A chiller fault diagnosis method and system for unbalanced data

Abstract

The present invention discloses a chiller fault diagnosis method and system for unbalanced data, which belongs to the field of chiller fault diagnosis, including: based on a disturbance reconstruction mechanism, using the original data to restore balanced fault data with the characteristics of the original data; using a spatiotemporal autoencoder to extract and model the temporal characteristics and spatial structure of the balanced fault data, and reconstruct the original data; using a dynamic topological temporal feature integration network to extract temporal features and topological features respectively and adopt a dynamic gated fusion mechanism to perform adaptive fusion; classifying the fused features to obtain fault diagnosis results. The present invention realizes the parallel generation of multiple types of fault samples, ensuring that each type of generated fault data has similar temporal and spatial characteristics to the original fault data, can adapt to the real-time changing working conditions of the chiller, realizes the collaborative modeling of temporal features and dynamic topological features in the complex multi-sensor system of the chiller, and can also achieve high-precision real-time diagnosis on edge devices.

Description

Technical Field

[0001] The present invention belongs to the technical field of chiller fault diagnosis, and in particular relates to a chiller fault diagnosis method and system oriented to imbalance data. Background Art

[0002] The statements herein merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Chillers precisely meet a building's cooling needs by efficiently coordinating the refrigeration cycle. Because chillers typically consist of multiple key components, including compressors, condensers, evaporators, valves, and sensors, and operate under multivariable coupling and time-varying loads for extended periods, abnormal operating conditions can arise from equipment aging, operating parameter mismatches, or sudden environmental changes. Continued operation under fault conditions can not only lead to decreased cooling efficiency and increased energy consumption, but can also result in insufficient cooling due to system control failures, compromising indoor temperature control and equipment safety. Therefore, it is necessary to develop an efficient and accurate method for diagnosing chiller faults.

[0004] Chillers have a variety of failure modes, including compressor failure, valve leakage, and heat exchanger efficiency degradation. This leads to significant differences in the amount of fault data for different fault types. Therefore, in chiller fault diagnosis, the unequal number of fault types leads to data imbalance, resulting in poor recognition accuracy of the trained model for small sample fault types. More specifically, data imbalance causes the model to learn prior information about the proportion of samples in the training set, resulting in a bias towards the majority class in actual predictions. Therefore, chiller fault diagnosis methods and systems need to consider data imbalance and adopt appropriate data generation methods to address this problem.

[0005] The inventors found that the data generated by the data generation method used in the existing chiller fault diagnosis method is not ideal in terms of data stability. The main problems are as follows: (1) Due to the large difference in the number of samples of different fault types, the data distribution is unbalanced, which may lead to deviations in the fault diagnosis results; and the traditional data generation method usually adopts a serial generation method, that is, generating data for a single fault type one by one, which has the problem of low generation efficiency. (2) In the fault data generation stage, because the traditional method fails to fully capture and integrate the temporal dynamics in the original fault data and the spatial topological characteristics between sensors, the generated fault data of a certain type has a large deviation from the original fault data in terms of temporal relationship and spatial structure. (3) The traditional fault diagnosis method only focuses on the temporal or static structure of the chiller operation data, and cannot collaboratively model the temporal characteristics and dynamic topological characteristics, thus lacking the ability to adapt to real-time changing working conditions, which often leads to poor performance of the fault diagnosis model in practical applications. Summary of the Invention

[0006] The purpose of the present invention is to overcome the deficiencies in the above-mentioned prior art and to provide a chiller fault diagnosis method and system for unbalanced data, which realizes the parallel generation of multiple types of fault samples, ensures that each type of generated fault data has similar spatiotemporal characteristics to the original fault data, can adapt to the real-time changing working conditions of the chiller, and realize the collaborative modeling of timing characteristics and dynamic topological characteristics in the complex multi-sensor system of the chiller, so as to achieve high-precision real-time diagnosis even on edge devices.

[0007] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0008] On the one hand, the technical solution of the present invention provides a chiller fault diagnosis method for imbalanced data, comprising:

[0009] Based on the disturbance reconstruction mechanism, the original data is used to restore the balanced fault data with the characteristics of the original data;

[0010] The spatiotemporal autoencoder is used to extract and model the temporal features and spatial structure of the balanced fault data, reconstruct the original data, and obtain low-dimensional features that retain the temporal dynamic features and spatial topological structure of the original data;

[0011] The dynamic topological temporal feature integration network is used to extract temporal features and topological features respectively, and a dynamic gated fusion mechanism is used for adaptive fusion.

