Industrial equipment fault diagnosis method, terminal device and computer readable storage medium

CN122839121APending Publication Date: 2026-09-29中电华创(苏州)电力技术研究有限公司 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610754010.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-05-28
Publication Date
2026-09-29

AI Technical Summary

Technical Problem

[0005]本发明旨在解决现有工业设备故障诊断方法中,无标注数据的自监督表征学习与设备物理机理割裂,导致在极少标注数据条件下难以获得物理一致性的设备状态表征的技术问题

Benefits of technology

[0016]有益效果:本发明通过在动态加权拓扑图的构建过程中融合物理连接权重与数据相关权重,使图结构既保留设备的物理本质又捕捉数据层面的潜在关联;在此基础上,将组件健康状态指数、拓扑结构完整性约束及故障模式区分度约束统一嵌入无标注数据的工业特征提取与表示优化过程,使节点特征恢复受物理统计特征约束、邻接关系重构受设备拓扑完整性约束、判别性优化受工业状态语义约束,从而在标注数据极度稀缺的条件下,利用大量无标注数据学习到兼具数据判别性与物理一致性的设备状态表征;进而结合少量标注样本训练故障分类模型,实现对工业设备故障的准确诊断。由此,本发明有效解决了无标注数据的自监督表征学习与工业物理机理割裂的技术瓶颈,突破了标注数据稀缺对模型性能的限制,显著提升了工业设备故障诊断的准确性与物理可解释性。

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122839121A_ABST
    Figure CN122839121A_ABST
Patent Text Reader

Abstract

The application discloses a kind of industrial equipment unattended chart representation fault diagnosis methods of fusion physical constraint. In view of the problem that data is scarce in industrial field and existing self-supervised representation is disconnected with equipment physical mechanism, first, based on the correlation between equipment physical connection and sensor data, a weighted topological graph dynamically evolving with working condition is constructed;Second, the health state index of component, the constraint of topological integrity and fault mode distinguish degree are fused, the node feature recovery, adjacent relationship reconstruction and discriminant optimization are carried out on the weighted topological graph, and the equipment state feature representation with data discriminant and physical consistency is obtained;Again, a small amount of labeled samples are used to train a fault classification model, and real-time data is constructed into a dynamic graph to input the model to obtain a diagnosis result.The application effectively breaks through the bottleneck of labeled scarcity, significantly improves the accuracy, real-time performance and physical interpretability of industrial fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology for industrial equipment. Specifically, it relates to a method for diagnosing industrial equipment faults, a terminal device, and a computer-readable storage medium. Background Technology

[0002] In modern chemical, energy, and heavy equipment manufacturing processes, key equipment such as rotating machinery, reactors, and heat exchangers are commonly equipped with multimodal sensor arrays, continuously generating high-dimensional time-series data. This data naturally possesses graph structure attributes: equipment components form nodes, and material transfer pipelines, control signal lines, and mechanical transmission chains form edges. However, industrial sites have long faced the dilemma of extremely scarce data annotation, meaning that the proportion of fault samples is usually less than 5%, and the annotation process relies on senior engineers to conduct manual judgment based on operating parameters and fault phenomena, which is costly and time-consuming.

[0003] Traditional supervised learning methods heavily rely on large amounts of labeled data, making convergence difficult in industrial scenarios. While unsupervised methods such as autoencoders and principal component analysis can handle unlabeled data, they can only extract shallow statistical features and cannot characterize the complex topological relationships and dynamic evolution of equipment operation. In recent years, the introduction of graph neural networks and self-supervised learning techniques has provided new avenues for utilizing unlabeled industrial data. However, existing methods generally suffer from a fundamental flaw: the self-supervised pre-training process is severely disconnected from the physical topology, operating mechanisms, and dynamic conditions of industrial equipment. Specifically, the design of tasks such as node masking, graph structure reconstruction, and contrastive learning is entirely based on data-driven statistical assumptions, failing to embed physical constraints such as mechanical transmission torque, temperature-pressure coupling, and vibration kurtosis into the representation learning process. This results in learned feature representations that are mathematically discriminative but physically lack interpretability and consistency, making it difficult to accurately reflect the true health status of the equipment.

[0004] Therefore, how to fully integrate the physical constraints of industrial equipment into the self-supervised learning process without labeled data under the condition of extremely scarce labeled data, and obtain a dynamic equipment state representation that has both data discrimination ability and physical consistency, is a technical problem that urgently needs to be solved in the field of industrial intelligent operation and maintenance. Summary of the Invention

[0005] This invention aims to solve the technical problem in existing industrial equipment fault diagnosis methods where self-supervised representation learning without labeled data is disconnected from the physical mechanism of the equipment, making it difficult to obtain physically consistent equipment state representations under conditions of very little labeled data.

[0006] To achieve the aforementioned objectives, this invention provides a method for fault diagnosis of industrial equipment, comprising: acquiring sensor time-series data of key components of the industrial equipment; constructing a weighted topology graph that dynamically evolves with the operating state of the equipment based on the physical connection relationships between the equipment components and the data correlation of the sensor time-series data; wherein, the nodes of the weighted topology graph correspond to equipment components or sensor measurement points, and the edge weights of the weighted topology graph are dynamically determined by fusing physical connection weights and data correlation weights; and, based on the weighted topology graph, performing multi-task industrial feature extraction and analysis by fusing component health status indices, equipment topology integrity constraints, and fault mode discrimination constraints. The process involves optimizing the representation to obtain a state feature representation of industrial equipment. The component health status index is determined by the physical statistical characteristics of the sensor time-series data. The multi-task industrial feature extraction and representation optimization includes: performing node feature recovery based on the component health status index; performing adjacency relationship reconstruction based on the equipment topology integrity; and performing discriminative feature optimization based on the fault mode discrimination index to aggregate similar industrial state features and separate dissimilar industrial state features. A fault classification model is trained using the industrial equipment state feature representation and a small number of labeled fault samples. Real-time collected equipment sensor data is used to construct a weighted topology graph, which is then input into the fault classification model to obtain fault diagnosis results.

