and let be the intermediate feature representation of the task. As a query .

[0162] Enhanced task features Input the category header specific to task t, then... After the activation function is applied, the final fault classification result of the component corresponding to task t is obtained. The calculation formula is as follows:

[0163]

[0164] in, and Learnable parameters for a task-specific classification header. For activation function, The component fault classification results for task t.

[0165] The classification results of all diagnostic tasks are concatenated to obtain a concatenated prediction vector. Based on this prediction vector and the corresponding true label vector, a joint loss function is calculated. All learnable parameters of the multi-task collaborative diagnostic model are updated through backpropagation to achieve collaborative optimization of all diagnostic tasks and improve the overall diagnostic performance of the model.

[0166] Furthermore, the progressive transfer learning module specifically includes:

[0167] Defining the data boundary between the source domain and the target domain - the source domain data is the complex equipment operation data under multiple working conditions in the pre-training stage, corresponding to the pre-trained model parameters; the target domain data is:

[0168]

[0169] wherein, is the total number of target domain samples, is the input feature of the target domain, is the corresponding real fault label;

[0170] Setting the loss function (used to quantify the difference between the prediction result and the real label), the balance coefficient λ (regulating the weight of the regularization term) and the learning rate of each stage.

[0171] Progressive transfer learning includes three stages;

[0172] The first stage, only unfreezing the classification head parameters, specifically including:

[0173] Freezing all level parameters of the physical structure fusion graph attention network (PSI-GAT) in the graph feature extraction step, while freezing the cross-attention layer parameters in the multi-task collaborative diagnosis step, only unfreezing the learnable parameters of the exclusive classification head corresponding to each diagnosis task in the multi-task collaborative diagnosis step, including the classification head weight parameters and bias parameters ;

[0174] Based on the target domain labeled data Training to let the classification head quickly adapt to the fault label distribution under the new working condition as the training target, and the optimization objective function used in the training process is:

[0175]

[0176] wherein, is the current parameter to be optimized, is the source domain pre-training parameter, is the regularization term, used to retain the effective features learned from the source domain; is the loss function, and λ is the balance coefficient;

[0177] The second stage, unfreezing the feature layer and cross-attention layer parameters, specifically including:

[0178] Unfreezing the last two layers of parameters of PSI-GAT in the graph feature extraction step (i.e. all learnable parameters of the second layer 4-head attention mechanism layer and the third layer single-head attention mechanism layer), while unfreezing the cross-attention layer parameters in the multi-task collaborative diagnosis step, keeping the learnable parameters of the first layer of PSI-GAT in the frozen state;

[0179] Continuing to use target domain labeled data Training, using the objective function, updates the unfrozen parameters by backpropagation; to let the feature extraction module and the cross-task information fusion module adapt to the signal features of the new working conditions (such as high-frequency vibration, fault features under special load) as the target.

[0180] The third stage is full-parameter end-to-end fine-tuning, specifically including:

[0181] Unfreeze all learnable parameters of the model, including the first layer parameters of PSI-GAT, all parameters of the cross-attention layer, the exclusive classification head parameters of each diagnostic task, and all learnable matrix and vector parameters in the model, and train them with a low learning rate;

[0182] Based on target domain labeled data Perform end-to-end fine-tuning, continuously optimize all parameters through the objective function, and ensure that global parameters adapt to new working conditions; to further improve the diagnostic accuracy and robustness of the model in new working conditions.

[0183] After training is completed, the final diagnostic model adapted to the new working conditions is saved, which reuses the effective feature extraction capability of the source domain pre-training, while adapting to the fault feature distribution of the target domain new working conditions, and can be directly used for real-time fault diagnosis of complex equipment in new working conditions under multiple working conditions.

[0184] Further, the system further comprises a diagnostic result output and application closed loop module, in which:

[0185] Deploy the complete diagnostic model optimized by progressive transfer learning to the target application scenario, input real-time running data of complex equipment under multiple working conditions;

[0186] Through the deployed model, the input data is processed, and after the global graph-level features are extracted by PSI-GAT, the cross-task information fusion and classification head operation of the multi-task collaborative diagnosis module, the fault diagnosis results of each key component are output, the results are the fault probability and corresponding classification label output by the σ activation function, directly reflecting the health status of each key component;

[0187] Based on the attention weight in PSI-GAT and the cross-attention mechanism calculation result of multi-task collaborative diagnosis, the propagation strength and interaction relationship of fault signals between each key component are represented, the fault source and related affected components are determined, and the explainability of the diagnostic result is improved.

[0188] Synchronously output the fault diagnosis results of each key component and the explainability representation information, provide clear component health status reference for equipment maintenance personnel, support maintenance decision making, realize accurate positioning and efficient disposal of faults.

[0189] In a third aspect, the present application also provides an electronic device, comprising at least one processor; and a memory connected with the at least one processor in communication; the memory stores computer program instructions executable by the at least one processor, and the computer program instructions are executed by the at least one processor to enable the at least one processor to execute the system-level composite diagnosis method for complex equipment under multiple working conditions.

[0190] In a fourth aspect, the present application also provides a computer readable storage medium, characterized in that the computer readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the system-level composite diagnosis method for complex equipment under multiple working conditions.

[0191] The present application has the beneficial effects that, by constructing a physical structure integrated graph attention network, the fault signal propagation and coupling effect between components are accurately captured, the composite fault recognition accuracy is improved, the maintenance personnel can quickly locate the fault source and identify the potential concurrent fault risk, by the hierarchical multi-task collaborative diagnosis mechanism, the multi-task collaborative diagnosis capability is strengthened, the synchronous and accurate evaluation of the health state of the key components of the bogie is realized, and the system-level diagnosis performance and fault recognition efficiency are significantly improved, by the progressive transfer learning strategy, the cross-working-condition adaptability of the model is enhanced, the diagnosis model can quickly adapt to different line conditions and load working conditions, an efficient self-learning and self-adaptive mechanism is constructed, the intelligent level of the diagnosis system is significantly improved, the dependence on artificial experience is reduced, and the dynamic change response under complex scenes is more timely. In addition, the present application has reliable design principle, simple structure, and very wide application prospect.

