Feature-level deep explainable air conditioner fault diagnosis method based on neural network model

Abstract

The application discloses a feature-level deep explainable air conditioner fault diagnosis method based on a neural network model, belongs to the technical field of computer systems based on specific calculation models, and comprises the following steps: S1, constructing fault training data; S2, converting the fault training data into graph data based on a node classification k-nearest neighbor graph construction strategy, and obtaining graph structure data; S3, obtaining high-order node representation, updating network parameters by using a loss function based on the high-order node representation, and completing a final fault diagnosis task; S4, analyzing a diagnosis process and a result of the graph neural network by using three feature-level explainable methods, and obtaining a quantitative influence law of the number of graph convolution layers, the number of channels and the learning rate on the performance of the graph convolution fault diagnosis network; and S5, obtaining explanation effects of the three feature-level explainable methods under a cold water unit fault experiment. The application solves the problem that high diagnosis precision and deep explainability of a model cannot be considered in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of computer systems based on specific computing models, and in particular to a feature-level deep interpretable air-conditioning fault diagnosis method based on a neural network model. Background Art

[0002] Existing air conditioning fault diagnosis systems use intelligent diagnosis models to automatically extract features and analyze them to diagnose faults. However, traditional intelligent diagnosis models have strong black box properties and cannot make the model's diagnosis process transparent.

[0003] For example, the Chinese invention patent with announcement number CN111723925B discloses a method, device, equipment and medium for diagnosing faults of intelligent train air-conditioning units in transit, including: constructing a standard event template library based on a standard electric power time series, which includes transient and steady-state standard event templates; obtaining corresponding transient and steady-state time subsequences based on historical electric power time series, and finding matching transient and steady-state standard event templates in the standard event template library, thereby obtaining corresponding historical electric power steady-state residual subsequences, and using them as historical samples to train an extreme learning machine to obtain a train air-conditioning unit fault prediction model; obtaining the real-time electric power time series of the train air-conditioning unit in transit, and obtaining the corresponding real-time electric power residual sequence according to the same method as above, and inputting it into the train air-conditioning unit fault prediction model, thereby predicting the fault type of the train air-conditioning unit in transit.

[0004] However, in the process of implementing the technical solutions of the embodiments of the present application, the present application discovered that the above technology has at least the following technical problems:

[0005] Traditional diagnostic approaches based on statistical laws or shallow machine learning methods suffer from strong black-box nature and insufficient feature interpretation. The lack of transparency in interpreting model diagnostic results can lead to a trust crisis when the model's diagnostic results differ from the user's perceived cause. Furthermore, interpretable methods lack in-depth feature-level analysis. Traditional graph neural networks cannot directly assess feature importance, thus failing to determine the contribution of features to diagnostic results. This makes it difficult for users to obtain valuable explanatory information, resulting in a struggle to balance high diagnostic accuracy with deep interpretability. Summary of the Invention

[0006] The embodiment of the present application solves the problem in the prior art that it is impossible to achieve both high diagnostic accuracy and deep interpretability of the model by providing a feature-level deeply interpretable air-conditioning fault diagnosis method based on a neural network model, thereby improving the transparency of the internal training mechanism and the operation process and improving the accuracy of fault cause analysis.

[0007] The embodiment of the present application provides a feature-level deep interpretable air-conditioning fault diagnosis method based on a neural network model, comprising the following steps: S1, obtaining an experimental data set and performing data preprocessing to construct fault training data; S2, converting the fault training data into graph data based on a k-nearest neighbor graph construction strategy for node classification, and obtaining graph structure data at the same time, wherein the graph data is used for feature extraction, information transmission and iterative training of a subsequent graph convolution fault diagnosis network; S3, constructing a graph convolution fault diagnosis network based on a node-level graph generation strategy, thereby aggregating and updating the node information and edge connection relationship in the graph structure data to obtain a high-order node representation, and based on the high-order node representation, The dot indicates that the loss function is used for back propagation to update the network parameters and complete the final fault diagnosis task; S4. The diagnostic process and results of the graph neural network are analyzed by three feature-level interpretable methods, and the quantitative influence of the number of graph convolution layers, the number of channels and the learning rate on the performance of the graph convolution fault diagnosis network is obtained; the three feature-level interpretable methods include gradient representation, reconstructed convolution and corrected back propagation; S5. A matching evaluation standard based on the three feature-level interpretable methods and the actual fault cause is constructed to obtain the matching degree of the three feature-level interpretable methods, thereby obtaining the explanation effect of the three feature-level interpretable methods in the chiller fault experiment.

[0008] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0009] 1. The feature-level deep interpretable air-conditioning fault diagnosis method based on the neural network model provided by the present invention constructs a graph convolution fault diagnosis network based on the node-level graph generation strategy and introduces three feature-level interpretable methods to analyze the diagnosis process and results, thereby obtaining the quantitative influence of the number of graph convolution layers, the number of channels and the learning rate on the performance of the graph convolution fault diagnosis network and the matching degree of the interpretation results, thereby achieving the improvement of the transparency of the internal training mechanism and the operation process and the improvement of the accuracy of the fault cause analysis, effectively solving the problem in the existing technology that it is impossible to take into account both the high diagnostic accuracy and deep interpretability of the model.

[0010] 2. The present invention uses a graph neural network fault diagnosis model to capture the potential characteristics of the mechanical structure of equipment components, combines node attribute characteristics and edge correlation information to achieve joint modeling of multi-dimensional information fusion, performs multi-hop information transmission based on the neighbor aggregation mechanism, dynamically updates node characteristics, and deeply mines the dependencies between nodes.

[0011] 3. This invention provides visualization analysis and feature importance evaluation for the graph neural network fault diagnosis model through three feature-level interpretable methods: gradient representation, reconstructed convolution, and modified back propagation, which improves the transparency of the network diagnosis process and alleviates the trust crisis of the model.

[0012] 4. The present invention obtains the analysis effect and accuracy of the interpretable method by analyzing the importance score vector, and calculates the matching degree between the method-based explanation and the experience-based explanation, thereby obtaining the most accurate interpretation method for the diagnostic results of the chiller fault data.