[0007] Preferably, the training of the fault classification model includes: constructing a tree-like hierarchical classifier, wherein the major category nodes of the tree-like hierarchical classifier correspond to the fault types of the equipment subsystem, the sub-category nodes of the tree-like hierarchical classifier correspond to the component-level fault types, and feature weights are dynamically allocated between different levels through a fault mode attention mechanism.

[0008] Preferably, the training fault classification model includes: embedding the physical law constraints of equipment operation as regularization terms into the loss function, and adopting a hierarchical parameter unfreezing strategy to gradually release pre-trained parameters from shallow to deep according to the correlation between the model layer and the fault diagnosis task.

[0009] Preferably, before constructing a weighted topology graph from the real-time collected device sensor data, the industrial equipment fault diagnosis method further includes: incrementally updating the weighted topology graph based on the rate of change of the sensor data and / or changes in the component connection relationships, recalculating only the weights of the changed nodes and their adjacent edges.

[0010] Preferably, obtaining the fault diagnosis result includes: obtaining the probability distribution of fault prediction through multiple random inactivation forward propagation, quantifying the random uncertainty and cognitive uncertainty of the diagnosis result, and obtaining an uncertainty index; when the uncertainty index exceeds a preset threshold, marking the fault diagnosis result as a low-confidence pending review state and triggering a manual review process; when the uncertainty index does not exceed the preset threshold, directly outputting the fault diagnosis result.

[0011] Preferably, the industrial equipment fault diagnosis method further includes: predicting the fault probability evolution trend of multiple future time steps based on time series graph features, calculating the fault deterioration index; and triggering hierarchical operation and maintenance early warning based on the fault deterioration index.

[0012] Preferably, before training the fault classification model using the state feature representation of the industrial equipment and a small number of labeled fault samples, the industrial equipment fault diagnosis method further includes: calculating a fault risk index based on the average repair time, average fault-free time, and economic loss coefficient of the fault type; and applying differential weighting to the small number of labeled samples based on the fault risk index, so that high-risk fault samples obtain higher optimization weights during training.

[0013] Preferably, after obtaining the fault diagnosis results, the industrial equipment fault diagnosis method further includes: using the manually confirmed online diagnostic feedback as a supervision signal, periodically recalculating the feature importance score and dynamically adjusting the feature enhancement coefficient, so that the fault classification model adapts to the long-term drift of equipment operating conditions.

[0014] The present invention also provides a terminal device, the terminal device including a processor and a memory connected to the processor, wherein the memory stores program instructions; the processor is used to execute the program instructions stored in the memory to implement the above-described industrial equipment fault diagnosis method.

[0015] The present invention also provides a computer-readable storage medium storing program instructions that, when executed, implement the above-described industrial equipment fault diagnosis method.

[0016] Beneficial Effects: This invention integrates physical connection weights and data relevance weights during the construction of a dynamically weighted topology graph, ensuring that the graph structure retains the physical essence of the equipment while capturing potential data-level relationships. Furthermore, it embeds component health status indices, topological integrity constraints, and fault mode discriminative constraints into the industrial feature extraction and representation optimization process using unlabeled data. This ensures that node feature recovery is constrained by physical statistical features, adjacency relationship reconstruction is constrained by equipment topological integrity, and discriminative optimization is constrained by industrial state semantics. Thus, even with extremely scarce labeled data, this invention learns equipment state representations that combine data discriminativity and physical consistency using a large amount of unlabeled data. Finally, it trains a fault classification model using a small number of labeled samples, achieving accurate diagnosis of industrial equipment faults. Therefore, this invention effectively solves the technical bottleneck of the disconnect between self-supervised representation learning with unlabeled data and industrial physical mechanisms, overcomes the limitations imposed by scarce labeled data on model performance, and significantly improves the accuracy and physical interpretability of industrial equipment fault diagnosis. Attached Figure Description

[0017] The above and other aspects, features, and advantages of embodiments of the present invention will become clearer from the following description taken in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of an industrial equipment fault diagnosis method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a terminal device according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer storage medium according to an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] As described in the background section, in existing industrial equipment fault diagnosis methods, the self-supervised representation learning process without labeled data fails to fully integrate the physical topological constraints and operating mechanism information of industrial equipment. Furthermore, the extraction and updating of node features by graph neural networks fails to deeply couple with the health status, topological integrity, and fault mode discrimination of equipment components. As a result, under conditions of extremely scarce labeled data, the equipment state representation learned by the model lacks physical consistency, leading to insufficient accuracy and reliability of fault diagnosis.

[0020] To address the aforementioned technical problems, an embodiment of the present invention provides a method for fault diagnosis of industrial equipment using unsupervised graph representation that integrates physical constraints. The main technical solution is as follows: a dynamic weighted topology graph integrating physical connection weights and data correlation weights is constructed; multi-task industrial feature extraction and representation optimization are performed on the dynamic graph based on component health status index, equipment topology integrity constraints, and fault mode discriminability constraints, thereby obtaining an industrial equipment state feature representation that combines data discriminability and physical consistency; a fault classification model is then trained using a small number of labeled fault samples; and finally, fault diagnosis is completed based on real-time graph data input.