[0192] It can be seen that, compared with the prior art, the present application has outstanding substantial characteristics and significant progress, and the beneficial effects of its implementation are also obvious. BRIEF DESCRIPTION OF DRAWINGS

[0193] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0194] Figure 1 A flowchart of a system-level composite diagnosis method for complex equipment under multiple working conditions is provided for the present embodiment.

[0195] Figure 2 A module schematic diagram of a system-level composite diagnosis system for complex equipment under multiple working conditions is provided for the present embodiment.

[0196] Wherein, 1 - system structure modeling module, 2 - graph feature extraction module, 3 - multi-task collaborative diagnosis module, 4 - progressive transfer learning module. DETAILED DESCRIPTION

[0197] The application will be described in detail below with reference to the accompanying drawings and specific embodiments, and the following embodiments are an explanation of the application, and the application is not limited to the following embodiments.

[0198] Embodiment 1

[0199] The embodiment provides a system-level composite diagnosis method for complex equipment under multiple working conditions, as shown in the figure, which specifically includes the following steps: Figure 1

[0200] Step S1, the step of system structure modeling, in which:

[0201] Based on the physical connection relationship of the complex equipment, a graph structure model is constructed; the key components are modeled as graph nodes of the graph structure model, the physical connection relationship between the key components is modeled as directed edges of the graph structure model, and the connection topological relationship of the graph structure model is formalized by an edge index matrix; and a structured input is provided for subsequent feature extraction;

[0202] Step S2, the step of graph feature extraction, in which:

[0203] Based on the graph structure model obtained by system structure modeling, a three-layer stacked physical structure fusion graph attention network (PSI-GAT) is used to extract node features layer by layer and optimize training stability by using a multi-head attention mechanism, inter-layer batch normalization and residual connection, and then node feature enhancement and attention pooling are performed to aggregate and generate global graph-level features, thereby providing unified basic feature support for subsequent multi-task collaborative diagnosis;

[0204] Step S3, the step of multi-task collaborative diagnosis, in which:

[0205] Based on the global graph-level features obtained by graph feature extraction, a multi-task collaborative diagnosis model with a hierarchical structure is constructed, the lower layer provides a unified basic feature representation through hard parameter sharing, the upper layer performs task inter soft parameter sharing and complementary information fusion through cross attention mechanism, and the feature enhancement and exclusive classification head operation output the fault diagnosis results of each key component, and all tasks are optimized through a joint loss function, thereby improving the accuracy and efficiency of composite fault diagnosis;

[0206] Step S4, the step of progressive transfer learning, in which:

[0207] ​For new operating conditions of complex equipment under multiple working conditions (such as new lines, load modes or component configurations), based on the pre-trained model of the source working condition, a phased parameter unfreezing strategy is adopted, combined with a regularization objective function and a low learning rate fine-tuning, to guide the model to gradually adapt to the target domain data from the classification layer to the feature extraction layer, effectively alleviate the problem of catastrophic forgetting and negative transfer, realize the efficient self-adaptation of the diagnosis model across working conditions, and ensure the accuracy of fault diagnosis under new working conditions.

[0208] Further, the method further comprises the steps of data acquisition and preprocessing, in which steps:

[0209] A plurality of types of sensors, including vibration sensors, current sensors and acoustic sensors, are deployed, and are respectively installed at specified positions of key components of the complex equipment (such as the drive end and fan end of the traction motor in the bogie system, the input and output shafts of the gear box, the shaft box end cover, etc.), and multi-channel original operation signals such as three-axis acceleration, three-phase current, rotational speed and sound intensity are collected; the sampling frequency is set to 25600Hz, and the synchronous collection of all sensor signals is realized through a multi-channel data collection system, so as to ensure that the data is aligned in the time dimension.

[0210] A sliding window with a fixed length of 1024 points is used to segment the collected continuous time series original signals, and each window is used as an independent analysis sample.

[0211] For each signal segment sample, 22 time domain and frequency domain statistical features are extracted, including mean, standard deviation, peak value, root mean square, kurtosis, skewness, waveform factor, pulse factor, etc. in the time domain, and spectral energy, spectral entropy, main frequency amplitude, and energy proportion of key frequency bands of 0-1000Hz, 1000-2000Hz, 2000-4000Hz, etc. in the frequency domain.

[0212] The Pearson correlation coefficient between each extracted feature and the fault label of each key component of the equipment is calculated, sorted according to the correlation score, and the features with strong correlation with the fault are selected to form the final feature set of each key component, so as to reduce the data dimension and remove redundant information.

[0213] The filtered features are subjected to Z-score standardization processing to eliminate the dimensional differences between different features; the standardization parameters are calculated based on the training set data only and are applied to the validation set and the test set simultaneously to avoid data leakage and ensure the stability of subsequent model training. After the processing, each key component corresponds to a standardized feature vector, which serves as the feature basis for the graph nodes in the system structure modeling stage.

[0214] Further, the step S1 specifically further comprises:

[0215] Based on the physical composition and functional core of the complex equipment under multiple working conditions, the key components of the complex equipment are screened and determined, and each key component is modeled as a node in a graph structure, and each node is associated with its corresponding physical attribute and operating characteristic dimension.

[0216] The complex equipment takes a bogie system as a typical application example, and the key component nodes specifically include a traction motor (node index 0), a gear box (node index 1), a left axle box (node index 2), and a right axle box (node index 3), and the characteristic basis of each node is a feature vector obtained by preprocessing a multi-dimensional physical signal collected during the operation of the complex equipment.

[0217] According to the actual physical connection relationship and signal propagation path between the key components of the complex equipment under multiple working conditions, the rigid connection and power transmission association between the components are modeled as edges in the graph structure. Taking the bogie system as an example, the physical connection relationship specifically includes: power transmission physical connection between the traction motor and the gear box, rigid support physical connection between the gear box and the left axle box, and rigid support physical connection between the gear box and the right axle box; each physical connection corresponds to two opposite directed edges in the graph structure to represent the bidirectional propagation characteristics of fault signals between the components of the complex equipment.

[0218] The directed edge connection relationship between the nodes is formally defined by an edge index matrix, the first row elements of the edge index matrix represent the starting node index of each directed edge, the second row elements represent the ending node index of each directed edge, and each column of the matrix corresponds to an independent directed edge. The edge index matrix is specifically:

[0219]

[0220] Wherein, 0 represents the traction motor; 1 represents the gear box, 2 represents the left axle box, and 3 represents the right axle box.

