A neural network-based parameter correlation fault diagnosis method and system

Abstract

The application relates to the technical field of fault diagnosis, and discloses a parameter correlation fault diagnosis method and system based on a neural network, aiming to solve the problem that a traditional diagnosis method relies on a single threshold and is difficult to capture implicit compound faults. The method models through space-time alignment of multi-source sensing data and a graph neural network, constructs a dynamic parameter correlation graph, extracts node embedding by training an encoder using contrastive learning, identifies an abnormal propagation path and locates a fault source through topological difference, and outputs the type, position and confidence in combination with a fault template library matching. The system comprises sensing acquisition, edge preprocessing, graph construction, graph coding, difference measurement, fault location, pattern matching and decision output modules, supports multi-fault concurrent diagnosis, self-calibration updating and sensor fault-tolerant compensation. The application realizes early and accurate identification of implicit faults, reduces labeling dependence, improves diagnosis robustness and maintenance decision efficiency, and adapts to equipment aging and complex working condition evolution.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and specifically to a parameter correlation fault diagnosis method and system based on neural networks. Background Technology

[0002] With the continuous improvement of the intelligence and automation of industrial equipment, the importance of fault diagnosis technology in ensuring stable system operation and reducing maintenance costs is becoming increasingly prominent. Traditional fault diagnosis methods mostly rely on expert experience or statistical models, judging the operating status of equipment through preset thresholds and fixed logic. Their core principle is based on the induction of static patterns of historical faults. However, modern complex systems exhibit highly nonlinear, strongly coupled, and dynamically correlated parameters during operation: there are implicit interactions between the parameters of various sensors inside the equipment, and external operating condition disturbances and load fluctuations can trigger real-time evolution of the parameter correlation structure. Static models, unable to capture the dynamic coupling mechanism between parameters, often lead to false alarms, missed alarms, or diagnostic lags, exacerbating the accumulation of hidden damage to equipment, control strategy mismatch, and other risks, significantly weakening system reliability. In addition, users have increasingly stringent requirements for diagnostic accuracy, response speed, and cross-scenario generalization capabilities (such as sudden fault early warning and multi-condition adaptive diagnosis). The rigid criterion system of traditional methods is difficult to meet the needs of high-dimensional parameter correlation modeling and real-time inference in complex industrial scenarios. Summary of the Invention

[0003] This invention provides a parameter correlation fault diagnosis method and system based on neural networks. It constructs a spatiotemporal alignment framework for multi-source heterogeneous sensor data to achieve high-density synchronous acquisition and structured characterization of key physical parameters during equipment operation. Based on this, a graph neural network modeling mechanism for nonlinear dynamic coupling relationships between parameters is established. Using historical data under normal operating conditions as training samples, the system learns the correlation strength distribution and propagation path of each parameter node under steady-state operating conditions. When real-time operating data is input, the system calculates the topological difference between the current parameter correlation graph and the baseline correlation graph to identify abnormal propagation paths and deviating nodes, thereby locating potential fault sources. Simultaneously, a fault mode matching engine is introduced to compare the identified abnormal correlation patterns with a pre-built fault feature library, outputting the fault type, location, and confidence score, ultimately forming a closed-loop diagnostic decision output.

[0004] As one embodiment of the present invention, the method includes: a multimodal sensor array deployed at key measuring points of industrial equipment, used to synchronously collect five types of physical parameters during equipment operation: vibration, temperature, pressure, current, and rotational speed. The sampling frequency is uniformly set to 1000 Hz, and the timestamp accuracy is not less than 1 microsecond. The collected raw data undergoes preprocessing operations via edge computing nodes. The preprocessing operations include zero-point drift correction, outlier removal, sliding window mean filtering, and normalization mapping to form standardized time-series data blocks. Each data block contains 500 consecutive sampling points, constituting a basic input unit for a diagnostic cycle. The standardized time-series data blocks are then fed into a parameter correlation graph construction module. Each physical parameter is used as a graph node, and the mutual information value between any two nodes within a sliding time window is used as the edge weight to construct a dynamically weighted undirected graph. The graph structure is updated in real time as the time window slides, with an update step size of 50 sampling points. The dynamically weighted undirected graph is input to a graph neural network encoder, which adopts a three-layer graph convolutional layer stacked structure. Each layer includes adjacency matrix normalization, linear transformation of node features, and nonlinear activation function operation, finally outputting the hidden layer embedding vector of each node, with a dimension of 128. The hidden layer embedding vector is sent to the association difference measurement module, which calculates the cosine distance matrix between the current graph structure and the reference graph structure in the node embedding space. For nodes in the matrix that exceed a preset threshold, the module performs a cosine distance measurement. Edges corresponding to elements with a value of 0.3 are marked to form an abnormal associated edge set. This abnormal associated edge set is input to the fault source localization module. This module calculates the fault contribution of each node based on the positional relationship of the nodes connected by the abnormal edges in the physical topology of the equipment, combined with the direction and magnitude of edge weight changes, using a weighted centrality algorithm. The node with the highest contribution is identified as a candidate fault source. The candidate fault source and its associated abnormal patterns are sent to the fault pattern matching engine. This engine has a pre-stored feature template library covering five typical faults: bearing wear, gear tooth breakage, shaft misalignment, lubrication failure, and motor inter-turn short circuit. Each template consists of a typical topological structure and edge weights of the parameter association graph under the corresponding fault state. The distribution range and node embedding vector cluster centers constitute the system. The matching process uses the structure-preserving embedding distance metric to calculate the similarity score between the current abnormal pattern and each template. The fault type corresponding to the highest score is the diagnosis result, and the confidence level of the result is also output. The confidence level is defined as the ratio of the embedding space distance between the current pattern and the best matching template to the distance of the second best template. When the ratio is less than 0.5, it is judged as a high-confidence diagnosis. The diagnosis result is output through the human-computer interaction interface, including the fault type name, the location coordinates, the confidence level value, and the suggested handling measures. At the same time, the equipment maintenance work order generation module is triggered to automatically create a maintenance task record including the diagnosis time, equipment number, fault description, and priority identifier.

