A method and system for intelligent product risk identification based on big data.

Abstract

This invention discloses a method and system for intelligent product risk identification based on big data. The method includes collecting monitoring data and related data of a preset wind turbine gearbox within a target industrial scenario; performing multi-source feature extraction on the monitoring data and related data to obtain a monitoring feature set and related feature levels; extracting key information from the monitoring feature set and related feature levels to obtain key monitoring information; performing consistency comparison based on the key monitoring information to obtain information contrast; and obtaining problematic data and fuzzy data. The method further involves extracting temporal anomaly relationship features and performing joint decomposition on the fuzzy data to obtain fuzzy features; performing risk identification and joint decomposition on the problematic data to obtain risk dimensionality reduction features; coupling the fuzzy features and risk dimensionality reduction features by sample unit to obtain sample risk features; and constructing an industrial intelligent monitoring product risk identification model based on a graph neural network based on the sample risk features, and outputting the identification results.

Description

Technical Field

[0001] This invention relates to the field of risk identification, and in particular to a method and system for identifying risks in intelligent products based on big data. Background Technology

[0002] With the advent of Industry 4.0 and the rapid development of intelligent manufacturing, key intelligent products such as wind turbine gearboxes and industrial robots have become core assets of modern industrial systems. Their operational status directly affects the stability, safety, and economic efficiency of the entire production system. Therefore, accurate and timely risk identification and health status monitoring of these intelligent products are of paramount importance for achieving predictive maintenance, avoiding catastrophic failures, and reducing operation and maintenance costs.

[0003] Currently, big data-based intelligent product risk identification technology has become a research hotspot in both industry and academia. Traditional methods mainly rely on setting fixed thresholds for alarms based on single types of monitoring data, or using shallow machine learning models to classify manually extracted features. These methods have significant limitations: First, they are difficult to effectively integrate multi-source heterogeneous monitoring and correlation data, resulting in insufficient information utilization; second, traditional methods lack sensitivity for ambiguous abnormal signals generated by early, weak, or coupled faults, easily leading to missed detections; third, most models treat monitoring units as independent entities, ignoring the structural connections between equipment and other units in complex industrial systems, thus failing to capture the transmission and diffusion effects of risks and making system-level risk assessment difficult.

[0004] In recent years, deep learning technology, especially graph neural networks, has provided new ideas for solving the above challenges. However, existing deep learning-based risk identification models still face many challenges: first, at the feature extraction level, how to adaptively extract highly discriminative key information from multi-source data and differentiate between deterministic and uncertain / fuzzy data; second, at the model construction level, how to combine the features of individual devices with the system's topology to build a dynamic model that can understand the risk transmission mechanism of complex systems. Therefore, there is an urgent need for a method that can deeply integrate multi-source information, intelligently distinguish and process different types of anomalies, and perform risk reasoning from a global system perspective to improve the accuracy, early detection, and robustness of risk identification in intelligent products. Summary of the Invention

[0005] The purpose of this invention is to provide a method for identifying risks in intelligent products based on big data.

[0006] To achieve the above objectives, the present invention is implemented according to the following technical solution:

[0007] This invention includes the following steps:

[0008] The system collects monitoring data and related data of a pre-set wind turbine gearbox within the target industrial scenario, preprocesses the monitoring data and related data, and establishes a synchronization time mechanism. The monitoring data includes vibration data, process data, and online oil data. The related data includes maintenance history data, environmental data, offline report data, inspection data, and image data.

[0009] This is used to perform multi-source feature extraction on the monitoring data and the associated data to obtain a monitoring feature set and an associated feature level, to extract key information from the monitoring feature set and the associated feature level to obtain key monitoring information, and to perform consistency comparison based on the key monitoring information to obtain information contrast. If the information contrast is higher than the consistency threshold, or lower than the consistency threshold and is obviously abnormal, then the output is problematic data; otherwise, the output is ambiguous data.

[0010] The fuzzy data is subjected to time-series anomaly relationship feature extraction and joint decomposition to obtain fuzzy features; the problem data is subjected to risk identification and joint decomposition to obtain risk dimensionality reduction features; and the fuzzy features and the risk dimensionality reduction features are coupled according to sample units to obtain sample risk features.

[0011] Based on the risk characteristics of the samples, a risk identification model for industrial intelligent monitoring products based on graph neural networks is constructed. The data to be identified is input into the risk identification model for industrial intelligent monitoring products, and the risk identification results are output.