[0012] The fused features are classified to obtain the fault diagnosis results.

[0013] In at least one embodiment, the disturbance reconstruction mechanism is specifically:

[0014] Gradually add noise to the original data to make it conform to the Gaussian distribution;

[0015] Through the denoising function, the noise is gradually removed and the balanced fault data with the characteristics of the original data is reconstructed.

[0016] In at least one embodiment, a label guidance layer is incorporated into the denoising process to generate data corresponding to specific labels, thereby enabling parallel generation of multiple types of data.

[0017] In at least one embodiment, the spatiotemporal autoencoder consists of a first convolutional layer, a recurrent neural network layer, a graph attention network layer, and a second convolutional layer;

[0018] The first convolutional layer is used to preprocess the original input data and extract local features; the recurrent neural network layer is used to capture the temporal dependencies in the time series and extract dynamically changing information; the graph attention network layer is used to use the self-attention mechanism to model the complex topological relationships between sensor nodes and effectively aggregate multi-hop information; the second convolutional layer is used to integrate and compress multi-scale features to generate high-quality hidden states.

[0019] In at least one embodiment, the dynamic topology timing feature integration network includes a dynamic topology extraction branch and a timing feature extraction branch; the dynamic topology extraction branch is used to capture the correlation between different sensors in the chiller that change with operating conditions; the timing feature extraction branch is used to model the long-term time dependence of sensor signals and predict fault evolution trends.

[0020] In at least one embodiment, the training process of the temporal feature extraction branch is:

[0021] Input the time series fault data enhanced by the diffusion model; then use 1D convolution with a stride of 2 to reduce the sequence length and extract local features; then input the local features extracted by convolution into the dynamic sparse attention block, use sparse attention and parameter sharing to model global time series dependencies, and output time series features.

[0022] In at least one embodiment, the training process of the dynamic topology extraction branch is:

[0023] The input is a multidimensional signal collected by the sensor at a fixed frequency and split with a sliding window. Each sensor channel is independently normalized to construct a dynamic graph, which is then input into a Chebyshev graph convolution with m layers for convolution processing. A temporal convolution layer is used to capture the multi-scale temporal dependencies of the sensor signal. Finally, the importance weight of each sensor node is calculated through a node attention pooling layer, and the weighted aggregation is aggregated into a global feature to output a topological feature.

[0024] In at least one embodiment, the adaptive fusion process is:

[0025] The topological features are copied in time steps to align with the timing features. , then generate dynamic gating weights and output gated fusion features.

[0026] In at least one embodiment, the fused features are input into a hierarchical classifier for classification to obtain a final fault diagnosis result.

[0027] On the other hand, the technical solution of the present invention further provides a chiller fault diagnosis system for imbalance data, comprising:

[0028] The balanced fault data parallel generation module is configured to: based on a disturbance reconstruction mechanism, use the original data to restore the balanced fault data with the characteristics of the original data;

[0029] The data enhancement module is configured to: extract and model the temporal features and spatial structure of the balanced fault data using a spatiotemporal autoencoder, reconstruct the original data, and obtain low-dimensional features that preserve the temporal dynamic features and spatial topological structure of the original data;

[0030] The feature extraction and fusion module is configured to: use a dynamic topological temporal feature integration network to extract temporal features and topological features respectively and use a dynamic gated fusion mechanism to perform adaptive fusion;

[0031] The fault diagnosis module is configured to classify the fused features and obtain a fault diagnosis result.

[0032] The beneficial effects of the technical solution of the present invention are as follows:

[0033] 1) This invention realizes the parallel generation of multiple types of fault samples, significantly improving the efficiency of data generation and shortening the generation time. At the same time, by balancing the data distribution, the quality of the generated data is improved, effectively alleviating the problem of imbalanced chiller fault data, and providing a more accurate and comprehensive training data foundation for subsequent fault diagnosis.

[0034] 2) The fault data generated by the present invention is highly consistent with the original data in terms of temporal relationship and spatial structure, which significantly reduces the generation deviation, improves the authenticity and reliability of the data, provides higher-quality data support for the fault diagnosis system, and further improves the accuracy of fault detection and diagnosis.

