Explanatable fault diagnosis method and system integrating graph network and knowledge driving

Abstract

The invention relates to the technical field of fault diagnosis, and provides a graph network and knowledge driven interpretable fault diagnosis method and system, and the method comprises the following steps: extracting intra-class and inter-class features for obtained operation data, and constructing a data driving graph; constructing a knowledge graph based on prior knowledge, constructing a knowledge-driven graph, and fusing the data-driven graph and the knowledge-driven graph into a unified graph structure; constructing a plurality of graph attention network sub-models, and identifying and fusing the obtained unified graph structures to obtain a diagnosis result; and performing interpretability analysis on the diagnosis result based on an input feature gradient method and a graph neural network interpreter to obtain a comprehensive diagnosis result. According to the method, more accurate, more stable and more explainable fault diagnosis and decision support is provided through data driving, knowledge driving, intra-class and inter-class feature joint extraction, multi-model fusion and fusion of an explainable analysis method.

Description

Technical Field

[0001] The present disclosure relates to the technical field of fault diagnosis, and more specifically, to an interpretable fault diagnosis method and system that integrates graph networks and knowledge-driven approaches. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] In the field of modern architecture, the heating, ventilation, and air conditioning (HVAC) system undertakes the important tasks of regulating indoor environmental comfort and energy efficiency. With the promotion of energy conservation, emission reduction, and green building concepts, the energy efficiency level and operation reliability of the HVAC system have gradually become important indicators for measuring building performance. On the one hand, the system efficiency is directly related to the overall energy consumption and carbon emission level of the building; on the other hand, abnormalities in the system operation state may affect the health and comfort of the occupants. Therefore, real-time monitoring and fault diagnosis of the HVAC system operation state have become a key technical path for improving building energy management capabilities and indoor environmental quality.

[0004] The inventors found in their research that current fault diagnosis methods still have many limitations. Mainstream technologies mostly rely on time series analysis, single modeling means, and strategies based on empirical rules and threshold settings. Such methods have certain practicality under single working conditions, but their performance is often limited when facing complex multi-source heterogeneous data and diverse fault scenarios. Specific problems include: (1) Data and knowledge are fragmented, and feature extraction is one-sided: Traditional methods usually rely on a single path of data-driven or knowledge-driven approaches, making it difficult to effectively integrate information from different sources, resulting in features that fail to take into account both intra-class temporal differences and inter-class discriminability; (2) The model is single, and the working condition adaptability is poor: The diagnostic mechanism based on a single model lacks flexibility, making it difficult to adapt to changes under different working conditions and lacking the ability to fuse and switch across scenarios; (3) The diagnosis is black-box, and the interpretability is insufficient: Many methods only output fault labels, which limits the decision-making efficiency of maintenance personnel and the intelligent closed-loop control ability of the system. Summary of the Invention

[0005] To solve the above problems, the present disclosure proposes an interpretable fault diagnosis method and system that integrates graph networks and knowledge-driven approaches, which is a systematic diagnosis method that combines data-driven and knowledge-driven, jointly extracts intra-class and inter-class features, fuses multiple models, and performs interpretability analysis, providing more accurate, robust, and interpretable fault diagnosis and decision-making support for systems (such as HVAC systems) for fault diagnosis.

[0006] To achieve the above object, the present disclosure adopts the following technical solutions:

[0007] One or more embodiments provide an interpretable fault diagnosis method that fuses graph networks with knowledge-driven, including the following steps:

[0008] Extract intra-class and inter-class features from the acquired operation data to construct a data-driven graph; construct a knowledge-driven graph based on prior knowledge to build a knowledge graph, and fuse the data-driven graph with the knowledge-driven graph into a unified graph structure;

[0009] Construct multiple graph attention network sub-models, respectively identify and fuse the obtained unified graph structure to obtain a diagnosis result;

[0010] Use the input feature gradient method to analyze the diagnosis result to obtain a diagnosis result heat map, process the unified graph structure based on the graph neural network to extract key sub-graphs, obtain a visualized fault propagation path, and fuse the propagation path and the diagnosis result heat map to obtain a comprehensive diagnosis result.

[0011] One or more embodiments provide an interpretable fault diagnosis system that fuses graph networks with knowledge-driven, including:

[0012] A graph construction module based on data-driven and knowledge-driven, configured to: extract intra-class and inter-class features from the acquired operation data to construct a data-driven graph; construct a knowledge-driven graph based on prior knowledge to build a knowledge graph, and fuse the data-driven graph with the knowledge-driven graph into a unified graph structure;

[0013] A fault diagnosis module based on multi-model graph attention network ensemble learning, configured to construct multiple graph attention network sub-models, respectively identify and fuse the obtained unified graph structure to obtain a diagnosis result;

[0014] A diagnosis result interpretability analysis module, configured to use the input feature gradient method to analyze the diagnosis result to obtain a diagnosis result heat map, process the unified graph based on the graph neural network to extract key sub-graphs, obtain a visualized fault propagation path, and fuse the propagation path and the diagnosis result heat map to obtain a comprehensive diagnosis result.

[0015] One or more embodiments provide an interpretable fault diagnosis system that fuses graph networks with knowledge-driven, including a data acquisition device and a processor;

[0016] The data acquisition device includes sensors and a device monitoring terminal, and is used to acquire the operation data of the fault diagnosis object;

[0017] The processor is configured to execute the steps in the above-mentioned interpretable fault diagnosis method that fuses graph networks with knowledge-driven.

[0018] An electronic device includes a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the steps in the above-mentioned method for fusing graph network and knowledge-driven interpretable fault diagnosis are completed.