[0005] In one embodiment of the present invention, the system includes: a multimodal sensing data acquisition unit, which consists of a vibration sensor, an infrared temperature sensor, a piezoresistive pressure sensor, a Hall current sensor, and a photoelectric encoder. Each sensor is connected to an edge computing node via an industrial Ethernet bus. The data acquisition cycle is uniformly triggered by a hardware timer to ensure that the data from each channel is strictly aligned on the time axis; an edge preprocessing unit, which is embedded in the firmware layer of the edge computing node, performs data cleaning and format conversion operations, and outputs standardized data blocks that conform to the neural network input specifications; a parameter correlation graph construction unit, which runs on an embedded graphics processor, uses a sliding window mechanism to dynamically calculate the mutual information between parameters, and constructs a weighted graph structure that is updated in real time; and a graph neural network encoding unit, which is deployed on the edge computing node. The system includes a field-programmable gate array (FPGA) accelerator card that enables parallel processing of graph convolution operations, capable of encoding 20 graph structures per second; an association difference measurement unit, implemented on a general-purpose CPU core, that performs batch calculations of the cosine distance matrix and threshold filtering; a fault source localization unit that uses an improved betweenness centrality algorithm, combined with prior knowledge of the equipment's mechanical structure, to perform spatial weighted aggregation of the abnormal edge set and output the fault source coordinates; a fault mode matching unit that runs on a cloud server, utilizing a distributed vector database to store fault feature templates and supporting millisecond-level similarity retrieval; and a diagnostic decision output unit that is integrated into the human-machine interface of the equipment monitoring terminal, presenting diagnostic results as a visual topology graph overlaid with anomaly markers, and supporting the automatic generation and export of diagnostic reports.

[0006] As one embodiment of the present invention, the training process of the graph neural network encoder includes: collecting historical data of the device running continuously for 30 days in a fault-free state, dividing the training samples according to the diagnostic cycle, each sample containing a complete parameter association graph and its corresponding node embedding label; the node embedding label is generated through a self-supervised contrastive learning strategy, specifically by performing data augmentation on the normal graph structure of the same device under different time windows, the augmentation methods including random edge discarding, node feature perturbation and time axis resampling, to construct positive sample pairs; negative sample pairs are composed of abnormal graph structures collected from different devices or the same device in fault injection experiments; the loss function adopts a contrastive loss function, the objective of which is to maximize the similarity of positive sample pairs in the embedding space while minimizing the similarity of negative sample pairs; the training process adopts the batch gradient descent method, the batch size is set to 32, the initial learning rate is 0.001, and it decays to one-tenth of the original value every ten training rounds, with a total of 500 training rounds; after training, the encoder parameters are solidified and burned into the non-volatile storage area of ​​the field-programmable gate array chip to ensure the real-time performance and stability of the inference process.

[0007] As one embodiment of the present invention, the method for constructing the fault feature template library includes: applying five preset fault modes to the target equipment in a laboratory environment, repeating each fault mode ten times, and continuously collecting complete sensor data for ten minutes before and after each fault occurrence; performing the same preprocessing and graph construction process as normal data on each fault experiment data to extract the parameter association graph sequence under the fault state; calculating the average graph structure of each fault sequence during the entire fault evolution process as the baseline template for that fault type; simultaneously recording the dynamic evolution trajectory of the graph structure within five diagnostic cycles before and after the fault occurrence time to form a fault evolution sub-template set; storing all templates according to fault type and attaching metadata tags, the metadata including fault severity level, typical triggering conditions, and historical maintenance record index; the template library supports an online update mechanism, and when the on-site diagnostic results are manually verified and confirmed, the corresponding abnormal mode is automatically archived and included in the template library for rapid matching of similar faults in the future.

[0008] As one embodiment of the present invention, the process of generating the abnormal associated edge set further includes: after calculating the cosine distance matrix, performing dual threshold screening on the matrix elements, the first threshold being 0.3 for initially marking deviation edges; the second threshold being 0.5 for identifying strong abnormal edges; the nodes connected to the strong abnormal edges are assigned higher fault weights; at the same time, a time continuity constraint is introduced, if an edge is marked as abnormal in three consecutive diagnostic cycles, its fault confidence is increased by 20%; if the abnormal edge set contains more than three edges connecting the same node, then that node is forcibly promoted to a fault source candidate point, regardless of whether its individual contribution is the highest.

[0009] As one embodiment of the present invention, the human-computer interaction interface is further configured with a diagnostic result traceability function. Users can click on any fault node in the diagnostic report, and the system will automatically trace back the correlation strength change curve, embedded vector trajectory, and mutual information evolution heat map of the node in the 20 diagnostic cycles before the fault occurred. At the same time, it supports historical comparison of fault modes. Users can select any historical fault record, and the system will automatically calculate the structural similarity between the current abnormal mode and the selected historical mode, and highlight the difference area to assist maintenance personnel in judging the fault evolution trend.

[0010] As one embodiment of the present invention, the maintenance work order generation module further includes: automatically assigning maintenance priorities according to the fault type and confidence level, with priorities divided into four levels: emergency, high, medium, and low; emergency priority corresponds to a confidence level higher than 0.8 and a fault type of short circuit between motor turns or broken gear teeth, in which the system automatically sends an SMS notification to the maintenance supervisor and locks the equipment control authority; high priority corresponds to a confidence level between 0.6 and 0.8 or a fault type of bearing wear, in which the system pops up an alert window on the monitoring screen and generates a work order for response within two hours; medium priority corresponds to a confidence level between 0.4 and 0.6, in which the system generates an inspection work order within twenty-four hours; low priority corresponds to a confidence level lower than 0.4 but the abnormality persists, in which the system records it as an event to be observed and generates a summary report weekly.