[0221] Through the edge index matrix, the physical connection topology and fault signal propagation path between the key components of the complex equipment under multiple working conditions are completely represented, and a structured input basis is provided for subsequent feature extraction of the graph neural network.

[0222] Further, the step S2 specifically further includes:

[0223] Step S201, the step of defining the PSI-GAT network structure, in which:

[0224] A three-layer stacked graph attention network (PSI-GAT) is constructed, and the number of attention heads in each layer and the optimization mechanism between layers are determined. In the first layer, 8 attention heads are used to capture the diversified interaction patterns of adjacent nodes. In the second layer, 4 attention heads are used to aggregate the correlation information of a wider neighborhood. In the third layer, a single attention head is used for final feature aggregation and refinement. Batch normalization and residual connection are introduced between layers to enhance the stability of model training. the output features of the nodes of the layer The calculation rule is:

[0225]

[0226] wherein, is the intermediate representation of node i, is the residual connection matrix of the lth layer, is the exponential linear unit activation function, is the batch normalization operation.

[0227] Step S202, the step of generating the intermediate representation of the node, in which:

[0228] For each node i and its feature hi in the graph structure model, in the kth attention head, the attention score of the neighbor node j is calculated The calculation formula is:

[0229]

[0230] wherein, and are learnable parameters, || represents vector splicing, and LeakyReLU is an activation function.

[0231] The attention score is normalized to obtain the attention weight The calculation formula is:

[0232]

[0233] wherein, denotes the neighbor set of node i, and exp is the exponential function.

[0234] Based on the normalized attention weight , the feature of the neighbor node j is weighted and summed to obtain the output representation of node i in the kth attention head The calculation formula is:

[0235]

[0236] ​wherein, is an activation function, is a learnable parameter;

[0237] The outputs of the K attention heads are then merged, and the intermediate representation of node i is obtained by performing linear transformation on the merged output through a learnable parameter and then activating it through a σ function. The calculation formula is:

[0238]

[0239] wherein, is a learnable parameter, is an activation function;

[0240] Step S203, the step of aggregating the global graph-level features, in which:

[0241] The attention score calculation, weight normalization and feature aggregation are completed based on the 8-head attention mechanism. After the intermediate representation is obtained by merging the outputs of the 8 attention heads, the output feature rule of l=1 is processed to obtain the first layer node feature of PSI-GAT .

[0242] The score calculation, normalization and aggregation process is repeated based on the 4-head attention mechanism. After the intermediate representation is obtained by merging the outputs of the 4 attention heads, the output feature rule of l attention is added with a residual connection to obtain the second layer node feature of PSI-GAT .

[0243] The final feature refinement is completed based on the single-head attention mechanism. After the intermediate representation is obtained by aggregation, the output feature rule of l single is added with a residual connection to obtain the third layer node feature of PSI-GAT .

[0244] The node feature output by the third layer of PSI-GAT is enhanced. First, a linear transformation is performed on the node feature through a learnable parameter and a bias, and then the enhanced node feature is obtained by performing ReLU activation and LayerNorm normalization , and the calculation formula is:

[0245]

[0246] wherein, is a learnable parameter, is a bias, and ReLU is an activation function, is the node feature output by the third layer of PSI-GAT;

[0247] The enhanced node feature is processed through an attention-specific multi-layer perception (MLP_attn), and the node attention weight is obtained by activating it through a σ function , and the calculation formula is:

[0248]

[0249] where σ is a σ activation function, is a multi-layer perception, is an enhanced node feature;

[0250] Based on the node attention weight, an element-wise multiplication and weighted average operation is performed on the enhanced features of all nodes to obtain global graph-level features , and the calculation formula is:

[0251]

[0252] where N is the total number of nodes, is the node attention weight, is the enhanced node feature.

[0253] Further, the step S3 further includes:

[0254] A hierarchical multi-task collaborative diagnosis model is constructed, the lower layer of the model adopts a hard parameter sharing manner, and all diagnosis tasks share global graph-level features as unified basic feature representation; the upper layer adopts a soft parameter sharing manner, and information interaction and fusion between different diagnosis tasks are realized through a cross-attention mechanism. The diagnosis tasks aim at fault diagnosis of key components of complex equipment under multiple working conditions, and in a typical application example, correspond to fault diagnosis tasks of traction motors, gearboxes, left axle boxes and right axle boxes of bogies, and the total number of tasks N is 4.

[0255] Based on the global graph-level features, a task-specific linear transformation is performed on each diagnosis task to generate an intermediate feature representation of the task, and the transformation formula is:

[0256]

[0257] where, and are learnable parameters of task t, is the intermediate feature representation of task t, is the global graph-level feature.

[0258] For each diagnosis task t, the intermediate feature representation of the task is taken as the query , and the intermediate feature representations of all other diagnosis tasks are taken as the key and the value , respectively, and the correlation weight between task t and other tasks s is calculated through a cross-attention mechanism, and the calculation formula is:

[0259]

[0260] wherein, is a key dimension, and the Softmax function is used to normalize the attention scores.

[0261] The output results corresponding to the relevance weights of all other tasks s and the current task t are averaged and aggregated, and after a dimension projection operation (Projection), the intermediate feature representation of the current task t is superimposed to obtain an enhanced task feature , and the calculation formula is:

[0262]

[0263] wherein, N is the total number of diagnosis tasks, is used to map the fused features to the same dimension as , and is the intermediate feature representation of the current task t, and the intermediate feature representation of the task is taken as the query .

[0264] The enhanced task feature is input into the classification head dedicated to task t, and after the activation function operation, the final component fault classification result corresponding to task t is obtained, and the calculation formula is:

[0265]

[0266] wherein, and are the learnable parameters of the task t dedicated classification head, is an activation function, is the component fault classification result corresponding to task t.

[0267] The classification results of all diagnosis tasks are spliced to obtain a spliced prediction vector; based on the prediction vector and the corresponding real label vector, a joint loss function is calculated, all learnable parameters of the multi-task collaborative diagnosis model are updated through back propagation, and the collaborative optimization of all diagnosis tasks is realized, and the overall diagnosis performance of the model is improved.