[0019] A computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the steps in the above-mentioned method for fusing graph network and knowledge-driven interpretable fault diagnosis are completed.

[0020] Compared with the prior art, the beneficial effects of the present disclosure are as follows:

[0021] The present disclosure breaks through the limitations of traditional methods based on a single data source or a single model, integrates the advantages of data-driven and knowledge-driven approaches, generates a system state representation with multi-layer semantics, and significantly enhances the ability to perceive complex interaction relationships of faults. On this basis, a flexible diagnosis suitable for multiple working conditions and various fault types is achieved through a multi-model graph attention network structure, greatly improving the robustness and accuracy of diagnosis. In addition, by combining feature gradient visualization and graph structure interpretability mechanisms, the diagnosis results not only have high accuracy but also are traceable and transparent, which helps to quickly locate the root cause of faults and assist maintenance personnel in making decisions.

[0022] The advantages of the present disclosure and the advantages of additional aspects will be described in detail in the following specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of the disclosure. The schematic embodiments and descriptions thereof of the disclosure are used to explain the disclosure and do not constitute a limitation to the disclosure. <Q

[0024] Figure 1 is a block diagram of a system for fusing graph network and knowledge-driven interpretable fault diagnosis according to Embodiment 1 of the present disclosure;

[0025] Figure 2 is a schematic diagram of the processing process of a graph construction module based on data-driven and knowledge-driven according to Embodiment 1 of the present disclosure;

[0026] Figure 3 is a schematic diagram of the processing process of a fault diagnosis module based on multi-model graph attention network integrated learning according to Embodiment 1 of the present disclosure;

[0027] Figure 4 is a schematic diagram of the processing process of an interpretability analysis module for diagnosis results according to Embodiment 1 of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure pertains.

[0030] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features in the present disclosure may be combined with each other. The embodiments will be described in detail below with reference to the drawings.

[0031] Embodiment 1

[0032] In the technical solutions disclosed in one or more embodiments, as Figures 1 to 4 shown, an interpretable fault diagnosis system integrating graph network and knowledge-driven includes:

[0033] A graph construction module based on data-driven and knowledge-driven, configured to: extract intra-class and inter-class features from the acquired operation data to construct a data-driven graph; construct a knowledge graph based on prior knowledge to construct a knowledge-driven graph, and fuse the data-driven graph and the knowledge-driven graph into a unified graph structure;

[0034] A fault diagnosis module based on multi-model graph attention network ensemble learning, configured to construct multiple graph attention network sub-models, respectively identify and fuse the obtained unified graph structure to obtain a diagnosis result;

[0035] A diagnostic result interpretability analysis module, configured to use the input feature gradient method to analyze the diagnostic result to obtain a diagnostic result heat map, process the unified graph based on a graph neural network to extract a key sub-graph, obtain a visualized fault propagation path, and fuse the propagation path and the diagnostic result heat map to obtain a comprehensive diagnostic result;

[0036] In this embodiment, the system breaks through the limitations of traditional single data source or single model, integrates the advantages of data-driven and knowledge-driven, generates a system state expression with multi-layer semantics, and significantly enhances the perception ability of complex fault interaction relationships. On this basis, a multi-model graph attention network structure is used to achieve flexible diagnosis adaptable to multiple working conditions and various fault types, greatly improving the robustness and accuracy of diagnosis. In addition, combined with feature gradient visualization and graph structure interpretability mechanism, the diagnosis results not only have high accuracy, but also have traceability and transparency, which helps to quickly locate the root cause of the fault and assist maintenance personnel in making decisions.

[0037] When performing fault diagnosis on a heating, ventilation, and air conditioning (HVAC) system, one often has to face multi-source sensor data, highly coupled equipment components, and complex and changeable operating environments. Due to the possible presence of noise, incomplete data, and multiple potential fault types in actual working conditions, it is difficult to comprehensively reveal the internal fault mechanism of the system by relying solely on a single information source or feature extraction method. Therefore, this embodiment adopts a graph construction strategy that combines data-driven and knowledge-driven, extracts intra-class and inter-class fault features respectively, and integrates prior knowledge and data features, thereby effectively reducing the interference of noise and redundant information and providing clear and multi-dimensional inputs for subsequent multi-model diagnosis.

[0038] In some embodiments, a graph construction module based on data-driven and knowledge-driven is configured to perform the following process:

[0039] Step 11: Obtain multi-source time-series operation data, use a deep learning model to extract intra-class features, and process prior knowledge through discriminant analysis methods to obtain inter-class features;

[0040] Specifically, collecting multi-source time-series operation data from the HVAC system may include data such as temperature, pressure, flow rate, and equipment start-stop records.

[0041] Optionally, the deep learning model can adopt a Transformer model for intra-class feature extraction. The Transformer model introducing a multi-head self-attention mechanism performs multi-dimensional attention weighting on the input time-series sequence {x1, x2,..., x T} and outputs a feature vector that can represent the subtle differences of the same fault type in different time intervals. The output of the Transformer is expressed as:

[0042] [head1, head2,..., head h W O (1)

[0043] where head i represents the calculation result of the i-th attention head, [·] is the concatenation operation, and W Oare learnable linear transformation parameters. Through the multi-head attention mechanism, the model can focus on different subspaces of the time series data under the same fault mode, so as to obtain intra-class features.