[0035] 3) The present invention realizes efficient analysis and accurate diagnosis of multi-source heterogeneous sensor data through the collaboration of dynamic topology extraction branches and timing feature extraction branches. It has high precision, low latency and strong generalization ability, can adapt to the real-time changing working conditions of the chiller, and realize the collaborative modeling of timing features and dynamic topology features in the complex multi-sensor system of the chiller, and can also achieve high-precision real-time diagnosis on edge devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1 This is a schematic diagram of a chiller fault diagnosis method and system for imbalanced data according to the present invention;

[0038] Figure 2This is a schematic diagram of a module for parallel generation of balanced fault data according to the present invention;

[0039] Figure 3 Schematic diagram of the data enhancement module of the present invention;

[0040] Figure 4 It is a flow chart of the feature extraction and fusion module and the fault diagnosis module of the present invention. DETAILED DESCRIPTION

[0041] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0042] As introduced in the background technology, the purpose of the present invention is to overcome the shortcomings of the above-mentioned existing technologies and provide a chiller fault diagnosis method and system for unbalanced data, which realizes the parallel generation of multiple types of fault samples, ensures that each type of generated fault data has similar temporal and spatial characteristics to the original fault data, can adapt to the real-time changing working conditions of the chiller, and realize the collaborative modeling of timing characteristics and dynamic topological characteristics in the complex multi-sensor system of the chiller, and can achieve high-precision real-time diagnosis even on edge devices.

[0043] Example 1

[0044] In a typical embodiment of the present invention, Figure 1 As shown, this embodiment discloses a chiller fault diagnosis method for imbalanced data, including:

[0045] S100. Based on the disturbance reconstruction mechanism, the original data is used to restore the balanced fault data with the characteristics of the original data;

[0046] S200. Extract and model the temporal features and spatial structure of the balanced fault data using a spatiotemporal autoencoder, reconstruct the original data, and obtain low-dimensional features that maintain the temporal dynamic features and spatial topological structure of the original data;

[0047] S300. Using the dynamic topology timing feature integration network to extract timing features and topology features respectively and using the dynamic gated fusion mechanism for adaptive fusion;

[0048] S400. Classify the fused features to obtain fault diagnosis results.

[0049] A chiller fault diagnosis method oriented to unbalanced data is described in detail below with reference to specific embodiments.

[0050] S100. Based on the disturbance reconstruction mechanism, the original data is used to restore the balanced fault data with the characteristics of the original data.

[0051] In chiller fault diagnosis, data imbalance can cause diagnostic models to favor the majority class, reducing their ability to identify minority class faults. To address this challenge, a perturbation reconstruction mechanism is introduced into the latent space of an autoencoder to generate balanced fault data. Based on the principle of Markov chains, this perturbation reconstruction mechanism employs a progressive denoising function to gradually transform noisy data into useful data, thereby restoring and generating fault data with the original characteristics.

[0052] Specifically, the perturbation reconstruction mechanism includes two main steps: first, gradually adding noise to the original data to make it conform to the Gaussian distribution; then, through the designed denoising function, gradually removing the noise and reconstructing the balanced fault data. At the same time, in order to improve the efficiency and quality of data generation, a label guidance layer is incorporated into the denoising process, enabling the model to generate data corresponding to specific labels, thereby achieving parallel generation of multiple types of data. Figure 2 shown.

[0053] S101. Gradually add noise to the original data to make it conform to the Gaussian distribution.

[0054] During the noise injection process, for the original data Given an initial distribution , Gaussian noise is gradually added to this distribution in several steps, the number of steps is expressed as T , this process will eventually cause the data to be completely submerged by noise. The positive noise injection process is defined as follows:

[0055] (1);

[0056] in and Respectively represent the time step and Noise samples at time ; is a hyperparameter for scaling Gaussian noise, which varies with t Increase and increase; I represents the identity matrix, N Represents a Gaussian distribution.

[0057] S102. Use the designed denoising function to gradually remove noise and reconstruct balanced fault data.

[0058] Noise removal process By optimizing To complete the denoising task, is the set of model parameters that parameterize all key functions. The reverse process is defined as:

[0059] (2);

[0060] in, and denote the mean and variance of the noise removal process respectively.

[0061] According to the properties of Markov chain, we can conclude that:

[0062] (3);

[0063] Based on the probability density of the standard Gaussian distribution and Bayes' theorem, the posterior distribution is obtained The mean and variance of :

[0064] (4);

[0065] in , ,and Indicates the prediction of U-net at step t Added noise. Noise is sampled from a standard Gaussian distribution. By introducing the random variable , which can increase the diversity of data generated by DDPM.