[0013] 5. The present invention improves the transparency of the internal training mechanism and operation process by combining the graph neural network fault diagnosis model with the feature-level deep interpretability method, provides basic knowledge and criteria for the fault cause analysis of air-conditioning chillers, and lays the foundation for explainable intelligent fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flow chart of a feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model provided in an embodiment of the present application;

[0015] Figure 2 A schematic diagram of the research content provided in the embodiments of this application;

[0016] Figure 3 A research flow chart provided for the embodiments of this application;

[0017] Figure 4 A working cycle diagram of the chiller system provided in an embodiment of the present application;

[0018] Figure 5 A flow chart of feature-level interpretation based on the gradient representation method provided in an embodiment of the present application;

[0019] Figure 6 A flowchart of feature-level interpretation based on the reconstruction convolution method provided in an embodiment of the present application;

[0020] Figure 7 A flowchart of feature-level interpretation based on the modified back-propagation method provided in an embodiment of the present application.

[0021] Figure 8 A diagram illustrating the effect of different numbers of graph convolution layers on the degree of interpretation matching provided in the embodiments of this application.

[0022] Figure 9 A diagram showing the effect of different numbers of channels and learning rates on the degree of interpretation matching provided in the embodiments of this application.

[0023] Figure 10 The average interpretation matching degree of the test samples under various fault types using different interpretation methods provided in the embodiments of the present application. DETAILED DESCRIPTION

[0024] The embodiment of the application provides a feature-level deep explainable air conditioner fault diagnosis method based on a neural network model, solves the problem that high diagnosis accuracy and deep explainability of a model cannot be considered in the prior art, and improves the internal training mechanism and operation process transparency and the fault reason analysis accuracy.

[0025] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the drawings of the specification and specific embodiments.

[0026] As Figure 1 shown, the feature-level deep explainable air conditioner fault diagnosis method flowchart provided by the embodiment of the application based on a neural network model, as Figure 2 shown, Figure 2 the research content schematic diagram provided by the embodiment of the application, as Figure 3 shown, Figure 3 the research flowchart provided by the embodiment of the application, as Figure 1 , Figure 2 and Figure 3 known, the experimental object is a 90-ton centrifugal water chiller system, and the method comprises the following steps: S1, acquiring an experimental data set and performing data preprocessing to construct fault training data; S2, converting the fault training data into graph data based on a node classification k-nearest neighbor graph construction strategy, and obtaining graph structure data, wherein the graph data is used for subsequent feature extraction, information transmission and iterative training of a graph convolution fault diagnosis network; S3, constructing a graph convolution fault diagnosis network based on a node-level graph generation strategy, thereby aggregating and updating node information and edge connection relationships in the graph structure data to obtain high-order node representations, updating network parameters based on the high-order node representations by using a loss function, and completing a final fault diagnosis task; S4, analyzing the diagnosis process and results of the graph neural network by using three feature-level explainable methods to obtain the quantitative influence law of the number of graph convolution layers, the number of channels and the learning rate on the performance of the graph convolution fault diagnosis network; the three feature-level explainable methods comprise gradient representation, reconstruction convolution and corrected back propagation; S5, constructing a matching degree evaluation standard based on the three feature-level explainable methods and the true fault reason to obtain the matching degrees of the three feature-level explainable methods, and thereby obtaining the explanation effects of the three feature-level explainable methods in the water chiller fault experiment.

[0027] In the embodiment, the water chiller is composed of an evaporator, a condenser, a compressor, an expansion valve and a heat exchanger, and the working cycle principle of the water chiller is as shown in Figure 4 , Figure 4A working cycle diagram of the chiller system provided in an embodiment of the present application. It should be noted that the evaporator and condenser are both flooded shell and tube heat exchangers, and the refrigerant is R134a. The expansion valve is a thermal expansion valve. The evaporator transfers heat from the chilled water to the refrigerant through the tube wall. The compressor provides the additional work required for heat transfer from low temperature to high temperature. The condenser transfers heat from the refrigerant to the cooling water. The expansion valve throttles and reduces the pressure of the high-pressure liquid refrigerant to a low-temperature, low-pressure gas-liquid mixture, which eventually returns to the evaporator to absorb heat and refrigerate.

[0028] It should be noted that the 11 characteristic variables of the air-conditioning chiller are specifically shown in Table 1. Table 1 is a table of 11 characteristic variables of the chiller.

[0029] Table 1 Eleven characteristic variables of chillers

[0030] Number Feature Feature Description 0 TEI Chilled water inlet temperature 1 TEO Chilled water outlet temperature 2 TCI Cooling water inlet temperature 3 TCO Cooling water outlet temperature 4 TRE Evaporation temperature 5 TRC Condensing temperature 6 Tsh_suc Suction superheat 7 TRC_sub Subcooling of refrigerant 8 P_in Compressor power consumption 9 PO_feed Oil feed pressure 10 TCA Condenser approach

[0031] As can be seen from Table 1, the 11 characteristic variables of the chiller are TEI, TEO, TCI, TCO, TRE, TRC, Tsh_suc, TRC_sub, P_in, PO_feed and TCA, which represent the chilled water inlet temperature, chilled water outlet temperature, cooling water inlet temperature, cooling water outlet temperature, evaporation temperature, condensing temperature, suction superheat, refrigerant subcooling, compressor power consumption, oil supply pressure and condenser temperature difference, respectively.

[0032] If the normal state is regarded as one of the fault types, there are a total of 8 fault types of the chiller, as shown in Table 2. Table 2 shows the 8 typical fault types of the chiller.

[0033] Table 2 Eight typical failure types of chillers

[0034] Number Fault Type Fault Description 0 no Normal state 1 fwc Cooling water deficiency 2 fwe Chilled water deficiency 3 eo Excess oil 4 ro Excess refrigerant 5 rl Refrigerant leakage 6 cf Condenser fouling 7 nc Non-condensable gas generation

[0035] As can be seen from Table 2, the eight typical fault types of chillers are excessive lubricating oil (eo), presence of non-condensable gas (nc), condenser scaling (cf), insufficient cooling water (fwc), insufficient chilled water (fwe), refrigerant leakage (rl), excessive refrigerant (ro) and normal state (no).

[0036] The chiller air conditioner was subjected to various fault condition experiments. The specific method was as follows: simulation tests were conducted on various fault types of the chiller system, including 8 typical fault types such as condenser scaling, excessive lubricating oil, insufficient chilled water, and refrigerant leakage.