[0021] The fault diagnosis method according to embodiments of the present invention can be divided into three main stages. The first stage is the unsupervised graphical representation learning stage, in which a physically consistent representation of the state features of industrial equipment is obtained by fusing multi-task feature extraction with physical constraints on a large amount of unlabeled industrial sensor data. The second stage is the supervised fine-tuning stage, in which the feature representation obtained in the first stage is fine-tuned using a small number of labeled fault samples to train a classification model for a specific fault diagnosis task. The third stage is the online fault diagnosis stage, in which the trained model is deployed in the industrial field to perform online inference on real-time collected equipment data, output fault diagnosis results, and trigger subsequent operation and maintenance responses.

[0022] The following describes in detail the industrial equipment fault diagnosis method according to an embodiment of the present invention with reference to the accompanying drawings. Figure 1 This is a flowchart of an industrial equipment fault diagnosis method according to an embodiment of the present invention.

[0023] Reference Figure 1 The industrial equipment fault diagnosis method according to an embodiment of the present invention includes the following steps: Step S1: Obtain sensor timing data for each key component of the industrial equipment.

[0024] In actual industrial settings, a multimodal sensor array is built to collect data from the target equipment. Taking a reaction vessel in a chemical plant as an example, two piezoelectric vibration sensors (model PCB352C65, range ±50g, sensitivity 10mV / g, sampling frequency set to 25.6kHz) are deployed at the top and bottom of the reaction vessel. Three platinum resistance temperature sensors (model PT100, range -50~200℃, accuracy ±0.1℃, sampling frequency 10Hz) are installed inside the vessel. Four pressure transmitters (model Rosemount 3051, range 0~1.6MPa, accuracy ±0.075% FS, sampling frequency 5Hz) are installed on the inlet and outlet pipes. One Hall current sensor (model ACS712, range 0~5A, sampling frequency 1kHz) is installed on the stirring motor. All sensors are connected to the data acquisition card via an RS-485 bus, such as the NI USB-6363 model acquisition card. The acquisition card communicates with the industrial computer via USB 3.0. The industrial computer can be configured with an Intel i7-10700K CPU, 32GB DDR4 memory, and an NVIDIA RTX 3090 GPU.

[0025] Data transmission can use the Modbus RTU protocol, with a baud rate of 9600bps, 8 data bits, 1 stop bit, and no parity check. The industrial computer runs a real-time data acquisition module, packaging data every 100ms to generate a triplet data frame containing a timestamp, sensor number, and measurement value.

[0026] Next, data collection and preprocessing will be carried out.

[0027] Let the data acquisition period be T. acp The unit is seconds. Time-series data from M sensors are acquired per cycle to form the original dataset X, where... In the formula, x m,t Let represent the measurement value of the m-th sensor at time t.

[0028] Noise suppression processing was performed on the collected raw data. For vibration signals, an improved wavelet threshold denoising algorithm was used.

[0029] Let the number of wavelet decomposition layers be J, and the threshold function be: ; In the formula, Let be the wavelet coefficient at position k in the j-th layer, where j = 1, 2, …, J; and k be the sequence index of the wavelet coefficients within that layer. Adaptive thresholding. The noise standard deviation σ and the signal length N sig Confirmed, the formula is as follows: .

[0030] In this example, a 6-level decomposition is performed using the db4 wavelet basis, i.e., J=6. Hard thresholding is applied to the high-frequency coefficients (levels 5-6), and soft thresholding is applied to the low-frequency coefficients (levels 1-4). The noise standard deviation σ is estimated using the wavelet coefficients of level 1, and the signal length N is... sig Take 8192.

[0031] Then, spatiotemporal alignment is performed. Asynchronous data from different sensors are synchronized in time, and missing values ​​are filled in using linear interpolation. Let the measurement values ​​of a certain sensor at times t1 and t2 be... and Then the interpolation result at time t is: .

[0032] For temperature and pressure signals, a sliding median filter is used with a window size of 5 to effectively eliminate spike interference.

[0033] Finally, feature extraction is performed. Time-domain and frequency-domain features are extracted from the vibration signal. Time-domain features include peak value, kurtosis, and root mean square (RMS), while frequency-domain features include dominant frequency and spectral centroid. These features are combined with the original values ​​of temperature and pressure to form a 128-dimensional nodal feature vector.

[0034] Step S2: Construct a weighted topology graph that dynamically evolves with the operating status of the equipment.

[0035] The target industrial system is divided into several nodes. Taking a reactor as an example, it can be divided into 15 nodes, including the reactor body, stirring motor, stirring shaft, inlet valve, outlet valve, temperature sensor group, pressure sensor group, heating device, cooling device, and control system. The key components of the equipment are represented as a set of nodes V in the graph. In the formula, N node The total number of nodes. Each node is assigned a unique number and physical attribute label, such as a rotating component or a static device.

[0036] When constructing an edge set, it's necessary to distinguish between physical connection edges and control relationship edges. Physical connection edges represent the mechanical or material transfer relationships between components, such as the connection between the stirring motor and the stirring shaft via a coupling, or the connection between the inlet valve and the cylinder via a pipe. Control relationship edges represent signal control relationships, such as the transmission of control commands between the control system and the heating device. In this example, there are 18 physical connection edges and 7 control relationship edges.

[0037] The calculation of edge weights integrates two aspects: physical connection strength and data correlation. Physical connection weight refers to the edge weight assigned based on the actual physical coupling strength between components; for example, the greater the torque transmission between mechanical drive shafts, the higher the weight. Data correlation weight refers to the edge weight assigned based on the statistical correlation between sensor measurement data.