[0044] Optionally, for inter-class feature discrimination, the Discriminative Slow Feature Analysis (DSFA) method can be adopted. By defining the optimization objectives of increasing the inter-class divergence and decreasing the intra-class divergence, discriminative features that can effectively distinguish different fault classes are extracted;

[0045] Let the input data be where N is the number of data samples, d is the initial dimension, the inter-class divergence and intra-class divergence can be expressed based on the linear mapping matrix W1, and the constructed optimization objective is:

[0046]

[0047] where S between and S within represent the inter-class divergence matrix and intra-class divergence matrix respectively; by solving the above optimization objective, inter-class features with stronger separability are obtained.

[0048] Step 12: Concatenate the intra-class features and inter-class features into a new feature vector, and adopt a non-linear metric method to generate a data-driven graph;

[0049] Specifically, the intra-class features output by the Transformer and the inter-class features output by the DSFA are concatenated into a new feature vector f, so as to take into account the time patterns within the same fault type and the differences between different fault types.

[0050] After obtaining the comprehensive feature vector f, in order to construct an adjacency relationship at the data level, a non-linear metric method such as the diffusion distance is adopted to generate a preliminary graph structure, and let the feature vector f i and the feature vector f j represent the feature vectors of the i-th and j-th samples respectively, then the adjacency weight can be defined according to the Gaussian kernel function:

[0051]

[0052] where σ is the bandwidth parameter;

[0053] When w ij is greater than the set threshold, an edge can be connected between sample i and sample j to form a preliminary graph structure obtained by data driving; the adjacency matrix A of the data-driven graph d is composed of the adjacency weights w ij between nodes;

[0054] Step 13: According to the corresponding fault diagnosis field, obtain prior knowledge, extract the triple information among equipment, faults, and operating parameters to construct a knowledge graph, define the causal relationship between the nodes of the graph, generate the adjacency matrix of the knowledge graph, and obtain the knowledge-driven graph.

[0055] To more comprehensively characterize the fault mechanism of the air-conditioning system, this embodiment proposes a triple mapping mechanism based on prior knowledge. The core of this mechanism is to extract the association relationship among "equipment - fault type - operating parameter" from the prior knowledge base and formalize it into a mathematical model. Suppose the knowledge base contains M prior rules, and each rule can be expressed as a triple:

[0056] R m =(E i , F j , P k ), m = 1, 2,..., M (4)

[0057] Among them, E i represents the equipment component; F j represents the fault type; P k represents the operating parameter. These triples capture the coupling relationship among equipment components, fault modes, and operating parameters.

[0058] For example, compressor - insufficient refrigeration - temperature is the content of a knowledge graph.

[0059] Based on the experience of domain experts and historical maintenance records, define the causal relationship strength w uv between node u and node v as follows:

[0060]

[0061] Among them, α is the expert confidence adjustment factor, and the default value is 0.7; S uv represents the co-occurrence probability of node u and node v in historical data.

[0062] For the above-defined causal relationship, the strong causal relationship directly supported by prior rules is given a higher weight, while the weak causal relationship based on statistical association is calculated by weighting.

[0063] According to the above causal relationship strength, generate the adjacency matrix A k of the knowledge graph:

[0064]

[0065] Among them, N represents the total number of nodes in the knowledge graph; w :,vRepresents the set of incoming edge weights for node v; max(w:,v) is used to normalize the weights to eliminate the dimension difference and quantify the influence degree of the fault propagation path.

[0066] Step 14: Merge the data-driven graph and the knowledge-driven graph into a unified graph structure by means of adjacency matrix splicing or weighting;

[0067] To achieve the unified representation of data-driven and knowledge-driven, in this embodiment, the above-mentioned preliminary graph structure and the knowledge graph associated graph structure are merged into a unified graph structure A by means of adjacency matrix splicing or weighted superposition c ; The basic process of merging the data-driven graph and the knowledge-driven graph can be described as:

[0068] A c = φ(A d , A k ) (7)

[0069] where A d represents the adjacency matrix of the data-driven graph, which is the adjacency matrix obtained according to the feature similarity; φ(·) is a splicing, weighting or other fusion strategy, which is used to comprehensively reflect the fault association at the data level and the prior knowledge level.

[0070] In this embodiment, Transformer and discriminant analysis (DSFA) are used for modeling to extract intra-class and inter-class features, and a knowledge graph is constructed by combining the prior knowledge in the air-conditioning field, and a knowledge-driven graph is generated accordingly. By means of adjacency matrix splicing, the data-driven graph and the knowledge-driven graph are fused into a unified graph structure, and the constructed unified graph structure can simultaneously reflect the temporal pattern within the same fault type, the differences between different fault categories, and the mechanism relationship within the air-conditioning system, laying a solid foundation for subsequent fault diagnosis and interpretability analysis.

[0071] In the first module, that is, the graph construction module based on data-driven and knowledge-driven, after obtaining the graph structure that fuses data-driven and knowledge-driven, the second module of this embodiment proposes a fault diagnosis mechanism based on multi-model graph attention network ensemble learning for the high complexity and diversity of multi-condition diagnosis of the heating, ventilation and air conditioning (HVAC) system, which can not only make full use of the advantages of different models in their respective sub-datasets or different working conditions, but also improve the robustness and accuracy of the overall diagnosis through probability fusion.

[0072] In some embodiments, for the fault diagnosis module based on multi-model graph attention network ensemble learning, first, multiple graph attention network sub-models are constructed, and the above unified graph structure is trained to obtain multiple groups of diagnosis results. Subsequently, the Bayesian update fusion method is used to fuse the outputs of multiple sub-models to generate the node-level fault diagnosis results for the air-conditioning system;

[0073] Specifically, the fault diagnosis module based on the integrated learning of the multi-model graph attention network is configured to perform the following process:

[0074] Step 21: Extract node features according to the unified graph structure. For each node, assign a graph attention network, calculate the association relationship between the node and its neighbor nodes through the attention mechanism, assign learnable attention weights to each node, and update the node features. Based on the updated features, obtain the fault probability distribution.