[0011] As one embodiment of the present invention, the system further includes a model self-calibration unit, which periodically collects the operating data of the equipment under known health conditions, recalculates the baseline correlation graph, and performs a structural consistency check with the currently used baseline graph; if the consistency metric is lower than a preset safety threshold of 0.9, the model retraining process is triggered, the graph neural network encoder parameters are updated using the latest health data, and the normal state reference template in the fault feature template library is refreshed simultaneously to ensure that the diagnostic benchmark dynamically adapts with the aging process of the equipment.

[0012] As one embodiment of the present invention, the multimodal sensing data acquisition unit further includes a sensor health self-check module, which performs zero-point verification and range verification on the output of each sensor before the start of each diagnostic cycle; if the output of a sensor deviates from the factory calibration value by more than 5%, the sensor data is marked as unreliable, and its corresponding node is blocked during the graph construction stage, while triggering a sensor replacement warning; the blocked node does not participate in the association calculation in subsequent diagnoses, and its missing data is compensated by the data of adjacent physical location sensors through spatial interpolation. The interpolation method adopts the inverse distance weighting method, and the weighting coefficient is inversely proportional to the sensor spacing.

[0013] As one embodiment of the present invention, the fault mode matching engine further supports concurrent diagnosis of multiple faults. When the set of abnormal associated edges covers multiple unconnected subgraph structures, the system performs independent template matching on each subgraph and outputs multiple fault diagnosis results. Each result is accompanied by an independent confidence score and spatial positioning coordinates. The system simultaneously calculates the spatiotemporal correlation between each fault result. If the time interval between two faults is less than five diagnosis cycles and the spatial distance is less than 10% of the total length of the equipment, they are marked as associated fault events, and a joint handling plan is suggested in the output report.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0015] This invention overcomes the limitations of traditional fault diagnosis methods that rely on single threshold alarms or fixed rule matching by constructing a dynamic neural network representation system based on parameter correlation graphs, achieving early detection and precise location of latent equipment faults. Its graph structure modeling of nonlinear coupling relationships between parameters effectively identifies complex fault modes caused by multi-parameter collaborative anomalies, solving the problem of missed diagnoses due to isolated parameter analysis in existing technologies. By introducing a self-supervised comparative learning mechanism to train the graph encoder, the dependence on labeled fault data is significantly reduced, enabling the system to maintain strong generalization capabilities even in scenarios lacking historical fault samples. The online update mechanism of the fault feature template library ensures continuous accumulation and iterative optimization of diagnostic knowledge, adapting to pattern drift caused by equipment aging and changing operating conditions. The confidence score and automatic maintenance priority allocation functions attached to the diagnostic results significantly improve the scientific nature and response efficiency of maintenance decisions. Sensor health self-checking and data compensation mechanisms ensure the robust operation of the system when some sensor units fail. The multi-fault concurrent diagnosis and correlation analysis functions provide system-level insights for the health management of complex industrial systems, effectively avoiding the problem of ignoring system-level cascading risks due to isolated handling of single faults. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical architecture of a parameter correlation fault diagnosis method and system based on neural networks proposed in this invention.

[0017] Figure 2 This is a schematic diagram of the core principle framework of the graph neural network modeling mechanism for the nonlinear dynamic coupling relationship between parameters in this invention;

[0018] Figure 3 This is a flowchart illustrating the main stages of the process from multi-source sensor data acquisition to fault diagnosis decision output in this invention. Detailed Implementation

[0019] This invention provides a parameter correlation fault diagnosis method and system based on neural networks. Its core lies in constructing a spatiotemporal alignment framework for multi-source heterogeneous sensor data to achieve high-density synchronous acquisition and structured characterization of key physical parameters during equipment operation. Based on this, a graph neural network modeling mechanism for nonlinear dynamic coupling relationships between parameters is established. Using historical data under normal operating conditions as training samples, the system learns the correlation strength distribution and propagation path of each parameter node under steady-state operating conditions. When real-time operating data is input, the system calculates the topological difference between the current parameter correlation graph and the baseline correlation graph to identify abnormal propagation paths and deviating nodes, thereby locating potential fault sources. Simultaneously, a fault mode matching engine is introduced to compare the identified abnormal correlation patterns with a pre-built fault feature library, outputting the fault type, location, and confidence score, ultimately forming a closed-loop diagnostic decision output.