[0037] The specific method for obtaining the experimental dataset is to store fault data through a data acquisition system. The data acquisition system consists of sensors, a data acquisition unit, storage devices, data processing software, and a network connection. Sensors are primarily used to detect and measure various components and parameters of the chiller. The data acquisition unit receives the electrical signals sent by the sensors and converts them into digital signals. The data storage device and data processing software are used to store, process, analyze, and display the collected data. The network connection is used to connect the data acquisition system to other systems or devices to enable remote data transmission and sharing.

[0038] After acquiring experimental data, preprocessing is required. Specific steps include data standardization, steady-state processing, and dataset partitioning. The specific method involves processing different data sets according to specific rules to bring them into a consistent standard and eliminate dimensional differences between them. Next, steady-state processing is used for filtering to eliminate biased samples. Finally, the standardized and steady-state processed dataset is partitioned into a training set (70% of the data) and a test set (30% of the data). The independent training and test sets facilitate accurate evaluation of the model's diagnostic performance in real-world scenarios.

[0039] Furthermore, a k-nearest neighbor graph construction strategy based on node classification converts the fault training data into graph data. The specific method is as follows: the fault training data is used as the fault sample features, and the collected fault sample features are used as nodes in the graph structure. The fault training data includes chilled water inlet temperature, chilled water outlet temperature, cooling water inlet temperature, cooling water outlet temperature, evaporation temperature, condensing temperature, suction superheat, refrigerant subcooling, compressor power consumption, oil supply pressure and condenser temperature difference; the pairwise distance between nodes is obtained through Euclidean distance analysis, and the k value is set, thereby completing the conversion of the fault training data into graph structure data.

[0040] Furthermore, the fault training data is converted into graph data, and graph structure data is obtained at the same time. The specific method is: using Euclidean distance to calculate the pairwise distance between nodes, thereby obtaining the Euclidean distance of each node with other nodes; based on the Euclidean distance of each node with other nodes, the k nodes closest to each other are selected as neighbors of the node, thereby obtaining k similar neighbors of each node, and generating edges representing the relationship between these neighbors; constructing a k-nearest neighbor global graph, which is an undirected graph that uses Euclidean distance to represent weighted neighbor relationships; based on the k-nearest neighbor relationships of nodes and the k-nearest neighbor global graph; the k-nearest neighbor graph generation method based on node classification connects all nodes according to the k-nearest neighbor relationships of nodes, thereby converting the fault training data into graph data, and completing the conversion of the fault training data into graph structure data; extracting the node connection relationship and edge weight information from the graph data, you can form graph structure data that only contains the node topology structure; each edge in the graph structure data represents the k-nearest neighbor relationship between a pair of nodes, and the weight of the edge is the Euclidean distance between the two nodes.

[0041] In the embodiment, it should be noted that the Euclidean distance is used to calculate the pairwise distance between samples, and the purpose of setting the k value is to determine k similar neighbors of each sample and generate edges representing the relationship between these neighbors. For example, k is set to 2.

[0042] The Euclidean distance between each node and every other node is obtained by constructing a k-nearest-neighbor global graph G = (V, E). The k-nearest-neighbor global graph is an undirected graph that uses Euclidean distance to represent weighted proximity relationships, where V represents the set of nodes and E represents the set of edges. Each node vϵV represents a data point in the graph structure, and the weight of each edge eϵE represents the k-nearest-neighbor relationship, where the edge weight is the Euclidean distance between the two nodes. The Euclidean distance formula can be expressed as:

[0043] ;

[0044] Where d(xi,xj) represents the Euclidean distance between node xi and node xj in the graph structure data, and xi and xj represent two nodes in the graph structure data respectively.

[0045] It should be noted that the specific method for constructing the k-nearest-neighbor global graph G = (V, E) is as follows: Using the k-nearest-neighbor graph construction algorithm, a single chiller fault sample xi is used as a node in the graph Vi. The similarity between sample xi and other samples xj is calculated based on the Euclidean distance. The k samples with the highest similarity are used as other nodes in the graph G. Each node is assigned a corresponding sample feature vector, and the fault type of each sample is used as the node label to construct the final graph structure G = (V, E).

[0046] For a given dataset D = {x1, x2, ..., xn}, where n represents the maximum number of nodes in the dataset, the adjacency matrix AK of the k-nearest-neighbor global graph is defined as follows:

[0047] ;

[0048] Where d(xi,xj) represents the Euclidean distance between nodes xi and xj in the graph structure data, and KNN(xj) represents the set of k nearest neighbor points of data point xj.

[0049] The k-nearest neighbor graph generation method based on node classification connects all nodes according to the k-nearest neighbor relationship of the nodes, thereby converting the fault training data into graph data, completing the conversion of the fault training data into graph structure data; extracting the node connection relationship and edge weight information from the graph data, you can form a graph structure data containing only the node topology structure.

[0050] Furthermore, a graph convolutional fault diagnosis network based on the node-level graph generation strategy is constructed. The specific method is as follows: in the graph structure data, the status data of the chiller is embedded into the node as the characteristic attribute of the fault sample, thereby treating each node as a fault sample; based on the graph neural network, the characteristics of each node and its neighboring nodes are used as input, and the aggregation representation including the node structure information is obtained through aggregation function analysis, thereby obtaining a graph convolutional fault diagnosis network based on the node-level graph generation strategy.

[0051] In the embodiment, information aggregation is achieved by combining the sample's own information with the neighbor information, thereby building a bridge for information transfer in the graph convolution operation.

[0052] Based on the graph neural network, the features of each node and its neighboring nodes are used as input, and an aggregate representation including node structure information is obtained through aggregation function analysis. Thus, a graph convolutional fault diagnosis network based on the node-level graph generation strategy is obtained. The specific method is: the aggregate representation together with the node's own features are input into the update function ("update function" refers to the nonlinear mapping function used to update the node representation in the graph neural network. Its function is to calculate and generate the new feature representation vector (i.e., hidden state) of the node based on the node's own features and neighbor aggregation information in each round of iteration) to obtain the updated representation of each node. After iterative information transmission and update, each node will obtain a feature vector representing its attribute information. This process can be expressed as:

[0053] ;

[0054] ;

[0055] Where hut represents the hidden layer representation of node u at the t-th step, euv represents the feature of the edge connecting nodes u and v, Mt represents the aggregation function at the t-th step, mut+1 represents the local structure expression of the node after aggregation, and Ut represents the update function at the t-th step.