[0012] Furthermore, the method for extracting key monitoring information from the monitoring feature set and the associated feature level includes:

[0013] Extracting differences in monitoring feature sets: For the same sample and the same feature data in the time series, the sliding window dynamic time warping algorithm is used to align the historical and current time series curves, quantify the self-offset, and obtain the degradation deviation rate through the current feature value, rolling baseline value and running time. The degradation rate is output as key monitoring information; when the deviation is less than 3%, it is marked as environmental interference, and when the deviation is greater than or equal to 8%, it is marked as equipment degradation signal.

[0014] For the same sample with different features in a time series, the joint difference between different data is calculated by combining the Claus-Leibler divergence and Jaccard similarity distance. A temperature and humidity correction factor is introduced to correct the cross-data table bias rate of the same data at different times. The weighted K-means clustering algorithm is used to cluster the cross-data, and the feature contribution weights of the cross-data are calculated.

[0015] ;

[0016] ;

[0017] in Let be the sum of the within-class variances of the i-th dimension feature. For hyperparameters, The number of feature dimensions. Let be the sum of the within-class variances of the j-th dimension feature. Let x be the i-th feature value. The contribution weight for the i-th dimension feature. The number of clusters in the clustering. For cluster index, For the k-th cluster, It is the i-th eigenvalue of the k-th cluster center;

[0018] Features contributing more than 61% of the joint differences are output as the core anomaly set, and features contributing less than 50% are output as the composite fault combination.

[0019] Key monitoring information includes degradation rate, core anomaly set, and complex fault combinations.

[0020] Furthermore, the method for obtaining information contrast by performing consistency comparison based on the key monitoring information includes:

[0021] Obtain the historical baseline of degradation rate, align the historical baseline and real-time degradation rate to calculate dynamic time warping offset, and use p-value hypothesis testing to test the dynamic time warping offset to obtain the tested time warping offset; the null hypothesis of the p-value hypothesis test is that the degradation rate is less than or equal to the fluctuation range of the historical baseline, and p is greater than 0.05; the alternative hypothesis is that degradation is significantly accelerated, and p is less than or equal to 0.05.

[0022] A multimodal association model is used to align the distribution of core anomaly sets and calculate the information similarity of core anomaly information:

[0023] ;

[0024] ;

[0025] in Core abnormal element P and Q Information contrast between the two P As the core abnormal element, Q As another core abnormal element, H Core abnormal element P and Q Average distribution between To extract from the core abnormal element P to the average distribution H The relative entropy divergence, To extract from the core abnormal element Q to the average distribution HThe relative entropy divergence;

[0026] Construct a fault contribution radar chart based on the composite fault combination elements to obtain the contribution of the composite fault combination elements;

[0027] Information contrast is calculated based on time warping offset, information similarity, and contribution:

[0028] ;

[0029] in For information contrast, This is the time warping offset. The contribution of the composite fault combination elements. The degradation rate weighting coefficient, The core anomaly weighting coefficient, For composite fault weighting coefficients.

[0030] Furthermore, the method for extracting temporal anomaly relationship features and jointly decomposing the fuzzy data includes:

[0031] By combining correlation differences with fuzzy computation to quantify anomaly sensitivity, fuzzy data is input, and an attention mechanism is used to calculate dynamic correlation weights.

[0032] The association strength between fuzzy data and the global sequence is learned through a self-attention mechanism. If the association weights of outliers are concentrated in the neighborhood, the local focus is extracted. The Jensen-Shannon divergence between the current input attention distribution and the normal attention distribution is calculated. When the Jensen-Shannon divergence is greater than 0.25, it is marked as an outlier and the global deviation is extracted. The local focus and global deviation are used as temporal outlier relationship features.

[0033] The causal weights of temporal anomaly relationship features are obtained. When the causal weight is greater than the causal threshold, a weighted fuzzy rule is used to enhance the fuzzy causal strength. When the Mahalanobis distance is greater than the distance threshold, fuzzy C-means cluster center distance and membership degree correction are used to enhance the fuzzy cluster outlier. When the path deviation is greater than the cost threshold, fuzzy dynamic time warping is used to align historical cycles to enhance the seasonal gradient pattern. The fuzzy relationship features include fuzzy causal strength, fuzzy cluster outlier, and seasonal gradient pattern.