[0038] Specifically, physical connection weights Determined by the torque transmitted between components, the calculation formula is as follows: In the formula, T i,j For the torque transmitted between nodes i and j, the denominator max(T) i,j This represents the maximum torque transmitted between all component pairs. Data-related weights. The maximum information coefficient (MIC) is used for calculation, and the formula is as follows: In the formula, f i and f j Let be the feature vectors of nodes i and j. The final edge weights are determined by a dynamic fusion of physical weights and data weights, using the following formula: .

[0039] In the formula, α is the balance coefficient, which ranges from 0 to 1. For mechanical transmission systems, it can be 0.7; for rotating parts, it can be 0.8; and for static devices, it can be 0.5.

[0040] The graph data is stored in an adjacency list format. Each graph sample contains a node feature matrix, an adjacency matrix, a timestamp, and a label indicating the equipment's operating condition, such as normal or heating phase.

[0041] To achieve the dynamic evolution of the graph structure, a time-sliding window graph sequence G is constructed. t Its expression is: In the formula, E t Let t be the edge set of the t-th window. The edge weight matrix corresponding to this window is set to S sampling periods. The dynamic changes in the device state are captured by a sliding window.

[0042] During equipment operation, the graph structure is not static. When the rate of change in sensor data exceeds a set threshold, or when component connections change, the weighted topology graph needs to be incrementally updated. This involves recalculating only the weights of the changed nodes and their adjacent edges, without reconstructing the entire graph. The real-time update formula for edge weights is: In the formula, β(t) is the dynamic weighting coefficient, and its calculation formula is: , where σ x The standard deviation is based on historical data and is adaptively adjusted as node characteristics change. In this example, when the rate of change of a sensor value exceeds 10% / s, or when a new sensor is added, a graph structure update is immediately triggered. The update uses an differential calculation method, recalculating only the weights of the changed nodes and their adjacent edges, with the update time controlled within 20ms. Simultaneously, a full edge weight update is performed every 5 minutes to ensure the overall consistency of the graph structure.

[0043] Step S3: Perform multi-task industrial feature extraction and representation optimization based on the weighted topology graph to obtain the state feature representation of industrial equipment.

[0044] Among them, the industrial equipment status feature representation refers to a low-dimensional vector obtained through multi-task learning. This vector encodes the health status information and topological association information of the equipment components, and can simultaneously reflect the discriminativeness at the data level and the consistency at the physical level.

[0045] The component health status index is a scalar indicator calculated from the physical statistical characteristics of sensor time-series data, used to quantify the health status of equipment components. For example, for vibration-related nodes, the health status index is calculated using the kurtosis of the vibration signal, with the following formula: In the formula, x i Let μ be the vibration signal sequence of node i. i Let be the mean value of the vibration signal at node i. This represents the expected value calculation. For temperature or pressure-related nodes, the health status index is calculated using a deviation-standardized formula, i.e. In the formula, σ is the average value under normal operating conditions. i This represents the standard deviation under normal operating conditions.

[0046] This step performs three interrelated feature learning tasks simultaneously within the same training framework.

[0047] The first task is node feature recovery, which is performed based on the component health status index. During training, sensor failure is simulated, and 20% of the nodes are randomly selected for feature occlusion. The occluded features of node i are contaminated with zero-mean Gaussian noise, i.e. In the formula, f i Let be the original feature vector of node i, and σ be set according to the noise level of historical fault data. The model needs to predict both the original feature values ​​and the health status index of the node simultaneously. The loss function for the node recovery task uses the weighted mean squared error, and the formula is: In the formula, β is the weighting coefficient, which is taken as 0.7 in this example. and These are the predicted values.

[0048] The second task is adjacency reconstruction, which is performed based on the integrity of the device topology. This task requires that after the input graph undergoes encoding and decoding, the output reconstructed adjacency matrix be as close as possible to the original adjacency matrix, thereby forcing the model to maintain its memory of the device's physical topology. The loss function used is binary cross-entropy, with the following formula: In the formula, A i,j The element in the i-th row and j-th column of the original adjacency matrix. To reconstruct the elements at corresponding positions in the adjacency matrix.

[0049] Meanwhile, based on the prediction of future graph structure evolution from historical graph sequences, the loss function is expressed as: In the formula, T represents the Frobenius norm. p To predict the step size, For the first The true edge weight matrix at time step (k) This is the predicted value of the edge weight matrix at that moment.

[0050] The third task is discriminative feature optimization, which performs feature aggregation of similar industrial states and feature separation of dissimilar industrial states based on fault mode discrimination. The aggregation of similar industrial state features and the separation of dissimilar industrial state features are essentially an industrial semantic contrastive learning mechanism. This mechanism generates positive samples by applying small perturbations to the original graph and generates negative samples by disrupting the graph structure, enabling the model to learn feature representations with fault discrimination capabilities. Specifically, small perturbations are applied to the original graph G to generate positive samples G0. + For example, adding ±5% Gaussian noise to the edge weights. Structurally disrupting the graph structure (e.g., randomly shuffling 30% of the edge connections) generates negative samples G. - The loss function adopts the InfoNCE (Information-Noise Contrast Estimation) form, and the formula is as follows: In the formula, These are the feature representations of the anchor sample, the positive sample, and the k-th negative sample, respectively. For cosine similarity, For temperature parameters, N neg This represents the number of negative samples.

[0051] The total loss function for the joint training of the three tasks is the weighted sum of the losses from each task, as shown in the following formula: , In the formula, the sum of the weights γ1, γ2, γ3, and γ4 is 1. In this example, γ2 is taken as 0.4.

[0052] This step uses a spatiotemporal graph convolutional network as the feature extraction backbone. The spatial convolutional layer employs an improved Chebyshev multinomial convolution, supporting weighted adjacency matrix input. The convolution formula is as follows: , In the formula, T k Let K be a Chebyshev polynomial of order k. poly It is the highest order of the polynomial. The normalized Laplace matrix is ​​calculated using the following formula: In the formula, L is the Laplace matrix. D is the degree matrix. Let I be the largest eigenvalue and I be the identity matrix. For convolution kernel parameters, Let be the feature matrix of the l-th layer at time t.