[0020] As one embodiment of the present invention, the method includes: a multimodal sensor array deployed at key measuring points of industrial equipment, used to synchronously collect five types of physical parameters during equipment operation: vibration, temperature, pressure, current, and rotational speed. The sampling frequency is uniformly set to 1000 Hz, and the timestamp accuracy is not less than 1 microsecond. The collected raw data undergoes preprocessing operations via edge computing nodes. The preprocessing operations include zero-point drift correction, outlier removal, sliding window mean filtering, and normalization mapping to form standardized time-series data blocks. Each data block contains 500 consecutive sampling points, constituting a basic input unit for a diagnostic cycle. The standardized time-series data blocks are then fed into a parameter correlation graph construction module. Each physical parameter is used as a graph node, and the mutual information value between any two nodes within a sliding time window is used as the edge weight to construct a dynamically weighted undirected graph. The graph structure is updated in real time as the time window slides, with an update step size of 50 sampling points. The dynamically weighted undirected graph is input to a graph neural network encoder, which adopts a three-layer graph convolutional layer stacked structure. Each layer includes adjacency matrix normalization, linear transformation of node features, and nonlinear activation function operation, finally outputting the hidden layer embedding vector of each node, with a dimension of 128. The hidden layer embedding vector is sent to the association difference measurement module, which calculates the cosine distance matrix between the current graph structure and the reference graph structure in the node embedding space. For nodes in the matrix that exceed a preset threshold, the module performs a cosine distance measurement. Edges corresponding to elements with a value of 0.3 are marked to form an abnormal associated edge set. This abnormal associated edge set is input to the fault source localization module. This module calculates the fault contribution of each node based on the positional relationship of the nodes connected by the abnormal edges in the physical topology of the equipment, combined with the direction and magnitude of edge weight changes, using a weighted centrality algorithm. The node with the highest contribution is identified as a candidate fault source. The candidate fault source and its associated abnormal patterns are sent to the fault pattern matching engine. This engine has a pre-stored feature template library covering five typical faults: bearing wear, gear tooth breakage, shaft misalignment, lubrication failure, and motor inter-turn short circuit. Each template consists of a typical topological structure and edge weights of the parameter association graph under the corresponding fault state. The distribution range and node embedding vector cluster centers constitute the system. The matching process uses the structure-preserving embedding distance metric to calculate the similarity score between the current abnormal pattern and each template. The fault type corresponding to the highest score is the diagnosis result, and the confidence level of the result is also output. The confidence level is defined as the ratio of the embedding space distance between the current pattern and the best matching template to the distance of the second best template. When the ratio is less than 0.5, it is judged as a high-confidence diagnosis. The diagnosis result is output through the human-computer interaction interface, including the fault type name, the location coordinates, the confidence level value, and the suggested handling measures. At the same time, the equipment maintenance work order generation module is triggered to automatically create a maintenance task record including the diagnosis time, equipment number, fault description, and priority identifier.

[0021] In the above method, a multimodal sensor array deployed at key measurement points of industrial equipment is used to synchronously collect five types of physical parameters during equipment operation: vibration, temperature, pressure, current, and rotational speed. The sampling frequency is uniformly set to 1000 Hz, and the timestamp accuracy is no less than 1 microsecond. This sensor array consists of vibration sensors, infrared temperature sensors, piezoresistive pressure sensors, Hall current sensors, and photoelectric encoders. Each sensor is connected to an edge computing node via an industrial Ethernet bus. The data acquisition cycle is uniformly triggered by a hardware timer to ensure that the data from each channel is strictly aligned on the time axis. Vibration sensors are mounted on the surface of bearing housings or gearbox housings to capture high-frequency vibration signals of the equipment's mechanical structure. Their sensitivity is no less than 10 millivolts per meter per second squared, and their frequency response range covers 0 to 5000 Hz. Infrared temperature sensors are non-contactly mounted on motor housings or bearing end caps to measure surface temperature changes. The temperature measurement range is -40 degrees Celsius to +200 degrees Celsius, with a resolution of 0.1 degrees Celsius. Piezoresistive pressure sensors are integrated into hydraulic or pneumatic pipelines to monitor system pressure fluctuations in real time. The range is set according to the equipment's rated working pressure, with a typical value of 0 to 10 MPa and an accuracy class of 0.5. Hall effect current sensors are connected in series in the motor's main power supply circuit to measure the effective value and instantaneous waveform of the three-phase current. The bandwidth is no less than 20 kHz, and the linearity error is less than 1%. Photoelectric encoders are coaxially mounted on the rotating shaft end. The output pulse signal is converted into a speed value by a counter, with a resolution of no less than 1000 pulses per revolution and a response time of less than 1 millisecond. Before the start of each diagnostic cycle, all sensors undergo zero-point calibration and range verification. If the output of a sensor deviates from the factory calibration value by more than 5%, the sensor data is marked as unreliable, and its corresponding node is blocked during the graph construction phase, while triggering a sensor replacement warning. Blocked nodes do not participate in correlation calculations in subsequent diagnostics, and their missing data is compensated by spatial interpolation from data of adjacent physical location sensors. The interpolation method adopts the inverse distance weighting method, and the weighting coefficient is inversely proportional to the sensor spacing.

[0022] In the above method, the acquired raw data undergoes preprocessing operations via edge computing nodes. These preprocessing operations include zero-point drift correction, outlier removal, sliding window mean filtering, and normalization mapping, forming standardized time-series data blocks. Each data block contains 500 consecutive sampling points, constituting the basic input unit for one diagnostic cycle. Zero-point drift correction is achieved by calculating the static output mean of each channel under no-load conditions and subtracting this mean from subsequent sampled data. Outlier removal uses the three-sigma criterion, discarding data points exceeding the mean plus or minus three standard deviations and replacing them with linear interpolation of adjacent points. The sliding window mean filtering window length is set to 20 sampling points to smooth high-frequency noise. Normalization mapping uses the minimum-maximum scaling method to linearly map each channel's data to the 0-1 interval. The mapping formula is: Here, xmin and xmax represent the global minimum and maximum values ​​of each channel in the device's historical operating data. These values ​​are derived from 30 days of fault-free operating data and stored during system initialization. The preprocessed data is arranged chronologically, with every 500 consecutive sampling points encapsulated into a data block. Overlapping data blocks is allowed, with an overlap length of 450 sampling points to ensure data continuity during sliding window updates. Each data block is sent as an independent input unit to the subsequent processing module. Its internal structure is a 5*500 two-dimensional matrix, with rows corresponding to five types of physical parameters and columns corresponding to time-series sampling points.