[0066] By using the reparameterization technique, the final simplified objective function can be derived:

[0067] (5);

[0068] By using the objective function Training U-net can achieve accurate noise prediction.

[0069] S103. Incorporating a label guidance layer into the denoising process enables the model to generate data corresponding to specific labels, thereby achieving parallel generation of multiple types of data.

[0070] When training the reverse process of DDPM, U-net receives not only the generated data but also the steps t The embedded code is taken as input. t The embedding code is defined as follows:

[0071] (6);

[0072] in d represents the length of the embedded code, i Indicates the length index. tThe embedding code of is input into U-net together with the noisy data, which can capture the temporal dependency of the noise.

[0073] To balance the problem of decreased generation quality caused by noise guidance, this embodiment uses a hybrid approach of conditional guidance and unconditional guidance when embedding fault labels into the denoising process to guide data generation, and imposes weight constraints on the conditional guidance. The formula for conditional guidance is as follows:

[0074] (7);

[0075] in W is the similarity hyperparameter, t represents the time step, c represents label encoding, It is in step t The original vibration signal with noise added.

[0076] By gradually introducing noise to make the original data close to Gaussian distribution, and then using denoising strategy to gradually restore the data, while incorporating label guidance, the parallel generation of multiple types of fault samples is achieved, which significantly improves the efficiency of data generation and shortens the generation time. At the same time, by balancing the data distribution, the quality of generated data is improved, effectively alleviating the problem of imbalanced chiller fault data, solving the problems of imbalanced chiller fault data and low efficiency of traditional serial generation methods in the existing technology, and providing a more accurate and comprehensive training data foundation for subsequent fault diagnosis.

[0077] S200. Extract and model the temporal features and spatial structure of the balanced fault data using a spatiotemporal autoencoder, reconstruct the original data, and obtain low-dimensional features that maintain the temporal dynamic features and spatial topological structure of the original data;

[0078] Given the significant temporal dependence and topological relevance of chiller data, this embodiment designs a spatiotemporal autoencoder framework. Recurrent neural network layers and graph attention network layers are introduced into the autoencoder to extract the temporal and spatial characteristics of the fault data, respectively. This ensures that the spatiotemporal characteristics of the hidden spatial data are retained, ensuring that the generated fault data remains similar to the original data in terms of spatiotemporal characteristics. This ensures that the extracted hidden states contain sufficient pattern features to support the reconstruction of chiller fault data. The spatiotemporal autoencoder can autonomously learn and extract representative intrinsic pattern features, utilizing low-dimensional hidden states to reconstruct chiller fault data.

[0079] In the encoder part, the spatiotemporal autoencoder consists of a first convolutional layer, a recurrent neural network layer, a graph attention network layer, and a second convolutional layer. Among them, the first convolutional layer is used to preprocess the original input data and extract local features; the middle recurrent neural network layer captures the temporal dependencies in the time series, thereby extracting dynamically changing information; the graph attention network layer that follows uses the self-attention mechanism to model the complex topological relationships between sensor nodes and effectively aggregate multi-hop information; the last second convolutional layer is used to further integrate and compress multi-scale features to generate high-quality hidden states. This design fully considers the characteristics of the chiller data with significant temporal dependence and topological correlation, using the convolutional layer to capture local features, the recurrent neural network layer to model the temporal dependencies of the time series, and the graph attention network layer to capture the complex topological relationships between nodes in the system. Figure 3 As shown, the decoder in this embodiment adopts a network structure symmetrical to that of the encoder to ensure efficient reconstruction of the original data.

[0080] S201. Extract and model the time series features of the balance fault data.

[0081] In view of the temporal dependency of the data, in the encoder, the information of each time step is first embedded through a 1×1 convolution to obtain For chiller fault data, this embodiment uses a recurrent neural network to gradually capture temporal dependencies and compress time steps. By flexibly setting the hidden state dimension and number of layers in each layer, the model can integrate information from all historical time steps while using as little memory as possible. For a time series x, the recursive computation using a multi-layer recurrent neural network unit R is formally defined as:

[0082] (8);

[0083] in Indicates the t The hidden state of time steps, is the ReLU activation function, and are the weight matrices from input to hidden layer and from hidden layer to hidden layer, is the bias term, and the output is .

[0084] S202. Extract and model the spatial structure of the balanced fault data.