[0053] Temporal convolutional layers use one-dimensional convolutional kernels to capture temporal dependencies, as shown in the following formula: , In this example, the one-dimensional convolution kernel size c ker We set a value of 3, a stride of 1, and padding of 1 to maintain the temporal length. Residual connections are added to each spatiotemporal convolutional block using the following formula: In the formula, The final feature after residual connection. These are intermediate features after convolution transformation. This is an identity mapping of the input from the previous time step, to alleviate the gradient vanishing problem in deep networks.

[0054] The network also includes an attention enhancement mechanism. The node importance weight α i The calculation formula is as follows: , In the formula, u is the learnable attention vector of the node. h is the node attention transformation matrix. i Let be the hidden features of node i.

[0055] Time step weight β t The calculation formula is: In the formula, v is the temporal attention learnable vector, U is the temporal attention transformation matrix, and H is the temporal attention transformation matrix. t Let be the graph feature matrix at time step t.

[0056] The final graph representation of Z is the weighted sum of the node features at each time step according to the attention weights, and its calculation formula is: .

[0057] Before model training, the Xavier initialization method is used to initialize the network weights. For the weight matrix... Its elements follow a uniform distribution, that is: In the formula, n in and n out These represent the number of input and output neurons, respectively. The bias parameters are initialized to 0.

[0058] The optimizer uses the momentum-driven Adam optimizer, with momentum parameters β1 set to 0.9 and β2 set to 0.999. The initial learning rate η is set to 0.001, and a cosine annealing strategy is used to adjust the learning rate. The learning rate adjustment formula is as follows: , In the formula, ηmin Take 10 -5 η max Take 0.001, T max t represents the total number of training rounds, and t represents the current round number.

[0059] During training, the graph sequence data is divided into training and validation sets in an 8:2 ratio. An early stopping strategy is employed to prevent overfitting, and a patience value of 10 is used. In each iteration, B graph sequence samples are randomly selected, with B set to 32, and the multi-task joint loss is calculated. The network parameters are updated through backpropagation. The total number of training epochs T max Set the value to 200, and save the model parameters every 10 rounds.

[0060] Step S4: Train a fault classification model using the state feature representation of industrial equipment and a small number of labeled fault samples.

[0061] Before commencing formal training, a small number of labeled samples were prepared using risk-weighted methods. In this example, eight typical fault types were selected from historical fault records, including agitator shaft imbalance, seal leakage, and heating element failure. 30-50 samples were collected for each fault type, along with 100 normal samples, for a total of 500 labeled samples. Sample labeling was conducted independently by two senior engineers, including fault type, occurrence time, and severity. An expert review meeting was organized to determine the final labels for dissenting samples.

[0062] To address the scarcity of labeled data, virtual samples are generated based on the failure mechanisms of industrial equipment for data augmentation. Taking gearbox tooth surface wear failure as an example, the vibration signal of the original failure sample is s(t), and the generated sample is: In the formula, k is the amplitude adjustment coefficient, ranging from 0.8 to 1.2. Δt is the time offset, ranging from -0.01s to 0.01s, used to simulate the time difference of fault development. Time-domain scaling and frequency-domain shifting are performed on the vibration fault samples to expand the number of samples for each type of fault to 80.

[0063] Furthermore, the Mean Time To Repair (MTTR) is determined based on the type of fault. c Mean Time Between Failures (MTBF) c and the economic loss coefficient c e Calculate the failure risk index r c The formula is: In the formula, MTTR c Mean Time To Repair (MTBF) for fault type c c c is the mean time between failures (MTBF). e This represents the economic loss coefficient caused by the malfunction, expressed in ten thousand yuan per hour. For example, the risk index for a heating element malfunction is 0.8, and the risk index for a minor valve leak is 0.3. Sample weights. The allocation is based on a risk index normalization formula: Through this differentiated weighting mechanism, high-risk fault samples receive higher weights in the training loss, thus enabling the model to prioritize the identification of high-risk faults during the optimization process.

[0064] In addition, the Isolation Forest algorithm is used to identify outliers in the labeled data. The formula for calculating the outlier score in Isolation Forest is: In the formula, h(x) is the path length of sample x, E[h(x)] is the expected path length, c(n) is the normalization constant, and n is the number of samples in the sample subset of the isolated forest. The calculation formula is as follows: In the formula, γ≈0.5772, which is Euler's constant. The outlier score threshold is set to 0.7, and samples with scores exceeding the threshold are removed.

[0065] Next, a dynamic classifier is constructed. The fault classification model adopts a tree-like hierarchical classifier structure to match the natural hierarchical characteristics of industrial faults. Let the number of fault levels be H. cls The lth layer has C l If there are 1 fault category, then the total parameters of the classifier are: In the formula, d l The input feature dimension for layer l is defined. In this example, the first-level classification has three main categories: mechanical faults, thermal faults, and control faults. The second-level classification has eight subcategories; for example, mechanical faults are further subdivided into agitator shaft imbalance, bearing wear, etc. Each level uses a two-layer multilayer perceptron structure with an input dimension of 256, a hidden layer dimension of 128, and an output layer dimension equal to the number of categories in that layer. The activation function is a combination of ReLU and Softmax.

[0066] A fault mode attention mechanism is added to the classifier output layer. This mechanism trains dedicated attention vectors for different fault categories and optimizes them using gradient descent to give higher weights to key feature dimensions relevant to that fault category. For the c-th fault category, its classification score... The calculation formula is: In the formula, D is the total dimension of the input feature vector, and z k Let k be the k-th dimension of the input feature. Attention weights. The calculation formula is: In the formula, Let be the attention parameter vector for the c-th type of fault.

[0067] Simultaneously, an uncertainty quantification mechanism is introduced. Dropout layers are added to the two hidden layers of the classifier, with a dropout rate of 0.1. N... mc The next random inactivation forward propagation, in this example N mc Taking 20, we obtain the predicted distribution. Uncertainty index u c The calculation formula is as follows: , In the formula, the first term represents random uncertainty, reflecting the prediction fluctuations caused by data noise; the second term represents cognitive uncertainty, reflecting the degree of knowledge gaps in the model regarding unknown failure modes.

[0068] Subsequently, tiered fine-tuning and physical constraint regularization were performed. A tiered parameter unfreezing strategy was adopted, dividing the pre-trained model into a low-level feature extraction layer, a mid-level structure perception layer, and a high-level task adaptation layer. The tiered parameter unfreezing strategy refers to gradually releasing pre-trained parameters according to the correlation between model layers and the fault diagnosis task, rather than releasing all parameters at once, to avoid catastrophic forgetting with limited labeled data. Specifically, in the initial stage, only the parameters of the high-level task adaptation layer were unfrozen, with a learning rate of 5×10⁻⁶. -5 When the validation set accuracy improves by less than 0.5% for five consecutive rounds, the parameters of the middle and top layers are unfrozen, and the learning rate is reduced to 3×10. -5 When the validation set accuracy improves by less than 0.5% for five consecutive rounds, all parameters are unfrozen, and the learning rate is further reduced to 10. -5 .

[0069] In the fine-tuning loss function, physical constraints related to equipment operation are introduced as regularization terms. Taking rotating machinery as an example, the correlation between vibration characteristics and rotational speed is constrained, and the calculation formula for the regularization loss term R is as follows: , In the formula, The vibration characteristics predicted by the model are used as rotational speed. The function f i For measured features, Take 0.01.

[0070] Taking a reactor as an example, constraints are added to the relationship between temperature and pressure. Under normal operating conditions, pressure P and temperature T satisfy a linear relationship: P = kT + b. This linearity coefficient is determined through fitting normal samples. The regularization loss term is: .

[0071] To constrain the relationship between the current I and the speed n of the stirring motor, under normal operating conditions, the current is proportional to the speed, and the regularization loss term is: , This represents the average value of the stirring motor current I across all samples under normal operating conditions. This represents the average motor speed n across all normal operating condition samples.

[0072] The fine-tuning phase employs a hybrid loss function, with a total loss of [missing information]. The calculation formula is as follows: , where λ ce This represents the balancing weight coefficient.

[0073] Cross-entropy loss Using a weighted form, the formula is: , where N l To label the total sample size, C f This represents the total number of fault categories. For the true label of sample i, To predict probabilities, The weights are for the failure risk index.

[0074] Comparative loss The calculation formula is as follows: , In the formula, s i,j =1 indicates that samples i and j are of the same type of fault; m is 1.0, which is a marginal parameter; z i and z j This represents the features of the sample.

[0075] The fine-tuning process uses a piecewise cosine annealing learning rate. The initial learning rate η0 is set to 5 × 10⁻⁶. -5 In the first 30% of the fine-tuning rounds, the temperature is linearly increased to η. max =2×10 -4 The cosine cooling occurs in the last 70% of the cycles.

[0076] The learning rate formula is: In the formula, T fine To fine-tune the total number of rounds, η min Take 5×10 -6 .

[0077] Furthermore, the fine-tuning process employs an early stopping strategy, with a patience value of 8. The N_{ens} models with the highest validation accuracy are saved; in this example, N... ens Taking 3, the final prediction result is a weighted fusion of the outputs of each model, and the formula is: In the formula, the model weight α k The calculation formula is: .

[0078] Among them, acc k Let be the validation accuracy of the k-th model. In this example, the validation accuracies of the three models are 93.2%, 92.8%, and 92.5%, respectively. After fusion with weights of 0.4:0.3:0.3, the accuracy is improved to 94.1%.

[0079] Step S5: Construct a weighted topology graph from the real-time collected device sensor data and input it into the fault classification model to obtain fault diagnosis results.

[0080] The trained fault classification model is deployed in industrial settings. The hardware architecture employs an edge computing and cloud-based collaborative model. At the edge, an industrial control computer is deployed, configured with an Intel i5-11400 CPU, 16GB of RAM, and an NVIDIA T4 GPU, running a real-time data processing program. This program completes data preprocessing and graph construction every 500ms, then calls the inference engine for forward computation, with single-sample inference time controlled within 280ms, meeting industrial real-time requirements. A server cluster is deployed in the cloud, responsible for model updates and historical data analysis, automatically synchronizing diagnostic results and human feedback from the edge every morning.

[0081] During real-time operation, let the real-time data stream be X. real (t), which is expressed as follows: , In the formula, t is a continuous timestamp.

[0082] The data is filtered in real time, and a moving average filter is used to eliminate high-frequency noise. The formula is as follows: In the formula, N win Let 5 be the size of the filter window, and Δt be the sampling interval.

[0083] The weighted topology graph is incrementally updated based on the rate of change of sensor data and changes in component connections. When the rate of change of a sensor value exceeds 10% / s, or when a new sensor is added, a graph structure update is immediately triggered. The update uses a differential calculation method, recalculating only the weights of the changed nodes and their adjacent edges, with update time controlled within 20ms. Simultaneously, a full edge weight update is performed every 5 minutes to ensure the overall consistency of the graph structure.