[0056] The vector is input into the fully connected layer for classification, and the node's corresponding label is output. Specifically, this layer maps the high-order representation vectors of the nodes extracted by the graph convolutional network (GCN) to the target classification space, thereby determining the label category (e.g., fault type) to which each node belongs. The error loss is calculated by calculating the difference between the predicted label category and the true label category. Gradient backpropagation based on this loss is then performed to update the parameters of the graph convolutional network.

[0057] The core idea of ​​graph convolutional networks is to aggregate the features of adjacent nodes using the information transfer mechanism so that the attributes of each node are represented by a feature vector of fixed length. The calculation formula of graph convolution can be expressed as:

[0058] ;

[0059] Where H(l) represents the node feature matrix of the lth layer, A represents the adjacency matrix, I represents the identity matrix, , W(l) represents the weight matrix of the lth layer, D is the degree matrix, represents the angle matrix, Represents the activation function.

[0060] It should be noted that in each layer of graph convolution, node features can propagate and aggregate information based on adjacency relationships, thereby updating their own attribute features.

[0061] It should be noted that the specific steps for building a graph neural network fault diagnosis model are: first, updating the node representation, then performing feature aggregation, nonlinear transformation, and stacking graph convolution layers. The specific method for updating the node representation is: for each node in the graph, its own features are aggregated with the features of neighboring nodes to form a new node representation. This process can be expressed as: Where Z(l+1) is the set of nodes to be updated, obtained by multiplying the adjacency matrix A and the node features H(l). To consider the relationship strength and topological structure between nodes, the features of neighboring nodes are weighted and summed, with the weights determined by the adjacency matrix and degree matrix of the graph structure. The feature aggregation process can be expressed as: Where, for The result is obtained after symmetric normalization. The activation function ReLu is used to perform nonlinear transformation on the updated node representation. The calculation process is:

[0062] ; ; Finally, the above process is iterated multiple times, with each layer using the output of the previous layer as input to learn more advanced node representations and complete the graph convolution layer stacking.

[0063] The node-based graph construction method is more suitable for air conditioning chiller fault data. Because the nodes can understand the key factors of the fault at the feature level, it helps improve the interpretability of the model.

[0064] Furthermore, a high-order node representation is obtained. The specific method is as follows: the aggregated representation together with the node's own features are input into the update function to obtain the updated representation of each node and the feature vector corresponding to each node; the feature vector is input into the fully connected layer for classification processing, and the predicted fault label is output; the true fault label is obtained, and the difference analysis between the true fault label and the predicted fault label is performed to obtain the error loss; gradient backpropagation is performed based on the error loss to update the parameters of the graph convolutional network; multi-layer graph convolution operations are performed based on the graph convolutional network to aggregate the node's own features with the features of neighboring nodes, and after nonlinear transformation and multiple iterations, a high-order node representation is obtained; the high-order node representation integrates the feature information of the node itself and its neighboring nodes, and is used to characterize the position and characteristics of the node in the graph structure.

[0065] Furthermore, the quantitative influence of the number of graph convolution layers, the number of channels and the learning rate on the performance of the graph convolution fault diagnosis network is obtained. The specific method is: based on the gradient representation method, the partial derivative of the input feature to the prediction result is calculated for each prediction node category, and the feature importance vector is generated to obtain the key influencing features; the key influencing features are used for subsequent matching and comparison with the fault discrimination features preset in the database to obtain the accuracy of the gradient representation method; based on the reconstruction convolution method, the feature map is visualized to the input space through the deconvolution operation, and the input features are activated through back propagation; based on the modified back propagation method, by modifying the gradient calculation rule of ReLU in the back propagation, only the positive gradient component is retained, thereby obtaining the explanatory contribution of the positive effective features; based on the gradient representation method, the reconstruction convolution method and the modified back propagation method under different combinations of graph convolution layers, channels and learning rates The performance of the proposed method is normalized and analyzed to obtain the diagnostic interpretation performance characteristics under different conditions; the feature importance vector is normalized according to the feature dimension, and its stability and consistency with the changes of the number of graph convolution layers, the number of channels, and the learning rate are compared and analyzed to obtain the quantitative influence of the number of graph convolution layers, the number of channels, and the learning rate on the diagnostic performance and matching degree; based on the gradient representation method, the partial derivative of the input feature to the prediction result is calculated for each prediction node category to generate the feature importance vector and obtain the key influencing features. The specific method is as follows: the pixels and graph structure in the input data are obtained and input into the scoring function, and the degree of influence of the pixels in the graph structure on the scoring function is analyzed to obtain the ranking order for the pixel group; based on the nonlinear relationship between the scoring function and the graph structure, the first-order Taylor expansion is used for approximation to obtain the importance of each feature to the model prediction result.

[0066] In the embodiment, the specific application of the three feature-level interpretable methods in the chiller fault diagnosis process is as follows: Figure 5 This is a flowchart of the feature-level interpretation based on the gradient representation method provided in an embodiment of the present application. Where v represents the input node, and each node is assigned a feature vector. S is the final fault diagnosis result obtained after the graph convolution operation and activation function. It represents the gradient of a specific fault type relative to the input, which is obtained by calculating the gradient of the input feature relative to the model prediction output. Taking the faulty condenser fouling (cf) as an example, the gradient of the input relative to the output of the corresponding category is first calculated and the absolute value is taken to obtain the node feature importance FI under this category. Then, the fault discrimination features are automatically located according to the FI value to explain the model results. Finally, the fault discrimination features corresponding to each fault are compared with the domain knowledge and maintenance experience to evaluate the matching degree of the explanation method. Figure 5As can be seen from the figure, for condenser fouling faults, the most important feature for the model's prediction results is the oil supply pressure (PO_feed). This indicates that oil supply pressure significantly influences the fault network's diagnostic results, which aligns with the fault discrimination features derived from domain knowledge. Insufficient oil supply pressure hinders the flow of lubricating oil, preventing the formation of an effective oil film on the surfaces of key moving components such as the compressor bearings and rotor. This leads to dry friction and abnormal wear, which can cause component seizure during long-term operation. Insufficient oil supply also reduces the compressor's oil seal effectiveness, increasing friction between the rotor and cylinder block and generating abnormal noise. Furthermore, insufficient oil supply pressure prevents lubricating oil from reaching required parts, leading to insufficient cooling and lubricating oil supply. This can affect the power supply to the energy conditioning device or trigger a low-pressure shutdown. Excessive oil supply pressure may indicate oil blockage or oil quality degradation, increasing oil pump load, energy consumption, and even damage to the pump or filter. Problems like a clogged oil filter and aging oil can directly lead to abnormal oil supply pressure, typically manifesting as a low-pressure alarm or a shutdown due to oil pressure differential protection. This can also be accompanied by unusual compressor noise, bearing overheating, and abnormally high energy consumption. During maintenance, focus on checking the oil filter, oil pump, solenoid valve, and oil quality to ensure that oil pressure parameters are within the unit's rated range.