[0075] Specifically, first, input the unified graph structure output by the first module into multiple graph attention network sub-models. The unified graph structure includes the adjacency matrix A c and the information of the node feature matrix X.

[0076] Assign learnable attention weights to each node through the graph attention network to highlight the neighbor information that is most critical for fault diagnosis during the message passing process. Let the features of node i and its neighbor node j be h i and h j respectively. Then, the attention coefficient of the graph attention network can be expressed as:

[0077] α ij = softmax j (LeakyRELU(a T [W2h i || W2h j )) (8)

[0078] where W2 is a learnable linear transformation matrix, a is an attention parameter vector, LeakyRELU(·) is a ReLU activation function with a negative slope, and || represents the vector concatenation operation.

[0079] After calculating the attention coefficient α ij , the updated feature of node i is determined by the following formula:

[0080]

[0081] where is the neighbor set of node i, and σ(·) is a non-linear activation function.

[0082] In the above solution, the node features are aggregated in parallel through the multi-head attention mechanism, and this embodiment can more finely capture the associations between complex faults in the air conditioning system.

[0083] To achieve the efficient integration of the multi-graph attention network model, this embodiment proposes a dynamic fusion mechanism based on Bayesian update rules. This mechanism combines the device operating state, historical fault data, and the diagnostic capabilities of each sub-model to gradually optimize the fault classification results.

[0084] Step 22: Iterate the output fault probabilities of the graph attention network model using the Bayesian update fusion method to obtain the fault probability distribution output by each graph attention network model. Calculate the deviation degree between the output of each graph attention network model and the reference distribution through the KL divergence, and determine the fusion weights to perform fusion to obtain the final diagnostic result.

[0085] Step 221: Each graph attention network serves as a sub-model and is trained to: For the unified HVAC system operation graph structure data, construct multiple graph attention network sub-models (GAT), and each model is trained for different working conditions (data subsets); the trained sub-models output the node-level fault probability distribution based on the unified graph structure output in Step 1.

[0086] Assume that the system contains k graph attention network sub-models, and each sub-model is trained for different working condition data sets, thus having different diagnostic capabilities. Define the node-level fault probability distribution output by the m-th sub-model as:

[0087] p m (y∣x) = (p m,1 , p m,2 ,..., p m,C ) (10)

[0088] where C represents the number of fault categories; p m,C represents the probability that the sub-model m determines that the node belongs to category C.

[0089] Step 222: Construct the initial prior according to the current operating state bias and the fault occurrence frequency in historical data:

[0090] Set the initial prior probability P (0) (y = c) based on the device operating state and historical data. The specific calculation method is as follows:

[0091]

[0092] where I(y init = c) is the initial state indicator function, which is used to reflect whether the current operating state of the device is biased towards a certain specific fault category; is the frequency of occurrence of the fault of category c in historical operations; λ is a weight parameter used to balance the influence of the initial state and historical data.

[0093] Step 223: Perform Bayesian iterative update. Every time the output probability of a sub-model m is received, it is regarded as a new observation. Use the Bayesian update rule to fuse the current posterior probability with the sub-model output to obtain a new posterior probability distribution. Iteratively execute to gradually fuse the diagnostic information of multiple sub-models.

[0094] By successively receiving the outputs of each sub-model, the posterior probability is gradually updated using the Bayesian update rule.

[0095] When the m-th sub-model outputs, the posterior probability is updated as:

[0096]

[0097] This process realizes the probability iteration from prior to posterior, ensuring that each step of the update fully considers the diagnostic results of the current sub-model and its consistency with the existing information.

[0098] Step 224: Calculate the deviation degree between the output of each graph attention network model and the reference distribution through KL divergence, determine the fusion weight, and perform weighted fusion on the diagnostic information output by each graph attention network model through a weighted method;

[0099] To highlight the dynamic collaboration ability of multiple models, the working condition matching degree factor β is introduced m Perform weighted fusion on the outputs of each sub-model (and each graph attention network model). The final fault classification result P * (y = c) is defined as:

[0100]

[0101] Among them, the weight β m is determined by calculating the deviation degree between the model output and the reference distribution through KL divergence:

[0102]

[0103] Among them, D KL (p m ||p base ) represents the KL divergence between the output distribution p m of the sub-model m and the reference distribution p base ; γ is the temperature coefficient, which is used to control the smoothness of weight allocation.

[0104] In this embodiment, through the above formula (14), the model weights can be dynamically adjusted according to the working condition matching degree, so as to achieve a more accurate multi-model integrated learning strategy for fault classification. Through the Bayesian update rule and the dynamic weight mechanism guided by the working condition matching degree factor, the gradual fusion and optimization of the diagnostic results of multiple graph attention sub-models are realized, enabling the model weights to be dynamically adjusted according to the adaptation degree of different working conditions, thereby improving the fault classification accuracy and robustness of the HVAC system under multiple working conditions. This embodiment can significantly improve the stability and accuracy of diagnosis under multiple working conditions.

[0105] Compared with the traditional single-model method, in this embodiment, multiple sub-models are set up, which can not only conduct specialized training and fault identification for sub-models according to different time windows or load patterns, but also effectively integrate the diagnostic advantages of each sub-model, reducing overfitting or bias to specific scenarios. At the same time, in the face of challenges such as noise interference and data imbalance, the multi-model fusion and Bayesian update mechanism further enhances the robustness and adaptive ability of the fault diagnosis process.