[0268] Further, the step S4 further includes:

[0269] Defining the boundary of source domain and target domain data - the source domain data is the complex equipment running data under multiple working conditions in the pre-training stage, corresponding to the pre-trained model parameters; the target domain data is:

[0270]

[0271] ​​wherein, is the total number of target domain samples, is the input feature of the target domain, is the corresponding real fault label;

[0272] The loss function for quantifying the difference between the prediction result and the real label, the balance coefficient λ (regulating the weight of the regularization term) and the learning rate of each stage.

[0273] Progressive transfer learning includes three stages;

[0274] The first stage only unfreezes the classification head parameters, specifically including:

[0275] Freezing all level parameters of the physical structure fusion graph attention network (PSI-GAT) in the graph feature extraction step, while freezing the cross-attention layer parameters in the multi-task collaborative diagnosis step, only unfreezing the learning parameters of the exclusive classification head corresponding to each diagnosis task in the multi-task collaborative diagnosis step, including the classification head weight parameters and bias parameters ;

[0276] Based on the target domain labeled data , the training is carried out to let the classification head quickly adapt to the fault label distribution under the new working condition as the training target. The optimization objective function used in the training process is:

[0277]

[0278] wherein, is the current parameter to be optimized, is the source domain pre-training parameter, is the regularization term, used to preserve the effective features learned from the source domain; is the loss function, and λ is the balance coefficient;

[0279] The second stage unfreezes the feature layer and the cross-attention layer parameters, specifically including:

[0280] Unfreeze the last two layers of parameters of PSI-GAT in the graph feature extraction step (i.e. all learning parameters of the second layer 4-head attention mechanism layer and the third layer single-head attention mechanism layer), while unfreezing the cross-attention layer parameters in the multi-task collaborative diagnosis step, keeping the learning parameters of the first layer of PSI-GAT in a frozen state;

[0281] Continue to use the target domain labeled data for training, follow the target function, and update the unfreezed parameters through back propagation; to let the feature extraction module and the cross-task information fusion module adapt to the signal features (such as high-frequency vibration, fault features under special load) under the new working condition as the target.

[0282] The third stage is full-parameter end-to-end fine-tuning, specifically including:

[0283] All learnable parameters of the unfreezing model are trained with a low learning rate, including the first layer parameters of PSI-GAT, all parameters across attention layers, exclusive classification head parameters of each diagnostic task, and all learnable matrix and vector parameters in the model.

[0284] Based on target domain labeled data End-to-end fine-tuning is performed to continuously optimize all parameters through the objective function, ensuring that global parameters are adapted to new conditions in coordination; and further improving the diagnostic accuracy and robustness of the model in new conditions.

[0285] After training, the final diagnostic model adapted to new conditions is saved, which reuses the effective feature extraction capability of the source domain pre-training while adapting to the fault feature distribution of the target domain new condition, and can be directly used for real-time fault diagnosis of complex equipment in new conditions under multiple conditions.

[0286] Further, the method further includes the step S5 of outputting the diagnosis result and applying a closed loop, in which:

[0287] The complete diagnostic model optimized by progressive transfer learning is deployed to the target application scenario, and real-time running data of the complex equipment under multiple conditions is inputted;

[0288] After processing the input data through the deployed model, the fault diagnosis results of each key component are outputted after the global graph-level features are extracted by PSI-GAT, the cross-task information fusion and classification head operation of the multi-task collaborative diagnosis module, the results being the fault probability and corresponding classification label outputted by the σ activation function, directly reflecting the health status of each key component;

[0289] Based on the attention weight in PSI-GAT and the cross-attention mechanism calculation result of multi-task collaborative diagnosis, the propagation strength and interaction relationship of fault signals between each key component are represented, the fault source and related affected components are determined, and the explainability of the diagnosis result is improved.

[0290] The fault diagnosis results and explainability representation information of each key component are outputted synchronously, providing clear component health status reference for equipment maintenance personnel, supporting maintenance decision making, and realizing accurate positioning and efficient disposal of faults.

[0291] Embodiment 2:

[0292] The embodiment provides a system-level composite diagnosis system for complex equipment under multiple conditions, as shown in Figure 2 The system includes a system structure modeling module 1, a graph feature extraction module 2, a multi-task collaborative diagnosis module 3, and a progressive transfer learning module 4.

[0293] The system structure modeling module 1, in which:

[0294] Based on the physical connection relationship of the complex equipment, a graph structure model is constructed; the key components are modeled as graph nodes of the graph structure model, the physical connection relationship between the key components is modeled as directed edges of the graph structure model, and the connection topological relationship of the graph structure model is formalized in the form of an edge index matrix; and a structured input is provided for subsequent feature extraction;

[0295] The graph feature extraction module 2, in which:

[0296] Based on the graph structure model obtained by system structure modeling, a three-layer stacked physical structure fusion graph attention network (PSI-GAT) is used to extract node features and optimize training stability layer by layer by using a multi-head attention mechanism, inter-layer batch normalization and residual connection, and then node feature enhancement and attention pooling are performed to aggregate and generate global graph-level features, thereby providing unified basic feature support for subsequent multi-task collaborative diagnosis;

[0297] The multi-task collaborative diagnosis module 3, in which:

[0298] Based on the global graph-level features obtained by graph feature extraction, a multi-task collaborative diagnosis model with a hierarchical structure is constructed, the lower layer provides a unified basic feature representation through hard parameter sharing, the upper layer performs task soft parameter sharing and complementary information fusion through cross-attention mechanism, feature enhancement, and exclusive classification head operation to output the fault diagnosis results of each key component, and all tasks are optimized through a joint loss function to improve the accuracy and efficiency of complex fault diagnosis;

[0299] The progressive transfer learning module 4, in which:

[0300] For new operating conditions (such as new lines, load modes or component configurations) of complex equipment under multiple operating conditions, based on the pre-trained model of the source operating condition, a phased parameter unfreezing strategy is adopted, combined with a regularization objective function and a low learning rate fine-tuning, to guide the model to gradually adapt to the target domain data from the classification layer to the feature extraction layer, effectively alleviate the catastrophic forgetting and negative transfer problems, realize efficient self-adaptation of the diagnosis model across operating conditions, and ensure the accuracy of fault diagnosis under new operating conditions.