[0067] The gradient representation method is constructed as the first interpretable method for the graph neural network fault diagnosis process of air conditioning chillers. By analyzing the contribution of each pixel in the input data to the network's prediction results, the network's attention area is revealed. This process can be expressed as: Where Sc(I) represents the score function for category c, c represents the specific category, the graph structure is represented as I, ωc is the model weight vector, and bc is the bias. The purpose of this calculation process is to assign a ranking order to the pixel groups in the graph structure I based on the degree to which the pixels affect the score Sc(I).

[0068] Since the category score Sc(I) is not linearly related to I in deep neural networks, for a given graph structure I, Sc(I) can be approximated using a first-order Taylor expansion in a local area near I0. The expression is: ; where ω is the derivative of Sc with respect to I at point I0 in the graph structure. The expression is: ; where b is the bias.

[0069] In the formula, the importance FI of each feature to the model prediction result can be expressed as: .

[0070] Furthermore, based on the reconstructed convolution method, the feature map is visualized to the input space through the deconvolution operation, and the input features are activated through back propagation. The specific method is: the feature map of a certain layer to be visualized is input into the deconvolution layer, and the feature map of this layer in the input space is activated through back propagation. Then, the feature heat map is constructed through the color mapping principle to obtain the degree of influence of different feature variables in each network layer on the final prediction result; the importance score of the feature is obtained according to the visualization result analysis of the feature map; the visualization result of the feature map is used to identify the key feature areas or variables in the prediction task.

[0071] In the embodiment, it should be noted that a reconstruction convolution method is constructed as the second interpretable method for the fault diagnosis process of the graph neural network of the air-conditioning chiller. The basic principle is to measure the importance of the input variable to the prediction result by calculating the gradient of the output relative to the input. First, the feature map of a certain layer to be visualized is input into the deconvolution layer, and the feature map of this layer in the input space is activated by back propagation. Then, a feature heat map is constructed by the color mapping principle to obtain the degree of influence of different feature variables in each network layer on the final prediction result. The deconvolution process can be expressed as: Where y is the feature map output after deconvolution. x is the feature map output by the graph convolution model. ω' is the deconvolution kernel. Finally, the feature importance score FI is obtained based on the visualization results of the feature map.

[0072] Figure 6 It is a feature-level interpretation process based on the reconstruction convolution method. Specifically, the principle of the reconstruction convolution method is that when it is necessary to visualize the feature map of a certain layer of the neural network, the feature map is input into the deconvolution layer to calculate the activation features in the input space through back propagation, and then the feature heat map is constructed through the color mapping principle to obtain the degree of influence of different feature variables in each network layer on the final prediction result. In the embodiment, taking the condenser scaling (cf) of the chiller as an example, the activation map of the input feature is first obtained by deconvolution, and then the FI of each variable is obtained by the color mapping principle. The color depth is proportional to the size of the data value. Then, after obtaining the FI, the fault discrimination features are automatically located according to the FI value and compared with the domain knowledge and maintenance experience to verify the matching degree of the interpretation method. Reference Figure 6As shown, the reconstructed convolution interpretable method identifies the oil supply pressure (PO_feed) as the fault-discriminating feature for condenser faults. The condenser temperature difference (TCA) can also be considered as a fault-discriminating feature, which aligns with domain knowledge. The condenser temperature difference will have a certain impact on the chiller's cooling efficiency, energy consumption, compressor, and lubrication system. When the temperature difference is too small, the condensing pressure decreases, the pressure difference between the condenser and evaporator is insufficient, the refrigerant circulation volume decreases, the cooling efficiency decreases, and the oil return power is insufficient. This may trigger low-pressure alarms, surge protection, motor overheat protection, and oil shortage alarms. When the temperature difference is too large, the condensing pressure increases, the compressor pressure ratio increases, power consumption increases, and exhaust temperature becomes abnormal, which exacerbates bearing and rotor wear. This may cause compressor high-pressure protection shutdown, oil carbonization failure, and oil degradation.

[0073] Furthermore, based on the modified back propagation method, by modifying the gradient calculation rules of ReLU in the back propagation, only the positive gradient components are retained, thereby obtaining the explanatory contribution of the positive effective features. The specific method is: extracting the output result of the node to be explained from the graph convolutional network, that is, the score of the fault category; the node to be explained is a certain operation sample of the chiller, and its category is a certain type of fault identified by the system; by tracing the reverse calculation path of the node from the output layer to the input layer, all the calculation layers that have passed the ReLU activation function are identified; in each layer of back propagation, the gradient passing through ReLU is screened and the ones that have passed through are eliminated. There is a gradient component less than zero; according to the corrected reverse path, the forward gradient is propagated from the output layer to the input layer layer by layer, thereby obtaining the reverse gradient contribution value of each feature dimension data in the input layer to form a feature importance vector; based on the feature importance vector, normalization processing is performed to obtain the influence degree of different features on the fault sample; the feature variables in the feature importance vector are sorted from high to low, and a preset number of feature variables in the database are obtained as the key judgment basis for the prediction result, so as to obtain the explanatory contribution of the positive effective features, thereby assisting the model to make the final fault diagnosis result and perform credibility verification.

[0074] In the embodiment, a modified backpropagation method is constructed as a third interpretable method for the graph neural network fault diagnosis process of air-conditioning chillers. This method is similar to the reconstructed convolution method and aims to evaluate the sensitivity and importance of different features by visualizing the contribution of input features to the output, thereby understanding the internal decision-making principles of the model. Unlike the reconstructed convolution method, the modified backpropagation interpretable method modifies the gradient calculation rules of ReLU in backpropagation and sets the activation output of negative values ​​to zero. Therefore, it only focuses on the propagation process of positive gradients and ignores the contribution of negative values.

