A class imbalance deep map learning method and system based on class association relationship and cost sensitivity

Abstract

The application belongs to the field of class imbalance feature learning, and discloses a class imbalance deep graph learning method and system based on class association relationship and cost sensitivity, which comprises the following steps: a signal collector collects device signal waveform data in real time and saves the data, divides the saved waveform data, gives the sample a class label according to a log file or manual labeling, and constitutes a final petrochemical dynamic device signal sample data set; the signal data with concurrent fault classes are subjected to graph learning to obtain an optimal graph structure; and through a graph convolution network and the fusion of cost-sensitive learning, classification of the unbalanced fault signal data is realized. Through the deep fusion of the innovative graph learning architecture and the cost-sensitive mechanism, the application effectively solves the class imbalance problem in concurrent fault classification, and provides a more reliable technical solution for fault diagnosis in the fields of industrial equipment monitoring and intelligent operation and maintenance.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of class-imbalanced feature learning technology, and particularly relates to a class-imbalanced deep graph learning method and system based on class association and cost sensitivity. Background Technology

[0002] In the process of acquiring signals from petrochemical equipment, the normal operating time of the equipment is much longer than the time spent in failure, and the probability of different failures varies significantly even for the same type of equipment. Therefore, the acquired signals from petrochemical equipment suffer from a severe class imbalance problem. Class imbalance causes the classifier to favor the majority class, reducing the accuracy of fault classification. Thus, class imbalance is a problem that must be addressed in fault signal classification. The core of class imbalance feature learning lies in solving the problem of classifier bias towards the class with more samples due to the inconsistent sample numbers between classes. Existing methods mainly address class imbalance from two levels: data balancing and feature learning.

[0003] Data balancing: Data balancing methods aim to reduce the disparity between imbalanced classes at the data level, thereby mitigating the impact of class imbalance on classifiers. These methods typically use sampling to ensure that the number of samples for all classes in the dataset is consistent. Currently, mainstream data balancing techniques include undersampling, oversampling, and combined sampling.

[0004] Undersampling methods mitigate majority-class bias by reducing the number of majority-class samples. To address the issue that random undersampling might remove potentially useful information from the majority class and degrade classifier performance, some approaches use the K-nearest neighbor rule to select cluster centers or their neighbors to represent the majority-class data. Other methods, targeting imbalanced data, design undersampling strategies based on class overlap to specifically extract samples crucial for classification. Still other approaches address the potential for undersampling to corrupt the distribution information of the sample set by proposing sensitivity-based multivariate undersampling methods to achieve class balance while maintaining data representativeness.

[0005] Oversampling methods can improve classification performance by increasing the number of minority class samples. The classic SMOTE algorithm generates new, non-repeating minority class samples by randomly selecting nearest neighbor samples of the same class for interpolation. To address the overgeneralization problem of SMOTE, improved schemes introduce selection weights when selecting nearest neighbor samples; another method divides imbalanced datasets into distance bands and uses adaptive variable neighborhood SMOTE to generate samples within the bands, improving specificity. When dealing with linearly inseparable datasets, one approach uses self-organizing maps to project the high-dimensional space onto two dimensions, synthesizing new samples within and between clusters; another approach utilizes the feature extraction invariance of adaptive subspace self-organizing maps to generate samples in the kernel subspace. Furthermore, the PCBoost algorithm constructs a decision tree weak classifier using information gain ratio: at the initial iteration, synthetic minority class samples are added to balance training information; after the subclassifier is formed, the "perturbation" is corrected by deleting incorrectly classified synthetic samples, improving model stability.

[0006] Combined sampling, by fusing undersampling and oversampling techniques, can construct class-balanced datasets. One approach first filters noisy and boundary sample data before using the SMOTE method to generate a class-balanced training sample set. Another approach proposes a three-step sequential intelligent hybrid sampling strategy specifically designed to address the overfitting problem that may occur during SMOTE oversampling, thereby ensuring minority class augmentation while avoiding excessive model reliance on dummy samples.

[0007] Ensemble learning methods combine multiple weak classifiers to construct a strong classifier. One approach proposes a bagging ensemble learning algorithm based on SMOTE and differential sampling rate, applying it to an imbalanced decision tree ensemble model for enterprise credit rating. Addressing the potential problems of upsampling positive samples or downsampling negative samples during classifier ensemble—random undersampling easily overlooks some information samples, while upsampling to add synthetic positive samples increases training complexity—another approach proposes a progressive augmentation algorithm. This algorithm, while generating different classifiers, gradually inserts irrelevant sample groups into the augmentation process to avoid information loss and improve the robustness of the ensemble classifier.

[0008] (a) The closest existing technology

[0009] One related approach is the "Cost-Sensitive Graph Neural Network (CSGNN)"—a structure that combines cost-sensitive learning with graph neural networks, primarily used to address class imbalance problems in graph data (such as fraud detection in telecommunications networks) ([arXiv][1]). Furthermore, more relevant to the field of fault diagnosis is the "COST-SENSITIVE DEEPBELIEF NETWORK" (ECS-DBN), which utilizes adaptive differential evolution to automatically optimize misclassification costs and integrates them into the training of a deep trust network (DBN), particularly applicable to imbalanced data classification tasks in tool condition monitoring ([arXiv][2]).

[0010] (II) Existing Technical Problems

[0011] While the CSGNN method cleverly combines GNN and cost-sensitive learning, its application scope is limited to fraud or anomaly detection of graph data. For industrial fault diagnosis scenarios such as vibration signals with strong time-series / time-frequency features and severe class imbalance, its graph structure construction, class relationship modeling, and time-series feature integration are difficult, and it fails to effectively adapt to the inter-sample relationships under complex mechanical operating conditions.

[0012] The ECS-DBN method automatically sets cost-sensitive parameters through evolutionary algorithms, reducing the difficulty of manual parameter tuning. However, it adopts a traditional DBN structure, lacks graph structure learning capabilities, cannot explicitly model the relationships between categories, and is also difficult to take into account graph convolution capabilities. In addition, DBN has limited ability to express high-dimensional temporal vibration signals and capture inter-class relationships, and may perform poorly in complex environments with multiple faults.

[0013] Although existing CSGNN and ECS-DBN methods have made innovations in addressing the class imbalance problem from the perspectives of graph structure or cost sensitivity, they have not yet formed a complete solution within a deep framework that integrates class association graph learning and cost-sensitive mechanisms. In particular, for graph convolutional network classification tasks involving vibration signals from petrochemical equipment and concurrent fault scenarios, further research and technological breakthroughs are still urgently needed. Summary of the Invention

[0014] To address the problems existing in the prior art, this invention provides a class-imbalanced deep graph learning method and system based on class association and cost sensitivity.

[0015] This invention is implemented as follows: a class-imbalanced deep graph learning method based on class association and cost sensitivity, specifically including:

[0016] S1: The signal acquisition device collects and saves the signal waveform data of the equipment in real time. The saved waveform data is segmented, and the samples are labeled with categories according to log files or manual annotations to form the final petrochemical equipment signal sample dataset.

[0017] S2: Perform graph learning on signal data with concurrent fault categories to obtain the optimal graph structure;

[0018] S3: Classification of imbalanced fault signal data is achieved by using graph convolutional networks and incorporating cost-sensitive learning.

[0019] Furthermore, in step S2, the implicit relationships between nodes are learned through a graph learning network to obtain the optimal graph representation, and the fault signal samples are regarded as nodes in the graph. Its corresponding label set is {x1, x2, ..., x N Graph learning networks obtain f(x) by constructing a two-layer neural network. i ,x j The first layer of this network is a low-dimensional embedding network, which uses weight parameters. The input fault signal features are reduced in dimensionality. The second layer is a mapping network used to further map the reduced features. Node x i and x j The relationship between the function f(x) i ,x j This can usually be expressed as:

[0020]

[0021] Where, x ' i =W T x i x j ' =W T x j ReLU(·) is the activation function;

[0022] This paper introduces the correlation between concurrent faults and their individual fault categories into graph learning, and designs a graph learning loss function that integrates fault correlations to simultaneously model the relationships between faults. Based on concurrent fault category information, a set C of signal sample pairs with correlations between fault categories is constructed, aiming for these correlated pairs to have large f(·) values. The weight vector of the second-layer mapping network is optimized by minimizing the improved loss function. The graph learning loss function integrating fault correlations is as follows:

[0023]

[0024] Lnew The third term is used to ensure the sparsity of graph S.

[0025] Furthermore, in S3, based on the optimal graph representation S of the fault signal data, a graph convolutional network is used to learn the potential feature representation for each graph node. The layer propagation rule of the graph convolutional network is as follows:

[0026]

[0027] Among them, H (l) Let A be the output of the l-th layer of the network, and let A be the adjacency matrix of graph S. ' =A+I, Based on the above propagation rules, θ (l) Let H be the weight matrix of the l-th layer. Assuming the graph convolutional network has L layers, the final feature output is H. (L) .

[0028] Furthermore, in S3, a cost-sensitive loss is introduced to increase the penalty for misclassification of minority class signal samples by the classification model. A cost-sensitive layer is set in the network, and the cost parameter is used as part of the network parameters for designing and jointly optimizing the classification loss function. The classification loss function fused with the cost-sensitive layer is defined as follows:

[0029]

[0030] Here, φ represents the cost parameter. To minimize the above loss function and find the optimal parameters (θ, φ), the cost parameter and network parameters are jointly optimized through alternating iterations.

[0031] Another objective of this invention is to provide a class-imbalanced deep graph learning system based on class association and cost sensitivity, the system specifically comprising:

[0032] The signal input module is used to acquire equipment signal waveform data and process it into a petrochemical dynamic equipment signal sample dataset.

[0033] The graph learning module is used to perform graph learning on signal data with concurrent fault categories to obtain the optimal graph structure;

[0034] The graph convolution module is used to classify unbalanced fault signal data by using graph convolutional networks and incorporating cost-sensitive learning.

[0035] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0036] First, this invention proposes a class-imbalanced deep graph learning method based on class association and cost sensitivity, providing an innovative solution to the class imbalance problem in concurrent fault signal data. This method organically combines graph learning techniques with a cost-sensitive mechanism, significantly improving the performance of concurrent fault classification in several aspects. Compared with existing technologies, the advantages of this method are mainly reflected in the following aspects:

[0037] (1) By introducing graph learning technology, this invention can effectively model the complex relationships between concurrent faults and single faults. Traditional methods often ignore the inherent connections between different fault categories when dealing with concurrent faults, resulting in limited classification performance. This invention innovatively constructs a graph structure representation of fault categories, where nodes represent different fault categories and edges represent the co-occurrence and dependency relationships between faults. The introduction of this graph structure enables the model to learn deep feature representations of single and concurrent faults simultaneously through graph convolutional networks, thereby more accurately capturing the complex patterns of concurrent faults.

[0038] (2) This invention innovatively designs a graph learning loss function that incorporates cost-sensitive parameters, effectively solving the classification bias problem caused by class imbalance. Unlike simple resampling methods that may disrupt the original data distribution, or traditional cost-sensitive learning methods that only adjust the decision boundary, this method embeds a cost-sensitive mechanism throughout the entire process of feature learning and classification optimization. By assigning adaptive weights to different fault categories, the model can automatically adjust the degree of attention given to samples of each category.

[0039] (3) This invention achieves synergistic optimization of fault correlation modeling and cost-sensitive learning, forming an end-to-end joint training framework. Traditional methods often treat correlation learning and class imbalance handling as independent modules, making it difficult to fully leverage their synergistic effect. However, this method, through a unified loss function, enables graph convolutional networks to simultaneously optimize graph structure representation and cost-sensitive classification, thereby learning more discriminative feature representations. This integrated design not only improves the model's generalization ability but also enables it to adapt to more complex real-world industrial scenarios.

[0040] In summary, this invention effectively solves the class imbalance problem in concurrent fault classification through the deep integration of an innovative graph learning architecture and a cost-sensitive mechanism. Compared with existing technologies, this invention has significant advantages in feature representation learning, classification robustness, and model generalization ability. Especially when processing concurrent fault data with high imbalance ratios and complex relationships, this invention demonstrates significant superiority, providing a more reliable technical solution for fault diagnosis in fields such as industrial equipment monitoring and intelligent operation and maintenance.

[0041] Secondly, by combining innovative deep graph learning with a cost-sensitive mechanism, this technology can effectively solve the problem of fault diagnosis caused by imbalanced data in industrial equipment monitoring, providing strong technical support for intelligent operation and maintenance in industries such as petrochemicals, power, and manufacturing. In practical applications, this technology can significantly improve the accuracy and timeliness of equipment fault diagnosis, enabling early fault warnings and precise fault location, thereby greatly reducing unplanned downtime, improving production efficiency, and lowering maintenance costs.

[0042] Existing deep learning methods, when dealing with class imbalance problems, mostly rely on traditional imbalance handling strategies (such as resampling or cost-sensitive learning), failing to fully utilize the characteristics of graph-structured data. This invention deeply integrates deep graph learning with cost-sensitive mechanisms, proposing a class imbalance deep graph learning method based on class association relationships and cost sensitivity, filling a technological gap in this field. When processing petrochemical equipment fault signal data, this method achieves significantly higher classification accuracy than traditional methods, providing a more reliable technical solution for industrial equipment monitoring and intelligent operation and maintenance.

[0043] Traditional methods for handling class imbalance primarily rely on data balancing techniques such as undersampling, oversampling, or combined sampling. While these methods can alleviate class imbalance to some extent, they have significant limitations: undersampling may lose important information from the majority class, oversampling easily leads to overfitting, and the effectiveness of combined sampling depends on the rationality of the sampling strategy. This invention abandons sampling strategies that rely solely on the data level, instead focusing on feature learning and model optimization. Through the joint design of graph learning techniques and cost-sensitive mechanisms, it avoids the destruction of the original data distribution by sampling methods, while significantly improving the model's classification performance.

[0044] Traditional cost-sensitive learning methods typically adjust the classifier's decision boundary by statically setting cost parameters. However, this static setting lacks flexibility and struggles to adapt to complex data distributions. This invention innovatively incorporates cost parameters as part of the network for dynamic optimization. By iteratively optimizing both network and cost parameters, the model can adaptively adjust its focus on minority class samples. This dynamic optimization mechanism overcomes the reliance on empirically derived parameters in traditional cost-sensitive learning, significantly improving the model's robustness and generalization ability.

[0045] Existing methods often treat concurrent failures as independent categories, neglecting the inherent correlation between concurrent failures and individual failures. This invention employs graph learning technology to explicitly model the relationships between failure categories as a graph structure, where nodes represent failure categories and edges represent co-occurrence relationships between failures. By designing a graph learning loss function that integrates failure correlations, the model can simultaneously optimize the similarity between nodes and the correlation between failure categories, thereby more accurately capturing the complex patterns of concurrent failures. This correlation modeling method overcomes the shortcomings of traditional methods that ignore the dependencies between failures. Attached Figure Description

[0046] Figure 1 This is a flowchart of the class-imbalanced deep graph learning method based on category association and cost sensitivity provided in this embodiment of the invention;

[0047] Figure 2 This is a waveform example diagram of a normal signal sample and three fault signal samples provided in an embodiment of the present invention;

[0048] Figure 3 This is a technical roadmap for a class-imbalanced deep graph learning system based on category association and cost sensitivity, provided by embodiments of the present invention. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0050] Under current technological conditions, petrochemical equipment generally relies on single signal feature extraction and traditional classifiers for fault diagnosis. However, this approach faces two prominent problems in practical industrial applications: First, multiple types of faults often occur concurrently during equipment operation, leading to complex coupling relationships between the original signals, which traditional methods struggle to fully reveal in the feature space. Second, the actual collected data often suffers from severe class imbalance, with a large number of normal or common fault samples masking a few critical fault samples, causing missed detections and misjudgments during model deployment, posing a significant risk to the safe production and maintenance management of the equipment. Therefore, how to effectively model the correlation structure of concurrent categories while maintaining the real-time requirements of industrial signals, and simultaneously solve the class imbalance problem, has become a pressing technical bottleneck that needs to be overcome in industrial applications.

[0051] To address the aforementioned issues, this method introduces a rigorous annotation mechanism during the signal acquisition and dataset construction stages. Multimodal waveform signals, including vibration, current, and acoustic signals, are acquired through a real-time acquisition device, and the samples are categorized using operational logs or manual annotations, forming a traceable dataset. This approach not only ensures data integrity but also enables subsequent analysis to be applied to specific working conditions, thus laying a solid foundation for graph structure modeling in complex scenarios. The dataset construction method emphasizes operability in industrial settings, avoiding the limitations of relying on manual experience for feature selection.

[0052] In the graph learning phase, this method maps signal samples to a graph structure, where nodes represent samples of different categories, and edges reflect the potential coupling relationships between categories. An optimization algorithm dynamically adjusts the edge weights and adjacency relationships to obtain the optimal graph structure. This design overcomes the limitations of traditional single-sample vectorization, enabling the model to not only capture single-point features but also reveal the contextual dependencies between fault categories. For example, when a certain bearing fault and motor overload often occur simultaneously, the graph structure can strengthen their correlation, thereby improving the sensitivity of identification.

[0053] In the feature propagation stage, a graph convolutional network is introduced to iteratively learn the graph structure. Each convolutional operation fuses the local features of a node with the weighted features of its neighboring nodes, resulting in multi-level aggregation of signal features within the graph structure. In this way, the model gradually learns the global semantic representations between different fault modes, no longer limited to local feature differences. This cross-node feature propagation is particularly crucial in complex operating conditions, ensuring that the model maintains its discriminative ability in the context of concurrent faults.

[0054] Considering the highly uneven distribution of industrial signal samples, this method introduces a cost-sensitive learning mechanism during the classification stage. By constructing a cost matrix or loss weighting function, a higher classification cost is assigned to minority class faults, thereby enhancing the model's ability to identify rare samples during optimization. This strategy avoids the drawback of overfitting to the majority class during traditional training, ensuring that critical small-sample faults receive sufficient attention in the model output. The effect is that it effectively reduces the probability of false alarms and false negatives during deployment, meeting the stringent safety and reliability requirements of industrial environments.

[0055] This method achieves closed-loop optimization from data acquisition and graph structure modeling to cost-sensitive classification, solving two long-standing problems in industrial applications: difficulty in modeling the correlation between concurrent faults and insufficient accurate diagnosis under class imbalance conditions. Its working principle is as follows: first, it ensures high-quality datasets and accurate labels; second, it constructs the correlation topology between faults through graph learning; third, it achieves layer-by-layer feature fusion and expression through graph convolution; and finally, it uses a cost-sensitive mechanism to offset the bias caused by class imbalance. The entire process is tightly integrated, possessing both theoretical innovation and meeting the engineering needs of field applications in the petrochemical industry.

[0056] like Figure 1 As shown, this embodiment of the invention provides a class-imbalanced deep graph learning method based on class association and cost sensitivity. The method specifically includes:

[0057] S1: The signal acquisition device collects and saves the signal waveform data of the equipment in real time. The saved waveform data is segmented, and the samples are labeled with categories according to log files or manual annotations to form the final petrochemical equipment signal sample dataset.

[0058] S2: Perform graph learning on signal data with concurrent fault categories to obtain the optimal graph structure;

[0059] S3: Classification of imbalanced fault signal data is achieved by using graph convolutional networks and incorporating cost-sensitive learning.

[0060] This invention acquires vibration waveform data of petrochemical equipment in real time using a signal acquisition device, and then segments and labels the data to construct a signal sample dataset containing normal and various fault states. During data construction, a uniform sampling length and number of sampling points are used to ensure temporal consistency of the samples, and class labels are accurately assigned to the samples through log files or manual annotation. Since the equipment operates far more frequently in normal conditions than in fault states, the resulting dataset inevitably exhibits significant class imbalance.

[0061] To address potential concurrent faults in signals, this invention introduces a graph learning mechanism. This mechanism constructs a graph structure based on the similarity between samples, capturing the latent correlations between categories. During graph learning, signal samples serve as nodes in the graph, and the edges between nodes are defined by the similarity or correlation between sample features. Optimizing the graph structure better represents complex equipment operating states and the relationships between multiple faults. This graph structure provides high-quality topological constraints for subsequent feature extraction and classification, enabling the model to learn the intrinsic connections between categories from a global perspective.

[0062] In the classification stage, this invention employs a graph convolutional network as the feature extraction and classification framework, and introduces a cost-sensitive mechanism into the loss function. By setting differentiated misclassification costs for different classes during training, the ability to identify minority class fault samples is significantly improved, avoiding the limitation of the model biased towards the majority class. Ultimately, this invention can obtain stable and accurate classification results on imbalanced datasets, and is particularly effective in detecting minority class fault signals, thereby improving the reliability of equipment fault diagnosis and the safety of the production process.

[0063] Table 1, in S1, describes three types of failures in petrochemical equipment.

[0064] Figure 2 The waveforms of the collected normal vibration signals and three types of fault vibration signals from the equipment are displayed. The saved waveform data is segmented, for example, a waveform data with a length of 1 second and containing 1024 sampling points is segmented into 1 sample. The initial feature dimension of the sample is 1024 dimensions. Then, according to the log file or manual annotation, the sample is labeled with a category, such as normal category or fault category, to form the final petrochemical dynamic equipment signal sample dataset.

[0065] Table 1 Three common types of petrochemical equipment failures

[0066]

[0067] Because the collected signal data from petrochemical equipment contains significantly fewer fault signals than normal signals, the sample data is prone to class imbalance, meaning there are fewer fault samples than normal samples. For example, a dataset consisting of 500 friction fault samples and 10,000 normal samples is an example of class imbalance. This makes fault signal classification difficult, as fault samples are easily classified as normal, seriously impacting petrochemical production safety.

[0068] Due to the coupling between components of petrochemical equipment, there are certain correlations between different fault categories, leading to concurrent faults. A concurrent fault refers to the simultaneous occurrence of two or more faults in a single piece of equipment; typically, there are fewer concurrent fault samples than single fault samples. Therefore, in class-imbalanced datasets containing concurrent fault samples, the class imbalance is more pronounced among normal categories, single faults, and concurrent faults.

[0069] S2, which utilizes graph convolutional network technology to learn the representation of petrochemical equipment fault signal data, presupposes representing the correlation structure between signal samples using a graph. GLCN is a framework that unifies graph learning and graph convolution, where the graph learning network can learn the implicit relationships between nodes, thereby obtaining the optimal graph representation. Fault signal samples are considered as nodes in a graph. Its corresponding label set is {x1, x2, ..., x N The goal of graph learning is to learn a graph S, in which nodes x i and x j The relationship between them can be expressed by a nonnegative function S. ij =f(x) i ,x j The graph learning network obtains f(x) by constructing a two-layer neural network. i ,x j The first layer of this network is a low-dimensional embedding network, which uses weight parameters. The first layer performs dimensionality reduction on the input fault signal features to overcome problems such as high-dimensional data computation and noise; the second layer is a mapping network used to further map the dimensionality-reduced features. Node x i and x j The relationship between the function f(x) i ,x j This can usually be expressed as:

[0070]

[0071] Where, x ' i =W T x i x j ' =W T x j ReLU(·) is the activation function to ensure that f(x) i ,x j The nonnegativity of ). The weight vector a is optimized by optimizing the following loss function:

[0072]

[0073] If x i and x j If the distance between them is large, that is, the correlation is small, then this loss function encourages the relationship function f(x) to be larger. i ,x j The smaller value is taken, and the second term in L is used to control the sparsity of graph S.

[0074] Concurrent faults typically develop from a single fault category they encompass; therefore, there is a certain correlation between the single fault category contained within a concurrent fault and the concurrent fault itself. As shown in Table 2, fault A is correlated with concurrent fault AB, and fault B is correlated with concurrent fault AB. This invention introduces the correlation between concurrent faults and their contained single fault categories into graph learning and designs a graph learning loss function that integrates fault correlations, enabling it to simultaneously model the correlations between faults. Specifically, based on concurrent fault category information, a set C of signal sample pairs with correlations between fault categories is constructed, and it is desired that the signal sample pairs with correlations between fault categories have a large f(·) value. The weight vector of the second-layer mapping network is optimized by minimizing the improved loss function. The graph learning loss function that integrates fault correlations is as follows:

[0075]

[0076] If x i and x j If the distance between them is large and the labels are not related, then this loss function encourages the relationship function f(x) to be more relevant. i ,x j Take the smaller value; if x i and x j If the distance between the labels is small and the labels are relevant, then this loss function encourages the relationship function f(x) to be more relevant. i ,x j Take the larger value, L new The third term is used to ensure the sparsity of graph S.

[0077] S3, based on the optimal graph representation S of the fault signal data, learns the potential feature representation for each graph node using a graph convolutional network. The layer propagation rule of the graph convolutional network is:

[0078]

[0079] Among them, H (l) Let A be the output of the l-th layer of the network, and let A be the adjacency matrix of graph S. Based on the above propagation rules, θ (l) Let H be the weight matrix of the l-th layer. Assuming the graph convolutional network has L layers, the final feature output is H. (L) .

[0080] To effectively classify the depth map features of imbalanced fault signal data, a cost-sensitive loss is introduced to increase the penalty for misclassification of minority class signal samples. The cost matrix is ​​the foundation for cost-sensitive learning. Taking signal data with four categories—normal N, fault A, fault B, and concurrent fault AB—as an example, the corresponding cost parameter matrix is ​​shown in Table 2.

[0081] Table 2 Cost Parameter Matrix

[0082]

[0083] `cost(x,y)` represents the penalty parameter when the classifier misclassifies a sample belonging to class x as class y. Its value can usually be initially determined based on the ratio of the number of samples in each class, and there exists a transformation relationship of `cost(x,y) = 1 / cost(x,y)`. Using the above cost parameter matrix, the cost-sensitive classification loss function for incorporating concurrent faults can be set as follows:

[0084]

[0085] Where θ represents the network's own parameters, N is the number of training samples, p(·) represents the true label distribution, and q(·) is the label distribution predicted by the classification network. i and q i Let L(·) represent the true label and the predicted label of the i-th sample, respectively, and let L(·) represent the specific classification loss function, such as mean squared error, cross-entropy, etc.

[0086] Cost parameters are typically determined based on the proportion of samples from each class, and thus have a certain degree of empirical basis. To enable cost-sensitive parameters to adaptively optimize during network training, a cost-sensitive layer is incorporated into the network. The cost parameters are then included as part of the network parameters in the design of the classification loss function and jointly optimized. The classification loss function incorporating the cost-sensitive layer is defined as follows:

[0087]

[0088] Here, φ represents the cost parameter. To minimize the above loss function and find the optimal parameters (θ, φ), the cost parameter and network parameters are jointly optimized through alternating iterations. Specifically, the cost parameter φ is kept constant while the network parameter θ is updated. After one round of data training, the network parameter θ is fixed, and the cost parameter φ is updated. This classification network uses stochastic gradient descent for error backpropagation.

[0089] like Figure 3 As shown, an embodiment of the present invention provides a class-imbalanced deep graph learning system based on class association and cost sensitivity, specifically including:

[0090] The signal input module is used to acquire equipment signal waveform data and process it into a petrochemical dynamic equipment signal sample dataset.

[0091] The graph learning module is used to perform graph learning on signal data with concurrent fault categories to obtain the optimal graph structure;

[0092] The graph convolution module is used to classify unbalanced fault signal data by using graph convolutional networks and incorporating cost-sensitive learning.

[0093] The signal input module is responsible for acquiring the raw vibration, acoustic, or current waveforms of petrochemical equipment and converting them into a standardized sample dataset. This process involves not only noise suppression, normalization, and feature preprocessing, but also segmenting continuous time-series signals into windowed samples suitable for training and inference, thus providing a reliable data foundation for subsequent graph learning. Through this processing, the multidimensional signal characteristics of complex equipment under concurrent fault conditions are preserved and unified.

[0094] The graph learning module transforms input signal samples into a graph structure representation. Since faults often exhibit potential correlations, the system models the similarity between samples using category relationships and determines the connection strength between nodes and edges through a graph structure optimization algorithm. In this way, the system not only preserves the local features of individual samples but also portrays the interdependencies between different fault categories, thus possessing stronger contextual representation capabilities during deep learning. This stage effectively completes the upgrade from traditional vectorized representation to structured graph representation.

[0095] The graph convolution module operates on the aforementioned graph structure, achieving cross-node propagation and aggregation of signal features through a Graph Convolutional Network (GCN). In each convolutional operation, nodes not only fuse their own features but also receive weighted features from neighboring nodes, thus forming a comprehensive semantic representation. By introducing class relationships, graph convolution can enhance the identification of key fault modes and reduce classification ambiguity caused by sample sparsity or noise. Through multi-layer graph convolution iterations, the system gradually extracts the high-order features most discriminative for classifying imbalanced fault modes.

[0096] A cost-sensitive learning mechanism is embedded in the classification stage to address the problem of uneven distribution of fault class samples. Traditional models easily achieve high accuracy in the majority class but are insufficient in identifying minority class faults. This system introduces different cost matrices or loss weighting methods to assign higher weights to minority class samples during training, thereby improving the detection capability of rare and critical faults. Overall, the system achieves closed-loop optimization from signal acquisition to graph learning to cost-sensitive classification, enabling it to maintain stable and accurate diagnostic performance even in complex conditions such as class imbalance and concurrent faults.

[0097] This invention is applicable to equipment health management in equipment-intensive industries such as petrochemicals, power, energy, and rail transportation. Specifically, this technology can be applied to typical scenarios such as vibration signal analysis of rotating machinery (e.g., compressors, steam turbines, generators), bearing wear monitoring, and gearbox fault diagnosis. Through real-time acquisition and analysis of equipment operating signals, it accurately identifies various single faults and concurrent fault modes.

[0098] By introducing graph learning techniques, the complex relationships between concurrent and single faults can be effectively modeled. The introduction of graph structures enables the model to learn deep feature representations of both single and concurrent faults simultaneously through graph convolutional networks, thereby more accurately capturing the complex patterns of concurrent faults.

[0099] By incorporating a cost-sensitive graph learning loss function, the classification bias problem caused by class imbalance is effectively solved. By assigning adaptive weights to different fault categories, the model can automatically adjust the level of attention given to samples of each category.

[0100] This method achieves collaborative optimization of fault correlation modeling and cost-sensitive learning, forming an end-to-end joint training framework.

[0101] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0102] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A class-imbalanced deep graph learning method based on class association and cost sensitivity, characterized in that, include: Step 1: Collect equipment vibration signal waveform data using a signal acquisition device, segment the waveform data and label it with category tags to form a petrochemical equipment signal sample dataset; Step 2: Perform graph learning on the signal data with concurrent fault categories to obtain the optimal graph structure representation; The category association refers to the relationship between a concurrent fault and the single fault category it contains; In the graph learning process, a set of signal sample pairs with correlations among fault categories is constructed, and a graph learning loss function that integrates the correlations among fault categories is used for optimization. For sample pairs of categories that do not have correlations, a penalty term based on Euclidean distance is introduced to reduce their similarity. For related category sample pairs, a similarity enhancement term based on Euclidean distance is introduced to improve their similarity; a norm-based regularization term is introduced to ensure that the generated graph structure is sparsity; the optimal graph structure representation is obtained by minimizing the loss function. Step 3: Based on the optimal graph structure representation, classify the imbalanced fault signal data by using a graph convolutional network and fusing cost-sensitive learning; A cost-sensitive layer is set up in the classification stage to introduce a cost parameter to increase the penalty for misclassification of minority class samples. The network parameters and cost parameters are jointly optimized by alternating iterations. The classification loss function calculates the error between the predicted result and the true label and is weighted by the cost parameter.

2. The method as described in claim 1, characterized in that, In step two, the graph learning network models the implicit relationships between nodes by constructing a two-layer neural network. The first layer is a low-dimensional embedding network used to reduce the dimensionality of the input features, and the second layer is a mapping network used to generate a node relationship representation function. The relationship representation function is obtained by calculating the dot product of the node feature difference vector and the weight vector, processing it with a nonlinear activation function, performing an exponential operation, and normalizing the results for all nodes.

3. The method as described in claim 1, characterized in that, In step three, the hierarchical propagation rule of the graph convolutional network is as follows: In each layer, the adjacency matrix is ​​first added to the identity matrix to obtain the extended adjacency matrix, then the degree matrix of the extended adjacency matrix is ​​normalized, and then multiplied with the feature matrix and the corresponding weight matrix of the previous layer. Finally, the feature representation of the layer is output through a nonlinear activation function.

4. A class-imbalanced deep graph learning system based on class association and cost sensitivity, characterized in that, The system is used to execute the class-imbalanced deep graph learning method based on class association and cost sensitivity as described in claim 1, comprising: The signal input module is used to acquire equipment signal waveform data and generate a petrochemical equipment signal sample dataset. The graph learning module is used to perform graph learning on signal data with concurrent fault categories to obtain the optimal graph structure; The graph convolution module is used to classify unbalanced fault signal data by using a graph convolutional network and incorporating cost-sensitive learning, based on the optimal graph structure.

5. The system as described in claim 4, characterized in that, The graph learning module includes a low-dimensional embedding submodule and a mapping submodule. The former is used to reduce the dimensionality of the input signal features, while the latter is used to generate a representation of the relationship between nodes and optimize the mapping weights by combining the correlation of fault categories.

6. The system as described in claim 4, characterized in that, The graph convolution module includes a cost-sensitive layer, which is used to introduce a cost parameter into the classification loss function to improve the classification accuracy of minority class samples.

7. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the method as claimed in any one of claims 1 to 3.

8. A depth map learning apparatus, comprising a processor and a memory, wherein the memory stores a program executable on the processor, the program, when executed, implementing the method as described in any one of claims 1 to 3.

Images