[0034] Fuzzy trend decomposition, fuzzy seasonal decomposition, and fuzzy residual decomposition are performed on the fuzzy relation features to obtain the trend component, seasonal component, and residual component, as expressed in the following expression:

[0035] ;

[0036] ;

[0037] ;

[0038] in Let be the fuzzy relation feature at time t. Let be the membership degree at time t. The intercept is... The linear drift coefficient is... For the ridge regression regularization strength, To set an upper limit for monitoring time, Let c be the frequency value of the c-th frequency component. , For the c-th triangular fuzzy number to be estimated, Let t be the order of the Fourier series, and t be the current time. Let be the fuzzy residual component at time t. Let be the seasonal component at time t. Let be the high-dimensional residual at time t. For the fuzzy data at time t, Let be the estimated value of the trend component at time t. For model parameter vectors, ;

[0039] The trend component, seasonal component, and residual component are output as fuzzy features, and the time series anomaly relationship features are labeled.

[0040] Furthermore, the method for risk identification and joint decomposition of the problem data includes:

[0041] Problem data is input into the physical mechanism and hybrid integral model. A knowledge graph is constructed based on the physical structure of the gearbox, component information, and corresponding topological relationships. The meshing frequency of each gear stage and the fault characteristic frequency of the bearing are calculated based on the mechanical model. The meshing frequency, fault characteristic frequency, and sideband theoretical calculation formulas are stored in the database. The spectrum of vibration signals in the problem data is analyzed. The physical mechanism and hybrid integral model use the knowledge graph to quickly locate the frequency components with abnormally high energy. The fault characteristic frequencies include the inner ring, outer ring, rolling elements, and cage.

[0042] Guided by the knowledge graph, the amplitude, kurtosis factor, and sideband energy related to the theoretical fault characteristic frequencies are extracted from the spectrum of the problem data to generate a preliminary list of fault hypotheses and calculate the physical confidence score of the hypotheses.

[0043] Meanwhile, the problem data is input into a pre-trained one-dimensional deep convolutional neural network model, which outputs a multi-class probability vector; where the multi-class probability vector represents the data-driven probability of belonging to each fault type.

[0044] The physical confidence score and data-driven probability are weighted and fused to generate fault diagnosis results, and the comprehensive confidence score is labeled.

[0045] Based on the fault diagnosis results, fault features are extracted, the risk exceedance multiple is calculated based on the fault features and the baseline threshold, the growth trend of fault features within the time window is analyzed, and the risk level is obtained by locating fault features in a two-dimensional risk matrix by combining the risk exceedance multiple and the growth trend. Fault features with a risk level greater than the risk threshold are taken as target features.

[0046] Vibration signals in multi-channel target features are decomposed into intrinsic mode functions (IMFs). Based on physical information, a hybrid multi-fault decomposition and diagnostic model is used to identify fault characteristic frequencies. IMF components are selected as effective components. The effective components are combined and blind source separation technology is used to obtain mutually independent source signals. The source signals are jointly decomposed and output as risk dimensionality reduction features. The source signals contain fault characteristic information.

[0047] Furthermore, the method for coupling the fuzzy features and the risk dimensionality reduction features on a sample unit basis includes:

[0048] For each sample, the fuzzy feature vector and the risk dimensionality reduction feature vector are batch normalized and mapped to similar distribution intervals. The dimensions of the fuzzy feature vector and the risk dimensionality reduction feature vector are unified through a fully connected layer to obtain the first fuzzy feature and the first risk feature.

[0049] Using the first fuzzy feature as the query and the first risk feature as the key and value, calculate the attention weight matrix to obtain the second fuzzy feature affected by the risk feature; using the first risk feature as the query and the first fuzzy feature as the key and value, calculate the attention weight to obtain the second risk feature affected by the fuzzy feature.

[0050] The second fuzzy feature and the second risk feature are concatenated, and then passed through a fully connected layer and a sigmoid activation function to generate a gate value between 0 and 1, expressed as:

[0051] ;

[0052] in This is the gate value. This is the second fuzzy feature. This is the second risk characteristic. This is the weight matrix. For bias terms;

[0053] The second fuzzy feature and the second risk feature are weighted and summed using the gate value to generate the sample risk feature, expressed as follows:

[0054] ;

[0055] in As a sample risk characteristic, This is for element-wise multiplication.

[0056] Furthermore, the method for constructing a risk identification model for industrial intelligent monitoring products based on graph neural networks according to the risk characteristics of the samples includes:

[0057] The entire industrial monitoring system is constructed as a heterogeneous graph. The node set includes monitoring unit nodes and subsystem nodes; the edge set represents the edge relationships between nodes; and the meta-path is a semantically defined meta-path based on node and edge relationships. Among them, the monitoring unit nodes are the sample risk characteristics of the monitored smart products, and the subsystem nodes are functional units. The edge relationships include physical connection relationships, functional dependency relationships, and membership relationships.