[0301] Further, the system further comprises a data acquisition and preprocessing module, in which:

[0302] A plurality of types of sensors, including vibration sensors, current sensors, and acoustic sensors, are deployed at designated positions of key components of the complex equipment (such as the drive end and fan end of the traction motor in the bogie system, the input and output shafts of the gearbox, the axle box end cover, etc.) to collect multi-channel raw operating signals such as three-axis acceleration, three-phase current, rotational speed, and sound intensity; the sampling frequency is set to 25600 Hz, and the synchronous collection of all sensor signals is realized through a multi-channel data collection system to ensure that the data is aligned in the time dimension.

[0303] A sliding window with a fixed length of 1024 points is used to segment the collected continuous time-series raw signals, and each window is used as an independent analysis sample.

[0304] For each signal segment sample, 22 time-domain and frequency-domain statistical features are extracted, including mean, standard deviation, peak value, root mean square, kurtosis, skewness, waveform factor, pulse factor, etc. in the time domain, and spectral energy, spectral entropy, main frequency amplitude, and energy proportion of key frequency bands of 0-1000 Hz, 1000-2000 Hz, and 2000-4000 Hz in the frequency domain.

[0305] The Pearson correlation coefficient between each extracted feature and the fault label of each key component of the equipment is calculated, sorted by correlation score, and the features with strong correlation with the fault are selected to form the final feature set of each key component, to reduce the data dimension and remove redundant information.

[0306] The filtered features are subjected to Z-score standardization processing to eliminate the dimensional differences between different features; the standardization parameters are calculated based on the training set data only and are applied to the validation set and test set simultaneously to avoid data leakage and ensure the stability of subsequent model training. After the processing, each key component corresponds to a standardized feature vector, which serves as the feature basis of the graph nodes in the system structure modeling stage.

[0307] Further, the system structure modeling module 1 specifically further includes:

[0308] Based on the physical composition and functional core of the complex equipment under multiple working conditions, the key components of the complex equipment are screened and determined, and each key component is modeled as a node in the graph structure, and each node is associated with its corresponding physical attribute and operating feature dimension.

[0309] The complex equipment takes the bogie system as a typical application example, and the key component nodes specifically include the traction motor (node index 0), the gearbox (node index 1), the left axle box (node index 2), and the right axle box (node index 3), and the feature basis of each node is the feature vector of the multi-dimensional physical signals collected during the operation of the complex equipment after preprocessing.

[0310] According to the actual physical connection relationship and signal propagation path between the key components of the complex equipment under multiple working conditions, the rigid connection, power transmission association and the like between the components are modeled as edges in the graph structure. Taking the bogie system as an example, the physical connection relationship specifically includes: the power transmission physical connection between the traction motor and the gear box, the rigid support physical connection between the gear box and the left axle box, and the rigid support physical connection between the gear box and the right axle box; each physical connection corresponds to two opposite directed edges in the graph structure to represent the bidirectional propagation characteristics of the fault signal between the components of the complex equipment.

[0311] The directed edge connection relationship between the nodes is formally defined by an edge index matrix, the first row elements of the edge index matrix represent the starting node index of each directed edge, the second row elements represent the ending node index of each directed edge, and each column of the matrix corresponds to an independent directed edge. The edge index matrix is specifically:

[0312]

[0313] Wherein, 0 represents the traction motor; 1 represents the gear box, 2 represents the left axle box, and 3 represents the right axle box;

[0314] Through the edge index matrix, the physical connection topology and the fault signal propagation path between the key components of the complex equipment under multiple working conditions are completely represented, providing a structured input basis for subsequent feature extraction of the graph neural network.

[0315] Further, the graph feature extraction module 2 specifically further includes:

[0316] A three-layer stacked graph attention network (PSI-GAT) is constructed, and the number of attention heads of each layer and the interlayer optimization mechanism are specified; wherein the first layer adopts an 8-head attention mechanism to capture the diversified interaction mode of adjacent nodes; the second layer adopts a 4-head attention mechanism to aggregate the associated information of a wider neighborhood; the third layer adopts a single-head attention mechanism for final feature aggregation and refinement; batch normalization and residual connection are introduced between layers to enhance the stability of model training. The output features of the nodes of the first layer The output features of the nodes of the first layer The calculation rule is:

[0317]

[0318] Wherein, is the intermediate representation of node i, is the residual connection matrix of the lth layer, is an exponential linear unit activation function, is a batch normalization operation.

[0319] For each node i and its feature hi in the graph structure model, in the kth attention head, the attention score of the node i and its neighbor node j is calculated , and the calculation formula is

[0320]

[0321] wherein, and are learnable parameters, || represents vector splicing, and LeakyReLU is an activation function.

[0322] The attention score is normalized to obtain the attention weight , and the calculation formula is

[0323]

[0324] wherein, N i represents the neighbor set of node i, and exp is an exponential function.

[0325] Based on the normalized attention weight , the feature of the neighbor node j is weighted and summed to obtain the output representation of node i in the kth attention head , and the calculation formula is

[0326]

[0327] wherein, is an activation function, is a learnable parameter;

[0328] The outputs of the K attention heads are then combined, and the learnable parameter W o is transformed through the sigma activation to obtain the intermediate representation of node i , and the calculation formula is

[0329]

[0330] wherein, is a learnable parameter, is an activation function;

[0331] Based on the 8-head attention mechanism, the attention score calculation, weight normalization and feature aggregation are completed, the intermediate representation is obtained after the outputs of the 8 attention heads are combined, and the output feature rule of l=1 is processed to obtain the first layer node feature of the PSI-GAT ;

[0332] The score calculation, normalization and aggregation process is repeated based on the 4 attention mechanisms, the intermediate representation is obtained by merging the outputs of the 4 attention heads, and the residual connection is added according to the output feature rules of the l attention to obtain the second layer node feature of the PSI-GAT ;

[0333] The final feature refinement is completed based on a single-head attention mechanism, and the residual connection is added according to the output feature rules of the l single after the intermediate representation is aggregated to obtain the third layer node feature of the PSI-GAT .