[0075] The three explainable methods of gradient-based representation, reconstructed convolution, and corrected backpropagation are used to analyze the diagnosis results of the chiller fault data, and the influence of different factors on the graph neural network and the explanation method is explored.

[0076] Figure 7 The feature-level explanation process based on the corrected backpropagation method is shown in FIG. 6. Similar to the reconstructed convolution method, the final input feature activation map is obtained by backpropagation, and the importance of each variable is obtained according to the color mapping. Referring to FIG. 6, the fault discrimination feature of the condenser fouling (cf) is still the oil supply pressure (PO_feed), which matches the domain knowledge. Figure 7

[0077] Further, the matching degrees of the three feature-level explainable methods are obtained. Specifically, the three feature-level explainable methods are based on the trained graph neural network model, the node sample of fault diagnosis is input, and the feature importance score vector of the node sample is generated; the feature importance score vector is used to represent the importance of each feature variable in the model judgment result in this diagnosis prediction; under multiple test samples and fault types, the average matching degrees of the three feature-level explainable methods are calculated, and a matching degree graph is drawn; for each feature importance score vector, a key feature subset is extracted by setting an importance threshold, and a domain knowledge feature label set corresponding to the fault type is constructed; the matching degrees between the fault types of the three feature-level explainable methods and the real labels are calculated; different network parameters are compared in the test, and the influence trend of the matching degrees is obtained; the different network parameters include the number of graph convolution layers, the number of channels, and the learning rate. It should be noted that the "trained graph neural network model" is obtained by forward reasoning based on the high-order node representation in step S3, and the network parameters are updated by backpropagation through the loss function, wherein the training is performed by using the fault training data (fault diagnosis node sample).

[0078] In an embodiment, a matching degree calculation formula is constructed to evaluate the effects of the three explainable methods. The matching degree reflects the ability of the model to identify faults and discriminate features. The higher the matching degree, the more accurate the analysis of the explanation method, and the closer it is to the real fault cause based on domain knowledge or expert experience. The matching degree calculation formula can be represented as: ; in the formula, M represents the matching degree, NGNN represents the number of fault discrimination features correctly identified by the model, and NExperience represents the number of preset fault discrimination features in the database.

[0079] The number of preset fault discrimination features in the database can be obtained based on the number of fault discrimination features derived from domain knowledge and expert experience.

[0080] ​The matching degree reflects the model's ability to identify faults and discriminate features. The higher the matching degree, the more accurate the analysis of the interpretation method is, and the closer it is to the actual cause of the fault based on domain knowledge or expert experience.

[0081] By analyzing the influence of parameters such as the number of graph convolution layers, the number of channels, and the learning rate, the evaluation results of the fault discrimination feature matching of three feature-level interpretability methods are obtained. Figure 8 The number of graph convolution layers affects the matching degree of the three feature-level interpretation methods. Specifically, as the number of graph convolution layers increases, the matching degree of the three interpretation methods—gradient representation, reconstructed convolution, and rectified backpropagation—decreases. This confirms that the interpretability of the model is inversely proportional to its complexity. With one graph convolution layer, the three methods can match 76%, 82%, and 82% of the fault discrimination features, respectively. However, with two graph convolution layers, the matching degree drops to 65%. With three graph convolution layers, the matching degree of gradient representation even drops to 53%. Compared with the gradient representation method, reconstructed convolution and rectified backpropagation provide more accurate interpretation results.

[0082] like Figure 9 As shown, for gradient representation methods, fewer channels and smaller learning rates generally result in a higher degree of fit. Therefore, the gradient representation method achieves the most accurate interpretation when using 64 channels and a learning rate of 0.001. For reconstructive convolution, the best interpretation is also achieved with 64 channels and a learning rate of 0.001. For rectified backpropagation, the learning rate significantly influences its interpretation fit. When the learning rate is 0.01, the fit drops significantly. Furthermore, the comparison results show that rectified backpropagation and reconstructive convolution consistently achieve better interpretations across various parameter combinations. Both the number of channels and the learning rate affect the interpretation results, with smaller learning rates being more beneficial for all three methods.

[0083] like Figure 10 As shown, Figure 10 The average interpretation matching degree of the test samples under various fault types for the different interpretation methods provided in the embodiments of the present application can be concluded from the image. The three interpretation methods of gradient representation, reconstruction convolution and layer graph convolution have different average matching degrees for the test samples under various fault types. Among them, the average matching degree of the corrected back propagation is the highest in the fault type ro (excessive refrigerant). By analyzing the average interpretation matching degree of the test samples under various fault types by different interpretation methods, important reference significance is provided for future fault troubleshooting.

[0084] Furthermore, it also includes updating the Euclidean distance between each node based on the importance score vector of each feature dimension in the gradient representation method, the reconstruction convolution method and the corrected back propagation method and its corresponding matching degree. The specific method is: obtaining the importance score vector of each feature dimension in the gradient representation method, the reconstruction convolution method and the corrected back propagation method and its corresponding matching degree; performing proportional analysis on the matching degree of each feature dimension in the gradient representation method, the reconstruction convolution method and the corrected back propagation method respectively, and obtaining the proportion of each feature dimension under each feature-level interpretable method; obtaining each feature proportion interval preset in the database and the feature influence proportion correction factor corresponding to each feature proportion interval, and matching them with the proportion under each feature-level interpretable method, thereby obtaining each feature dimension in The feature influence ratio correction factor under each feature-level interpretable method is marked as the first correction factor of each feature dimension under each interpretation method; each importance score vector interval preset in the database and the reference second correction factor corresponding to each importance score vector interval are obtained, and matched with the importance score vector to obtain the reference second correction factor; based on the first correction factor and the reference second correction factor of each feature dimension under each interpretation method, multiplicative coupling processing is performed to obtain the comprehensive correction factor of each feature dimension under each interpretation method; based on the comprehensive correction factor of each feature dimension under each interpretation method, the database is matched to obtain the Euclidean distance weight factor of each feature dimension; based on the Euclidean distance weight factor of each feature dimension, the Euclidean distance in the graph structure data is updated.