[0085] In view of the topological correlation of the data, this embodiment adds a correlation learning module - graph attention network in both the encoder and decoder to aggregate multi-hop information. Adaptive adjacency matrix A and each layer hidden state As input, dynamic weighted aggregation of information is achieved through the self-attention mechanism.

[0086] Specifically, for each node i For example, the next hidden state is defined as:

[0087] (9);

[0088] in, Representation node i The neighbor set of For the l The weight matrix of the layer, is the ReLU activation function; attention coefficient To measure the node j For Node i The impact of is calculated as follows:

[0089] (10);

[0090] in, is a learnable attention vector, || represents the concatenation operation of the vector, and the output is .

[0091] The encoder compresses the original chiller fault data of length T into a hidden state sequence through multi-layer RNN and GAT , usually, use the last hidden state As the representation of the input sequence, it is then passed through another 1×1 convolution to output the hidden state of the intrinsic pattern h In this embodiment, the decoder adopts a network structure symmetrical to that of the encoder to ensure efficient reconstruction of the original data.

[0092] In order to ensure that the generated data is accurate and the potential space distribution is close to the Gaussian distribution, the mean square error (MSE) and KL divergence are used as loss functions in this embodiment:

[0093] (11);

[0094] (12);

[0095] Where, represents the reconstruction loss value of the spatiotemporal autoencoder, Represents the original input data of the encoder, represents the reconstructed data generated by the decoder; represents the KL divergence loss value, represents the potential variable in the latent space i The mean of the dimension, Represents the latent variable i The standard deviation of the dimension.

[0096] By introducing a recurrent neural network layer and a graph attention network layer into the autoencoder, the temporal features and spatial structure of the data are deeply extracted and dynamically modeled, respectively, ensuring that the data in the latent space faithfully reflects the temporal and spatial characteristics of the original data. This design addresses the existing problem of deviations between the temporal relationship and spatial structure of generated fault data and the original data. The generated fault data is highly consistent with the original data in terms of temporal relationship and spatial structure, significantly reducing generation deviations and improving the authenticity and reliability of the data. This provides higher-quality data support for the fault diagnosis system and further enhances the accuracy of fault detection and diagnosis.

[0097] S300. Using the dynamic topology timing feature integration network to extract timing features and topology features respectively and using the dynamic gated fusion mechanism for adaptive fusion;

[0098] In order to adapt to the real-time changing working conditions of the chiller and realize the collaborative modeling of the timing characteristics and dynamic topological characteristics of the chiller, this embodiment proposes a dynamic topological timing feature integration network to perform real-time fault diagnosis of the chiller. The dynamic topological timing feature integration network is mainly divided into two branches: the dynamic topology extraction branch and the timing feature extraction branch. The dynamic topology extraction branch can handle the dynamic changes of the graph structure and is used to capture the correlation between different sensors in the chiller as the working conditions change. The timing feature extraction branch models the long-term time dependence of the sensor signal and predicts the fault evolution trend. The fusion of the two branches is essentially the synergistic enhancement of spatial topological modeling and timing pattern capture, forming a complementary advantage of "1+1>2" in industrial fault diagnosis. The specific model structure is as follows Figure 4 shown.

[0099] The joint training of the dynamic topology extraction branch and the temporal feature extraction branch is carried out in stages. First, the two branch networks are trained independently until convergence. Then, the backbone network is frozen and only the fusion gating weights and classifier are trained. The following describes the training process of the temporal feature extraction branch and the dynamic topology extraction branch respectively.

[0100] S301. Training of temporal feature extraction branch.

[0101] like Figure 4 As shown on the left, the time series feature extraction branch first inputs the time series fault data enhanced by the spatiotemporal autoencoder ,in L is the time step, D is the feature dimension. Then a 1D convolution with a stride of 2 is used to reduce the sequence length and extract local features:

[0102] (13);

[0103] in, is the weight parameter of the convolution kernel, It is a learnable position code. Output features , the sequence length is given by L Compress to L / 2, the feature dimension is D Expand to d model The convolution kernel size of 1D convolution is 3, and the sequence length is halved by step size 2. t The output is:

[0104] (14) ;

[0105] Subsequently, the local features extracted by the sliding window operation of the 1D convolution kernel in the time dimension are input into the dynamic sparse attention block. Each dynamic sparse attention block consists of a grouped query attention and a depthwise separable feedforward network (DS-FFN).

[0106] Grouped query attention is an improved self-attention mechanism that is mainly used to optimize the computational efficiency of the model when processing large-scale inputs. Its core idea is to reduce computational complexity by grouping queries. First, based on the input features of each time step, , projection generates query (Q), key (K), value (V) matrices. A total of 4 attention heads are used, and the 4 query heads are divided into 2 groups, each group sharing the key matrix and value matrix:

[0107] (15);

[0108] Among them, the weight matrix , ; , , each group contains two query heads.

[0109] Then, through the sparse attention mask, only the attention of each query and its five preceding and following keys is allowed to be calculated:

[0110] (16);

[0111] The goal of sparse attention masking is to limit the range of interactions between queries and keys by setting a mask. The mask is usually a binary matrix whose values indicate whether a query is allowed to interact with a key. Sparse attention masking can reduce unnecessary calculations, thereby improving the efficiency of the model when processing large-scale data. Thus, the single-head attention output is obtained:

[0112] (17);

[0113] like , using the key value of group 1: .like , using the key value of group 2: .

[0114] Then, the outputs of all heads are concatenated to obtain the output of the grouped query attention:

[0115] (18);

[0116] Among them, the parameter matrix This design significantly reduces the computational resource requirements while maintaining the expressive power of multi-head attention through group-shared key-value projection and local sparse attention.

[0117] Using a residual connection, the output of the previous layer is added to the residual of the input:

[0118] (19);

[0119] This connection allows gradients to flow directly through the network during backpropagation, thereby alleviating the vanishing gradient problem.

[0120] The summed output is normalized and then fed into the depthwise separable feedforward network layer, which consists of two main parts: the depthwise convolutional layer and the pointwise convolutional layer. The depthwise convolutional layer is responsible for performing spatial convolution operations on each channel of the input features to extract local features; the pointwise convolutional layer uses 1D convolution to linearly transform the features in the channel dimension, integrating information between channels. The entire process can be expressed as:

[0121] (20);

[0122] Using the residual connection again, the output of the previous layer is added to the residual of the input:

[0123] (twenty one);

[0124] In order to reduce the number of model parameters, a cross-layer parameter sharing mechanism is adopted, and odd layers and even layers share the key / value projection matrix and convolution kernel parameters respectively.

[0125] Hypothesis n After processing the dynamic sparse attention blocks of the stacked layers, the final output features of the model are , which is then used for subsequent fusion with the output of the node attention pooling layer of the dynamic topology extraction branch.

[0126] S302. Training of dynamic topology extraction branch.

[0127] like Figure 4 As shown on the right, in the training process of the dynamic topology extraction branch, the multi-dimensional sensor signal is first input , the signal comes from the sensor and is collected at a fixed frequency.

[0128] Then split it with sliding window:

[0129] (twenty two);

[0130] w is the window length, which indicates the number of consecutive time steps contained in each window; t is the current time step, and the end position of the window is ; D is the sensor feature dimension; From the time step t−w arrive t −1 data matrix.

[0131] The sliding window ensures that when new data arrives, the oldest data is removed and the latest data is added. Each sensor channel is then normalized independently:

[0132] (twenty three);

[0133] Then, a dynamic graph is constructed. Specifically, a graph structure is constructed, each sensor corresponds to a graph node, and the Pearson correlation coefficient is used to calculate the dynamic adjacency matrix:

[0134] (twenty four);

[0135] In-window sensor i The mean of .

[0136] The graph is sparsely processed, and only the most relevant node is retained for each node K Edge (such as K =3). Dynamic graph every The adjacency matrix is updated once every step. Using a double buffering mechanism, the background thread calculates the new adjacency matrix, and the foreground thread uses the old matrix.

[0137] Normalize the adjacency matrix symmetrically to enhance numerical stability:

[0138] (25);

[0139] in, is the degree matrix. Then, the dynamic graph is input to the common m In the Chebyshev graph convolution layer, the ReLU activation function is used, and the definition of graph convolution is as follows:

[0140] (26);

[0141] is the scaled normalized Laplacian matrix, is the original Laplace matrix, is the Chebyshev polynomial, is the trainable weight matrix of the k-th order polynomial in the l-th layer.

[0142] After the dynamic graph convolution, a temporal convolution layer (1D dilated convolution) is used to further capture the multi-scale temporal dependencies of the sensor signals. Finally, the node attention pooling layer is used to calculate the importance weight of each sensor node, and the weighted aggregation is converted into a global feature that can reflect the global influence of the sensor node in the structure:

[0143] (27);

[0144] (28);

[0145] is the importance weight of a single sensor node, is the node feature of a single sensor node, N is the number of sensor nodes.

[0146] S400. Classify the fused features to obtain fault diagnosis results.

[0147] After the dynamic topology extraction branch network and the temporal feature extraction branch network are independently trained to converge, the dynamic gated fusion mechanism is used to adaptively fuse the temporal features extracted by the two branches. and topological features .

[0148] First, the topological features are copied in time steps to align with the temporal features output by the temporal feature extraction branch. ,in, Indicates the length of the sequence after processing; N represents the number of sensor nodes; Represents the dimension of the feature vector.

[0149] Then perform dynamic gating weight generation:

[0150] (29);

[0151] is the 1D convolution kernel weight, is the bias term, the activation function is the Sigmoid function, and the output gate value The dynamic gated fusion mechanism ensures that the fusion weights are learned independently at each time step and each feature dimension.

[0152] Gating fuses the features output by two models:

[0153] (30);

[0154] when When it approaches 1, the feature of this time step mainly depends on the temporal feature extraction branch; when When it approaches 0, the structural features of the branch are extracted based on the dynamic topology.

[0155] The fused features are finally input into the hierarchical classifier for classification to obtain the final fault diagnosis results:

[0156] (31);

[0157] In general, this time series dynamic integration method achieves efficient analysis and accurate diagnosis of multi-source heterogeneous sensor data through the collaboration of dynamic topology extraction branches and time series feature extraction branches. It has high precision, low latency and strong generalization capabilities, providing a reliable edge intelligence solution for predictive maintenance of industrial equipment.

[0158] By introducing a fault diagnosis method based on a dynamic topological temporal feature integration network, a temporal feature extraction branch network is used to model global temporal dependencies using sparse attention and parameter sharing. At the same time, a dynamic topology extraction branch network is used to construct a graph structure according to real-time sensor correlation and perform Chebyshev graph convolution to capture the dynamic topological changes of the chiller sensors. Finally, the features of the two models are fused through a dynamic gated fusion mechanism to output the fault diagnosis results. High-precision real-time diagnosis can also be achieved on edge devices, which can adapt to the real-time changing working conditions of the chiller and realize the collaborative modeling of temporal features and dynamic topological features in the complex multi-sensor system of the chiller, solving the problem that traditional fault diagnosis methods lack the ability to adapt to real-time changing working conditions.

[0159] Example 2

[0160] In a typical embodiment of the present invention, this embodiment discloses a chiller fault diagnosis system for imbalanced data, including:

[0161] The balanced fault data parallel generation module is configured to: based on a disturbance reconstruction mechanism, use the original data to restore the balanced fault data with the characteristics of the original data;

[0162] The data enhancement module is configured to: extract and model the temporal features and spatial structure of the balanced fault data using a spatiotemporal autoencoder, reconstruct the original data, and obtain low-dimensional features that preserve the temporal dynamic features and spatial topological structure of the original data;

[0163] The feature extraction and fusion module is configured to: use a dynamic topological temporal feature integration network to extract temporal features and topological features respectively and use a dynamic gated fusion mechanism to perform adaptive fusion;

[0164] The fault diagnosis module is configured to classify the fused features and obtain a fault diagnosis result.

[0165] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A chiller fault diagnosis method for unbalanced data, characterized in that: include: Based on a perturbation reconstruction mechanism, the original data is used to restore balanced fault data with the characteristics of the original data. The perturbation reconstruction mechanism specifically includes: gradually adding noise to the original data to make it conform to the noise of Gaussian distribution; using a denoising function, the noise is gradually removed to reconstruct balanced fault data with the characteristics of the original data; a label guidance layer is incorporated into the denoising process to generate data corresponding to specific labels, thereby achieving parallel generation of multiple types of data; the specific labels are fault labels; A spatiotemporal autoencoder is used to extract and model the temporal features and spatial structure of the balance fault data, reconstruct the original data, and obtain low-dimensional features that retain the temporal dynamic features and spatial topological structure of the original data. The spatiotemporal autoencoder consists of a first convolutional layer, a recurrent neural network layer, a graph attention network layer, and a second convolutional layer. The first convolutional layer is used to preprocess the original input data and extract local features. The recurrent neural network layer is used to capture the temporal dependencies in the time series and extract information about dynamic changes. The graph attention network layer is used to model the complex topological relationships between sensor nodes using the self-attention mechanism and effectively aggregate multi-hop information; the second convolutional layer is used to integrate and compress multi-scale features to generate high-quality hidden states; A dynamic topology temporal feature integration network is used to extract temporal features and topological features respectively and a dynamic gated fusion mechanism is used to perform adaptive fusion; the dynamic topology temporal feature integration network includes a dynamic topology extraction branch and a temporal feature extraction branch; The dynamic topology extraction branch is used to capture the correlation between different sensors in the chiller as the operating conditions change; the time series feature extraction branch is used to model the long-term time dependence of sensor signals and predict fault evolution trends; The adaptive fusion process is as follows: The topological features are copied in time steps to align with the timing features. ,in, represents the length of the sequence after processing, N represents the number of sensor nodes, Represents the dimension of the feature vector; then generates dynamic gating weights through 1D convolution and Sigmoid function, and outputs the gated fusion features; The fused features are classified to obtain the fault diagnosis results.

2. The method for diagnosing chiller faults based on unbalanced data according to claim 1, wherein: The training process of the temporal feature extraction branch is as follows: Input the time series fault data enhanced by the diffusion model; then use 1D convolution with a stride of 2 to reduce the sequence length and extract local features; then input the local features extracted by convolution into the dynamic sparse attention block, use sparse attention and parameter sharing to model global time series dependencies, and output time series features.

3. The method for diagnosing chiller faults based on unbalanced data according to claim 1, wherein: The training process of the dynamic topology extraction branch is: The input is a multidimensional signal collected by the sensor at a fixed frequency and split with a sliding window. Each sensor channel is independently normalized to construct a dynamic graph, which is then input into a Chebyshev graph convolution with m layers for convolution processing. A temporal convolution layer is used to capture the multi-scale temporal dependencies of the sensor signal. Finally, the importance weight of each sensor node is calculated through a node attention pooling layer, and the weighted aggregation is aggregated into a global feature to output a topological feature.

4. The method for diagnosing chiller faults based on unbalanced data according to claim 1, wherein: The fused features are input into the hierarchical classifier for classification to obtain the final fault diagnosis results.

5. A chiller fault diagnosis system for imbalanced data, characterized in that: include: The balanced fault data parallel generation module is configured to: restore balanced fault data with the characteristics of the original data from the original data based on a perturbation reconstruction mechanism; the perturbation reconstruction mechanism specifically includes: gradually adding noise to the original data to transform it into noise that conforms to a Gaussian distribution; using a denoising function to gradually remove the noise and reconstruct balanced fault data with the characteristics of the original data; and incorporating a label guidance layer into the denoising process to generate data corresponding to specific labels, thereby achieving parallel generation of multiple types of data; the specific labels are fault labels; The data enhancement module is configured to: use a spatiotemporal autoencoder to extract and model the temporal features and spatial structure of the balance fault data, reconstruct the original data, and obtain low-dimensional features that retain the temporal dynamic features and spatial topological structure of the original data; the spatiotemporal autoencoder is composed of a first convolutional layer, a recurrent neural network layer, a graph attention network layer, and a second convolutional layer; the first convolutional layer is used to preprocess the original input data and extract local features; the recurrent neural network layer is used to capture the temporal dependencies in the time series and extract information about dynamic changes; The graph attention network layer is used to model the complex topological relationships between sensor nodes using the self-attention mechanism and effectively aggregate multi-hop information; the second convolutional layer is used to integrate and compress multi-scale features to generate high-quality hidden states; The feature extraction and fusion module is configured to: use a dynamic topology temporal feature integration network to extract temporal features and topological features respectively and use a dynamic gated fusion mechanism to perform adaptive fusion; the dynamic topology temporal feature integration network includes a dynamic topology extraction branch and a temporal feature extraction branch; The dynamic topology extraction branch is used to capture the correlation between different sensors in the chiller as the operating conditions change; the time series feature extraction branch is used to model the long-term time dependence of sensor signals and predict fault evolution trends; The adaptive fusion process is as follows: The topological features are copied in time steps to align with the timing features. ,in, represents the length of the sequence after processing, N represents the number of sensor nodes, Represents the dimension of the feature vector; then generates dynamic gating weights through 1D convolution and Sigmoid function, and outputs the gated fusion features; The fault diagnosis module is configured to classify the fused features and obtain a fault diagnosis result.

Images