[0334] The node feature output by the third layer of the PSI-GAT is enhanced, first through linear transformation by a learnable parameter and a bias, then activated by ReLU, and then normalized by LayerNorm to obtain the enhanced node feature , the calculation formula is:

[0335]

[0336] wherein, is a learnable parameter, is a bias, ReLU is an activation function, is the node feature output by the third layer of the PSI-GAT;

[0337] The enhanced node feature is processed by an attention-specific multi-layer perception (MLP_attn), and the node attention weight is obtained by a σ activation function , the calculation formula is:

[0338]

[0339] wherein, σ is a σ activation function, is a multi-layer perception, is the enhanced node feature;

[0340] Based on the node attention weight, element-wise multiplication and weighted average operation are performed on the enhanced features of all nodes to obtain the global graph-level feature , the calculation formula is:

[0341]

[0342] wherein, N is the total number of nodes, is the node attention weight, is the enhanced node feature.

[0343] Further, the multi-task cooperative diagnosis module 3 specifically further comprises:

[0344] A hierarchical multi-task collaborative diagnosis model is constructed, wherein a hard parameter sharing manner is used in the lower layer, and all diagnosis tasks share global graph-level features as unified basic feature representation; a soft parameter sharing manner is used in the upper layer, and information interaction and fusion between different diagnosis tasks are realized through a cross-attention mechanism. The diagnosis tasks aim at fault diagnosis of key components of complex equipment under multiple working conditions. In a typical application example, the fault diagnosis tasks of four components of a bogie system, i.e., traction motor, gear box, left axle box and right axle box, are performed, and the total number N of tasks is 4.

[0345] Based on the global graph-level features, a task-specific linear transformation is performed on each diagnosis task to generate the intermediate feature representation of the task, and the transformation formula is as follows:

[0346]

[0347] wherein, and are learnable parameters of task t, is the intermediate feature representation of task t, is the global graph-level feature.

[0348] For each diagnosis task t, the intermediate feature representation of the task is taken as the query , and the intermediate feature representations of all other diagnosis tasks are taken as the keys and the values , respectively. The correlation weight between task t and other tasks s is calculated through a cross-attention mechanism, and the calculation formula is as follows:

[0349]

[0350] wherein, is the dimension of the key , and a Softmax function is used to normalize the attention score.

[0351] The output results corresponding to the correlation weights of all other tasks s and the current task t are averaged and aggregated, and after a dimension projection operation, the enhanced task feature is obtained by superimposing the intermediate feature representation of the current task t, and the calculation formula is as follows:

[0352]

[0353] wherein, N is the total number of diagnosis tasks, is used to map the fused feature to the same dimension as , is the intermediate feature representation of the current task t, and the intermediate feature representation of the task As a query .

[0354] The enhanced task features are input into the classification head specific to task t, and after an activation function operation, the final component failure classification result corresponding to task t is obtained, and the calculation formula is:

[0355]

[0356] wherein, and are the learnable parameters of the classification head specific to task t, is an activation function, is the component failure classification result corresponding to task t.

[0357] The classification results of all diagnostic tasks are spliced to obtain a spliced prediction vector; based on the prediction vector and the corresponding real label vector, a joint loss function is calculated, all learnable parameters of the multi-task collaborative diagnosis model are updated through back propagation, the collaborative optimization of all diagnostic tasks is realized, and the overall diagnostic performance of the model is improved.

[0358] Further, the progressive transfer learning module 4 specifically further includes:

[0359] Defining the boundary of source domain and target domain data - source domain data is complex equipment running data under multiple working conditions in the pre-training stage, corresponding to the pre-training model parameters; target domain data is:

[0360]

[0361] wherein, is the total number of target domain samples, is the target domain input feature, is the corresponding real failure label;

[0362] Setting the loss function (used to quantify the difference between the prediction result and the real label), the balance coefficient λ (regulating the weight of the regularization term) and the learning rate of each stage.

[0363] The progressive transfer learning includes three stages.

[0364] The first stage, only unfreezing the classification head parameters, specifically includes:

[0365] Freezing all level parameters of the physical structure fusion graph attention network (PSI-GAT) in the feature extraction step, while freezing the cross-attention layer parameters in the multi-task collaborative diagnosis step, only unfreezing the learnable parameters of the exclusive classification head corresponding to each diagnostic task in the multi-task collaborative diagnosis step, including the classification head weight parameters and bias parameters​ ;

[0366] based on the target domain labeled data The training is performed to quickly adapt the classification head to the fault label distribution under the new working condition as a training target. An optimization objective function used in the training process is as follows:

[0367]

[0368] wherein, is a current parameter to be optimized, is a source domain pre-training parameter, is a regularization term, used to retain effective features learned from the source domain; is a loss function, and λ is a balance coefficient;

[0369] In the second stage, the parameters of the feature layer and the cross-attention layer are unfrozen, specifically including:

[0370] The last two layers of parameters of PSI-GAT in the feature extraction step (i.e., all learnable parameters of the second layer of 4-head attention mechanism layer and the third layer of single-head attention mechanism layer) are unfrozen, and the parameters of the cross-attention layer in the multi-task collaborative diagnosis step are also unfrozen, while the learnable parameters of the first layer of PSI-GAT are kept in a frozen state;

[0371] continue to use the target domain labeled data for training, and the target function is followed to update the unfrozen parameters through back propagation; so as to adapt the feature extraction module and the cross-task information fusion module to the signal features (such as high-frequency vibration and fault features under special load) under the new working condition.

[0372] In the third stage, all parameters are end-to-end fine-tuned, specifically including:

[0373] All learnable parameters of the model are unfrozen, including the first layer parameters of PSI-GAT, all parameters of the cross-attention layer, the dedicated classification head parameters of each diagnostic task, and all learnable matrix and vector parameters in the model, and a low learning rate is used for training;

[0374] based on the target domain labeled data end-to-end fine-tuning is performed, all parameters are continuously optimized through the target function, and the global parameters are collaboratively adapted to the new working condition; to further improve the diagnostic accuracy and robustness of the model under the new working condition.

[0375] After the training is completed, the final diagnostic model adapted to the new working condition is saved, which reuses the effective feature extraction capability of the source domain pre-training, and simultaneously adapts to the fault feature distribution of the target domain new working condition, and can be directly used for real-time fault diagnosis of complex equipment under the new working condition in multiple working conditions.