[0085] In an embodiment, the Euclidean distance in the graph structure data is updated in order to realize an adjacency graph construction method that is more in line with actual diagnostic goals in the graph neural network. By introducing the analysis results from the interpretability method, differentiated weights are given to the distance calculation between features, thereby improving the structural representation quality and discrimination ability of the diagnostic graph in the graph structure construction stage.

[0086] This method obtains a first correction factor by introducing matching degree analysis, obtains a second correction factor by analyzing the importance score vector, and obtains a comprehensive correction factor by combining the first correction factor and the second correction factor. This method obtains the Euclidean distance weight factor of each feature dimension, and updates the Euclidean distance in the graph structure data based on the Euclidean distance weight factor of each feature dimension, thereby weighting the feature dimensions with discriminant advantages and guiding the connections between nodes to be closer to the correlation under domain knowledge. Graph convolution relies on the feature propagation of neighboring nodes. When the construction of adjacency relationships is more reasonable, information aggregation will be more focused on directions that have a strong contribution to prediction. By incorporating the interpretability dimension in the graph initialization stage, the model training stage is more controllable, and the interpretation output results are more consistent with the judgment logic. Therefore, the present invention realizes the dynamic optimization of the input graph structure of the graph neural network, thereby improving the semantic relevance of the connections between nodes and enhancing the graph structure's ability to express the chiller failure mode. At the same time, it effectively utilizes the feature significance analysis results under three types of interpretation paths: gradient representation, reconstructed convolution, and corrected backpropagation, so that the graph construction no longer relies on static features, but is adaptively optimized in combination with the diagnostic target, thereby improving the diagnostic accuracy and network convergence efficiency. By constructing a distance adjustment mechanism driven by explanatory factors, the present invention realizes closed-loop feedback between the construction of graph neural network structure and feature-level interpretation results, enhances the interpretability of the model and the credibility of the diagnostic results, and reduces the dependence on the number of training samples and deep network configuration.

[0087] In summary, this embodiment constructs a graph convolutional fault diagnosis network based on a node-level graph generation strategy and introduces three feature-level interpretable methods to analyze the diagnosis process and results, thereby obtaining the quantitative influence of the number of graph convolutional layers, the number of channels and the learning rate on the performance of the graph convolutional fault diagnosis network and the matching degree of the interpretation results, thereby achieving improved transparency of the internal training mechanism and the operation process and improved accuracy of fault cause analysis, effectively solving the problem in the existing technology that it is impossible to take into account both high diagnostic accuracy and deep interpretability of the model.

[0088] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0089] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0090] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0091] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0092] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0093] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model, characterized by: The following steps are involved: S1. Obtain the experimental data set and perform data preprocessing to construct fault training data; S2. A k-nearest neighbor graph construction strategy based on node classification converts the fault training data into graph data and obtains graph structure data. The graph data is used for feature extraction, information transfer, and iterative training of the subsequent graph convolutional fault diagnosis network. S3. Build a graph convolutional fault diagnosis network based on a node-level graph generation strategy. This aggregates and updates the node information and edge connectivity in the graph structure data to obtain a high-order node representation. Backpropagation of the loss function based on the high-order node representation updates the network parameters and completes the final fault diagnosis task. S4. Analyze the diagnostic process and results of graph neural networks using three feature-level interpretability methods, and obtain the quantitative impact of the number of graph convolutional layers, number of channels, and learning rate on the performance of graph convolutional fault diagnosis networks. The three feature-level interpretability methods include gradient representation, reconstructed convolution, and modified backpropagation; S5. Construct a matching evaluation standard based on the three feature-level interpretability methods and the actual fault causes, obtain the matching degree of the three feature-level interpretability methods, and thus obtain the interpretation effect of the three feature-level interpretability methods in the chiller fault experiment; The k-nearest neighbor graph construction strategy based on node classification converts fault training data into graph data and obtains graph structure data at the same time. The specific method is as follows: The fault training data is used as the fault sample features, and the collected fault sample features are used as nodes in the graph structure. The fault training data includes chilled water inlet temperature, chilled water outlet temperature, cooling water inlet temperature, cooling water outlet temperature, evaporation temperature, condensing temperature, suction superheat, refrigerant subcooling, compressor power consumption, oil supply pressure and condenser temperature difference; The pairwise distances between nodes are obtained through Euclidean distance analysis, and the k value is set to convert the fault training data into graph data and obtain graph structure data. The matching degree of the three feature-level interpretable methods is obtained as follows: The three feature-level interpretability methods are all based on the trained graph neural network model, and the fault diagnosis node samples are input to generate the feature importance score vectors of the node samples respectively; The feature importance score vector is used to represent the importance of each feature variable to the model judgment result in this diagnosis prediction; Under multiple test samples and fault types, the average matching degree of the three feature-level interpretability methods is calculated and a matching degree graph is drawn; For each feature importance score vector, a subset of key features is extracted by setting an importance threshold, thereby constructing a set of domain knowledge feature labels corresponding to the fault type; Compute the matching degree between the fault type and the true label of the three feature-level interpretability methods; Comparative tests were conducted on different network parameters to obtain the changing trend of the impact of different network parameters on matching degree; The different network parameters include the number of graph convolution layers, the number of channels, and the learning rate.

2. The feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model as claimed in claim 1, characterized in that: The method for converting the fault training data into graph data and obtaining graph structure data is as follows: The Euclidean distance is used to calculate the pairwise distance between nodes, thereby obtaining the Euclidean distance between each node and other nodes; Based on the Euclidean distance between each node and other nodes, the k nodes closest to it are selected as the neighbors of the node, thereby obtaining k similar neighbors of each node and generating edges representing the relationship between these neighbors; Constructing a k-nearest-neighbor global graph, wherein the k-nearest-neighbor global graph is an undirected graph that uses Euclidean distance to represent weighted neighbor relationships; Node-based k-nearest neighbor relationships and k-nearest neighbor global graph; The k-nearest neighbor graph generation method based on node classification connects all nodes according to the k-nearest neighbor relationship of the nodes, thereby converting the fault training data into graph data, completing the conversion of the fault training data into graph structure data; Extracting node connection relationships and edge weight information from graph data can form graph structure data containing only node topology structure; Each edge in the graph structure data represents a k-nearest neighbor relationship between a pair of nodes, and the weight of the edge is the Euclidean distance between the two nodes.