[0058] The risk identification model for industrial intelligent monitoring products includes a node internal and external feature enhancement module, a system dynamic representation module, a feature fusion and heterogeneous graph relationship aggregation module, and risk discrimination and model training.

[0059] The node internal and external feature enhancement module learns attention vectors for each node type, maps nodes to the corresponding attention vectors according to node type, and adaptively weights the original features of the nodes.

[0060] The system dynamic characterization module takes the sample risk characteristics of multiple time steps within each monitoring unit as input, uses the Transformer encoder to capture long-term dependencies in the sequence, introduces a temporal attention layer, automatically focuses on the key time step with the greatest risk evolution, and generates a global dynamic characterization.

[0061] The feature fusion and heterogeneous graph relationship aggregation module includes feature fusion and heterogeneous graph relationship aggregation. The heterogeneous graph relationship aggregation includes node-level attention and semantic-level attention. Feature fusion concatenates static enhanced features and dynamic temporal features and passes them through a fully connected layer to generate the initial embedding of the node. Node-level attention uses a graph attention network to calculate the attention coefficients of the node on the neighbors of the metapath and aggregates them to obtain the metapath-specific embedding of the node. Semantic-level attention introduces a semantic-level attention vector, calculates the global importance weight of the metapath in the risk identification task, and weights and fuses all metapath-specific embeddings to obtain the final embedding.

[0062] Risk discrimination and model training include a risk discriminator and a loss function; the risk discriminator takes the final embedding of the node as input to a multilayer perceptron, connects it to a sigmoid activation function, and outputs the risk probability of the monitoring unit; the loss function adopts focus loss.

[0063] Secondly, a big data-based intelligent product risk identification system includes:

[0064] The data acquisition and processing module is used to acquire monitoring data and related data of a preset wind turbine gearbox in the target industrial scenario, preprocess the monitoring data and related data, and establish a synchronization time mechanism. The monitoring data includes vibration data, process data, and online oil data. The related data includes maintenance history data, environmental data, offline report data, inspection data, and image data.

[0065] Risk classification module: used to perform multi-source feature extraction on the monitoring data and the associated data to obtain monitoring feature set and associated feature level, extract key information from the monitoring feature set and the associated feature level to obtain key monitoring information, and perform consistency comparison on the key monitoring information to obtain information contrast. If the information contrast is higher than the consistency threshold, or lower than the consistency threshold and obviously abnormal, the output is problem data; otherwise, the output is fuzzy data.

[0066] Dimensionality reduction coupling module: used to extract time-series anomaly relationship features and jointly decompose the fuzzy data to obtain fuzzy features, to identify risks and jointly decompose the problem data to obtain risk dimensionality reduction features, and to couple the fuzzy features and the risk dimensionality reduction features according to sample units to obtain sample risk features;

[0067] Modeling output module: used to construct a risk identification model for industrial intelligent monitoring products based on graph neural networks according to the risk characteristics of the samples, input the data to be identified into the risk identification model for industrial intelligent monitoring products, and output the risk identification results.

[0068] The beneficial effects of this invention are:

[0069] This invention is a method and system for intelligent product risk identification based on big data. Compared with existing technologies, this invention has the following technical advantages:

[0070] This invention, through preprocessing, multi-source feature extraction, consistency comparison, obtaining fuzzy features, obtaining risk dimensionality reduction features, feature coupling, and model building steps, and through intelligent fusion and classification of multi-source data, combined with graph neural network system modeling, significantly improves the accuracy of risk identification and early warning capabilities, realizing a leap from individual monitoring to system-level risk assessment, and effectively ensuring the safe and stable operation of industrial intelligent products. Attached Figure Description

[0071] Figure 1 This is a flowchart illustrating the steps of a big data-based intelligent product risk identification method according to the present invention. Detailed Implementation

[0072] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0073] The present invention provides a method and system for intelligent product risk identification based on big data, comprising the following steps:

[0074] like Figure 1 As shown, this embodiment includes the following steps:

[0075] The system collects monitoring data and related data of a pre-set wind turbine gearbox within the target industrial scenario, preprocesses the monitoring data and related data, and establishes a synchronization time mechanism. The monitoring data includes vibration data, process data, and online oil data. The related data includes maintenance history data, environmental data, offline report data, inspection data, and image data.

[0076] In practical evaluation, preprocessing includes using the IQR criterion or density-based clustering algorithm to remove abnormal sensor readings; reconstructing missing data based on generative adversarial networks or variational autoencoders; and standardizing vibration and oil temperature heterogeneous data using Min-Max or Z-score.