[0084] The time series diagram was generated using an overlapping sliding window, with a window step size S set to 0.5T. win T win The window length is used to ensure the continuity of the graph sequence.

[0085] Real-time acquired device sensor data is preprocessed to construct a weighted topology graph, which is then input into the fault classification model. The input graph data undergoes a spatiotemporal graph convolutional network to extract 256-dimensional features. Cross-window feature fusion is performed between adjacent windows to mitigate boundary effects. Let z be the features of windows t and t+1. t and z t+1 , fusion feature z fusion The calculation formula is: The formula for calculating the fusion coefficient γ is as follows: Where IoU is the intersection-union ratio of a graph structure, defined as the size of the intersection of two window edge sets divided by the size of their union, i.e.: .

[0086] Fault-sensitive features are enhanced through an attention mechanism. Feature importance scores (s) are calculated. k It is represented as: In the formula, To classify and output the partial derivative with respect to the k-th dimension feature, C f Let D be the total number of fault categories, and D be the total dimension of the features. Enhanced features. The calculation formula is as follows: , Where, λ s We set 0.5 as the strengthening coefficient.

[0087] The initial fault type prediction result is obtained. The probability distribution of the fault prediction is acquired through multiple random inactivation forward propagations, quantifying the random and cognitive uncertainties of the diagnostic results, and obtaining the total uncertainty index u. c The so-called multiple random inactivation forward propagation refers to randomly masking the hidden layer neurons of the classifier multiple times at a preset inactivation rate during the inference phase, calculating multiple sets of prediction results for each, and then statistically analyzing their distribution characteristics. When the uncertainty index exceeds the preset threshold of 0.08, the diagnostic result is marked as low confidence and awaiting review, triggering a manual review process where on-site engineers confirm the fault type. When the uncertainty index is below the threshold and the maximum prediction probability exceeds 0.85, the diagnostic result is output directly.

[0088] Predicting the evolution trend of failure probability at multiple future time steps based on time series graph features. A Long Short-Term Memory (LSTM) network is used to process the most recent L... seq The feature sequence of a window, in this example L seq Take 10 and predict the future T. p = Failure probability at 5 time steps. The prediction formula is: .

[0089] in, =1, 2,…, T p The Long Short-Term Memory (LSTM) network has 128 hidden layer dimensions, 2 layers, and a Dropout rate of 0.2.

[0090] Calculate the fault deterioration index d c The formula is: .

[0091] When d c Greater than the threshold θ d An alert is triggered at a certain time. In this example, when d... c A level 3 alert is triggered when the value is >0.06, prompting maintenance personnel to pay attention. When d... c A level 2 warning is triggered when the value is >0.1, and maintenance is prepared. When d... cA level 1 warning is triggered when the value exceeds 0.2, and an emergency shutdown and inspection are recommended.

[0092] Furthermore, by integrating classification results, uncertainty indicators, and trend prediction information, Dempster-Shafer evidence theory is employed for decision-level fusion. Let the basic probability distributions of each evidence source be m1, m2, and m3, and the fusion rule be: , where K ds The conflict coefficient is calculated using the following formula: Finally, the fault type with the highest confidence level is selected as the diagnostic result. Here, A represents the final hypothesis after fusion, that is, the fault type output by the diagnosis; A1 represents the hypothesis supported by the first source of evidence (such as the fault classifier); A2 represents the hypothesis supported by the second source of evidence (such as the uncertainty quantification module); and A3 represents the hypothesis supported by the third source of evidence (such as the trend prediction module).

[0093] The diagnostic results are visualized through an industrial monitoring interface. The interface includes a device topology diagram, with node colors gradually changing from green to yellow to red based on health status. It also includes feature trend graphs, displaying real-time and predicted curves for key features such as vibration kurtosis and temperature deviation. The system automatically generates a diagnostic report, including the fault type, confidence level, and recommended actions.

[0094] Establish a diagnostic accuracy feedback mechanism. Engineers use an interface to mark diagnostic results as correct, incorrect, or suspected. The system uses human confirmation feedback as a supervisory signal, recalculates feature importance scores weekly, and dynamically adjusts the feature enhancement coefficient λ. s The formula is adjusted as follows: .

[0095] Where η is 0.1, and is an adjustment coefficient, acc recent This represents the accuracy of the most recent 100 diagnoses. This mechanism enables the fault classification model to adapt to long-term drift in equipment operating conditions and maintain stable performance throughout the equipment's service life.

[0096] Through the steps described above, this invention achieves high-precision and real-time fault diagnosis in industrial scenarios where labeled data is scarce. A three-month validation study on a reactor in a chemical plant showed that the utilization rate of unlabeled data reached 95%, only 500 labeled data samples were required, the average diagnostic accuracy reached 94.3%, the high-risk fault identification rate reached 98.7%, the false alarm rate was less than 1.2%, the single-sample processing time was 280ms, it ran continuously for 90 days without failure, and the model drift was less than 3%.

[0097] In summary, this invention first dynamically integrates the physical connection strength of devices with the correlation of sensor data at the graph construction level, enabling the graph structure to retain the physical essence of the industrial system while evolving in real time with operating conditions. Second, in the representation learning stage without labeled data, component health status index, topological integrity constraints, and fault mode discrimination constraints are uniformly embedded into three collaborative tasks: node recovery, adjacency reconstruction, and discriminative optimization. This ensures that the device state representation learned by the model is mathematically discriminative while physically conforming to the objective laws of device operation. Finally, layered fine-tuning is performed by combining a small number of labeled samples with physical regularization, and the trained model is deployed on edge computing nodes to perform online inference and incremental updates.

[0098] Therefore, in industrial scenarios where labeled data is extremely scarce, this invention effectively solves the technical bottleneck of the disconnect between unlabeled self-supervised representation learning and equipment physical mechanisms, enabling the model to learn dynamic equipment state representations that combine data discriminability and physical consistency using massive amounts of unlabeled data, thereby achieving accurate industrial fault diagnosis.

[0099] To implement the industrial equipment fault diagnosis method in the above embodiments, this application also provides a terminal device 300, for details please refer to [link to details]. Figure 2 The terminal device 300 in this application embodiment includes a processor 31, a memory 32, an input / output device 33, and a bus 34.

[0100] The processor 31, memory 32, and input / output device 33 are respectively connected to the bus 34. The memory 32 stores program data, and the processor 31 is used to execute the program data to implement the industrial equipment fault diagnosis method described in the above embodiments.

[0101] In this embodiment, processor 31 can also be referred to as a CPU (Central Processing Unit). Processor 31 may be an integrated voltage control system chip with signal processing capabilities. Processor 31 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor, or processor 31 can be any conventional processor.

[0102] This application also provides a computer storage medium; please refer to the following: Figure 3 , Figure 3 This is a schematic diagram of a computer storage medium according to an embodiment of the present application. The computer storage medium 40 stores program data 41, which is used to implement the industrial equipment fault diagnosis method of the above embodiment when executed by the processor.

[0103] When the embodiments of this application are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0104] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for diagnosing faults in industrial equipment, characterized in that, include: Acquire sensor timing data for key components of industrial equipment; Based on the physical connection relationships between device components and the data correlation of the sensor time-series data, a weighted topology graph that dynamically evolves with the device operating status is constructed; wherein, the nodes of the weighted topology graph correspond to device components or sensor measurement points, and the edge weights of the weighted topology graph are dynamically determined by the fusion of physical connection weights and data correlation weights; Based on the weighted topology graph, the component health status index, equipment topology integrity constraints, and fault mode discrimination constraints are integrated to perform multi-task industrial feature extraction and representation optimization on the weighted topology graph to obtain an industrial equipment state feature representation. The component health status index is determined by the physical statistical characteristics of the sensor time-series data. The multi-task industrial feature extraction and representation optimization includes: node feature recovery based on the component health status index; adjacency relationship reconstruction based on the equipment topology integrity; and discriminative feature optimization based on the fault mode discrimination to aggregate similar industrial state features and separate dissimilar industrial state features. A fault classification model is trained using the state feature representation of the industrial equipment and a small number of labeled fault samples. The real-time collected device sensor data is used to construct a weighted topology graph, which is then input into the fault classification model to obtain fault diagnosis results.

2. The industrial equipment fault diagnosis method according to claim 1, characterized in that, The trained fault classification model includes: A tree-structured hierarchical classifier is constructed, where the major category nodes of the tree-structured hierarchical classifier correspond to the fault types of the device subsystems, and the sub-category nodes of the tree-structured hierarchical classifier correspond to the fault types of the components. Feature weights are dynamically allocated between different levels through a fault mode attention mechanism.

3. The industrial equipment fault diagnosis method according to claim 1, characterized in that, The trained fault classification model includes: The physical laws governing equipment operation are used as regularization terms embedded in the loss function, and a hierarchical parameter unfreezing strategy is adopted to gradually release pre-trained parameters from shallow to deep according to the correlation between the model layer and the fault diagnosis task.

4. The industrial equipment fault diagnosis method according to claim 1, characterized in that, Before constructing a weighted topology map from the real-time collected device sensor data, the industrial equipment fault diagnosis method further includes: Based on the rate of change of sensor data and / or changes in component connection relationships, the weighted topology graph is incrementally updated, and only the weights of the changed nodes and their adjacent edges are recalculated.

5. The industrial equipment fault diagnosis method according to claim 1, characterized in that, The process of obtaining fault diagnosis results includes: The probability distribution of fault prediction is obtained by multiple random inactivation forward propagations, the random uncertainty and cognitive uncertainty of the diagnostic results are quantified, and uncertainty indicators are obtained. When the uncertainty index exceeds the preset threshold, the fault diagnosis result is marked as low confidence pending review and a manual review process is triggered; when the uncertainty index does not exceed the preset threshold, the fault diagnosis result is directly output.

6. The industrial equipment fault diagnosis method according to claim 1 or 5, characterized in that, The industrial equipment fault diagnosis method also includes: Based on the characteristics of time series graphs, predict the evolution trend of failure probability at multiple future time steps and calculate the failure deterioration index; The fault deterioration index triggers a tiered maintenance early warning system.

7. The industrial equipment fault diagnosis method according to claim 1, characterized in that, Before training a fault classification model using the state feature representation of the industrial equipment and a small number of labeled fault samples, the industrial equipment fault diagnosis method further includes: The failure risk index is calculated based on the average repair time, average time without failure, and economic loss coefficient of the failure type. The labeled samples are differentially weighted based on the fault risk index, so that high-risk fault samples can obtain higher optimization weights during training.

8. The industrial equipment fault diagnosis method according to claim 1, characterized in that, After obtaining the fault diagnosis results, the industrial equipment fault diagnosis method further includes: The online diagnostic feedback confirmed by humans is used as a supervision signal. The feature importance score is periodically recalculated and the feature enhancement coefficient is dynamically adjusted so that the fault classification model can adapt to the long-term drift of equipment operating conditions.

9. A terminal device, characterized in that, The terminal device includes a processor and a memory connected to the processor, wherein... The memory stores program instructions; The processor is used to execute program instructions stored in the memory to implement the industrial equipment fault diagnosis method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed, implement the industrial equipment fault diagnosis method as described in any one of claims 1 to 8.