3. The feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model as claimed in claim 1, characterized in that: The specific method for constructing a graph convolution fault diagnosis network based on a node-level graph generation strategy is as follows: In the graph structure data, the status data of the chiller is embedded into the node as the characteristic attribute of the fault sample, so each node is regarded as a fault sample; Based on the graph neural network, the features of each node and its neighboring nodes are taken as input, and the aggregate representation including the node structure information is obtained through aggregation function analysis, thereby obtaining a graph convolutional fault diagnosis network based on the node-level graph generation strategy.

4. The feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model as claimed in claim 3, characterized in that: The specific method for obtaining the high-order node representation is as follows: The aggregated representation and the node’s own features are input into the update function to obtain the updated representation of each node and the feature vector corresponding to each node; The feature vector is input into the fully connected layer for classification processing, and the predicted fault label is output; Obtain the true fault label and perform difference analysis between the true fault label and the predicted fault label to obtain the error loss; Perform gradient backpropagation based on the error loss to update the parameters of the graph convolutional network; Based on the graph convolution network, multi-layer graph convolution operations are performed to aggregate the features of the node itself with the features of neighboring nodes. After nonlinear transformation and multiple iterations, high-order node representation is obtained; The high-order nodes represent feature information of the fused node itself and its neighboring nodes, and are used to characterize the position and characteristics of the nodes in the graph structure.

5. The feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model as claimed in claim 3, characterized in that: The quantitative influence of the number of graph convolution layers, the number of channels, and the learning rate on the performance of the graph convolution fault diagnosis network is obtained as follows: Based on the gradient representation method, the partial derivative of the input feature to the prediction result is calculated for each prediction node category, and the feature importance vector is generated to obtain the key influencing features; The key influencing features are subsequently used for matching and comparison with the fault discrimination features preset in the database to obtain the accuracy of the gradient representation method; Based on the reconstruction convolution method, the feature map is visualized into the input space through deconvolution operation, and the input features are activated through back propagation; Based on the modified back propagation method, by modifying the gradient calculation rule of ReLU in back propagation, only the positive gradient component is retained, thereby obtaining the explanatory contribution of positive effective features; Based on the performance of the gradient representation method, the reconstructed convolution method, and the modified backpropagation method under different numbers of graph convolution layers, channels, and learning rates, a normalized analysis was conducted to obtain the diagnostic interpretation performance characteristics under different conditions; The feature importance vectors are normalized according to the feature dimension, and their stability and consistency are compared and analyzed with the number of graph convolution layers, number of channels, and learning rate. The quantitative influence of the number of graph convolution layers, number of channels, and learning rate on the explanation matching degree is obtained; The gradient representation method is based on calculating the partial derivative of the input feature to the prediction result for each prediction node category, generating a feature importance vector, and obtaining the key influencing features. The specific method is as follows: Obtain the pixels and graph structure in the input data and input them into the scoring function. Analyze the influence of the pixels in the graph structure on the scoring function and assign a ranking order to the pixel groups. Based on the nonlinear relationship between the score function and the graph structure, the first-order Taylor expansion is used for approximation to obtain the importance of each feature to the model prediction results.

6. The feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model as claimed in claim 5, characterized in that: The reconstruction convolution method visualizes the feature map into the input space through deconvolution operation and activates the input features through back propagation. The specific method is: The feature map of a certain layer that needs to be visualized is input into the deconvolution layer, and the feature map in the input space is activated through back propagation. Then, a feature heat map is constructed through the color mapping principle to obtain the degree of influence of different feature variables in each network layer on the final prediction result; The importance score of the feature is obtained based on the visualization results of the feature map; The visualization result of the feature map is used to identify key feature regions or variables in the prediction task.

7. The feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model as claimed in claim 5, characterized in that: The modified back propagation method modifies the gradient calculation rule of ReLU in back propagation and retains only the positive gradient component, thereby obtaining the explanatory contribution of the positive effective features. The specific method is: Extract the output of the node to be explained from the graph convolutional network, that is, the score of the fault category; The node to be explained is a certain operation sample of the chiller, and its category is a certain type of fault identified by the system; By tracing the reverse computation path of the nodes from the output layer to the input layer, all computational layers that pass through the ReLU activation function are identified; In each layer of back propagation, the gradient passing through ReLU is screened and all gradient components less than zero are removed; According to the corrected reverse path, the forward gradient is propagated from the output layer to the input layer layer by layer, thereby obtaining the reverse gradient contribution value of each feature dimension in the input layer to form the feature importance vector; Normalization is performed based on the feature importance vector to obtain the ranking of the influence of different features on the fault samples; By sorting the feature variables in the feature importance vector from high to low and obtaining a preset number of feature variables in the database as the key judgment basis for the prediction result, the explanatory contribution of the positive and effective features can be obtained, thereby assisting the model to make the final fault diagnosis result and perform credibility verification.

8. The feature-level deep interpretable air conditioning fault diagnosis method based on a neural network model as claimed in claim 1, characterized in that: It also includes the importance score vectors of each feature dimension in the gradient representation method, the reconstructed convolution method, and the modified backpropagation method, as well as their corresponding matching degrees, thereby updating the Euclidean distance between each node. The specific method is: Obtain the importance score vector of each feature dimension in the gradient representation method, the reconstructed convolution method, and the modified backpropagation method, as well as its corresponding matching degree; The matching degree of each feature dimension in the gradient representation method, the reconstructed convolution method and the modified back propagation method is analyzed proportionally, and the proportion of each feature dimension under each feature-level interpretable method is obtained; Obtain each feature proportion interval preset in the database and the feature influence proportion correction factor corresponding to each feature proportion interval, and match them with the proportions under each feature-level interpretable method, thereby obtaining the feature influence proportion correction factor of each feature dimension under each feature-level interpretable method, and mark it as the first correction factor of each feature dimension under each interpretation method; Obtaining each importance score vector interval preset in the database and a reference second correction factor corresponding to each importance score vector interval, and matching them with the importance score vector to obtain the reference second correction factor; Based on the first correction factor of each feature dimension under each interpretation method and the reference second correction factor, a multiplicative coupling process is performed to obtain the comprehensive correction factor of each feature dimension under each interpretation method; Based on the comprehensive correction factors of each feature dimension under each interpretation method, the database is matched to obtain the Euclidean distance weight factor of each feature dimension; The Euclidean distance in the graph structure data is updated based on the Euclidean distance weight factor of each feature dimension.

Images