[0106] In some embodiments, the interpretability analysis module is configured to perform the following process:

[0107] Step 31: Based on the obtained fault diagnosis results, using the input feature gradient method, quantify and visualize the importance of each input feature x input to the graph attention network at the feature level, and obtain the contribution degree of each node input feature x to the fault decision-making.

[0108] Step 32: For the obtained unified graph structure, extract the local subgraph of the target node at the structure level through the graph neural network interpreter, mark the interaction relationship of key nodes and edges, and obtain the fault propagation path.

[0109] Step 33: Integrate the contribution degree of each node feature to the fault decision-making with the fault propagation path to obtain a multi-dimensional interpretation report of the fault diagnosis process, that is, the comprehensive diagnosis result.

[0110] In this embodiment, the combination of the two visualization results of the contribution degree of each node feature to the fault decision-making and the fault propagation path forms a multi-dimensional interpretation report of the fault diagnosis process, which can not only clarify the contribution degree of sensor variables, but also show the propagation mechanism of faults in the graph structure, thereby providing a more intuitive and scientific basis for the diagnostic decision-making of the HVAC system. The following is a specific description.

[0111] After completing the fault diagnosis of the multi-model graph attention network, traditional methods usually only output a single "normal" or "fault type" label, making it difficult to provide in-depth fault tracing information for operation and maintenance personnel. Facing the huge sensor network and component topology of the HVAC system, just knowing that "a certain node has a fault" is far from enough to guide subsequent repairs or optimization adjustments; especially in scenarios with large data scales and complex mechanisms, engineers are more in need of understanding questions such as "why did the diagnostic process reach this conclusion" and "which features or nodes play key roles in it", so as to improve the efficiency and accuracy of fault troubleshooting.

[0112] Therefore, this embodiment introduces an interpretability analysis strategy for fault diagnosis results. Through the multi-level combination of the input feature gradient method and the graph neural network interpreter, the diagnostic process and results are no longer a "black box", but can clearly present the causes and propagation paths of faults, and provide visualization and interpretation means for the interpretability requirements of diagnostic results at the feature level and the structure level respectively. The process of module 3 is as Figure 4 shown:

[0113] In step 31, at the feature level, the input feature gradient method is used to measure the contribution of the sensor variable, that is, the input feature x, to the fault decision.

[0114] Let the fault prediction result output by the diagnostic model for the target node be be the fused diagnostic result output by the fault diagnosis module based on the multi-model graph attention network ensemble learning, and the input node feature is x. When performing backpropagation on to obtain the gradient , the contribution value of each feature element to the prediction result can be obtained by calculating the following product:

[0115]

[0116] where, ⊙ represents the element-wise multiplication operation. When a certain feature or sensor signal plays a crucial role in the fault prediction, the gradient of this feature and its corresponding contribution value will be significantly higher than other features.

[0117] Furthermore, map the contribution degree value of each node feature x to the fault decision into a visual heat map or importance ranking. Operation and maintenance personnel can quickly locate the crucial sensor parameters based on this, so as to conduct further troubleshooting or improvement.

[0118] In step 32, through the graph neural network interpreter, the neighborhood relationship and its influence path of fault propagation are revealed from the graph structure level. For the fault prediction result of node v, the graph neural network interpreter explores a set of subgraphs such that this subgraph maintains high prediction consistency while removing unnecessary edges and nodes. Its objective can be written as:

[0119]

[0120] Among them, is the consistency measure of the prediction result, and Ω(·) is the regularization term of the subgraph scale or complexity. Through iterative optimization, the graph neural network interpreter finds the set of nodes and edges most relevant to the fault prediction, effectively eliminating the information irrelevant to or interfering with the fault. Finally, the extracted key subgraph is visualized, showing the neighborhood of the fault point in the system, the key edges and their weights, obtaining the visualized fault propagation path, so as to show "how the fault leads to the diagnostic conclusion through the interaction of various components or features".

[0121] In step 33, in order to achieve comprehensive interpretability from the feature level and the structure level to the fault location, the contribution degree of each node feature output by the input feature gradient method to the fault decision and the fault propagation path are fused and analyzed. Specifically:

[0122] Step 331: For the nodes or parameters with high feature contribution degree, mark and display them in the key subgraph that constitutes the fault propagation path;

[0123] Specifically, for the sensor nodes or operating parameters with prominent importance, the corresponding nodes can be highlighted in the subgraph;

[0124] Step 332: Superimpose the gradient contribution value of the corresponding feature in the feature contribution degree heat map for the fault propagation path and nodes in the key subgraph, so as to make the feature contribution degree more prominent.

[0125] By using the above-mentioned interpretability analysis module of this embodiment, when the fault result occurs, the operation and maintenance personnel can not only know "what kind of fault has occurred in the system", but also clearly know "which key parameters and interactions have caused this fault", so as to formulate a more efficient and accurate treatment plan. This interpretability analysis module provides visual and transparent support for the fault diagnosis of the HVAC system, greatly enhancing the operability and reliability of the system in this embodiment in practical engineering applications.

[0126] Embodiment 2

[0127] Based on Embodiment 1, this embodiment provides an interpretable fault diagnosis method integrating graph network and knowledge-driven, including the following steps:

[0128] Step 1: Extract intra-class and inter-class features from the acquired operation data to construct a data-driven graph; construct a knowledge graph based on prior knowledge to construct a knowledge-driven graph, and fuse the data-driven graph and the knowledge-driven graph into a unified graph structure;

[0129] Step 2: Construct multiple graph attention network sub-models, respectively identify and fuse the obtained unified graph structure to obtain a diagnosis result;

[0130] Step 3: Use the input feature gradient method to analyze the diagnostic results to obtain a diagnostic result heat map. Process the unified graph based on a graph neural network to extract key subgraphs, obtain a visualized fault propagation path, and fuse the propagation path and the diagnostic result heat map to obtain a comprehensive diagnostic result.

[0131] In this embodiment, based on the complexity of multi-source heterogeneous data and its potential fault propagation characteristics during the operation of a building heating, ventilation, and air conditioning (HVAC) system, the method uses a graph structure to uniformly model the operating state. First, the operation data is used to extract intra-class stable features and inter-class discriminant features through time series and difference modeling, thereby constructing a data-driven graph expressing the current operating state of the system. On the other hand, a knowledge graph reflecting equipment topology, causal association, and functional hierarchy is constructed based on expert knowledge and physical principles to obtain a knowledge-driven graph containing prior logic. After fusing the two, a unified graph structure is generated, taking into account both data distribution and physical logic.

[0132] Build multiple graph attention network sub-models on the unified graph structure. Each sub-model is trained under specific working conditions or data subsets respectively to capture fault patterns in their respective fields. Summarize the outputs of each model through a fusion strategy (such as Bayesian inference or weighted averaging) to generate a robust fault diagnosis result. To achieve interpretability analysis, use the input feature gradient method to generate an influence degree diagram of each variable on the diagnostic result to assist in identifying key influencing measurement points; further extract key subgraphs from the unified graph structure through a graph neural network interpreter to reveal the interaction relationships between key nodes and edges in the system and the fault propagation path. Finally, overlay the heat map and the propagation path to output an intuitive and multi-level comprehensive diagnostic result.

[0133] This implementation method breaks through the limitations of traditional methods based on a single data source or a single model, integrates the advantages of data-driven and knowledge-driven, generates a system state expression with multi-layer semantics, and significantly enhances the ability to perceive complex fault interaction relationships. On this basis, a flexible diagnosis suitable for multiple working conditions and various fault types is realized through a multi-model graph attention network structure, greatly improving the robustness and accuracy of the diagnosis. In addition, by combining feature gradient visualization and graph structure interpretability mechanisms, the diagnostic result not only has high accuracy but also has traceability and transparency, which helps to quickly locate the root cause of the fault and assist maintenance personnel in making decisions.

[0134] In step 1, the method for determining the unified graph structure includes the following steps:

[0135] Step 11: Obtain multi-source time-series operation data, use a deep learning model to extract intra-class features, and process prior knowledge through a discriminant analysis method to obtain inter-class features;

[0136] Specifically, the deep learning model can adopt the Transformer model;

[0137] Step 12: Concatenate the intra-class features and inter-class features into a new feature vector, and adopt a non-linear metric method to generate a data-driven graph;

[0138] Step 13: According to the corresponding fault diagnosis field, obtain prior knowledge and extract the triple information among the equipment, faults, and operating parameters to construct a knowledge graph, define the causal relationship between the graph nodes, generate the adjacency matrix of the knowledge graph, and obtain a knowledge-driven graph;

[0139] Step 14: Merge the data-driven graph and the knowledge-driven graph into a unified graph structure by means of adjacency matrix concatenation or weighting;

[0140] In order to balance data-driven and knowledge-driven to enrich the fault feature expression, this embodiment adopts a feature extraction strategy that combines the Transformer and discriminant analysis technologies, which not only finely models the time series distribution (intra-class features) of the same fault type, but also effectively discriminates between different fault types (inter-class features). By fusing the knowledge graph constructed by prior knowledge and the preliminary graph structure generated by data-driven, the domain knowledge such as the physical topology and fault mechanism of the air-conditioning system is efficiently docked with the sensor data, and a fusion graph structure that can comprehensively reflect the complex interaction relationship and fault features of the system is generated. This scheme effectively reduces the interference of noise and redundant information, and enhances the depth and accuracy of the description of various fault mechanisms.

[0141] In step 2, the method of constructing multiple graph attention network sub-models, respectively identifying and fusing the obtained unified graph structure to obtain the diagnosis result includes the following steps:

[0142] Step 21: Extract node features according to the unified graph structure, allocate a graph attention network for each node, calculate the association relationship between the node and its neighbor nodes through the attention mechanism, allocate learnable attention weights for each node and update the node features, and obtain the fault probability distribution based on the updated features;

[0143] Step 22: Adopt the Bayesian update fusion method to iterate the output fault probability of the graph attention network model, obtain the fault probability distribution output by each graph attention network model, dynamically update the fusion weights of each graph attention network model, and fuse the outputs of each graph attention network model to obtain the final diagnosis result;

[0144] To cope with the multi-condition switching of HVAC systems, local environmental changes, and equipment load differences, in this embodiment, multiple graph attention network sub-models are constructed and integrated diagnosis is achieved through Bayesian update fusion. Multi-model training can learn specialized fault features for different sub-datasets such as different time windows and different external conditions respectively. During fusion, the outputs of each model are comprehensively evaluated, thus significantly improving the reliability and robustness of the overall diagnosis results. In the face of noise interference or extreme operating conditions, ensemble learning can also effectively reduce the misdiagnosis or missed diagnosis probability of a single model and has stronger generalization ability.

[0145] Furthermore, the divergence between the output of each graph attention network model and the reference distribution is calculated by KL divergence, and the fusion weights of each graph attention network model are dynamically updated;

[0146] Step 221, Multi-submodel training and output: For the unified HVAC system operation graph structure data, multiple graph attention network sub-models (GAT) are constructed, and each sub-model is trained for different operating conditions (data subsets); the trained sub-models output the node-level fault probability distribution based on the unified graph structure output in Step 1;

[0147] Step 222, Construct an initial prior according to the current operating state bias and the fault occurrence frequency in historical data:

[0148] Step 223, Use the Bayesian update fusion method to iterate the output fault probabilities of the graph attention network models for Bayesian iterative update: Each time the output probability of a sub-model m is received, it is regarded as a new observation; the current posterior probability is fused with the sub-model output using the Bayesian update rule to obtain a new posterior probability distribution; Iteratively execute to gradually fuse the diagnostic information of multiple sub-models. Among them, the sub-model is the graph attention network model.

[0149] Step 224, Calculate the divergence between the output of each graph attention network model and the reference distribution by KL divergence, determine the fusion weights, and perform weighted fusion on the diagnostic information output by each graph attention network model through a weighted method;

[0150] In Step 3, the input feature gradient method is used to analyze the diagnostic results to obtain a diagnostic result heat map. The unified graph is processed based on the graph neural network to extract key sub-graphs, and the visualized fault propagation path is obtained. The propagation path and the diagnostic result heat map are fused to obtain a comprehensive diagnostic result, including the following steps:

[0151] Step 31, Based on the obtained fault diagnosis results, use the input feature gradient method to quantify and visualize the importance of each feature x input to the graph attention network at the feature level, and obtain the contribution degree of each node feature x to the fault decision;

[0152] Step 32: For the obtained unified graph structure, extract the local subgraph of the target node at the structural level through a graph neural network interpreter, mark the interaction relationships of key nodes and edges, and obtain the fault propagation path.

[0153] Step 33: Integrate the contribution degree of each node feature to the fault decision-making with the fault propagation path to obtain a multi-dimensional interpretation report of the fault diagnosis process, that is, the comprehensive diagnosis result.

[0154] To improve the multi-dimensional interpretability of the fault diagnosis result to facilitate fault tracing and operation and maintenance decision-making, this embodiment introduces two interpretability techniques, the input feature gradient method and the graph neural network interpreter, to analyze the fault diagnosis process from two aspects: feature contribution and key subgraphs of the graph structure. The input feature gradient method can quantify and visualize the influence degree of each sensor variable on the diagnostic output, assisting the operation and maintenance personnel to quickly locate the most critical features or measurement points; the graph neural network interpreter reveals the interaction relationships and propagation paths of key nodes and edges leading to the fault determination at the graph structure level, providing an intuitive insight into the fault mechanism. This multi-level interpretation effectively avoids the limitation of the "black box" diagnosis that it is difficult to locate the fault source, and provides a transparent basis for quick decision-making and precise maintenance.

[0155] In summary, through the systematic solution of "data and knowledge fusion + multi-model integrated learning + multi-dimensional interpretability", this embodiment has greatly improved the accuracy, adaptability and transparency of HVAC fault diagnosis in complex operating environments, and has significant engineering application value and promotion prospects.

[0156] It should be noted here that each step in this embodiment corresponds one by one to the implementation methods of each module in Embodiment 1, and the specific implementation processes are the same. The detailed implementation processes will not be repeated here.

[0157] Embodiment 3

[0158] Based on Embodiment 2, this embodiment provides an interpretable fault diagnosis system integrating graph network and knowledge-driven, including a data acquisition device and a processor;

[0159] The data acquisition device includes sensors and a device monitoring terminal for collecting the operation data of the fault diagnosis object;

[0160] The processor is configured to execute the steps in an interpretable fault diagnosis method integrating graph network and knowledge-driven described in Embodiment 2.

[0161] For the HVAC system, the data acquisition device includes, but is not limited to, the following sensors:

[0162] Temperature sensors: Distributed at key nodes such as the return air outlet, supply air outlet, and condenser outlet to monitor the air temperature changes in the system.

[0163] Humidity sensor: Deployed in the air-conditioned room and the air supply duct to evaluate the air humidity.

[0164] Pressure sensor: Used to detect the working pressures at the inlet and outlet of the compressor in the refrigerant circuit and identify whether the compression state is abnormal.

[0165] Flow sensor: Monitor the changes in the cooling water or refrigerant flow rate to determine whether there is a flow blockage or pipeline leakage.

[0166] Current / voltage sensor: Monitor the energy consumption of electrical components such as compressors, fans, and control modules to evaluate their load status and energy efficiency performance.

[0167] Vibration / noise sensor: Installed at the bearing positions of the compressor and the fan to capture potential mechanical fault signals.

[0168] Among them, the device monitoring terminal can be communicatively connected to the device or set inside the operating device to obtain the operating data of the device.

[0169] Embodiment 4

[0170] This embodiment provides an electronic device, including a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the steps in a method for interpretable fault diagnosis integrating a graph network and knowledge-driven in Embodiment 2 are completed.

[0171] Embodiment 5

[0172] This embodiment provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by the processor, the steps in a method for interpretable fault diagnosis integrating a graph network and knowledge-driven in Embodiment 2 are completed.

[0173] The above are only the preferred embodiments of the present disclosure and are not used to limit the present disclosure. For those skilled in the art, the present disclosure can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present disclosure shall be included within the protection scope of the present disclosure.