[0023] In the above method, the standardized time-series data block is fed into a parameter correlation graph construction module. This module uses each physical parameter as a graph node and the mutual information value between any two nodes within a sliding time window as the edge weight to construct a dynamically weighted undirected graph. The graph structure is updated in real time as the time window slides, with an update step size of 50 sampling points. The mutual information is calculated using the kernel density estimation method. For any two parameter sequences X and Y, their mutual information I(X;Y) is defined as:

[0024]

[0025] Where p(x,y) is the joint probability density, and p(x) and p(y) are the marginal probability densities. The probability density function is estimated using a Gaussian kernel function, and the bandwidth parameter is automatically determined by the Silverman rule. The calculated mutual information values ​​are exponentially normalized and compressed to the interval between 0 and 1, serving as the weights of the corresponding edges in the graph. The initial graph structure is constructed when the first data block is input, containing five nodes and ten undirected edges. Subsequently, for each new data block received, the sliding window advances 50 sampling points, recalculates the mutual information matrix of the data within the window, and updates the graph edge weights. The number of graph nodes is fixed at five, and the node identifiers are bound one-to-one with the physical parameter types, remaining unchanged over time. The graph structure is stored in the form of an adjacency matrix, where the matrix element Aij represents the edge weight between node i and node j, and the diagonal elements are always zero.

[0026] In the above method, the dynamically weighted undirected graph is input to a graph neural network encoder, which employs a three-layer stacked graph convolutional layer structure. Each layer includes adjacency matrix normalization, linear transformation of node features, and nonlinear activation function operations, ultimately outputting the hidden layer embedding vector of each node, with a dimension of 128. The graph convolutional layer operations follow the following formula:

[0027] ,

[0028] in For the first Layer node feature matrix, For learnable weight matrix, To add a self-loop adjacency matrix, for The degree matrix, To correct the activation function of the linear unit, the first layer input features are statistical feature vectors of the original parameter sequences of the nodes, with a dimension of ten, including mean, variance, skewness, kurtosis, maximum value, minimum value, zero-crossing rate, energy, spectral centroid, and spectral entropy; the second layer output has a dimension of 64; and the third layer output has a dimension of 128. Batch normalization is performed after each layer operation to accelerate convergence. The encoder parameters are optimized during the training phase using a self-supervised contrastive learning strategy. The training samples are derived from 30 consecutive days of historical data under fault-free conditions. Each sample contains a complete parameter correlation graph and its corresponding node embedding label. Positive sample pairs are constructed through data augmentation, including a random edge drop probability of 0.1, a node feature perturbation standard deviation of 0.05, and a time axis resampling scaling factor of 0.9 to 1.1. Negative sample pairs are composed of experimental data injected from different devices or faults. The contrastive loss function is used, aiming to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs. Training employs batch gradient descent with a batch size of 32, an initial learning rate of 0.001, and a decay to one-tenth of the original value every ten rounds, for a total of 500 training rounds. After training, the parameters are permanently burned into the non-volatile memory area of ​​the field-programmable gate array chip.

[0029] In the above method, the hidden layer embedding vector is fed into the association difference measurement module. This module calculates the cosine distance matrix between the current graph structure and the reference graph structure in the node embedding space. Edges corresponding to elements in the matrix that exceed a preset threshold of 0.3 are marked, forming a set of abnormal association edges. The cosine distance calculation formula is:

[0030]

[0031] Where u and v are the embedding vectors of the same node in the current graph and the baseline graph, respectively. The baseline graph structure consists of the set of embedding vectors of each node obtained during the device's most recent health status calibration, and is stored in non-volatile memory. The calculated distance matrix is ​​a 5x5 symmetric matrix, with diagonal elements ignored; a dual threshold screening is performed on off-diagonal elements, with a first threshold of 0.3 used for initial marking of deviation edges and a second threshold of 0.5 used for identifying strong abnormal edges; nodes connected by strong abnormal edges are assigned higher fault weights; at the same time, a time continuity constraint is introduced: if an edge is marked as abnormal in three consecutive diagnostic cycles, its fault confidence is increased by 20%; if the set of abnormal edges contains more than three edges connecting the same node, then that node is forcibly promoted to a candidate fault source point, regardless of whether its individual contribution is the highest.

[0032] In the above method, the set of abnormal associated edges is input to the fault source localization module. This module calculates the fault contribution of each node using a weighted centrality algorithm, based on the positional relationship of the nodes connected by the abnormal edges in the device's physical topology, combined with the direction and magnitude of edge weight changes. The node with the highest contribution is identified as a candidate fault source. The weighted centrality algorithm defines the fault contribution Ci of node i as:

[0033] ,

[0034] Where N(i) is the set of neighboring nodes connected to node i, Wij is the current edge weight, ΔWij is the absolute difference between the current weight and the baseline weight, and dij is the Euclidean distance between nodes i and j in the physical space of the device, which is pre-determined by the sensor installation location coordinates. The contribution calculation result is normalized to the interval between 0 and 1, and the one with the largest value is selected as the candidate point for the fault source. If there are multiple candidate points with a contribution difference of less than 0.1, all candidate points are retained and enter the subsequent matching stage.

[0035] In the above method, the candidate fault sources and their associated anomaly patterns are fed into a fault mode matching engine. This engine pre-stores a feature template library covering five typical faults: bearing wear, gear tooth breakage, shaft misalignment, lubrication failure, and inter-turn short circuit in motors. Each template consists of a typical topological structure of the parameter association graph under the corresponding fault state, the distribution range of edge weights, and the cluster center of the node embedding vector. The matching process uses a structure-preserving embedding distance metric to calculate the similarity score between the current anomaly pattern and each template. The fault type corresponding to the highest score is the diagnosis result, and the confidence level of this result is output. The confidence level is defined as the ratio of the embedding space distance between the current pattern and the best matching template to the distance of the second-best template. A ratio less than 0.5 is considered a high-confidence diagnosis. The structure-preserving embedding distance metric aligns the current graph structure with the template graph structure in the embedding space and then calculates the Frobenius norm distance. The alignment process uses Protodyakonov analysis to eliminate translation, rotation, and scaling differences. The template library construction method includes: applying five preset fault modes to the target equipment in a laboratory environment, with each fault mode repeated ten times, and continuously collecting complete sensor data for ten minutes before and after each fault occurrence; performing the same preprocessing and graph construction process as normal data on each fault experiment data to extract parameter association graph sequences under fault conditions; calculating the average graph structure of each fault sequence throughout the entire fault evolution process as the baseline template for that fault type; simultaneously recording the dynamic evolution trajectory of the graph structure within five diagnostic cycles before and after the fault occurrence time to form a fault evolution sub-template set; storing all templates according to fault type and attaching metadata tags, with the metadata including fault severity level, typical triggering conditions, and historical maintenance record index; the template library supports an online update mechanism, and when the on-site diagnostic results are manually reviewed and confirmed, the corresponding abnormal mode is automatically archived and included in the template library.