[0077] Time synchronization technology: Deploy hardware timestamps at the sensor end to ensure millisecond-level synchronization of high-speed data such as vibration and temperature; use dynamic time warping algorithms to align low-frequency correlated data such as maintenance records and offline reports;

[0078] Using a wind turbine gearbox in an industrial setting as the research object, monitoring data and related data were collected in April 2023. Monitoring data included: vibration data showing an increase in peak axial vibration acceleration from 0.5g to 1.8g; process data showing an increase in gearbox oil temperature from 65°C to 78°C; online oil data showing an increase in wear particle concentration from 50ppm to 120ppm according to ferrography analysis; related data showing a bearing replacement 3 months prior; environmental data showing an ambient temperature of 25°C and humidity of 60%; and image data showing a localized hotspot temperature of 85°C in the gearbox, as indicated by an infrared thermogram.

[0079] This is used to perform multi-source feature extraction on the monitoring data and the associated data to obtain a monitoring feature set and an associated feature level, to extract key information from the monitoring feature set and the associated feature level to obtain key monitoring information, and to perform consistency comparison based on the key monitoring information to obtain information contrast. If the information contrast is higher than the consistency threshold, or lower than the consistency threshold and is obviously abnormal, then the output is problematic data; otherwise, the output is ambiguous data.

[0080] In the actual assessment, key monitoring information included: a vibration peak deviation rate of 260%, marked as a signal of equipment degradation; a joint difference of 0.72, a vibration characteristic contribution weight of 0.68, and an oil temperature of 0.45°C; the core anomaly set was the vibration peak value; and the composite fault combination was the oil temperature anomaly.

[0081] Consistency comparison: The time warping offset is 0.42, the core anomaly set information similarity is 0.85, the composite fault contribution is 0.45, the CCI is 0.95, the consistency threshold is 0.8, and the vibration anomaly is obvious, so it is judged as problematic data;

[0082] Problem data risk identification and joint decomposition: The knowledge graph locates the bearing outer ring fault characteristic frequency as 85Hz. Spectral analysis shows an increase in amplitude at 85Hz and significant sideband energy. The probability of bearing outer ring fault is 0.82, and the probability of gear wear is 0.2. The weighted fusion result is outer ring fault (overall confidence level 0.84). The risk exceedance factor is 2.25, and the risk matrix identifies high risk (growth trend + exceedance factor > 2). Joint decomposition is EMD decomposition of the vibration signal, selecting the IMF3 component (containing the 85Hz component). After blind source separation, the risk dimensionality reduction feature vector is obtained [0.72, 0.31, 0.58].

[0083] The fuzzy data is subjected to time-series anomaly relationship feature extraction and joint decomposition to obtain fuzzy features; the problem data is subjected to risk identification and joint decomposition to obtain risk dimensionality reduction features; and the fuzzy features and the risk dimensionality reduction features are coupled according to sample units to obtain sample risk features.

[0084] In the actual evaluation, the fuzzy feature extraction results are as follows: local focus is 0.62, global deviation is 0.41, fuzzy causality is 0.33, seasonal gradient pattern is 0.21, trend component is 0.28, seasonal component is 0.15, and residual component is 0.09; the fuzzy feature vector is [0.33, 0.21, 0.28, 0.15, 0.09].

[0085] Feature coupling: fuzzy features are [0.33, 0.21, 0.28, 0.15, 0.09]; risk dimensionality reduction features are [0.72, 0.31, 0.58]; the gating value is 0.67; the sample risk features are [0.48, 0.27, 0.45].

[0086] Based on the risk characteristics of the samples, a risk identification model for industrial intelligent monitoring products based on graph neural networks is constructed. The data to be identified is input into the risk identification model for industrial intelligent monitoring products, and the risk identification result is output.

[0087] In actual assessment, the monitoring unit node in the risk identification model of industrial intelligent monitoring products represents the sample risk characteristics, the subsystem node represents the gearbox transmission chain, and the edge relationship represents the physical connection (gear-bearing). The meta-path is monitoring unit → transmission chain → gearbox, with a risk probability of 0.89. The identification result is high risk, and it is recommended to check the outer ring of the bearing immediately.

[0088] In this embodiment, the method for extracting key monitoring information from the monitoring feature set and the associated feature level includes:

[0089] Extracting differences in monitoring feature sets: For the same sample and the same feature data in the time series, the sliding window dynamic time warping algorithm is used to align the historical and current time series curves, quantify the self-offset, and obtain the degradation deviation rate through the current feature value, rolling baseline value and running time. The degradation rate is output as key monitoring information; when the deviation is less than 3%, it is marked as environmental interference, and when the deviation is greater than or equal to 8%, it is marked as equipment degradation signal.