[0174] Although the specific implementation manners of the present disclosure are described above in conjunction with the accompanying drawings, it is not a limitation to the protection scope of the present disclosure. Those skilled in the art should understand that based on the technical solutions of the present disclosure, various modifications or deformations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. An interpretable fault diagnosis method integrating graph network and knowledge-driven, characterized in that It includes the following steps: Extract intra-class and inter-class features from the obtained operation data, and construct a data-driven graph; construct a knowledge-driven graph based on prior knowledge to build a knowledge graph, and fuse the data-driven graph and the knowledge-driven graph into a unified graph structure; Construct multiple graph attention network sub-models, respectively identify and fuse the obtained unified graph structure to obtain a diagnosis result; Use the input feature gradient method to analyze the diagnosis result to obtain a diagnosis result heat map, process the unified graph structure based on the graph neural network to extract the key sub-graph, obtain the visualized fault propagation path, and fuse the propagation path and the diagnosis result heat map to obtain a comprehensive diagnosis result.

2. The interpretable fault diagnosis method integrating graph network and knowledge-driven according to claim 1, characterized in that The method for determining the unified graph structure includes the following steps: Obtain multi-source time-series operation data, use a deep learning model to extract intra-class features, and process the prior knowledge through a discriminant analysis method to obtain inter-class features; Concatenate the intra-class features and the inter-class features into a new feature vector, and use a non-linear metric method to generate a data-driven graph; According to the corresponding fault diagnosis field, obtain prior knowledge and extract the triple information between the equipment, faults, and operation parameters to construct a knowledge graph, define the causal relationship between the graph nodes, generate the adjacency matrix of the knowledge graph, and obtain a knowledge-driven graph; Merge the data-driven graph and the knowledge-driven graph into a unified graph structure by means of adjacency matrix concatenation or weighting.

3. The interpretable fault diagnosis method integrating graph network and knowledge-driven according to claim 1, wherein: The method for constructing multiple graph attention network sub-models, respectively identifying and fusing the obtained unified graph structure to obtain a diagnosis result includes the following steps: Extract node features according to the unified graph structure, allocate a graph attention network for each node, calculate the correlation relationship between the node and its neighbor nodes through the attention mechanism, assign a learnable attention weight to each node and update the node features, and obtain a fault probability distribution based on the updated features; Use the Bayesian update fusion method to iterate the output fault probability of the graph attention network model, obtain the fault probability distribution output by each graph attention network model, dynamically update the fusion weights of each graph attention network model, and fuse the outputs of each graph attention network model to obtain the final diagnosis result.

4. The interpretable fault diagnosis method integrating graph network and knowledge-driven according to claim 3, wherein: Calculate the deviation degree between the output of each graph attention network model and the benchmark distribution through the KL divergence, and dynamically update the fusion weights of each graph attention network model.

5. The interpretable fault diagnosis method integrating graph network and knowledge-driven according to claim 3, characterized in that: Use the Bayesian update fusion method to iterate the output fault probability of the graph attention network model, perform Bayesian iterative update. Each time the output probability of a sub-model m is received, it is regarded as a new observation; use the Bayesian update rule to fuse the current posterior probability with the sub-model output to obtain a new posterior probability distribution; iterate and execute to gradually fuse the diagnostic information of multiple sub-models; among them, the sub-model is the graph attention network model.

6. The interpretable fault diagnosis method integrating graph network and knowledge-driven according to claim 1, characterized in that: The method for determining the comprehensive diagnosis result includes the following steps: Based on the obtained fault diagnosis result, use the input feature gradient method to quantify and visualize the importance of each node feature x input to the graph attention network at the feature level, and obtain the contribution degree of each node feature x to the fault decision; For the obtained unified graph structure, at the structural level, a local subgraph of the target node is extracted through a graph neural network interpreter, the interaction relationships of key nodes and edges are marked, and a fault propagation path is obtained. The contribution degree of each node feature to the fault decision is fused with the fault propagation path to obtain a multi-dimensional interpretation report of the fault diagnosis process, that is, a comprehensive diagnosis result.

7. An interpretable fault diagnosis system integrating graph network and knowledge-driven, characterized in that, Including: A graph construction module based on data-driven and knowledge-driven, which is configured to: extract intra-class and inter-class features from the acquired operation data and construct a data-driven graph. Construct a knowledge-driven graph based on prior knowledge, and fuse the data-driven graph and the knowledge-driven graph into a unified graph structure. A fault diagnosis module based on multi-model graph attention network ensemble learning, which is configured to construct multiple graph attention network sub-models, respectively identify and fuse the obtained unified graph structure, and obtain a diagnosis result. An interpretability analysis module for diagnosis results, which is configured to use the input feature gradient method to analyze the diagnosis results to obtain a heat map of the diagnosis results, extract key subgraphs from the unified graph based on the graph neural network, obtain a visualized fault propagation path, and fuse the propagation path and the heat map of the diagnosis results to obtain a comprehensive diagnosis result.

8. An interpretable fault diagnosis system integrating graph network and knowledge-driven, characterized in that: Including a data acquisition device and a processor; The data acquisition device includes a sensor and a device monitoring terminal for collecting operation data of the fault diagnosis object. The processor is configured to execute the steps in an interpretable fault diagnosis method that fuses a graph network and knowledge-driven according to any one of claims 1-6.

9. An electronic device, characterized in that, Including a memory, a processor, and computer instructions stored on the memory and running on the processor. When the computer instructions are run by the processor, the steps in an interpretable fault diagnosis method that fuses a graph network and knowledge-driven according to any one of claims 1-6 are completed.

10. A computer-readable storage medium, characterized in that, For storing computer instructions, when the computer instructions are executed by the processor, the steps in an interpretable fault diagnosis method that fuses a graph network and knowledge-driven according to any one of claims 1-6 are completed.

Images