Water chilling unit fault diagnosis method based on interpretable deep learning model

Abstract

The invention discloses a water chilling unit fault diagnosis method based on an interpretable deep learning model, and the method comprises the following steps: S1, data preprocessing and spatial feature embedding: collecting various sensor data of a water chilling unit, and carrying out the cleaning, filtering and Min-Max normalization preprocessing of the collected data; the explanation mechanism is deeply integrated into the model architecture, decision logic is presented in a visual form, the model explanation ability is improved, meanwhile, spatial feature embedding is performed according to the physical flow path of the water chilling unit, parameter spatial relevance is reserved, the understanding ability of the model to the system state is improved, high precision and practicability are achieved, and the method is suitable for popularization and application. Operation and maintenance can be assisted to quickly position faults, and the operation cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of chiller unit fault diagnosis technology, and in particular to a chiller unit fault diagnosis method based on an interpretable deep learning model. Background Technology

[0002] Chillers are core energy-consuming equipment in building HVAC systems, and their operating status directly affects the energy efficiency, reliability, and operating costs of the entire system. Fault detection and diagnosis technology is crucial to ensuring the safe, efficient, and stable operation of chillers. Early fault diagnosis methods mainly relied on expert experience and knowledge-driven methods based on physical models. These methods heavily depend on domain experts, are difficult to scale, and have poor adaptability. With the development of sensor technology and the Internet of Things, data-driven methods have gradually become mainstream, especially deep learning technology. Due to its powerful automatic feature extraction and complex pattern recognition capabilities, it has shown significant advantages in chiller fault diagnosis. Models such as convolutional neural networks and recurrent neural networks are widely used, and diagnostic accuracy is continuously improving.

[0003] In the field of chiller unit fault diagnosis, traditional machine learning-based methods have been widely used. Support Vector Machines (SVMs) distinguish fault types by constructing an optimal classification hyperplane, demonstrating good performance when processing small sample data. Random Forest algorithms use multiple decision trees for ensemble learning, effectively handling data classification problems in high-dimensional feature spaces. Extreme Gradient Boosting algorithms gradually optimize model prediction performance through a gradient boosting framework, showing significant advantages in feature importance assessment. The development of deep learning technology has provided new solutions for fault diagnosis. One-dimensional convolutional neural networks automatically extract local feature patterns from sensor signals through the sliding operation of convolutional kernels on time-series data. Gated recurrent units (ROUs) utilize the gating mechanism of update and reset gates to effectively model long-term dependencies in time-series data. Some studies combine self-attention mechanisms with recurrent neural networks, enhancing the model's ability to focus on features at key time steps by calculating the importance weights of features at different positions in the sequence.

[0004] However, most existing deep learning-based chiller unit fault diagnosis models are like "black boxes," with opaque internal decision-making logic, making it difficult for maintenance personnel to understand and trust the model's diagnostic results, thus hindering their deployment and application in actual engineering projects. In addition, existing methods typically flatten sensor data into vector inputs, destroying the inherent spatial correlation and topological structure of parameters in the chiller unit's physical system, limiting the model's ability to understand the overall state of the system. In view of the above, this application proposes a chiller unit fault diagnosis method based on an interpretable deep learning model. Summary of the Invention

[0005] Based on the technical problems existing in the background technology, this invention proposes a fault diagnosis method for chiller units based on an interpretable deep learning model.

[0006] This invention proposes a fault diagnosis method for chiller units based on an interpretable deep learning model, comprising the following steps:

[0007] S1: Data preprocessing and spatial feature embedding: Collect data from various sensors of the chiller unit, and perform preprocessing such as cleaning, filtering and Min-Max normalization on the collected data;

[0008] S2: Construct and train the xMHAC-PDCNN chiller unit fault diagnosis model: The xMHAC-PDCNN chiller unit fault diagnosis model is constructed by concatenating the multi-head self-attention mechanism module MHAC guided by multi-scale convolution and the parallel dilated convolutional neural network module PDCNN. The multi-head self-attention mechanism module MHAC dynamically learns the importance of features, and the parallel dilated convolutional neural network module PDCNN extracts multi-scale contextual information. Finally, the fault category probability distribution is output through a fully connected layer and a softmax function.

[0009] S3: Model Interpretive Analysis: The inter-layer correlation propagation technique is used to calculate the contribution of input features to diagnostic decisions and visualize the heat map. The attention weight matrix of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution is extracted and the feature relationship interpretation matrix is ​​obtained after averaging, normalization and max pooling.

[0010] S4: Generation of Diagnostic Results and Interpretive Reports: By combining the classification results of the integrated model, the heatmap of the importance of inter-layer correlation propagation features, and the attention feature relationship diagram, a comprehensive fault diagnosis and interpretation report is generated.

[0011] Preferably, in step S1, the collected data includes 64 features, including but not limited to temperature, pressure, flow rate, and power.

[0012] Preferably, the specific logical steps for processing the data in step S1 are as follows:

[0013] S101: Data cleaning: Remove transient data during system startup and shutdown, as these data change drastically, do not reflect steady-state operating characteristics, and will interfere with model learning;

[0014] S102: Data Normalization: Min-Max normalization is performed on each collected feature data, scaling it to the [0,1] interval to eliminate the influence of numerical scale on model training. The formula used for normalization is: ,in Represents the original value of a certain feature. This represents the minimum value of the feature across the entire training dataset. x represents the maximum value of this feature across the entire training dataset. norm Represents the normalized eigenvalues;

[0015] S103: Spatial Feature Embedding: The normalized 64-dimensional one-dimensional feature vector is systematically rearranged into an 8×8 two-dimensional matrix according to the topology of the five physical flow paths of the chiller unit experimental device. Finally, a tensor with dimensions B×1×8×8 is output, where B is the batch size, 1 is the number of input channels, and 8×8 is the spatial dimension.

[0016] Preferably, in S103, the five physical flow paths include the evaporator side, the condenser side, the compressor and lubrication oil circuit, the refrigerant circuit, and the auxiliary system;

[0017] The principle for arranging an 8×8 two-dimensional matrix is ​​as follows: sensors belonging to the same subsystem are arranged adjacent to each other in the matrix; sensors that are physically close or functionally closely related are also kept spatially adjacent in the matrix.

[0018] Preferably, in S2, the multi-head self-attention mechanism module MHAC guided by multi-scale convolution is used to dynamically learn the importance of each position in the input feature map, and to enhance key features and suppress noise;

[0019] The implementation process of the MHAC (Multi-Head Self-Attention Mechanism Module) guided by multi-scale convolution is as follows:

[0020] S2011: Uses three different convolutional layers to generate query Q, key K, and value V tensors from the same input, respectively. ;

[0021] S2012: The query Q, key K, and value V are split into H heads along the channel dimension, and then the scaled dot product attention weights are calculated using the following formula: ;

[0022] S2013: The outputs of multiple heads are concatenated and passed through a linear layer, then residually connected to the original input X, and a learnable scaling factor γ is introduced, the formula of which is: ;

[0023] Where X represents the input tensor, The number of input channels is represented by h×w, which represents the spatial dimension. This represents a two-dimensional convolutional layer with a kernel size of k; H represents the number of convolution channels; H represents the number of attention heads. This is a scaling factor to prevent the gradient from vanishing due to an excessively large dot product result; Represents the attention weight matrix; Y represents a learnable scalar parameter, initialized to 0; Y represents the final output of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution.

[0024] Preferably, in S2, the parallel dilated convolutional neural network module PDCNN is used to efficiently extract multi-scale contextual information from the feature maps weighted by the MHAC module. This module contains n=3 parallel convolutional branches. The first layer of each branch uses traditional convolutional kernels of different sizes to capture local features of different granularities. In each branch, subsequent layers use dilated convolutions. Dilated convolutions expand the receptive field by inserting spaces between convolutional kernel elements without increasing the number of parameters. Finally, the feature maps output by each parallel branch are concatenated along the channel dimension using the following formula: ; ,in This represents the value of the output feature map at position (i,j); x represents the input feature map. This represents the weight of the convolution kernel at position (m,n); Represents the expansion rate, when When =1, it is a standard convolution; Indicates the bias term; Feature maps representing different branches.

[0025] Preferably, in step S2, the formula used to output the fault category probability distribution through the fully connected layer and the Softmax function is: ,in Let represent the i-th output element of the fully connected layer, and K represent the total number of fault categories.

[0026] Preferably, in step S3, when calculating the contribution of input features to the diagnostic decision using inter-layer correlation propagation technology, the final output decision of the model needs to be backpropagated to the input layer. Different propagation rules are used for different layers in the model, including LRP-ε and LRP-γ, and the formula used in the calculation is as follows: ,in This represents the contribution of neuron i in layer l to neuron j in layer l+1. This represents the value of the i-th feature in layer l. This represents the weights from layer l to layer l+1. This represents the correlation score of layer l+1.

[0027] Preferably, in step S3, the attention weight matrix of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution is extracted, and the feature relation interpretation matrix is ​​obtained by averaging, normalizing, and max pooling. The formula used is as follows:

[0028] ;

[0029] ;

[0030] in This represents the weight matrix of an attention head. This represents the normalized average attention weight matrix. This indicates that the final result is a feature relation interpretation matrix representing a certain fault category.

[0031] Compared with existing technologies, the beneficial effects of this invention are:

[0032] 1. By deeply integrating inter-layer correlation propagation technology, attention weight analysis and other interpretation mechanisms into the xMHAC-PDCNN model architecture, instead of adding interpretation tools outside the model, the contribution of input features to diagnostic decisions can be accurately calculated and visualized. This allows operation and maintenance personnel to clearly "see through" the internal decision-making logic of the model, understand the basis for the generation of diagnostic results, effectively eliminate concerns about trust in the model, overcome the obstacle of the "black box" problem to actual engineering deployment, and promote the technology from theoretical application to engineering implementation.

[0033] 2. By using a spatial feature embedding method based on the physical flow path topology of the chiller unit, the one-dimensional feature vector is rearranged into an 8×8 two-dimensional matrix, which keeps sensors in the same subsystem, with close physical location or functional relationship in spatial proximity, and fully preserves the inherent spatial correlation and topology of the parameters. This input method is more in line with the actual physical system of the chiller unit, helps the model understand the operating state from a "system-level" perspective, avoids the state understanding bias caused by feature fragmentation, and significantly improves the model's ability to perceive and analyze the state of complex systems.

[0034] This invention solves the "black box" problem of existing models by deeply integrating the interpretation mechanism into the model architecture and presenting the decision logic in a visual form. It eliminates the trust concerns of operation and maintenance personnel and helps with engineering deployment. At the same time, it embeds spatial features based on the physical flow path of the chiller unit, preserves the spatial correlation of parameters, improves the model's ability to understand the system state, and has high accuracy, real-time performance and practicality. It can help operation and maintenance personnel quickly locate faults and reduce operating costs. Attached Figure Description

[0035] Figure 1 This is a flowchart of a chiller unit fault diagnosis method based on an interpretable deep learning model proposed in this invention. Detailed Implementation

[0036] The present invention will be further explained below with reference to specific embodiments.

[0037] Example

[0038] Reference Figure 1This embodiment proposes a fault diagnosis method for chiller units based on an interpretable deep learning model, including the following steps:

[0039] S1: Data preprocessing and spatial feature embedding: Collect data from various sensors of the chiller unit, and perform preprocessing such as cleaning, filtering and Min-Max normalization on the collected data;

[0040] The collected data includes 64 features, including but not limited to temperature, pressure, flow rate, and power.

[0041] The specific logical steps for processing the data are as follows:

[0042] S101: Data cleaning: Remove transient data during system startup and shutdown, as these data change drastically, do not reflect steady-state operating characteristics, and will interfere with model learning;

[0043] S102: Data Normalization: Min-Max normalization is performed on each collected feature data, scaling it to the [0,1] interval to eliminate the influence of numerical scale on model training. The formula used for normalization is: ,in Represents the original value of a certain feature. This represents the minimum value of the feature across the entire training dataset. x represents the maximum value of this feature across the entire training dataset. norm Represents the normalized eigenvalues;

[0044] S103: Spatial Feature Embedding: The normalized 64-dimensional one-dimensional feature vector is systematically rearranged into an 8×8 two-dimensional matrix according to the topology of the five physical flow paths of the chiller unit experimental device. Finally, a tensor with dimensions B×1×8×8 is output, where B is the batch size, 1 is the number of input channels, and 8×8 is the spatial dimension.

[0045] The five physical flow paths include the evaporator side, the condenser side, the compressor and lubrication oil circuit, the refrigerant circuit, and the auxiliary system;

[0046] The principle for arranging an 8×8 two-dimensional matrix is ​​as follows: sensors belonging to the same subsystem are arranged adjacent to each other in the matrix; sensors that are physically close or functionally closely related are also kept spatially adjacent in the matrix.

[0047] S2: Construct and train the xMHAC-PDCNN chiller unit fault diagnosis model: The xMHAC-PDCNN chiller unit fault diagnosis model is constructed by concatenating the multi-head self-attention mechanism module MHAC guided by multi-scale convolution and the parallel dilated convolutional neural network module PDCNN. The multi-head self-attention mechanism module MHAC dynamically learns the importance of features, and the parallel dilated convolutional neural network module PDCNN extracts multi-scale contextual information. Finally, the fault category probability distribution is output through a fully connected layer and a softmax function.

[0048] The multi-head self-attention mechanism module MHAC, guided by multi-scale convolution, is used to dynamically learn the importance of each position in the input feature map and enhance key features while suppressing noise.

[0049] The implementation process of the MHAC (Multi-Head Self-Attention Mechanism Module) guided by multi-scale convolution is as follows:

[0050] S2011: Uses three different convolutional layers to generate query Q, key K, and value V tensors from the same input, respectively. ;

[0051] S2012: The query Q, key K, and value V are split into H heads along the channel dimension, and then the scaled dot product attention weights are calculated using the following formula: ;

[0052] S2013: The outputs of multiple heads are concatenated and passed through a linear layer, then residually connected to the original input X, and a learnable scaling factor γ is introduced, the formula of which is: ;

[0053] Where X represents the input tensor, The number of input channels is represented by h×w, which represents the spatial dimension. This represents a two-dimensional convolutional layer with a kernel size of k; H represents the number of convolution channels; H represents the number of attention heads. This is a scaling factor to prevent the gradient from vanishing due to an excessively large dot product result; Represents the attention weight matrix; Y represents a learnable scalar parameter, initialized to 0; Y represents the final output of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution.

[0054] The Parallel Dilated Convolutional Neural Network (PDCNN) module is used to efficiently extract multi-scale contextual information from the weighted feature maps of the MHAC module. This module contains n=3 parallel convolutional branches. The first layer of each branch uses traditional convolutional kernels of different sizes to capture local features of varying granularity. Within each branch, subsequent layers use dilated convolutions, which expand the receptive field by inserting spaces between kernel elements without increasing the number of parameters. Finally, the feature maps output from each parallel branch are concatenated along the channel dimension using the following formula: ; ,in This represents the value of the output feature map at position (i,j); x represents the input feature map. This represents the weight of the convolution kernel at position (m,n); Represents the expansion rate, when When =1, it is a standard convolution; Indicates the bias term; Feature maps representing different branches;

[0055] The formula used to output the fault category probability distribution through a fully connected layer and the Softmax function is: ,in This represents the i-th output element of the fully connected layer, and K represents the total number of fault categories;

[0056] S3: Model Interpretive Analysis: The inter-layer correlation propagation technique is used to calculate the contribution of input features to diagnostic decisions and visualize the heat map. The attention weight matrix of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution is extracted and the feature relationship interpretation matrix is ​​obtained after averaging, normalization and max pooling.

[0057] When using inter-layer correlation propagation techniques to calculate the contribution of input features to diagnostic decisions, the model's final output decision needs to be backpropagated to the input layer. Different propagation rules are used for different layers in the model, including LRP-ε and LRP-γ, and the formula used in the calculation is as follows: ,in This represents the contribution of neuron i in layer l to neuron j in layer l+1. This represents the value of the i-th feature in layer l. This represents the weights from layer l to layer l+1. The relevance score represents layer l+1;

[0058] The attention weight matrix of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution is extracted, and the feature relation interpretation matrix is ​​obtained by averaging, normalizing, and max pooling. The formula used is as follows:

[0059] ;

[0060] ;

[0061] in This represents the weight matrix of an attention head. This represents the normalized average attention weight matrix. This indicates that the final feature relation interpretation matrix representing a certain fault category is obtained;

[0062] S4: Generation of Diagnostic Results and Explanatory Reports: By combining the classification results of the integrated model, the heatmap of the importance of inter-layer correlation propagation features, and the attention feature relationship diagram, a comprehensive fault diagnosis and explanation report is generated. This report not only gives the fault type, but also clearly points out the diagnostic basis (key parameters), and cross-validates with domain knowledge, greatly improving the credibility and practicality of the diagnostic results.

[0063] This embodiment deeply integrates the interpretation mechanism into the model architecture, presents the decision-making logic in a visual form, solves the "black box" problem of existing models, eliminates the trust concerns of operation and maintenance personnel, and assists in engineering deployment. At the same time, it embeds spatial features based on the physical flow path of the chiller unit, preserves the spatial correlation of parameters, improves the model's ability to understand the system state, and has high accuracy, real-time performance and practicality. It can help operation and maintenance personnel quickly locate faults and reduce operating costs.

[0064] Case 1:

[0065] The specific steps for diagnosing chiller unit faults are as follows:

[0066] S01: Data Preparation: The ASHRAE RP-1043 dataset was used. This dataset originated from an experiment of a 316kW centrifugal chiller unit and includes 7 typical faults (such as condenser scaling, refrigerant leakage, excessive lubricating oil, reduced condensate flow, reduced evaporator water flow, non-condensable gases and excessive refrigerant charge) and normal conditions. Data was collected under 27 different operating conditions and includes 64 sensing and calculation parameters.

[0067] S02: Data preprocessing and embedding: Transient data during startup and shutdown are removed, and 64 steady-state parameters are normalized by Min-Max. At the same time, based on the drawings of the five physical flow paths (evaporator side, condenser side, compressor lubrication oil path, etc.) and the sensor locations, the 64 features are carefully arranged into an 8×8 two-dimensional grid as model input.

[0068] S03: Model Training and Validation: The xMHAC-PDCNN model was constructed with the following parameter settings: The MHAC multi-head self-attention mechanism module, guided by multi-scale convolution, uses 8 attention heads, with Q / K convolution kernels of 7×7 and V convolution kernels of 5×5. The parallel dilated convolutional neural network module PDCNN has 3 parallel branches with convolution kernel sizes of 7×7, 5×5, and 3×3, respectively, and uses dilated convolutions with dilation rates of 2 and 4. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio. The model was trained using cross-entropy loss and the Adam optimizer with a learning rate of 0.001. On the test set, the model achieved an F1 score of 1.0 for diagnosing faults such as "reduced condensate flow," demonstrating excellent overall performance.

[0069] S04: Interpretive Analysis: Taking the "condenser scaling" fault as an example: The LRP thermogram shows that the correlation scores of the lubricating oil supply pressure and the lubricating oil inlet and outlet pressure difference are the highest (1.0 and 0.79, respectively), which is completely consistent with the physical mechanism of the fault leading to a decrease in heat exchange efficiency and affecting the operation of the lubrication system; The reconstructed attention map shows that the above two key oil circuit parameters are strongly correlated with the evaporator cooling capacity and other characteristics, revealing the system-level linkage effect captured by the model;

[0070] S05: Result Output: The system finally outputs the diagnostic result as "condenser scaling", along with the LRP thermogram and analysis description from S4, providing maintenance personnel with a clear and reliable action guide.

[0071] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of chiller units based on an interpretable deep learning model, characterized in that, Includes the following steps: S1: Data preprocessing and spatial feature embedding: Collect data from various sensors of the chiller unit, and perform preprocessing such as cleaning, filtering and Min-Max normalization on the collected data; S2: Construct and train the xMHAC-PDCNN chiller unit fault diagnosis model: The xMHAC-PDCNN chiller unit fault diagnosis model is constructed by concatenating the multi-head self-attention mechanism module MHAC guided by multi-scale convolution and the parallel dilated convolutional neural network module PDCNN. The multi-head self-attention mechanism module MHAC dynamically learns the importance of features, and the parallel dilated convolutional neural network module PDCNN extracts multi-scale contextual information. Finally, the fault category probability distribution is output through a fully connected layer and a softmax function. S3: Model Interpretive Analysis: The inter-layer correlation propagation technique is used to calculate the contribution of input features to diagnostic decisions and visualize the heat map. The attention weight matrix of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution is extracted and the feature relationship interpretation matrix is ​​obtained after averaging, normalization and max pooling. S4: Generation of Diagnostic Results and Interpretive Reports: By combining the classification results of the integrated model, the heatmap of the importance of inter-layer correlation propagation features, and the attention feature relationship diagram, a comprehensive fault diagnosis and interpretation report is generated.

2. The method for fault diagnosis of chiller units based on an interpretable deep learning model according to claim 1, characterized in that, In S1, the collected data includes 64 features, including but not limited to temperature, pressure, flow rate, and power.

3. The method for fault diagnosis of chiller units based on an interpretable deep learning model according to claim 1, characterized in that, The specific logical steps for processing the data in S1 are as follows: S101: Data cleaning: Remove transient data during system startup and shutdown, as these data change drastically, do not reflect steady-state operating characteristics, and will interfere with model learning; S102: Data Normalization: Min-Max normalization is performed on each collected feature data, scaling it to the [0,1] interval to eliminate the influence of numerical scale on model training. The formula used for normalization is: ,in Represents the original value of a certain feature. This represents the minimum value of the feature across the entire training dataset. x represents the maximum value of this feature across the entire training dataset. norm Represents the normalized eigenvalues; S103: Spatial Feature Embedding: The normalized 64-dimensional one-dimensional feature vector is systematically rearranged into an 8×8 two-dimensional matrix according to the topology of the five physical flow paths of the chiller unit experimental device. Finally, a tensor with dimensions B×1×8×8 is output, where B is the batch size, 1 is the number of input channels, and 8×8 is the spatial dimension.

4. The method for fault diagnosis of chiller units based on an interpretable deep learning model according to claim 3, characterized in that, In S103, the five physical flow paths include the evaporator side, the condenser side, the compressor and lubrication oil circuit, the refrigerant circuit, and the auxiliary system; The principle for arranging an 8×8 two-dimensional matrix is ​​as follows: sensors belonging to the same subsystem are arranged adjacent to each other in the matrix; sensors that are physically close or functionally closely related are also kept spatially adjacent in the matrix.

5. The method for fault diagnosis of chiller units based on an interpretable deep learning model according to claim 1, characterized in that, In S2, the multi-head self-attention mechanism module MHAC guided by multi-scale convolution is used to dynamically learn the importance of each position in the input feature map, and to enhance key features and suppress noise; The implementation process of the MHAC (Multi-Head Self-Attention Mechanism Module) guided by multi-scale convolution is as follows: S2011: Uses three different convolutional layers to generate query Q, key K, and value V tensors from the same input, respectively. ; S2012: The query Q, key K, and value V are split into H heads along the channel dimension, and then the scaled dot product attention weights are calculated using the following formula: ; S2013: The outputs of multiple heads are concatenated and passed through a linear layer, then residually connected to the original input X, and a learnable scaling factor γ is introduced, the formula of which is: ; Where X represents the input tensor, The number of input channels is represented by h×w, which represents the spatial dimension. This represents a two-dimensional convolutional layer with a kernel size of k; H represents the number of convolution channels; H represents the number of attention heads. This is a scaling factor to prevent the gradient from vanishing due to an excessively large dot product result; Represents the attention weight matrix; Y represents a learnable scalar parameter, initialized to 0; Y represents the final output of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution.

6. The method for fault diagnosis of chiller units based on an interpretable deep learning model according to claim 1, characterized in that, In S2, the parallel dilated convolutional neural network module PDCNN is used to efficiently extract multi-scale contextual information from the weighted feature map of the MHAC module. This module contains n=3 parallel convolutional branches. The first layer of each branch uses traditional convolutional kernels of different sizes to capture local features of different granularities. In each branch, subsequent layers use dilated convolutions. Dilated convolutions expand the receptive field by inserting spaces between convolutional kernel elements without increasing the number of parameters. Finally, the feature maps output by each parallel branch are concatenated along the channel dimension using the following formula: ; ,in This represents the value of the output feature map at position (i,j); x represents the input feature map. This represents the weight of the convolution kernel at position (m,n); Represents the expansion rate, when When =1, it is a standard convolution; Indicates the bias term; Feature maps representing different branches.

7. The method for fault diagnosis of chiller units based on an interpretable deep learning model according to claim 1, characterized in that, In S2, the formula used to output the fault category probability distribution through the fully connected layer and the Softmax function is: ,in Let represent the i-th output element of the fully connected layer, and K represent the total number of fault categories.

8. The method for fault diagnosis of chiller units based on an interpretable deep learning model according to claim 1, characterized in that, In step S3, when calculating the contribution of input features to the diagnostic decision using inter-layer correlation propagation technology, the final output decision of the model needs to be backpropagated to the input layer. Different propagation rules are used for different layers in the model, including LRP-ε and LRP-γ, and the formula used in the calculation is as follows: ,in This represents the contribution of neuron i in layer l to neuron j in layer l+1. This represents the value of the i-th feature in layer l. This represents the weights from layer l to layer l+1. This represents the correlation score of layer l+1.

9. The method for fault diagnosis of chiller units based on an interpretable deep learning model according to claim 1, characterized in that, In step S3, the attention weight matrix of the multi-head self-attention mechanism module MHAC guided by multi-scale convolution is extracted, and the feature relation interpretation matrix is ​​obtained by averaging, normalizing, and max pooling. The formula used is as follows: ； ； in This represents the weight matrix of an attention head. This represents the normalized average attention weight matrix. This indicates that the final result is a feature relation interpretation matrix representing a certain fault category.

Images