[0090] For the same sample with different features in a time series, the joint difference between different data is calculated by combining the Claus-Leibler divergence and Jaccard similarity distance. A temperature and humidity correction factor is introduced to correct the cross-data table bias rate of the same data at different times. The weighted K-means clustering algorithm is used to cluster the cross-data, and the feature contribution weights of the cross-data are calculated.

[0091] ;

[0092] ;

[0093] in Let be the sum of the within-class variances of the i-th dimension feature. For hyperparameters, The number of feature dimensions. Let be the sum of the within-class variances of the j-th dimension feature. Let x be the i-th feature value. The contribution weight for the i-th dimension feature. The number of clusters in the clustering. For cluster index, For the k-th cluster, It is the i-th eigenvalue of the k-th cluster center;

[0094] Features contributing more than 61% of the joint differences are output as the core anomaly set, and features contributing less than 50% are output as the composite fault combination.

[0095] Key monitoring information includes degradation rate, core anomaly set, and complex fault combinations.

[0096] In this embodiment, the method for obtaining information contrast by performing consistency comparison based on the key monitoring information includes:

[0097] Obtain the historical baseline of degradation rate, align the historical baseline and real-time degradation rate to calculate dynamic time warping offset, and use p-value hypothesis testing to test the dynamic time warping offset to obtain the tested time warping offset; the null hypothesis of the p-value hypothesis test is that the degradation rate is less than or equal to the fluctuation range of the historical baseline, and p is greater than 0.05; the alternative hypothesis is that degradation is significantly accelerated, and p is less than or equal to 0.05.

[0098] A multimodal association model is used to align the distribution of core anomaly sets and calculate the information similarity of core anomaly information:

[0099] ;

[0100] ;

[0101] in Core abnormal element P and Q Information contrast between the two P As the core abnormal element, Q As another core abnormal element, H Core abnormal element P and Q Average distribution between To extract from the core abnormal element P to the average distribution H The relative entropy divergence, To extract from the core abnormal element Q to the average distribution H The relative entropy divergence;

[0102] Construct a fault contribution radar chart based on the composite fault combination elements to obtain the contribution of the composite fault combination elements;

[0103] Information contrast is calculated based on time warping offset, information similarity, and contribution:

[0104] ;

[0105] in For information contrast, This is the time warping offset. The contribution of the composite fault combination elements. The degradation rate weighting coefficient, The core anomaly weighting coefficient, For composite fault weighting coefficients.

[0106] In this embodiment, the method for extracting temporal anomaly relationship features and jointly decomposing the fuzzy data includes:

[0107] By combining correlation differences with fuzzy computation to quantify anomaly sensitivity, fuzzy data is input, and an attention mechanism is used to calculate dynamic correlation weights.

[0108] The association strength between fuzzy data and the global sequence is learned through a self-attention mechanism. If the association weights of outliers are concentrated in the neighborhood, the local focus is extracted. The Jensen-Shannon divergence between the current input attention distribution and the normal attention distribution is calculated. When the Jensen-Shannon divergence is greater than 0.25, it is marked as an outlier and the global deviation is extracted. The local focus and global deviation are used as temporal outlier relationship features.

[0109] The causal weights of temporal anomaly relationship features are obtained. When the causal weight is greater than the causal threshold, a weighted fuzzy rule is used to enhance the fuzzy causal strength. When the Mahalanobis distance is greater than the distance threshold, fuzzy C-means cluster center distance and membership degree correction are used to enhance the fuzzy cluster outlier. When the path deviation is greater than the cost threshold, fuzzy dynamic time warping is used to align historical cycles to enhance the seasonal gradient pattern. The fuzzy relationship features include fuzzy causal strength, fuzzy cluster outlier, and seasonal gradient pattern.

[0110] Fuzzy trend decomposition, fuzzy seasonal decomposition, and fuzzy residual decomposition are performed on the fuzzy relation features to obtain the trend component, seasonal component, and residual component, as expressed in the following expression:

[0111] ;

[0112] ;

[0113] ;

[0114] in Let be the fuzzy relation feature at time t. Let be the membership degree at time t. The intercept is... The linear drift coefficient is... For the ridge regression regularization strength, To set an upper limit for monitoring time, Let c be the frequency value of the c-th frequency component. , For the c-th triangular fuzzy number to be estimated, Let t be the order of the Fourier series, and t be the current time. Let be the fuzzy residual component at time t. Let be the seasonal component at time t. Let be the high-dimensional residual at time t. For the fuzzy data at time t, Let be the estimated value of the trend component at time t. For model parameter vectors, ;

[0115] The trend component, seasonal component, and residual component are output as fuzzy features, and the time series anomaly relationship features are labeled.