[0036] In the above method, the diagnostic results are output via a human-machine interface, including the fault type name, location coordinates, confidence level, and suggested remedial measures. Simultaneously, it triggers the equipment maintenance work order generation module, automatically creating a maintenance task record containing the diagnostic time, equipment number, fault description, and priority identifier. The human-machine interface is further configured with a diagnostic result traceability function. Users can click on any fault node in the diagnostic report, and the system automatically traces back the correlation strength change curve, embedded vector trajectory, and mutual information evolution heatmap of that node in the 20 diagnostic cycles prior to the fault occurrence. It also supports historical comparison of fault modes; users can select any historical fault record, and the system automatically calculates the structural similarity between the current abnormal mode and the selected historical mode, highlighting the areas of difference. The maintenance work order generation module automatically assigns maintenance priorities based on fault type and confidence level. Priorities are divided into four levels: emergency, high, medium, and low. Emergency priority corresponds to situations with a confidence level higher than 0.8 and fault types such as inter-turn short circuit in motors or broken gear teeth. The system automatically sends an SMS notification to the maintenance supervisor and locks equipment control permissions. High priority corresponds to situations with a confidence level between 0.6 and 0.8 or fault types such as bearing wear. The system displays an alert window on the monitoring screen and generates a work order for response within two hours. Medium priority corresponds to situations with a confidence level between 0.4 and 0.6. The system generates an inspection work order within 24 hours. Low priority corresponds to situations with a confidence level lower than 0.4 but where the abnormality persists. The system records these as events to be observed and generates a summary report weekly.

[0037] In the above method, the system further includes a model self-calibration unit, which periodically collects operational data of the equipment under known health conditions, recalculates the baseline correlation graph, and performs a structural consistency check with the currently used baseline graph. If the consistency metric is lower than a preset safety threshold of 0.9, the model retraining process is triggered. The graph neural network encoder parameters are updated using the latest health data, and the normal state reference templates in the fault feature template library are simultaneously refreshed to ensure that the diagnostic baseline dynamically adapts as the equipment ages. The consistency metric uses the Kendall rank correlation coefficient to calculate the consistency between the current baseline graph and the newly calculated baseline graph in terms of edge weight ranking; a coefficient lower than 0.9 is considered a significant drift. The retraining process reuses the original training framework but only uses the latest health data, reducing the number of training rounds to one hundred to accelerate convergence.

[0038] In the above method, the fault mode matching engine further supports concurrent diagnosis of multiple faults. When the set of abnormal associated edges covers multiple unconnected subgraph structures, the system performs independent template matching on each subgraph and outputs multiple fault diagnosis results. Each result is accompanied by an independent confidence score and spatial location coordinates. The system simultaneously calculates the spatiotemporal correlation between each fault result. If two faults are less than five diagnostic cycles apart in time and less than 10% of the total length of the equipment in space, they are marked as associated fault events, and a joint handling plan is suggested in the output report. The spatiotemporal correlation calculation is based on the time difference of the fault occurrence and the spatial Euclidean distance between the fault source coordinates. The joint handling plan is generated by matching a preset rule base, which contains the handling priority and resource allocation strategy for typical fault combinations.

[0039] This invention overcomes the limitations of traditional fault diagnosis methods that rely on single threshold alarms or fixed rule matching by constructing a dynamic neural network representation system based on parameter correlation graphs, achieving early detection and precise location of latent equipment faults. Its graph structure modeling of nonlinear coupling relationships between parameters effectively identifies complex fault modes caused by multi-parameter collaborative anomalies, solving the problem of missed diagnoses due to isolated parameter analysis in existing technologies. By introducing a self-supervised comparative learning mechanism to train the graph encoder, the dependence on labeled fault data is significantly reduced, enabling the system to maintain strong generalization capabilities even in scenarios lacking historical fault samples. The online update mechanism of the fault feature template library ensures continuous accumulation and iterative optimization of diagnostic knowledge, adapting to pattern drift caused by equipment aging and changing operating conditions. The confidence score and automatic maintenance priority allocation functions attached to the diagnostic results significantly improve the scientific nature and response efficiency of maintenance decisions. Sensor health self-checking and data compensation mechanisms ensure the robust operation of the system when some sensor units fail. The multi-fault concurrent diagnosis and correlation analysis functions provide system-level insights for the health management of complex industrial systems, effectively avoiding the problem of ignoring system-level cascading risks due to isolated handling of single faults.

[0040] The above content is only a preferred embodiment of the present invention. For those skilled in the art, many changes can be made in the specific implementation and application scope based on the ideas of the present invention. As long as these changes do not depart from the concept of the present invention, they all fall within the protection scope of this patent.

Claims

1. A parameter correlation fault diagnosis method based on neural networks, characterized in that, include: The multimodal sensor array deployed at key measurement points of industrial equipment synchronously collects five types of physical parameters during equipment operation: vibration, temperature, pressure, current, and rotational speed. The sampling frequency is uniformly set to 1000 Hz, and the timestamp accuracy is not less than 1 microsecond. Zero-point drift correction, outlier removal, sliding window mean filtering, and normalization mapping are performed on the collected raw data to form standardized time-series data blocks, each containing 500 consecutive sampling points; Using each physical parameter as a graph node and the mutual information value between any two nodes within the sliding time window as the edge weight, a dynamic weighted undirected graph is constructed. The graph structure is updated in real time as the time window slides, with an update step of 50 sampling points. The dynamically weighted undirected graph is input into the graph neural network encoder, which adopts a three-layer graph convolutional layer stacked structure. Each layer includes adjacency matrix normalization operation, node feature linear transformation and nonlinear activation function operation, and outputs a 128-dimensional hidden layer embedding vector for each node. Calculate the cosine distance between the embedding vectors of each node in the current graph structure and the corresponding embedding vectors of each node in the reference graph structure in the node embedding space, and construct a cosine distance matrix. Mark the edges corresponding to the elements in the matrix that exceed the preset threshold of 0.3 to form a set of abnormally associated edges. Based on the positional relationship of the nodes connected by the abnormal edge in the physical topology of the device, and combined with the direction and magnitude of the edge weight change, the weighted centrality algorithm is used to calculate the fault contribution of each node, and the node with the highest contribution is determined as a candidate fault source. The candidate fault sources and their associated abnormal patterns are input into the fault pattern matching engine. The fault pattern matching engine has a pre-stored feature template library covering five typical faults: bearing wear, gear tooth breakage, shaft misalignment, lubrication failure, and motor inter-turn short circuit. Each template consists of the typical topological structure of the parameter association graph under the corresponding fault state, the edge weight distribution range, and the node embedding vector clustering center. The similarity score between the current anomalous pattern and each template is calculated using the structure-preserving embedding distance metric. The fault type corresponding to the highest score is the diagnostic result, and the confidence score of this result is also output. The confidence score is defined as the ratio of the embedding space distance between the current pattern and the best matching template to the distance between the second-best template; specifically including: After aligning the current graph structure with the template graph structure in the embedding space, the Frobenius norm distance is calculated. The alignment process eliminates differences in translation, rotation, and scaling through Protodyakonov analysis. The template library construction method includes: applying five preset fault modes to the target device in a laboratory environment, repeating each fault mode 10 times, and continuously collecting complete sensor data for ten minutes before and after each fault occurrence. Perform the same preprocessing and graph construction process as for normal data on each fault experiment data, and extract the parameter correlation graph sequence under fault conditions; Calculate the average graph structure of each fault sequence throughout the entire fault evolution process, and use it as a baseline template for that fault type; Record the dynamic evolution trajectory of the diagram structure within five diagnostic cycles before and after the fault occurrence time to form a fault evolution sub-template set; All templates are stored according to fault type and are attached with metadata tags. The metadata includes fault severity level, typical triggering conditions and historical maintenance record index. The template library supports an online update mechanism. Once the on-site diagnostic results are manually reviewed and confirmed, the corresponding abnormal patterns are automatically archived and included in the template library. The diagnostic results are output through the human-machine interface, including the fault type name, location coordinates, confidence level, and suggested handling measures. At the same time, the equipment maintenance work order generation module is triggered to automatically create a maintenance task record that includes the diagnosis time, equipment number, fault description, and priority identifier.

2. The parameter correlation fault diagnosis method based on neural networks according to claim 1, characterized in that, The collected raw data undergoes zero-point drift correction, outlier removal, sliding window mean filtering, and normalization mapping to form standardized time-series data blocks, including: Calculate the mean static output of each channel under no-load conditions, and subtract this mean from the subsequent sampled data to complete the zero-point drift correction; The three sigma criterion was used to remove data points that exceeded the mean plus or minus three standard deviations, and these data points were replaced by linear interpolation of adjacent points. Perform sliding window mean filtering with a window length of 20 sampling points to smooth high-frequency noise; The minimum-maximum scaling method is used to linearly map the data of each channel to the interval between 0 and 1. The mapping formula is as follows: , where x min With x max These are the global minimum and maximum values ​​for each channel in the device's historical operating data.

3. The parameter correlation fault diagnosis method based on neural networks according to claim 2, characterized in that, Using each physical parameter as a graph node and the mutual information value between any two nodes within a sliding time window as the edge weight, a dynamically weighted undirected graph is constructed, including: The mutual information I(X;Y) of any two parameter sequences X and Y is calculated using the kernel density estimation method, as follows: ， Where p(x,y) is the joint probability density, p(x) and p(y) are the marginal probability densities, the probability density function is estimated by the Gaussian kernel function, and the bandwidth parameter is automatically determined by the Silverman rule; The calculated mutual information values ​​are compressed to the range of 0 to 1 by exponential normalization and used as the weights of the corresponding edges in the graph; The initial graph structure contains 5 nodes and 10 undirected edges. Subsequently, each time a new data block is received, the sliding window advances by 50 sampling points, the mutual information matrix of the data within the window is recalculated, and the graph edge weights are updated.

4. The parameter correlation fault diagnosis method based on neural networks according to claim 3, characterized in that, Inputting the dynamically weighted undirected graph into the graph neural network encoder includes: The first layer of input features is a 10-dimensional statistical feature vector of the original parameter sequence of the nodes, including mean, variance, skewness, kurtosis, maximum value, minimum value, zero crossing rate, energy, spectral centroid, and spectral entropy. The second layer outputs 64 dimensions, and the third layer outputs 128 dimensions. Batch normalization is performed after each layer of computation; The graph convolutional layer operation follows the formula: ， in For the first Layer node feature matrix For learnable weight matrix, To add a self-loop adjacency matrix, for The degree matrix, To modify the activation function of the linear unit.

5. The parameter correlation fault diagnosis method based on neural networks according to claim 4, characterized in that, Calculate the cosine distance matrix between the current graph structure and the reference graph structure in the node embedding space, including: The formula for calculating cosine distance is: ， Where u and v are the embedding vectors of the same node in the current graph and the reference graph, respectively; For off-diagonal elements, a dual threshold screening is performed: a first threshold of 0.3 is used to initially mark off-center edges, and a second threshold of 0.5 is used to identify strong anomalous edges. If an edge is marked as abnormal in three consecutive diagnostic cycles, its fault confidence is increased by 20%. If the set of abnormal edges contains more than three edges connecting the same node, then that node is forcibly promoted to a candidate fault source.

6. The parameter correlation fault diagnosis method based on neural networks according to claim 5, characterized in that, The weighted centrality algorithm is used to calculate the fault contribution of each node, including: Define the fault contribution C of node i. i for: ， Where N(i) is the set of neighboring nodes connected to node i, and W ij Let ΔW be the current edge weight. ij d is the absolute difference between the current weight and the benchmark weight. ij Let be the Euclidean distance between nodes i and j in the device's physical space; The contribution calculation results are normalized to the range of 0 to 1, and the one with the largest value is the candidate point of the fault source. If there are multiple candidate points with a contribution difference of less than 0.1, then all candidate points are retained and proceed to the subsequent matching stage.

7. The parameter correlation fault diagnosis method based on neural networks according to claim 6, characterized in that, The module that triggers the equipment maintenance work order generation includes: Repair priorities are automatically assigned based on fault type and confidence level, with four priority levels: emergency, high, medium, and low. Emergency priority corresponds to situations where the confidence level is higher than 0.8 and the fault type is short circuit between motor turns or broken gear teeth. The system will automatically send an SMS notification to the maintenance supervisor and lock the equipment control permissions. High priority corresponds to situations with a confidence level between 0.6 and 0.8 or a fault type of bearing wear. The system will pop up an alert window on the monitoring screen and generate a work order for response within two hours. For medium priority cases with a confidence level between 0.4 and 0.6, the system generates an inspection work order for the next 24 hours. Low priority corresponds to situations where the confidence level is below 0.4 but the anomaly persists. The system records these as events to be observed and generates a summary report weekly.

8. A parameter correlation fault diagnosis system based on a neural network, executing the parameter correlation fault diagnosis method based on a neural network as described in claim 1, characterized in that, include: The multimodal sensing data acquisition unit is used to synchronously collect five types of physical parameters during the operation of industrial equipment through a multimodal sensor array deployed at key measuring points of the equipment. The sampling frequency is uniformly set to 1000 Hz, and the timestamp accuracy is not less than 1 microsecond. The edge preprocessing unit is used to perform zero-point drift correction, outlier removal, sliding window mean filtering and normalization mapping on the acquired raw data to form standardized time series data blocks, each data block containing 500 consecutive sampling points; The parameter association graph construction unit is used to construct a dynamic weighted undirected graph with each physical parameter as a graph node and the mutual information value between any two nodes within the sliding time window as the edge weight. The graph structure is updated in real time as the time window slides, with an update step of 50 sampling points. The graph neural network encoding unit is used to input the dynamically weighted undirected graph into the graph neural network encoder. The graph neural network encoder adopts a three-layer graph convolutional layer stacked structure. Each layer includes adjacency matrix normalization operation, node feature linear transformation and nonlinear activation function operation, and outputs a 128-dimensional hidden layer embedding vector for each node. The correlation difference measurement unit is used to calculate the cosine distance matrix between the current graph structure and the reference graph structure in the node embedding space, and to mark the edges corresponding to the elements in the matrix that exceed the preset threshold of 0.3, forming a set of abnormal correlation edges; The fault source localization unit is used to calculate the fault contribution of each node based on the positional relationship of the nodes connected by the abnormal edge in the physical topology of the device, combined with the direction and magnitude of the edge weight change, and using a weighted centrality algorithm. The node with the highest contribution is determined as a candidate fault source. The fault mode matching unit is used to input the candidate fault source points and their associated abnormal patterns into the fault mode matching engine. The fault mode matching engine has a feature template library that covers five typical faults: bearing wear, gear tooth breakage, shaft misalignment, lubrication failure, and motor inter-turn short circuit. Each template consists of the typical topology of the parameter association graph under the corresponding fault state, the edge weight distribution range, and the node embedding vector clustering center. The diagnostic decision output unit is used to calculate the similarity score between the current abnormal pattern and each template using the structure-preserving embedding distance metric. The fault type corresponding to the highest score is the diagnostic result. At the same time, the confidence score of the result is output. The confidence score is defined as the ratio of the embedding space distance between the current pattern and the best matching template to the distance between the second best template. The diagnostic results are output through the human-machine interface, including the fault type name, location coordinates, confidence level, and suggested handling measures. At the same time, the equipment maintenance work order generation module is triggered to automatically create a maintenance task record that includes the diagnosis time, equipment number, fault description, and priority identifier.

9. The parameter correlation fault diagnosis system based on neural networks according to claim 8, characterized in that, The edge preprocessing unit is used for: Calculate the mean static output of each channel under no-load conditions, and subtract this mean from the subsequent sampled data to complete the zero-point drift correction; The three sigma criterion was used to remove data points that exceeded the mean plus or minus three standard deviations, and these data points were replaced by linear interpolation of adjacent points. Perform sliding window mean filtering with a window length of 20 sampling points to smooth high-frequency noise; The minimum-maximum scaling method is used to linearly map the data of each channel to the interval between 0 and 1. The mapping formula is as follows: , where X min With X max These are the global minimum and maximum values ​​for each channel in the device's historical operating data.

Images