[0376] Further, the system further comprises a diagnostic result output and application closed loop module, in which:

[0377] The complete diagnostic model optimized by the progressive transfer learning is deployed to a target application scenario, and real-time operation data of a complex device under multiple working conditions is inputted;

[0378] The input data is processed by the deployed model, global graph-level features are extracted by the PSI-GAT, cross-task information fusion and classification head operation are performed by the multi-task cooperative diagnosis module, and then fault diagnosis results of each key component are outputted, the results being fault probabilities and corresponding classification labels outputted by the sigma activation function, directly reflecting the health status of each key component;

[0379] Based on the attention weight in the PSI-GAT and the cross-attention mechanism calculation result of the multi-task cooperative diagnosis, the propagation strength and interaction relationship of fault signals between each key component are represented, the fault source and related affected components are determined, and the explainability of the diagnostic result is improved.

[0380] The fault diagnosis results of each key component and the explainability representation information are synchronously outputted, providing clear component health status reference for equipment maintenance personnel, supporting maintenance decision making, and realizing accurate positioning and efficient disposal of faults.

[0381] In a third aspect, an electronic device is provided, including at least one processor, and a memory communicatively connected with the at least one processor, the memory storing computer program instructions executable by the at least one processor, and the computer program instructions are executed by the at least one processor to enable the at least one processor to execute the system-level composite diagnosis method for a complex device under multiple working conditions.

[0382] In a fourth aspect, a computer readable storage medium is provided, characterized in that the computer readable storage medium stores computer instructions, and the computer instructions enable the computer to execute the system-level composite diagnosis method for a complex device under multiple working conditions.

[0383] The above disclosure is only the preferred embodiments of the present application, but the present application is not limited thereto, any non-creative changes and several improvements and refinements made by those skilled in the art without departing from the principles of the present application should fall within the protection scope of the present application.

Claims

1. A system-level composite diagnostic method for complex equipment under multiple operating conditions, characterized in that, Includes the following steps: Step S1, the system structure modeling step, in which: Based on the physical connection relationships of complex devices, a graph structure model is constructed; key components are modeled as graph nodes of the graph structure model, and the physical connection relationships between key components are modeled as directed edges of the graph structure model. The connection relationships of the graph structure model are then formally represented by an edge index matrix. Step S2, the graph feature extraction step, in which: Based on the graph structure model obtained from system structure modeling, a three-layer stacked PSI-GAT is used to extract node features layer by layer and optimize training stability through multi-head attention mechanism, inter-layer batch normalization and residual connection. Then, after node feature enhancement and attention pooling, global graph-level features are aggregated to generate global graph features. Step S3, the multi-task collaborative diagnosis step, in which: Based on global graph-level features obtained from graph feature extraction, a hierarchical multi-task collaborative diagnostic model is constructed. The lower layer provides unified basic features through hard parameter sharing, while the upper layer performs soft parameter sharing and complementary information fusion between tasks through cross-attention mechanism. After feature enhancement and dedicated classification head operation, the fault diagnosis results of each key component are output, and all tasks are collaboratively optimized through joint loss function. Step S4, the progressive transfer learning step, in which: Based on the pre-trained model under the source conditions, a phased parameter unfreezing strategy is adopted, combined with a regularized objective function and low learning rate fine-tuning, to guide the model to gradually adapt to the target domain data from the classification layer to the feature extraction layer.

2. The method according to claim 1, characterized in that, Step S1 further includes: Based on the physical composition and functional core of complex equipment under multiple working conditions, the key components of complex equipment are screened and determined. Each key component is modeled as a node in a graph structure, and each node is associated with its corresponding physical attributes and operational feature dimensions. Based on the actual physical connection relationship and signal propagation path between key components of complex equipment under multiple working conditions, the rigid connection and power transmission relationship between components are modeled as directed edges in a graph structure model; The directed edge connection relationship between the nodes is formally defined by the edge index matrix. The first row of the edge index matrix represents the starting node index of each directed edge, the second row represents the ending node index of each directed edge, and each column of the matrix corresponds to an independent directed edge.

3. The method according to claim 2, characterized in that, Step S2 further includes: A three-layer stacked PSI-GAT is constructed, and the configuration of the number of attention heads in each layer and the inter-layer optimization mechanism are defined. The first layer adopts an 8-head attention mechanism, the second layer adopts a 4-head attention mechanism, and the third layer adopts a single-head attention mechanism. Batch normalization and residual connections are introduced between layers. No. Layer nodes Output characteristics The calculation rules are as follows: in, Let i be the intermediate representation of node i. Let l be the residual connection matrix of the l-th layer. It is the activation function of the exponential linear unit. This is for batch normalization operations.

4. The method according to claim 3, characterized in that, Step S2 further includes: For each node i and its features in the graph structure model In the k-th attention head, calculate its attention score with its neighbor node j. The calculation formula is: in, and Here are the learnable parameters, || denotes vector concatenation, and LeakyReLU is the activation function; Attention score Normalization is performed to obtain attention weights. ; Features of neighbor node j We perform a weighted summation to obtain the output representation of node i at the k-th attention head. The calculation formula is: in, For activation function, These are learnable parameters; The outputs of K attention heads are combined and processed by learnable parameters. After transformation, the intermediate representation of node i is obtained through σ activation. The calculation formula is: In the formula, For learnable parameters, This is the activation function.

5. The method according to claim 4, characterized in that, Step S2 further includes: The node features output from the third layer of PSI-GAT are enhanced by performing a linear transformation using learnable parameters and biases, followed by ReLU activation and then LayerNorm normalization to obtain the enhanced node features. The calculation formula is: in, For learnable parameters, For bias, ReLU is the activation function. The node features are output from the third layer of PSI-GAT; The enhanced node features are processed using an attention-specific multilayer perceptron, and the node attention weights are obtained through a σ activation function. The calculation formula is: Where σ is the σ activation function, It is a multilayer perceptron. Enhanced node features; Based on node attention weights, element-wise multiplication and weighted averaging are performed on the enhanced features of all nodes to obtain global graph-level features. The calculation formula is: Where N is the total number of nodes. For node attention weights, This refers to the enhanced node features.

6. The method according to claim 5, characterized in that, Step S3 further includes: A hierarchical multi-task collaborative diagnostic model is constructed. Based on global graph-level features, a task-specific linear transformation is performed on each diagnostic task to generate an intermediate feature representation for that task. The transformation formula is as follows: in, and Let be the learnable parameters for task t. Let be the intermediate feature representation of task t. This is a global graph-level feature; For each diagnostic task t, the intermediate features of that task are represented as... As a query Represented by intermediate features of all other diagnostic tasks Each as a key Sum The relevance weights between task t and other tasks s are calculated using a cross-attention mechanism, and the calculation formula is as follows: in, For key In terms of dimensions, the Softmax function is used to normalize the attention score; The outputs corresponding to the relevance weights of all other tasks s with the current task t are averaged and aggregated. After dimensionality projection, these aggregated outputs are superimposed with the intermediate feature representation of the current task t to obtain the enhanced task features. The calculation formula is: Where N is the total number of diagnostic tasks. Map the fused features to... Consistent dimensions Let be the intermediate feature representation of the current task t, and let be the intermediate feature representation of the task. As a query ; Enhanced task features Input the category header specific to task t, then... After the activation function is applied, the final fault classification result of the component corresponding to task t is obtained. The calculation formula is as follows: in, and Learnable parameters for a task-specific classification header. For activation function, The component fault classification results corresponding to task t; The classification results of all diagnostic tasks are concatenated to obtain the concatenated prediction vector; based on this prediction vector and the corresponding true label vector, the joint loss function is calculated.

7. The method according to claim 6, characterized in that, Step S4 further includes: Progressive transfer learning consists of three stages; In the first stage, all hierarchical parameters of the physical structure fusion graph attention network in the graph feature extraction step are frozen, and the cross-attention layer parameters in the multi-task collaborative diagnosis step are also frozen. Only the learnable parameters of the dedicated classification head corresponding to each diagnostic task in the multi-task collaborative diagnosis step are unfrozen, including the classification head weight parameters. With bias parameters ; Based on target domain labeled data The training process uses the following objective function for optimization: in, These are the parameters to be optimized. Pre-trained parameters for the source domain, This is a regularization term used to preserve the effective features learned from the source domain; Let λ be the loss function and λ be the balance coefficient. In the second stage, the parameters of the last two layers of PSI-GAT in the feature extraction step are unfrozen, and the cross-attention layer parameters in the multi-task collaborative diagnosis step are unfrozen, while keeping the learnable parameters of the first layer of PSI-GAT in a frozen state. Continue using target domain labeled data Training is performed using the same objective function, updating the unfrozen parameters via backpropagation; In the third stage, all learnable parameters of the model are unfrozen, including the first layer parameters of PSI-GAT, all parameters across the attention layer, the specific classification head parameters for each diagnostic task, and all learnable matrix and vector parameters in the model, and training is performed using a low learning rate. Based on target domain labeled data Perform end-to-end fine-tuning, continuously optimizing all parameters through the objective function; after training is complete, save the final diagnostic model adapted to the new working conditions.

8. A system-level composite diagnostic system for complex equipment under multiple operating conditions, comprising a system structure modeling module, a graph feature extraction module, a multi-task collaborative diagnostic module, and a progressive transfer learning module; The system structure modeling module, in which: Based on the physical connection relationships of complex devices, a graph structure model is constructed; key components are modeled as graph nodes of the graph structure model, and the physical connection relationships between key components are modeled as directed edges of the graph structure model. The connection relationships of the graph structure model are then formally represented by an edge index matrix. The graph feature extraction module, in which: Based on the graph structure model obtained from system structure modeling, a three-layer stacked PSI-GAT is used to extract node features layer by layer and optimize training stability through multi-head attention mechanism, inter-layer batch normalization and residual connection. Then, after node feature enhancement and attention pooling, global graph-level features are aggregated to generate global graph features. The multi-task collaborative diagnosis module, in which: Based on global graph-level features obtained from graph feature extraction, a hierarchical multi-task collaborative diagnostic model is constructed. The lower layer provides unified basic features through hard parameter sharing, while the upper layer performs soft parameter sharing and complementary information fusion between tasks through cross-attention mechanism. After feature enhancement and dedicated classification head operation, the fault diagnosis results of each key component are output, and all tasks are collaboratively optimized through joint loss function. The progressive transfer learning module, in which: For new operating conditions of complex equipment under multiple operating conditions, a phased parameter unfreezing strategy is adopted based on the source operating condition pre-trained model. Combined with regularized objective function and low learning rate fine-tuning, the model is guided to gradually adapt to the target domain data from the classification layer to the feature extraction layer.

9. The system according to claim 8, characterized in that, The system structure modeling module further includes: Based on the physical composition and functional core of complex equipment under multiple working conditions, the key components of complex equipment are screened and determined. Each key component is modeled as a node in a graph structure, and each node is associated with its corresponding physical attributes and operational feature dimensions. Based on the actual physical connection relationship and signal propagation path between key components of complex equipment under multiple working conditions, the rigid connection and power transmission relationship between components are modeled as directed edges in a graph structure model; The directed edge connection relationship between the nodes is formally defined by the edge index matrix. The first row of the edge index matrix represents the starting node index of each directed edge, the second row represents the ending node index of each directed edge, and each column of the matrix corresponds to an independent directed edge.

10. The system according to claim 9, characterized in that, The graph feature extraction module further includes: Construct a three-layer stacked PSI-GAT, clarifying the attention head configuration of each layer and the inter-layer optimization mechanism; Layer nodes Output characteristics The calculation rules are as follows: in, Let i be the intermediate representation of node i. Let l be the residual connection matrix of the l-th layer. It is the activation function of the exponential linear unit. For batch normalization operations; For each node i and its features in the graph structure model In the k-th attention head, calculate its attention score with its neighbor node j. The calculation formula is: in, and Here are the learnable parameters, || denotes vector concatenation, and LeakyReLU is the activation function; Attention score Normalization is performed to obtain attention weights. Features of neighbor node j We perform a weighted summation to obtain the output representation of node i at the k-th attention head. ; Combine the outputs of K attention heads, and then use learnable parameters After transformation Activate to obtain the intermediate representation of node i. The calculation formula is: in, For learnable parameters, For activation functions; The node features output from the third layer of PSI-GAT are enhanced to obtain the enhanced node features. The enhanced node features are processed using a dedicated attention-based multilayer perceptron, and the node attention weights are obtained through a σ activation function. Element-wise multiplication and weighted averaging are performed on the enhanced features of all nodes to obtain global graph-level features. The calculation formula is: Where N is the total number of nodes. For node attention weights, This refers to the enhanced node features.

Images