[0116] In this embodiment, the method for risk identification and joint decomposition of the problem data includes:

[0117] Problem data is input into the physical mechanism and hybrid integral model. A knowledge graph is constructed based on the physical structure of the gearbox, component information, and corresponding topological relationships. The meshing frequency of each gear stage and the fault characteristic frequency of the bearing are calculated based on the mechanical model. The meshing frequency, fault characteristic frequency, and sideband theoretical calculation formulas are stored in the database. The spectrum of vibration signals in the problem data is analyzed. The physical mechanism and hybrid integral model use the knowledge graph to quickly locate the frequency components with abnormally high energy. The fault characteristic frequencies include the inner ring, outer ring, rolling elements, and cage.

[0118] Guided by the knowledge graph, the amplitude, kurtosis factor, and sideband energy related to the theoretical fault characteristic frequencies are extracted from the spectrum of the problem data to generate a preliminary list of fault hypotheses and calculate the physical confidence score of the hypotheses.

[0119] Meanwhile, the problem data is input into a pre-trained one-dimensional deep convolutional neural network model, which outputs a multi-class probability vector; where the multi-class probability vector represents the data-driven probability of belonging to each fault type.

[0120] The physical confidence score and data-driven probability are weighted and fused to generate fault diagnosis results, and the comprehensive confidence score is labeled.

[0121] Based on the fault diagnosis results, fault features are extracted, the risk exceedance multiple is calculated based on the fault features and the baseline threshold, the growth trend of fault features within the time window is analyzed, and the risk level is obtained by locating fault features in a two-dimensional risk matrix by combining the risk exceedance multiple and the growth trend. Fault features with a risk level greater than the risk threshold are taken as target features.

[0122] Vibration signals in multi-channel target features are decomposed into intrinsic mode functions (IMFs). Based on physical information, a hybrid multi-fault decomposition and diagnostic model is used to identify fault characteristic frequencies. IMF components are selected as effective components. The effective components are combined and blind source separation technology is used to obtain mutually independent source signals. The source signals are jointly decomposed and output as risk dimensionality reduction features. The source signals contain fault characteristic information.

[0123] In this embodiment, the method for coupling the fuzzy features and the risk dimensionality reduction features on a sample unit basis includes:

[0124] For each sample, the fuzzy feature vector and the risk dimensionality reduction feature vector are batch normalized and mapped to similar distribution intervals. The dimensions of the fuzzy feature vector and the risk dimensionality reduction feature vector are unified through a fully connected layer to obtain the first fuzzy feature and the first risk feature.

[0125] Using the first fuzzy feature as the query and the first risk feature as the key and value, calculate the attention weight matrix to obtain the second fuzzy feature affected by the risk feature; using the first risk feature as the query and the first fuzzy feature as the key and value, calculate the attention weight to obtain the second risk feature affected by the fuzzy feature.

[0126] The second fuzzy feature and the second risk feature are concatenated, and then passed through a fully connected layer and a sigmoid activation function to generate a gate value between 0 and 1, expressed as:

[0127] ;

[0128] in This is the gate value. This is the second fuzzy feature. This is the second risk characteristic. This is the weight matrix. For bias terms;

[0129] The second fuzzy feature and the second risk feature are weighted and summed using the gate value to generate the sample risk feature, expressed as follows:

[0130] ;

[0131] in As a sample risk characteristic, This is for element-wise multiplication.

[0132] In this embodiment, the method for constructing an industrial intelligent monitoring product risk identification model based on graph neural networks according to the sample risk characteristics includes:

[0133] The entire industrial monitoring system is constructed as a heterogeneous graph. The node set includes monitoring unit nodes and subsystem nodes; the edge set represents the edge relationships between nodes; and the meta-path is a semantically defined meta-path based on node and edge relationships. Among them, the monitoring unit nodes are the sample risk characteristics of the monitored smart products, and the subsystem nodes are functional units. The edge relationships include physical connection relationships, functional dependency relationships, and membership relationships.

[0134] The risk identification model for industrial intelligent monitoring products includes a node internal and external feature enhancement module, a system dynamic representation module, a feature fusion and heterogeneous graph relationship aggregation module, and risk discrimination and model training.

[0135] The node internal and external feature enhancement module learns attention vectors for each node type, maps nodes to the corresponding attention vectors according to node type, and adaptively weights the original features of the nodes.

[0136] The system dynamic characterization module takes the sample risk characteristics of multiple time steps within each monitoring unit as input, uses the Transformer encoder to capture long-term dependencies in the sequence, introduces a temporal attention layer, automatically focuses on the key time step with the greatest risk evolution, and generates a global dynamic characterization.

[0137] The feature fusion and heterogeneous graph relationship aggregation module includes feature fusion and heterogeneous graph relationship aggregation. The heterogeneous graph relationship aggregation includes node-level attention and semantic-level attention. Feature fusion concatenates static enhanced features and dynamic temporal features and passes them through a fully connected layer to generate the initial embedding of the node. Node-level attention uses a graph attention network to calculate the attention coefficients of the node on the neighbors of the metapath and aggregates them to obtain the metapath-specific embedding of the node. Semantic-level attention introduces a semantic-level attention vector, calculates the global importance weight of the metapath in the risk identification task, and weights and fuses all metapath-specific embeddings to obtain the final embedding.

[0138] Risk discrimination and model training include a risk discriminator and a loss function; the risk discriminator takes the final embedding of the node as input to a multilayer perceptron, connects it to a sigmoid activation function, and outputs the risk probability of the monitoring unit; the loss function adopts focus loss.

[0139] Secondly, a big data-based intelligent product risk identification system includes:

[0140] The data acquisition and processing module is used to acquire monitoring data and related data of a preset wind turbine gearbox in the target industrial scenario, preprocess the monitoring data and related data, and establish a synchronization time mechanism. The monitoring data includes vibration data, process data, and online oil data. The related data includes maintenance history data, environmental data, offline report data, inspection data, and image data.

[0141] Risk classification module: used to perform multi-source feature extraction on the monitoring data and the associated data to obtain monitoring feature set and associated feature level, extract key information from the monitoring feature set and the associated feature level to obtain key monitoring information, and perform consistency comparison on the key monitoring information to obtain information contrast. If the information contrast is higher than the consistency threshold, or lower than the consistency threshold and obviously abnormal, the output is problem data; otherwise, the output is fuzzy data.

[0142] Dimensionality reduction coupling module: used to extract time-series anomaly relationship features and jointly decompose the fuzzy data to obtain fuzzy features, to identify risks and jointly decompose the problem data to obtain risk dimensionality reduction features, and to couple the fuzzy features and the risk dimensionality reduction features according to sample units to obtain sample risk features;

[0143] Modeling output module: used to construct a risk identification model for industrial intelligent monitoring products based on graph neural networks according to the risk characteristics of the samples, input the data to be identified into the risk identification model for industrial intelligent monitoring products, and output the risk identification results.

[0144] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying risks in intelligent products based on big data, characterized in that, Includes the following steps: The system collects monitoring data and related data of a pre-set wind turbine gearbox within the target industrial scenario, preprocesses the monitoring data and related data, and establishes a synchronization time mechanism. The monitoring data includes vibration data, process data, and online oil data. The related data includes maintenance history data, environmental data, offline report data, inspection data, and image data. This is used to perform multi-source feature extraction on the monitoring data and the associated data to obtain a monitoring feature set and an associated feature level, extract key information from the monitoring feature set and the associated feature level to obtain key monitoring information, and perform consistency comparison based on the key monitoring information to obtain information contrast. If the information contrast is higher than the consistency threshold or lower than the consistency threshold and is obviously abnormal, the output is problematic data; otherwise, the output is ambiguous data. The fuzzy data undergoes temporal anomaly relationship feature extraction and joint decomposition to obtain fuzzy features; the problematic data undergoes risk identification and joint decomposition to obtain risk dimensionality reduction features; and the fuzzy features and risk dimensionality reduction features are coupled by sample unit to obtain sample risk features; including: Problem data is input into the physical mechanism and hybrid integral model. A knowledge graph is constructed based on the physical structure of the gearbox, component information, and corresponding topological relationships. The meshing frequency of each gear stage and the fault characteristic frequency of the bearing are calculated based on the mechanical model. The meshing frequency, fault characteristic frequency, and sideband theoretical calculation formulas are stored in the database. The spectrum of vibration signals in the problem data is analyzed. The physical mechanism and hybrid integral model use the knowledge graph to quickly locate the frequency components with abnormally high energy. The fault characteristic frequencies include the inner ring, outer ring, rolling elements, and cage. Guided by the knowledge graph, the amplitude, kurtosis factor, and sideband energy related to the theoretical fault characteristic frequencies are extracted from the spectrum of the problem data to generate a preliminary list of fault hypotheses and calculate the physical confidence score of the hypotheses. