Fault diagnosis method and device for heat exchange module

Abstract

The present invention relates to the technical field of heat exchange modules, and discloses a fault diagnosis method and device for a heat exchange module, wherein the method comprises: constructing a finite element model of the heat exchange module and collecting temperature signals, pressure signals and flow signals, dividing them into a source domain data set and a target domain data set; performing time-series spatial feature enhancement processing to obtain a sample enhanced feature image; constructing a feature extraction network; performing training and error reconstruction to obtain a pre-trained diagnosis model; fixing shallow feature extraction parameters, and inputting the target domain data set into the pre-trained diagnosis model, performing enhanced training on the fault features, and obtaining a trained fault diagnosis model; inputting real-time operation data into the trained fault diagnosis model to calculate the probability distribution of the fault type and evaluate the fault degree, and generating a fault diagnosis report. The present invention improves the feature fusion effect between multi-source signals of the heat exchange module, and enhances the ability to identify different types of faults.

Description

Technical Field

[0001] The present invention relates to the technical field of heat exchange modules, and in particular to a fault diagnosis method and device for a heat exchange module. Background Art

[0002] The operating status of the heat exchange module directly affects the energy efficiency and safety of the entire system, so accurate fault diagnosis is of great significance. Traditional fault diagnosis methods mainly rely on expert experience and simple signal processing technology, which is difficult to cope with fault diagnosis tasks under complex working conditions.

[0003] In recent years, deep learning technology has made significant progress in the field of fault diagnosis, but it still faces many challenges in the fault diagnosis of heat exchange modules. The primary problem is that the data distribution under different working conditions varies greatly, which makes it difficult to migrate the model trained under data-sufficient conditions to data-sufficient conditions. Secondly, existing fault diagnosis methods often only focus on the extraction of a single feature, ignoring the spatiotemporal correlation between multi-source signals such as temperature, pressure, and flow. In addition, due to the diversity and complexity of heat exchange module fault types, traditional feature extraction methods are difficult to capture the dynamic evolution of faults, especially in the extraction of weak features in the early stages of faults. Summary of the invention

[0004] The present invention provides a method and device for diagnosing a fault of a heat exchange module, which improves the feature fusion effect between multi-source signals of the heat exchange module and enhances the ability to identify different types of faults.

[0005] In a first aspect, the present invention provides a method for diagnosing a fault of a heat exchange module, the method for diagnosing a fault of a heat exchange module comprising:

[0006] Construct a finite element model of the heat exchange module and collect temperature signals, pressure signals, and flow signals to obtain source domain data sets and target domain data sets;

[0007] Performing Markov transition field transformation and temporal spatial feature enhancement processing on the source domain data set and the target domain data set to obtain a sample enhanced feature image;

[0008] Constructing a Siamese diagnostic network with a spatiotemporal attention mechanism according to the sample enhanced feature image, embedding an encoder-decoder structure and a channel attention module in the Siamese diagnostic network to obtain a feature extraction network;

[0009] Inputting the source domain data set into the feature extraction network for training and error reconstruction to obtain a pre-trained diagnostic model;

[0010] Fixing the shallow feature extraction parameters of the pre-trained diagnostic model, inputting the target domain data set into the pre-trained diagnostic model, performing enhanced training on the fault features, and obtaining a trained fault diagnosis model;

[0011] The real-time operation data of the heat exchange module is input into the trained fault diagnosis model to calculate the probability distribution of fault types and evaluate the fault degree, and generate a fault diagnosis report.

[0012] In a second aspect, the present invention provides a heat exchange module fault diagnosis device, the heat exchange module fault diagnosis device comprising:

[0013] An acquisition module is used to construct a finite element model of the heat exchange module and acquire temperature signals, pressure signals, and flow signals, and divide them into a source domain data set and a target domain data set;

[0014] A transformation module, used for performing Markov transition field transformation and temporal spatial feature enhancement processing on the source domain data set and the target domain data set to obtain a sample enhanced feature image;

[0015] A construction module is used to construct a Siamese diagnostic network with a spatiotemporal attention mechanism according to the sample enhanced feature image, and embed an encoder-decoder structure and a channel attention module in the Siamese diagnostic network to obtain a feature extraction network;

[0016] An error reconstruction module, used for inputting the source domain data set into the feature extraction network for training and error reconstruction to obtain a pre-trained diagnostic model;

[0017] An enhanced training module, used to fix the shallow feature extraction parameters of the pre-trained diagnostic model, input the target domain data set into the pre-trained diagnostic model, perform enhanced training on the fault features, and obtain a trained fault diagnosis model;

[0018] A generation module is used to input the real-time operation data of the heat exchange module into the trained fault diagnosis model to calculate the probability distribution of fault types and evaluate the fault degree, and generate a fault diagnosis report.

[0019] A third aspect of the present invention provides a fault diagnosis device for a heat exchange module, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the fault diagnosis device for the heat exchange module executes the above-mentioned fault diagnosis method for the heat exchange module.

[0020] A fourth aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, which, when executed on a computer, enable the computer to execute the above-mentioned method for diagnosing faults of a heat exchange module.

[0021] In the technical solution provided by the present invention, by constructing a feature enhancement method combining a Markov transition field and a recursive graph, the temporal dynamic features in multi-source signals such as temperature, pressure, and flow are effectively extracted, and the expression ability of fault features is enhanced; the Siamese diagnostic network structure and the spatiotemporal attention mechanism are adopted to realize the effective migration of fault features under different working conditions and solve the problem of data imbalance; the encoder-decoder structure and the channel attention module are introduced to improve the feature fusion effect between multi-source signals and enhance the model's recognition ability for different types of faults; a dynamic softmax loss function with an angular margin penalty is designed to optimize the classification boundary of the fault features and improve the generalization performance of the model; by fixing shallow features and migrating deep features, the knowledge transfer from the source domain to the target domain is realized, and the diagnostic effect of the model under insufficient data conditions is enhanced.

[0022] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention are realized and obtained by the structures particularly pointed out in the description, claims and drawings.

[0023] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of an embodiment of a method for diagnosing a fault of a heat exchange module according to an embodiment of the present invention;

[0025] Figure 2 A schematic diagram of an embodiment of a fault diagnosis device for a heat exchange module according to an embodiment of the present invention;

[0026] Figure 3 Schematic diagram of an embodiment of a fault diagnosis device for a heat exchange module in an embodiment of the present invention. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The terms "including" and "having" and any variations thereof mentioned in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device end including a series of steps or units is not limited to the listed steps or units, but may optionally include other steps or units that are not listed, or may optionally include other steps or units that are inherent to these processes, methods, products or device ends.

[0029] To facilitate understanding of this embodiment, a method for diagnosing a fault of a heat exchange module disclosed in an embodiment of the present invention is first described in detail. Figure 1 As shown, the method comprises the following steps:

[0030] 101. Construct a finite element model of the heat exchange module and collect temperature signals, pressure signals, and flow signals to obtain a source domain data set and a target domain data set;

[0031] It is understandable that the execution subject of the present invention may be a fault diagnosis device for a heat exchange module, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking a server as the execution subject.

[0032] Specifically, a three-dimensional grid model is established on a computer according to the specific structural dimensions, internal flow channel distribution and working parameters (such as working fluid type, inlet and outlet temperature, inlet and outlet pressure, volume flow, heat exchange requirements, material properties and operating environment conditions) of the heat exchange module. The three-dimensional model is subjected to reasonable geometric processing and a suitable grid division strategy is selected, such as the use of unstructured grids or multi-scale grid technology, so that it can not only encrypt the grid in key parts to improve local accuracy, but also moderately reduce the grid density in relatively uniform areas to control the calculation cost. After the grid division is completed, the finite element model of the heat exchange module is obtained. The finite element model is used to perform stress analysis and thermal flow field analysis on the heat exchange module to determine the key areas that are most sensitive to the system operation status and can reflect the fault characteristics. Through the numerical simulation of stress distribution and temperature field, pressure field and flow field, the target monitoring point coordinate set that is crucial to fault identification is extracted; these coordinate points are the reference positions for the subsequent sensor layout. According to the analysis results, high-precision temperature sensors, pressure sensors and flow sensors are installed at the corresponding positions of the physical device, and the acquisition parameters are set according to specific measurement requirements (such as data acquisition frequency, sampling interval, sensor resolution and noise filtering parameters) to achieve real-time data acquisition of the operating status of the heat exchange module. The heat exchange module is placed under various working conditions for experiments and data acquisition, including standard normal operating conditions and abnormal operating conditions under various known fault modes. The original data sets of temperature signals, pressure signals and flow signals are obtained through long-term stable sampling of the sensor array under different working conditions. These data show different dynamic characteristics in time series, thus reflecting the differences in the internal heat transfer and flow mechanisms of the system under healthy and faulty conditions. The original data set is segmented according to the set time window to form multiple segmented data matrices, which helps to extract more stable statistical features and time features in subsequent analysis and reduce the impact of outliers or noise on overall feature extraction. Data preprocessing is performed on the segmented data matrix, including signal denoising, normalization, filtering, missing value completion, and outlier detection and elimination operations to ensure that the data input to the model has good numerical stability and consistency, and a unified preprocessed data matrix is ​​obtained. According to the data volume distribution of the working condition type and the needs of the actual problem, the preprocessed data matrix is ​​divided into a source domain data set and a target domain data set. The source domain data set comes from relatively standard and fully labeled laboratory working conditions or simulation scenarios with known fault types and degrees. These data lay the foundation for subsequent model pre-training and feature learning. The target domain data set comes from the operating condition data in the actual application environment. The fault type distribution of this type of data is insufficient or there is a certain distribution offset, which constitutes an object that needs to be adapted through migration and fine-tuning.

[0033] 102. Performing Markov transition field transformation and temporal spatial feature enhancement processing on the source domain data set and the target domain data set to obtain a sample enhanced feature image;

[0034] Specifically, the temperature signal, pressure signal and flow signal obtained from the source domain data set and the target domain data set are divided into states, and the numerical range of the signal is processed in segments. By dividing the appropriate state interval, the continuous signal is converted into a discrete Markov state value. The discretization process can retain the core change characteristics of the signal and facilitate the subsequent Markov analysis and statistical modeling. The state transition frequency of the signal is counted based on the Markov state value, that is, the number of times each state is transferred from the current time point to the next time point is calculated. These frequencies are converted into a state transition probability matrix after standardization, thereby fully describing the dynamic evolution pattern of the signal in the time series. According to the state transition probability matrix, a two-dimensional image mapping relationship is constructed, and the transition probability of each state pair is mapped to the corresponding grayscale value to generate the Markov transition field feature image. The Markov transition field feature image is a visual representation method in the form of a two-dimensional matrix, which can intuitively show the dynamic change law of the signal from the texture and grayscale distribution of the image. In order to capture the implicit temporal correlation and local characteristics in the feature image, the Euclidean distance of the feature vector of each time point of the Markov transition field feature image is calculated. By calculating the Euclidean distance of the feature vectors between different time points, a distance matrix is ​​generated, which reflects the similarity and difference of the time series data at different time points. Based on this distance matrix, by setting a reasonable threshold condition, the point pairs with a distance greater than the threshold are assigned a value of 1, and the point pairs with a distance less than the threshold are assigned a value of 0, and the binary matrix of the recursive graph is generated. The recursive graph can capture the periodic or self-similar patterns in the signal. Therefore, by analyzing the structure of the recursive graph, the temporal pattern and periodic features are extracted. These features contain the core dynamic information of the signal changing over time. The generated Markov transition field feature image and the extracted temporal spatial features are used as input and passed to the low-rank representation network for feature fusion. The low-rank representation network has good dimensionality reduction and feature integration performance, and can convert the input multi-dimensional feature representation into a structured and high-information-density fusion feature, thereby more comprehensively characterizing the intrinsic characteristics and external manifestations of the fault signal. After dimensional reconstruction and normalization, the fusion feature is converted into a sample enhanced feature image. Dimension reconstruction further reduces feature redundancy, while normalization ensures the consistency of feature value range and improves its efficiency and stability in subsequent models. Sample enhanced feature images are obtained.

[0035] 103. Construct a Siamese diagnosis network with a spatiotemporal attention mechanism based on the sample enhanced feature image, embed an encoder-decoder structure and a channel attention module in the Siamese diagnosis network to obtain a feature extraction network;

[0036] Specifically, the sample enhanced feature image is taken as input and passed to the Siamese diagnostic network, which consists of two weight-sharing Siamese branch networks. The main structure of each branch network includes 4 convolutional layers and 4 maximum pooling layers to extract multi-scale feature representations. Among them, the convolution kernel size of the first convolutional layer is set to 7×7, the step size is 2, and the number of output channels is 64, which is used to capture the preliminary feature pattern of the input image. Deeper features are extracted through three convolutional layers with a convolution kernel size of 3×3 and a step size of 1, and the number of output channels is 128, 256, and 512 respectively. The maximum pooling layer is connected after each convolutional layer to ensure that the features are reduced layer by layer in spatial resolution while retaining important local information. After being processed by this structure, the sample Siamese feature map is generated. The sample Siamese feature map is encoder processed to obtain higher-level encoded features. The encoder consists of three encoding blocks, each of which contains two convolutional layers and one batch normalization layer. The convolution operation with a kernel size of 3×3 and a stride of 2 is used to extract more abstract feature representations. The batch normalization layer effectively accelerates the network training process and enhances its generalization ability. After each convolution operation, the ReLU activation function is used to introduce nonlinear transformation to ensure the richness of the features. After the encoder processing is completed, the generated encoding features enter the decoder module for feature reconstruction. The decoder structure consists of three decoding blocks, each of which includes two deconvolution layers and a batch normalization layer. The deconvolution operation restores the spatial dimension of the features through a reverse convolution with a kernel size of 3×3 and a stride of 2, while retaining the important information in the encoding features. Similar to the encoder, the ReLU activation function is applied after each deconvolution layer to ensure the continuity and expressiveness of the decoding features and generate sample decoding features. On this basis, the decoding features are used to construct a channel attention module to dynamically adjust the channel weights of the features. The sample decoding features are average pooled and max pooled to obtain two different global feature descriptors, which can extract global statistical information in each channel. The two descriptors are processed by a shared multi-layer perceptron to generate corresponding sample channel attention weights, which are used to measure the importance of each channel on different features. The sample channel attention weights are element-wise dot-producted with the sample decoding features to obtain the sample channel enhancement features, thereby highlighting the decoding features in the channel dimension. In order to capture the time series pattern of the sample channel enhancement features, a time series segmentation operation is performed on them to generate N sample time series segments. The sample time series segments are input into the time attention module and processed by the self-attention mechanism. The self-attention mechanism calculates the correlation of each time segment in the global time series to mine the long-term dependencies and dynamic patterns in the time series and obtain the sample time attention features. After the time attention features are fused with the sample channel enhancement features, the sample spatiotemporal enhancement features with comprehensive expression capabilities are generated.Perform global average pooling on the sample spatiotemporal enhancement features to convert their high-dimensional features into global feature vectors of fixed length. Perform feature transformation through two fully connected layers to extract and classify related feature representations, and finally complete the entire feature extraction network.

[0037] 104. Input the source domain data set into the feature extraction network for training and error reconstruction to obtain a pre-trained diagnostic model;

[0038] Specifically, the source domain dataset is input into the feature extraction network, which extracts the key features of the data through the encoder, decoder structure and attention module designed in advance. After the feature extraction process, the generated feature vector has high dimension and inconsistent distribution. It is normalized to map the feature vector to a unified numerical range, thereby improving the stability and reliability of subsequent similarity calculation and loss function optimization, and obtaining a normalized feature vector. According to the normalized feature vector, the cosine similarity matrix is ​​calculated. Cosine similarity is a measurement method that can evaluate the similarity between features by calculating the cosine value of the angle between vectors. The cosine similarity matrix is ​​multiplied by a scaling factor to obtain a scaled similarity matrix. Based on the scaled similarity matrix, the angle margin loss of sample pairs within the class and the angle margin loss of sample pairs between classes are calculated. The angle margin loss of sample pairs within the class is calculated by analyzing the angle distribution of sample pairs within the same class, and the angle margin loss of sample pairs between classes is calculated by comparing the angles between sample pairs of different classes. These two losses together constitute the first angle margin loss, which is used to constrain the angular distribution of features, making intra-class features more compact and inter-class features more dispersed, thereby improving the discrimination ability of the classifier. The category weight coefficient is calculated according to the distribution of samples in each category. These weight coefficients reflect the imbalance of sample distribution of different categories in the source domain dataset, thereby giving different weights to the angle margin losses of different categories. By multiplying the category weight coefficient with the first angle margin loss, the second angle margin loss is generated to optimize the model's sensitivity and discrimination to features of different categories. At the same time, in the feature extraction network, the reconstruction function of the encoder and decoder is optimized. The reconstruction loss of the network is obtained by calculating the mean square error between the reconstructed features and the input features. The mean square error measures the accuracy of the network in reconstructing the input data, thereby helping the network to better learn the low-dimensional representation of the input data. After the angle margin loss and the reconstruction loss are calculated, the two are weighted and summed according to a certain weight ratio to obtain the total loss function of the network. Based on the total loss function, the feature extraction network is back-propagated and parameter optimized. By calculating the gradient and adjusting the network parameters, the loss function gradually converges, and the network can more accurately extract and reconstruct the key features of the input data. The entire training process is carried out in an iterative optimization manner. After each round of iteration, the network parameters are updated so that the total loss gradually decreases. When the total loss function in the training process reaches a convergence state, it indicates that the network's feature extraction and reconstruction capabilities have reached an ideal level, and the final pre-trained diagnostic model is obtained.

[0039] 105. Fix the shallow feature extraction parameters of the pre-trained diagnosis model, input the target domain data set into the pre-trained diagnosis model, perform enhanced training on the fault features, and obtain a trained fault diagnosis model;

[0040] Specifically, the structure of the Siamese branch network is extracted from the pre-trained diagnostic model, and the parameters of the four convolutional layers and four maximum pooling layers in the network are locked. The locking operation fixes the network weights of the shallow feature extraction, ensuring that these shallow parameters are no longer updated when training on the target domain dataset, thereby retaining the basic feature expression ability learned in the source domain pre-training stage. The locked network is the Siamese basic network, whose function is to process the input target domain dataset through the fixed shallow feature extraction module to generate the basic features of the target domain. The basic features are the low-level representations of the target domain data after preliminary conversion, which can effectively characterize the local patterns and texture information of the input data. The basic features of the target domain are input into the encoder and decoder modules for feature reconstruction operations. The encoder compresses and abstracts the basic features of the target domain through multi-layer convolution operations to extract key high-level semantic information, while the decoder restores these high-level features to reconstructed features similar to the basic features through layer-by-layer deconvolution to obtain the reconstructed features of the target domain. The reconstructed features of the target domain are optimized by constructing a channel attention module. The reconstructed features of the target domain are subjected to average pooling and maximum pooling operations respectively to obtain two global descriptors, which summarize the overall information and local peak information of each channel feature respectively. The two descriptors are processed by a shared multi-layer perceptron to calculate the attention weight of each channel. The attention weight is subjected to a channel-by-channel weighted dot product operation with the reconstructed features of the target domain to generate the channel-enhanced features of the target domain, which highlight the important features related to the fault and suppress the interference information. The channel-enhanced features of the target domain are temporally segmented and processed by the self-attention mechanism. The temporal segmentation operation divides the target domain features into multiple time segments in order to analyze their time-varying patterns. The self-attention mechanism generates the spatiotemporal enhancement features of the target domain by calculating the similarity and importance of each time segment in the global temporal sequence. These features contain the dynamic spatiotemporal structure information of the target domain data. The spatiotemporal enhancement features of the target domain are input into the fault classifier, which predicts the probability distribution of different fault types and evaluates the severity of each fault. In order to optimize the fault diagnosis performance of the target domain, the classification loss and regression loss are calculated based on the diagnosis results of the target domain. The classification loss is used to evaluate the accuracy of fault type prediction, while the regression loss reflects the deviation of fault degree estimation. These two parts of loss together constitute the target domain optimization objective function, which is used as a guiding target for model parameter update throughout the training process. Through the back propagation algorithm, the gradient information of the objective function is passed layer by layer to the updateable parameter part of the network, especially the fault classifier module, and its parameter settings are gradually optimized. As the training iterations continue, the overall loss value of the network will gradually decrease. When the loss value of the objective function reaches a convergence state, it indicates that the network's feature extraction and fault classification capabilities for the target domain data have reached the optimal state, and the final trained fault diagnosis model is obtained.

[0041] 106. The real-time operation data of the heat exchange module is input into the trained fault diagnosis model to calculate the probability distribution of the fault type and evaluate the fault degree, and generate a fault diagnosis report.

[0042] Specifically, the real-time temperature signal, real-time pressure signal and real-time flow signal of the heat exchange module during operation are obtained through the real-time sensor network. These signals constitute the core monitoring data of the system operation status, and these signals are processed in segments through a unified time window. The continuous real-time data sequence is divided into multiple small operable units to facilitate subsequent data analysis and feature extraction. On the basis of segmentation, the real-time data is standardized to eliminate the dimensional differences and measurement noise between different signals and generate standardized real-time data. The standardized real-time data is divided into Markov state values ​​and the state transition probability is calculated. By dividing the numerical range of real-time data, the continuous real-time signal is converted into discrete Markov state values, and the state transition probability is calculated by counting the transition frequency between each state. These probabilities are visualized in the form of images to generate real-time enhanced feature images. Real-time enhanced feature images can intuitively reflect the dynamic patterns and key characteristics in real-time data. The real-time enhanced feature map is input into the Siamese diagnosis network of the trained fault diagnosis model. The pre-trained 4 convolutional layers and 4 maximum pooling layers are used to extract multi-scale features to obtain the real-time Siamese feature map, which captures the local structural information and basic features of the fault mode in the real-time data. In order to extract deeper semantic information, the real-time Siamese feature map is input into the encoder module of the model and encoded through 3 encoding blocks in sequence. Each encoding block compresses and abstracts the input features through convolution operations and batch normalization layers to generate real-time decoding features. Average pooling and maximum pooling operations are performed on the real-time decoding features to obtain two real-time feature descriptors. The real-time feature descriptors are input into a shared multi-layer perceptron to calculate the real-time channel attention weights of each feature channel. These weights are used to perform channel-by-channel dot product operations with the real-time decoding features to generate real-time channel enhanced features. The real-time channel enhanced features are divided into N equal parts in time series to generate N real-time time series segments. The real-time time series segments are input into the self-attention module. By calculating the correlation and importance between different time segments, real-time time attention features are generated to capture the global dynamic patterns and local dependencies in the time series, and the features are fused with the real-time channel enhancement features to form the final comprehensive feature representation. The fused features are reduced in dimension through the global average pooling operation, and feature mapping is performed through two fully connected layers to generate the target feature vector. The target feature vector is input into the fault classifier, and the probability distribution of different fault types is calculated by the classifier, and the severity of the fault is evaluated to obtain the fault type probability distribution and fault severity score. According to the preset fault level threshold, the fault severity score is mapped to the corresponding fault level, so as to intuitively evaluate the impact of the fault on the system operation. Combined with the fault type probability distribution, fault severity score and fault level, a comprehensive fault diagnosis report is generated.

[0043] In an embodiment of the present invention, by constructing a feature enhancement method combining a Markov transition field and a recursive graph, the temporal dynamic features in multi-source signals such as temperature, pressure, and flow are effectively extracted, and the expression ability of fault features is enhanced; the Siamese diagnostic network structure and the spatiotemporal attention mechanism are adopted to realize the effective migration of fault features under different working conditions and solve the problem of data imbalance; the encoder-decoder structure and the channel attention module are introduced to improve the feature fusion effect between multi-source signals and enhance the model's ability to recognize different types of faults; a dynamic softmax loss function with an angle margin penalty is designed to optimize the classification boundary of the fault features and improve the generalization performance of the model; by fixing shallow features and migrating deep features, the knowledge transfer from the source domain to the target domain is realized, and the diagnostic effect of the model under insufficient data conditions is enhanced.

[0044] In a specific embodiment, the process of executing step 101 may specifically include the following steps:

[0045] A three-dimensional grid model is established based on the structural parameters and working parameters of the heat exchange module, and the three-dimensional grid model is meshed to obtain a finite element model of the heat exchange module;

[0046] Perform stress analysis and thermal flow field analysis on the finite element model to obtain the target monitoring point coordinate set, and install temperature sensors, pressure sensors and flow sensors at the physical positions corresponding to the target monitoring point coordinate set, and set sensor acquisition parameters;

[0047] Perform multi-condition data collection on the heat exchange module according to the sensor acquisition parameters, record the temperature signal, pressure signal and flow signal under normal working conditions and fault conditions, and obtain the original data set;

[0048] The original data set is segmented according to the time window to obtain a segmented data matrix, and the segmented data matrix is ​​preprocessed to obtain a preprocessed data matrix;

[0049] The preprocessed data matrix is ​​divided into source domain data set and target domain data set according to the data volume of the working condition type.

[0050] Specifically, by collecting the structural parameters of the heat exchange module (such as the flow channel diameter , heat exchanger length , wall thickness , Material properties and etc.) and operating parameters (such as working fluid density , flow rate , inlet temperature , outlet pressure Based on these parameters, the flow channel and wall structure of the heat exchange module are constructed to ensure that the geometric model truly reflects the physical characteristics of the actual equipment. In order to meet the calculation requirements of finite element analysis, the three-dimensional model is meshed. Assume that the mesh unit type is tetrahedral unit and its size is It should be determined according to the regional characteristics. For example, a smaller mesh size is selected at the inlet and curved part of the heat exchanger flow channel to improve the calculation accuracy, while a larger mesh is selected in the straight pipe section to reduce the calculation cost. The goal of meshing is to generate a finite element model, in which the node coordinates are marked as , the number of units is , the number of nodes is . Perform stress analysis and thermal flow field analysis on the finite element model. Stress analysis uses the equilibrium equation:

[0051] ;

[0052] in is the stress tensor, is the body density vector, Represents the spatial gradient operator. Thermal flow field analysis is based on the steady-state heat conduction equation:

[0053] ;

[0054] in is the thermal conductivity of the material, is the temperature field, is the body heat source density. By numerically solving these equations, the distribution data of the temperature field and stress field in the heat exchange module are obtained. According to the calculation results, the areas sensitive to faults are selected as target monitoring points. These points are determined by identifying areas with large temperature gradients or stress concentrations, and their coordinates are expressed by the set Indicates that is the number of monitoring points. Temperature sensors, pressure sensors and flow sensors are installed at the physical locations corresponding to these target monitoring points to record the temperature ,pressure and flow To ensure the accuracy and frequency of data collection, set the sensor collection parameters according to the working conditions, such as sampling frequency. , Resolution and sampling time window After the hardware layout is completed, multi-condition data collection is performed. By controlling the operating parameters of the heat exchange module, normal working conditions (such as smooth fluid flow) and fault conditions (such as flow channel blockage, leakage, heat exchange performance degradation, etc.) are simulated, and temperature, pressure and flow signals are collected under each working condition to form an original data set. Assume that the total data sampling time is , the sampling frequency is , then the total number of sampling points for each signal is The collected raw data is expressed in matrix form as follows:

[0055] ;

[0056] Each column corresponds to the time series of a sensor, and each row corresponds to data at different time points. The original data set is segmented according to the time window. Assume that the time window length is , then the data matrix obtained after segmentation is expressed as ,in is the number of time windows. Perform data preprocessing on the segmented data matrix, including denoising (e.g., applying a low-pass filter ), normalization (such as using maximum value normalization ), and missing value filling (such as linear interpolation). The preprocessed data matrix is ​​used According to the amount of data of the working condition type, the preprocessed data matrix is ​​divided and the normal working condition data is used as the source domain data set. , fault condition data is used as the target domain dataset Assume that the number of samples under normal conditions is , the number of fault condition samples is ,but Taking pipeline blockage as an example, assuming that the flow signal under the blockage condition is There is a significant downward trend, The included features significantly reflect the failure mode, thus providing effective data support for subsequent analysis.

[0057] In a specific embodiment, the process of executing step 102 may specifically include the following steps:

[0058] The temperature signal, pressure signal and flow signal in the source domain data set and the target domain data set are divided into states according to the numerical range to obtain the Markov state value;

[0059] The state transition frequency is counted based on the Markov state value, and the state transition frequency is standardized to obtain the state transition probability matrix;

[0060] According to the state transition probability matrix, a two-dimensional image mapping relationship is constructed, and the state transition probability is mapped to a gray value to obtain a Markov transition field feature image;

[0061] The Euclidean distance is calculated for the feature vector at each time point in the Markov transition field feature image to obtain a distance matrix;

[0062] Based on the distance matrix, the threshold condition is set, and the point pairs with a distance greater than the threshold are assigned a value of 1, and the point pairs with a distance less than the threshold are assigned a value of 0, and the recursive Figure 2value matrix, and according to the recursion Figure 2 The value matrix extracts the time series pattern and periodic features to obtain the time series spatial features;

[0063] The Markov transition field feature image and the temporal spatial feature are input into the low-rank representation network to obtain the fused feature, and the fused feature is dimensionally reconstructed and normalized to obtain the sample enhanced feature image.

[0064] Specifically, temperature signals are extracted from the source domain dataset and the target domain dataset respectively. , pressure signal and flow signal The time series data of these signals are continuously changing. To facilitate the subsequent state analysis based on the Markov chain, these signals are discretized according to the value range. Assume that the range of signal values ​​is , dividing the range evenly into intervals, each interval has a width of For each signal value , through the formula:

[0065] ;

[0066] Map it to a discrete Markov state value . With flow signal For example, assuming its value range is [0,100], it is divided into intervals, the width of each interval is 10. If the flow rate at a certain moment is 37, then its corresponding status value is After completing the state division, according to the definition of Markov chain, the transition frequency between states is counted. and , if the status values ​​are and , then the state is transferred The frequency of .matrix Elements Indicates from the state Transfer to state After the statistics are completed, Normalize to get the state transition probability matrix , which is defined as:

[0067] ;

[0068] in Indicates from the state The sum of all the transition times that have been set out. According to the state transition probability matrix Construct a two-dimensional image mapping relationship Mapped to grayscale values. Assuming the grayscale value range is [0,255], through the linear mapping formula:

[0069] ;

[0070] Convert the probability matrix into a Markov transition field feature image This image can intuitively reflect the transition strength between different states and is used for subsequent feature analysis. After generating the Markov transition field feature image, analyze the similarity between its time points. Calculate Euclidean distance and define distance matrix The elements of are:

[0071] ;

[0072] in and Respectively indicate time points and No. By comparing the distances between different time points, the temporal variation of the signal is measured. , set a threshold To generate recursion Figure 2 Value Matrix . Judge by conditions:

[0073] ;

[0074] Mark the point pairs whose distance is greater than the threshold as 1, and the point pairs whose distance is less than or equal to the threshold as 0. Recursion Figure 2 The value matrix reflects the temporal self-similarity and periodicity of the signal through its image pattern. After extracting the temporal pattern and periodic features of the recurrence graph, the Markov transition field feature image and timing characteristics Input the low-rank representation network to extract fusion features. The low-rank representation network compresses the input high-dimensional features into low-dimensional representations through matrix decomposition technology while retaining key information. Suppose the fusion feature matrix is , whose elements are obtained through the optimization problem:

[0075] ;

[0076] in is the input feature matrix, is the weight matrix, is the nuclear norm, which is used to constrain the low rank. Perform dimension reconstruction and normalization to obtain the sample enhanced feature image . Feature scaling and standardization are used to ensure that the input data has a stable distribution and adaptability.

[0077] In a specific embodiment, the process of executing step 103 may specifically include the following steps:

[0078] The sample enhanced feature image is input into the Siamese diagnostic network to obtain the sample Siamese feature map. The Siamese diagnostic network includes two weight-sharing Siamese branch networks. Each Siamese branch network contains 4 convolutional layers and 4 maximum pooling layers. The convolution kernel size of the first convolutional layer is 7×7, the step size is 2, and the number of output channels is 64. The convolution kernel size of the second to fourth convolutional layers is 3×3, the step size is 1, and the number of output channels is 128, 256, and 512 respectively.

[0079] The sample Siamese feature map is processed by the encoder to obtain the encoded features. The encoder contains 3 encoding blocks, each of which consists of two convolutional layers and one batch normalization layer. The convolution kernel size of the convolutional layer is 3×3, the step size is 2, and the activation function uses the ReLU function.

[0080] The decoder processes the encoded features to obtain sample decoding features. The decoder contains three decoding blocks. Each decoding block consists of two deconvolution layers and one batch normalization layer. The convolution kernel size of the deconvolution layer is 3×3, the step size is 2, and the activation function uses the ReLU function.

[0081] A channel attention module is constructed based on the sample decoding features, and the sample decoding features are averaged and max-pooled to obtain two sample feature descriptors. The two sample feature descriptors are processed separately using a shared multi-layer perceptron to obtain the sample channel attention weights.

[0082] Perform a dot product operation on the sample channel attention weight and the sample decoding feature to obtain the sample channel enhanced feature, and perform time series segmentation on the sample channel enhanced feature to obtain N sample time series segments;

[0083] Construct a temporal attention module, perform self-attention calculation on N sample time series segments, obtain sample temporal attention features, and fuse the sample temporal attention features with the sample channel enhancement features to obtain sample spatiotemporal enhancement features;

[0084] The sample spatiotemporal enhanced features are globally averaged pooled, and feature transformation is performed through two fully connected layers to obtain a feature extraction network.

[0085] Specifically, the preprocessed sample is enhanced with feature images Input into the Siamese diagnostic network. The Siamese network consists of two weight-sharing Siamese branch networks, the purpose of which is to extract deep features of the input image. Each Siamese branch network contains 4 convolutional layers and 4 maximum pooling layers to extract features layer by layer. Assume that the size of the input feature image is (height, width and number of channels), in the first convolutional layer, the convolution kernel size is , the step size is 2, and the number of output channels is set to 64. The convolution operation is performed by the formula:

[0086] ;

[0087] in is the output feature map Position in The value of the channel, is the weight matrix of the convolution kernel, is the bias term, is the pixel value corresponding to the input image. The convolution layer outputs a The feature map of . Through three convolution kernels with size , the convolutional layers with a step size of 1 gradually extract deeper features. The number of output channels of the second convolutional layer is 128, the third layer is 256, and the fourth layer is 512. Each convolution is followed by a maximum pooling operation to halve the size of the feature map. Assume that the feature map output by the fourth convolutional layer is , whose size is This feature map is a sample Siamese feature map that records the deep features of the input image. The input is sent to the encoder for encoding. The encoder consists of three encoding blocks, each of which contains two convolutional layers and one batch normalization layer. The convolution kernel size of each convolution layer is , the step size is 2, and the activation function uses the ReLU function. Each convolution operation of the encoder is performed by the formula:

[0088] ;

[0089] Calculate the output feature map ,in is the convolution kernel weight, is the bias, ReLU is the activation function. After 3 encoding blocks, the encoding features output by the encoder are The size of the encoding feature is further reduced. Input to the decoder for feature decoding. The decoder contains 3 decoding blocks, each of which consists of two deconvolution layers and one batch normalization layer. The deconvolution operation restores the spatial dimension of the feature map through inverse convolution. The calculation formula is:

[0090] ;

[0091] in is the deconvolution kernel weight. The decoder gradually restores the feature map to the sample decoding features that match the input size. Based on the sample decoding features, a channel attention module is constructed to enhance the feature expression capability. The decoded features are subjected to global average pooling and maximum pooling operations to obtain two feature descriptors respectively. and , which is defined as:

[0092] ;

[0093] in Represents the channel index. These two feature descriptors are input into a shared multi-layer perceptron to calculate the attention weight of each channel :

[0094] ;

[0095] in is the Sigmoid activation function, is the weight matrix of the multi-layer perceptron. Through the channel-by-channel dot product operation, the attention weight is multiplied by the decoded feature to obtain the channel enhancement feature. In order to capture the temporal dynamic characteristics, the channel enhancement feature is Equal time division, generation The global temporal correlation of these fragments is calculated through the self-attention mechanism. The self-attention calculation formula is:

[0096] ;

[0097] in are query, key, and value matrices respectively, is the dimension of the key. Through the self-attention mechanism, the temporal attention feature is obtained and fused with the channel enhancement feature to generate the sample spatiotemporal enhancement feature. The spatiotemporal enhancement feature is globally averaged pooled and compressed into a fixed-length vector. Feature mapping is performed through two fully connected layers to generate the final feature extraction network output .

[0098] In a specific embodiment, the process of executing step 104 may specifically include the following steps:

[0099] Input the source domain data set into the feature extraction network for feature extraction and feature vector normalization to obtain a normalized feature vector;

[0100] Calculate the cosine similarity matrix based on the normalized eigenvectors, and multiply the cosine similarity matrix by the scaling factor to obtain a scaled similarity matrix;

[0101] Calculate the angular margin loss of intra-class sample pairs and the angular margin loss of inter-class sample pairs based on the scaled similarity matrix to obtain a first angular margin loss;

[0102] Calculate the category weight coefficient according to the distribution of samples in each category, and multiply the category weight coefficient by the first angle margin loss to obtain the second angle margin loss;

[0103] The mean square error is calculated for the reconstructed features output by the encoder and decoder in the feature extraction network and the input features to obtain the reconstruction loss, and the second angle margin loss and the reconstruction loss are weighted summed to obtain the total network loss function;

[0104] Based on the total network loss function, the feature extraction network is back-propagated and parameter optimized to obtain updated network parameters, and the updated network parameters are iteratively trained. When the loss function converges, a pre-trained diagnostic model is obtained.

[0105] Specifically, the source domain dataset Input feature extraction network, where Indicates samples, represents its category label, is the number of categories, is the number of source domain data samples. The feature extraction network consists of an encoder and a decoder. The encoder converts the input data into Convert to high-dimensional feature vector In order to ensure the numerical stability of the features in the similarity calculation, Perform normalization. The normalized eigenvector is defined as:

[0106] ;

[0107] in Represents the bi-norm of the vector, normalized , so that all eigenvectors are mapped to the unit sphere. According to the normalized eigenvector Calculate the cosine similarity matrix , whose elements Defined as:

[0108] ;

[0109] in and They are vectors and No. Quantity, Represents the feature vector and To enhance the sensitivity to feature distribution, Multiply by the scaling factor Get the scaled similarity matrix:

[0110] ;

[0111] The scaling factor is usually set to a large positive value (such as 30) to amplify the impact of similarity on the loss function. In order to optimize the network's ability to discriminate different categories, based on the scaled similarity matrix Calculate the angular margin loss of intra-class sample pairs and the angular margin loss of inter-class sample pairs. Define intra-class sample pairs as sample pairs of the same category, and inter-class sample pairs as sample pairs of different categories. Intra-class angular margin loss Defined as:

[0112] ;

[0113] in is the set of all sample pairs within a class, is the threshold of the intra-class angle. Inter-class angle margin loss Defined as:

[0114] ;

[0115] in is the set of all inter-class sample pairs, Is the minimum threshold of the angle between classes. Combining the two parts of loss, we get the first angle margin loss:

[0116] ;

[0117] Considering the uneven distribution of samples of different categories, the category weight coefficient is introduced To adjust the sensitivity of the loss to different categories. The category weight coefficient is calculated as:

[0118] ;

[0119] in It is The number of class samples. The final weighted angle margin loss It is expressed as:

[0120] ;

[0121] in Indicates At the same time, in order to enhance the network's ability to reconstruct the input data, the mean square error (MSE) between the output features of the encoder and decoder and the input features is calculated as the reconstruction loss :

[0122] ;

[0123] in is the decoder's reconstructed input sample . Combining the above loss functions, the total loss function of the network is defined as:

[0124] ;

[0125] in and is the loss weight, which controls the relative importance of angle loss and reconstruction loss. Based on the total loss function, the gradient is calculated by the back propagation algorithm to optimize the network parameters. Assume that the network weight is , the optimization goal is:

[0126] ;

[0127] The optimization process uses the gradient descent method, and the parameter update formula is:

[0128] ;

[0129] in is the learning rate, is the gradient of the total loss function with respect to the parameters. Through multiple rounds of iterative training, the loss function gradually converges and the network parameters tend to be stable. Taking the flow signal of the heat exchange module as an example, the source domain data contains the flow time series under normal operation and blockage faults. After the model is trained, it can accurately extract the features of different fault types and maintain robustness in the target domain migration, thus obtaining a pre-trained high-performance diagnostic model.

[0130] In a specific embodiment, the process of executing step 105 may specifically include the following steps:

[0131] The network parameters of the four convolutional layers and four maximum pooling layers of the Siamese branch network in the pre-trained diagnostic model are locked to obtain the Siamese basic network with locked parameters, and the target domain data set is input into the Siamese basic network with locked parameters for feature extraction to obtain the basic features of the target domain;

[0132] The target domain basic features are input into the encoder and decoder for feature reconstruction to obtain the target domain reconstructed features, and a channel attention module is constructed based on the target domain reconstructed features. Average pooling and maximum pooling operations are performed on the target domain reconstructed features, and the channel weights are calculated through a multi-layer perceptron to obtain the target domain channel enhanced features.

[0133] Perform temporal segmentation and self-attention calculation on the target domain channel enhancement features to obtain the target domain spatiotemporal enhancement features, and input the target domain spatiotemporal enhancement features into the fault classifier to calculate the fault type probability distribution and evaluate the fault degree to obtain the target domain diagnosis results;

[0134] The classification loss and regression loss are calculated according to the target domain diagnosis results to obtain the target domain optimization objective function. The parameters of the fault classifier are updated based on the target domain optimization objective function until the loss value converges to obtain a trained fault diagnosis model.

[0135] Specifically, a Siamese branch network is extracted from the pre-trained diagnostic model. The four convolutional layers and four maximum pooling layers of the network have been trained on the source domain dataset and have learned stable feature extraction capabilities. The network parameters of these layers are locked, and their weights and biases are fixed to prevent them from being updated during the target domain training process, forming a Siamese basic network with locked parameters. In this way, the Siamese basic network can maintain the generalization ability of source domain feature extraction while providing reliable initial conditions for target domain feature extraction. Assume that the target domain dataset is ,in Indicates target domain samples, is the sample size, Input the Siamese basic network with locked parameters and extract deep features layer by layer. The samples passed The feature representation after the convolutional layer is , the convolution operation is expressed as:

[0136] ;

[0137] in and They are The weight and bias of the layer convolution kernel, * indicates the convolution operation, Is the activation function. Through multiple layers of convolution and pooling operations, the basic features of the target domain are generated. ,in It is The basic features of the target domain are represented by The input is sent to the encoder and decoder for feature reconstruction. The encoder consists of multiple convolutional blocks, which compress the spatial dimensions of the features and extract high-level semantic information. Assume that the output of the encoder is , the decoder uses the inverse convolution operation to convert Reconstructed into the reconstructed features in the original feature space The reconstruction process is expressed as:

[0138] ;

[0139] Where Encoder (·) and Decoder (·) represent the operations of the encoder and decoder respectively. In order to optimize the target domain feature expression, the reconstructed feature The channel attention module is constructed based on . Global average pooling and maximum pooling operations are performed on each reconstructed feature to generate a global descriptor AvgPool and MaxPool , and its calculation formula is:

[0140] ;

[0141] ;

[0142] in and are the height and width of the feature map, respectively. Represents the channel index. These two descriptors are input into a shared multi-layer perceptron (MLP) to calculate the channel attention weight :

[0143] ;

[0144] in and is the weight matrix of the MLP, is the Sigmoid activation function. The channel enhancement feature is obtained by calculating the channel-by-channel dot product with the reconstructed feature:

[0145] ;

[0146] In order to capture the spatiotemporal characteristics of the target domain, the channel enhancement features are The time series is divided into equal parts to generate time series segments. Each time series segment is processed by the self-attention module to capture the global time dependency. The self-attention calculation formula is:

[0147] ;

[0148] in are query, key, and value matrices respectively, is the dimension of the key, and softmax is used to normalize the weights. By fusing the temporal attention features and channel enhancement features, the spatiotemporal enhancement features of the target domain are generated. The spatiotemporal enhancement features are input into the fault classifier to calculate the probability distribution of the fault type and evaluate the fault degree of the target domain data. The output of the fault classifier is the fault probability distribution of each sample and fault severity rating According to the classification label of the target domain and the true fault level, the classification loss is defined and regression loss :

[0149] ;

[0150] ;

[0151] in It is The true category labels of samples, is the true score of the fault degree. The overall objective function is:

[0152] ;

[0153] in and is the weight parameter. The parameters of the fault classifier are optimized by back propagation until the loss function Convergence completes the training of the target domain fault diagnosis model.

[0154] In a specific embodiment, the process of executing step 106 may specifically include the following steps:

[0155] Acquire the real-time temperature signal, real-time pressure signal and real-time flow signal of the heat exchange module to obtain real-time operation data, and segment and pre-process the real-time operation data to obtain standardized real-time data;

[0156] Perform Markov state value division and state transition probability calculation on the standardized real-time data to obtain a real-time enhanced feature image;

[0157] The real-time enhanced feature image is input into the Siamese diagnosis network of the trained fault diagnosis model for feature extraction to obtain a real-time Siamese feature map, and the real-time Siamese feature map is encoded sequentially through three encoding blocks in the encoder to obtain a real-time decoding feature;

[0158] Perform average pooling and maximum pooling operations on the real-time decoding features to obtain two real-time feature descriptors, and input them into a shared multi-layer perceptron to calculate the real-time channel attention weights. Perform a dot product operation on the real-time channel attention weights and the real-time decoding features to obtain real-time channel enhancement features, and perform N equal time series segmentation on the real-time channel enhancement features to obtain N real-time time series segments;

[0159] Perform self-attention calculation on N real-time time series segments to obtain real-time time attention features, fuse the real-time time attention features with the real-time channel enhancement features, and obtain the target feature vector through global average pooling and two fully connected layers;

[0160] The target feature vector is input into the fault classifier to classify the fault type and evaluate the fault severity, and the fault type probability distribution and fault severity score are obtained;

[0161] The fault level is divided according to the preset fault level threshold to obtain the fault level, and a fault diagnosis report is generated based on the fault type probability distribution, fault degree score and fault level.

[0162] Specifically, through the sensor network arranged on the heat exchange module, key operating parameters, including temperature signals, are collected in real time. , pressure signal and flow signal These signals are stored in time series to form a real-time running data matrix ,in Indicates the time point. Indicates the number of sensors (i.e. correspond , and ). Due to the noise and scale inconsistency of real-time data, the data is segmented and preprocessed. The signal sequence is segmented, assuming that each segment contains time points, the segmented data is ,in , . Perform denoising on each segment of data (e.g. using a sliding average filter MA ) and normalization (normalized to the interval [0,1]):

[0163] ;

[0164] in Indicates In the data segment time point, The standardized value of the signal. The standardized real-time data is divided into Markov state values ​​and the state transition probability is calculated. According to the value range of the signal, each signal is discretized into a finite Markov state value. For example, the signal is divided into state intervals, and the width of each state interval is For each value of the standardized data , through the formula:

[0165] ;

[0166] Get the discrete Markov state value . Count the state transition frequencies between adjacent time points and generate a state transition frequency matrix , whose elements are defined as:

[0167] ;

[0168] in The frequency matrix is ​​converted into a state transition probability matrix by normalization:

[0169] ;

[0170] And map the probability matrix into a grayscale image to generate a real-time enhanced feature image . Enhance the feature image in real time Input the trained fault diagnosis model into the Siamese diagnosis network, extract features layer by layer through its 4 convolutional layers and 4 maximum pooling layers, and obtain the real-time Siamese feature map .Will The three encoding blocks in the input encoder are compressed to obtain real-time decoding features. The encoding process is expressed as:

[0171] ;

[0172] Each convolution operation of the encoder uses a convolution kernel size of , with a step size of 2, and with the ReLU activation function and batch normalization layer. Perform average pooling and maximum pooling operations to obtain two feature descriptors respectively:

[0173] ;

[0174] ;

[0175] in are the height, width and channel index of the feature map respectively. Input the descriptor into the shared multi-layer perceptron to calculate the channel attention weight:

[0176] ;

[0177] in and is the weight matrix of the multilayer perceptron, is the Sigmoid function. By combining the attention weights with the decoded features through channel-by-channel dot products, channel-enhanced features are generated. Equal time division, generation The real-time time series segments are captured through self-attention calculations. The temporal attention features are combined with the channel enhancement features to generate the target feature vector through global average pooling and two fully connected layers. .Will Input the fault classifier and calculate the probability distribution of fault types and fault severity rating According to the preset fault level threshold, the fault degree score is graded to obtain the fault level . Combined , and , generate a fault diagnosis report containing fault type, probability distribution, severity and level.

[0178] The above describes the fault diagnosis method of the heat exchange module in the embodiment of the present invention. The following describes the fault diagnosis device of the heat exchange module in the embodiment of the present invention. Figure 2 , an embodiment of a fault diagnosis device for a heat exchange module in an embodiment of the present invention comprises:

[0179] The acquisition module 201 is used to construct a finite element model of the heat exchange module and acquire temperature signals, pressure signals and flow signals, and divide them into a source domain data set and a target domain data set;

[0180] The transformation module 202 is used to perform Markov transition field transformation and temporal spatial feature enhancement processing on the source domain data set and the target domain data set to obtain a sample enhanced feature image;

[0181] A construction module 203 is used to construct a Siamese diagnosis network with a spatiotemporal attention mechanism according to the sample enhanced feature image, and embed an encoder-decoder structure and a channel attention module in the Siamese diagnosis network to obtain a feature extraction network;

[0182] The error reconstruction module 204 is used to input the source domain data set into the feature extraction network for training and error reconstruction to obtain a pre-trained diagnostic model;

[0183] Enhanced training module 205, used to fix the shallow feature extraction parameters of the pre-trained diagnosis model, input the target domain data set into the pre-trained diagnosis model, perform enhanced training on the fault features, and obtain a trained fault diagnosis model;

[0184] The generation module 206 is used to input the real-time operation data of the heat exchange module into the trained fault diagnosis model to calculate the probability distribution of fault types and evaluate the fault degree, and generate a fault diagnosis report.

[0185] Through the collaborative cooperation of the above components, by constructing a feature enhancement method that combines Markov transition field and recursive graph, the temporal dynamic features in multi-source signals such as temperature, pressure, and flow are effectively extracted, and the expressive ability of fault features is enhanced; the Siamese diagnostic network structure and spatiotemporal attention mechanism are adopted to realize the effective migration of fault features under different working conditions and solve the problem of data imbalance; the encoder-decoder structure and channel attention module are introduced to improve the feature fusion effect between multi-source signals and enhance the model's ability to recognize different types of faults; a dynamic softmax loss function with angular margin penalty is designed to optimize the classification boundary of fault features and improve the generalization performance of the model; through the method of shallow feature fixation and deep feature migration, the knowledge transfer from the source domain to the target domain is realized, and the diagnostic effect of the model under conditions of insufficient data is enhanced.

[0186] above Figure 2 The fault diagnosis device of the heat exchange module in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The fault diagnosis device of the heat exchange module in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0187] Figure 3: is a structural schematic diagram of a heat exchange module fault diagnosis device provided by an embodiment of the present invention. The heat exchange module fault diagnosis device 300 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage device terminals) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be short-term storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the heat exchange module fault diagnosis device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the heat exchange module fault diagnosis device 300 to implement the steps of the above-mentioned heat exchange module fault diagnosis method.

[0188] The heat exchange module fault diagnosis device 300 may also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will appreciate that Figure 3 The structure of the fault diagnosis device for the heat exchange module shown does not constitute a limitation on the fault diagnosis device for the heat exchange module provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0189] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the fault diagnosis method for the heat exchange module.

[0190] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0191] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0192] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing a fault of a heat exchange module, characterized in that: The method comprises: Construct a finite element model of the heat exchange module and collect temperature signals, pressure signals and flow signals, and divide them into a source domain data set and a target domain data set; specifically, the method includes: establishing a three-dimensional grid model based on the structural parameters and working parameters of the heat exchange module, and gridding the three-dimensional grid model to obtain a finite element model of the heat exchange module; performing stress analysis and thermal flow field analysis on the finite element model to obtain a target monitoring point coordinate set, and installing temperature sensors, pressure sensors and flow sensors at physical positions corresponding to the target monitoring point coordinate set, and setting sensor acquisition parameters; performing multi-condition data acquisition on the heat exchange module according to the sensor acquisition parameters, recording temperature signals, pressure signals and flow signals under normal conditions and fault conditions, and obtaining an original data set; segmenting the original data set according to the time window to obtain a segmented data matrix, and performing data preprocessing on the segmented data matrix to obtain a preprocessed data matrix; dividing the preprocessed data matrix into a source domain data set and a target domain data set according to the data volume of the working condition type; The source domain data set and the target domain data set are subjected to Markov transition field transformation and time-series spatial feature enhancement processing to obtain a sample enhanced feature image; specifically, the method comprises: dividing the temperature signal, pressure signal and flow signal in the source domain data set and the target domain data set into states according to the numerical range to obtain a Markov state value; counting the state transition frequency based on the Markov state value, and normalizing the state transition frequency to obtain a state transition probability matrix; constructing a two-dimensional image mapping relationship according to the state transition probability matrix, mapping the state transition probability into a grayscale value, and obtaining a Markov state value. A Markov transition field feature image; calculating the Euclidean distance of the feature vector of each time point in the Markov transition field feature image to obtain a distance matrix; setting a threshold condition based on the distance matrix, assigning a value of 1 to a point pair whose distance is greater than the threshold and a value of 0 to a point pair whose distance is less than the threshold, to obtain a recursive graph binary matrix, and extracting time series patterns and periodic features according to the recursive graph binary matrix to obtain time series space features; inputting the Markov transition field feature image and the time series space features into a low-rank representation network to obtain fused features, and performing dimensional reconstruction and normalization processing on the fused features to obtain a sample enhanced feature image; Constructing a Siamese diagnostic network with a spatiotemporal attention mechanism according to the sample enhanced feature image, embedding an encoder-decoder structure and a channel attention module in the Siamese diagnostic network to obtain a feature extraction network; Inputting the source domain data set into the feature extraction network for training and error reconstruction to obtain a pre-trained diagnostic model; Fixing the shallow feature extraction parameters of the pre-trained diagnostic model, inputting the target domain data set into the pre-trained diagnostic model, performing enhanced training on the fault features, and obtaining a trained fault diagnosis model; The real-time operation data of the heat exchange module is input into the trained fault diagnosis model to calculate the probability distribution of fault types and evaluate the fault degree, and generate a fault diagnosis report.

2. The fault diagnosis method of the heat exchange module according to claim 1, characterized in that: The Siamese diagnostic network with a spatiotemporal attention mechanism is constructed according to the sample enhanced feature image, an encoder-decoder structure and a channel attention module are embedded in the Siamese diagnostic network to obtain a feature extraction network, including: Inputting the sample enhanced feature image into a Siamese diagnostic network to obtain a sample Siamese feature map, wherein the Siamese diagnostic network includes two weight-shared Siamese branch networks, each Siamese branch network includes 4 convolutional layers and 4 maximum pooling layers, wherein the convolution kernel size of the first convolutional layer is 7×7, the step size is 2, and the number of output channels is 64, and the convolution kernel size of the second to fourth convolutional layers is 3×3, the step size is 1, and the number of output channels is 128, 256, and 512, respectively; Performing encoder processing on the sample Siamese feature map to obtain encoded features, the encoder includes 3 encoding blocks, each encoding block consists of two convolutional layers and one batch normalization layer, wherein the convolution kernel size of the convolutional layer is 3×3, the step size is 2, and the activation function adopts the ReLU function; Performing decoder processing on the encoded features to obtain sample decoding features, the decoder includes 3 decoding blocks, each decoding block consists of two deconvolution layers and one batch normalization layer, wherein the convolution kernel size of the deconvolution layer is 3×3, the step size is 2, and the activation function adopts the ReLU function; Constructing a channel attention module according to the sample decoding features, performing average pooling and maximum pooling operations on the sample decoding features to obtain two sample feature descriptors, and using a shared multi-layer perceptron to process the two sample feature descriptors respectively to obtain sample channel attention weights; Performing a dot product operation on the sample channel attention weight and the sample decoding feature to obtain a sample channel enhancement feature, and performing time series segmentation on the sample channel enhancement feature to obtain N sample time series segments; Constructing a time attention module, performing self-attention calculation on the N sample time series segments to obtain sample time attention features, and performing feature fusion of the sample time attention features with the sample channel enhancement features to obtain sample spatiotemporal enhancement features; A global average pooling operation is performed on the spatiotemporal enhancement features of the sample, and feature transformation is performed through two fully connected layers to obtain a feature extraction network.

3. The fault diagnosis method of the heat exchange module according to claim 2, characterized in that: The step of inputting the source domain data set into the feature extraction network for training and error reconstruction to obtain a pre-trained diagnostic model includes: Inputting the source domain data set into the feature extraction network to perform feature extraction and feature vector normalization processing to obtain a normalized feature vector; Calculating a cosine similarity matrix based on the normalized feature vector, and multiplying the cosine similarity matrix by a scaling factor to obtain a scaled similarity matrix; Calculating the angular margin loss of intra-class sample pairs and the angular margin loss of inter-class sample pairs based on the scaled similarity matrix to obtain a first angular margin loss; Calculating a category weight coefficient according to the distribution of samples of each category, and multiplying the category weight coefficient by the first angle margin loss to obtain a second angle margin loss; Calculating a mean square error between the reconstructed features output by the encoder and the decoder in the feature extraction network and the input features to obtain a reconstruction loss, and performing a weighted summation of the second angle margin loss and the reconstruction loss to obtain a total network loss function; The feature extraction network is back-propagated and parameter optimized based on the total network loss function to obtain updated network parameters, and the updated network parameters are iteratively trained. When the loss function converges, a pre-trained diagnostic model is obtained.

4. The fault diagnosis method of the heat exchange module according to claim 3, characterized in that: The shallow feature extraction parameters of the pre-trained diagnostic model are fixed, and the target domain data set is input into the pre-trained diagnostic model to perform enhanced training on the fault features to obtain a trained fault diagnosis model, including: Locking the network parameters of the four convolutional layers and the four maximum pooling layers of the Siamese branch network in the pre-trained diagnostic model to obtain a Siamese basic network with locked parameters, and inputting the target domain data set into the Siamese basic network with locked parameters to extract features to obtain target domain basic features; The target domain basic features are input into the encoder and decoder for feature reconstruction to obtain target domain reconstructed features, and a channel attention module is constructed based on the target domain reconstructed features, average pooling and maximum pooling operations are performed on the target domain reconstructed features, and channel weights are calculated through a multi-layer perceptron to obtain target domain channel enhanced features; Performing time series segmentation and self-attention calculation on the target domain channel enhancement features to obtain target domain spatiotemporal enhancement features, and inputting the target domain spatiotemporal enhancement features into a fault classifier to calculate the probability distribution of fault types and evaluate the degree of faults to obtain target domain diagnosis results; The classification loss and regression loss are calculated according to the target domain diagnosis result to obtain the target domain optimization objective function, and the parameters of the fault classifier are updated based on the target domain optimization objective function until the loss value converges to obtain a trained fault diagnosis model.

5. The fault diagnosis method of the heat exchange module according to claim 4, characterized in that: The step of inputting the real-time operation data of the heat exchange module into the trained fault diagnosis model to calculate the probability distribution of fault types and evaluate the degree of faults, and generating a fault diagnosis report includes: Acquire the real-time temperature signal, real-time pressure signal and real-time flow signal of the heat exchange module to obtain real-time operation data, and segment and pre-process the real-time operation data to obtain standardized real-time data; Performing Markov state value division and state transition probability calculation on the standardized real-time data to obtain a real-time enhanced feature image; Inputting the real-time enhanced feature image into the Siamese diagnosis network of the trained fault diagnosis model to extract features, obtaining a real-time Siamese feature map, and encoding the real-time Siamese feature map in sequence through three encoding blocks in the encoder to obtain real-time decoding features; Performing average pooling and maximum pooling operations on the real-time decoding features to obtain two real-time feature descriptors, and inputting them into a shared multi-layer perceptron to calculate real-time channel attention weights, performing a dot product operation on the real-time channel attention weights and the real-time decoding features to obtain real-time channel enhancement features, and performing N equal time series segmentation on the real-time channel enhancement features to obtain N real-time time series segments; Performing self-attention calculation on the N real-time time series segments to obtain real-time time attention features, performing feature fusion on the real-time time attention features and the real-time channel enhancement features, and obtaining a target feature vector through global average pooling and two fully connected layers; Inputting the target feature vector into a fault classifier to classify the fault type and evaluate the fault severity, thereby obtaining a fault type probability distribution and a fault severity score; Fault levels are divided according to preset fault level thresholds to obtain fault levels, and a fault diagnosis report is generated based on the fault type probability distribution, the fault degree score and the fault level.

6. A heat exchange module fault diagnosis device, characterized in that: The device is used to perform a fault diagnosis method for a heat exchange module according to any one of claims 1 to 5, the device comprising: An acquisition module is used to construct a finite element model of the heat exchange module and acquire temperature signals, pressure signals, and flow signals, and divide them into a source domain data set and a target domain data set; A transformation module, used for performing Markov transition field transformation and temporal spatial feature enhancement processing on the source domain data set and the target domain data set to obtain a sample enhanced feature image; A construction module is used to construct a Siamese diagnostic network with a spatiotemporal attention mechanism according to the sample enhanced feature image, and embed an encoder-decoder structure and a channel attention module in the Siamese diagnostic network to obtain a feature extraction network; An error reconstruction module, used for inputting the source domain data set into the feature extraction network for training and error reconstruction to obtain a pre-trained diagnostic model; An enhanced training module, used to fix the shallow feature extraction parameters of the pre-trained diagnostic model, input the target domain data set into the pre-trained diagnostic model, perform enhanced training on the fault features, and obtain a trained fault diagnosis model; A generation module is used to input the real-time operation data of the heat exchange module into the trained fault diagnosis model to calculate the probability distribution of fault types and evaluate the fault degree, and generate a fault diagnosis report.

7. A heat exchange module fault diagnosis device, characterized in that: The heat exchange module fault diagnosis device comprises: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instruction in the memory so that the fault diagnosis device of the heat exchange module executes the fault diagnosis method of the heat exchange module according to any one of claims 1-5.

8. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, a fault diagnosis method for a heat exchange module according to any one of claims 1 to 5 is implemented.

Images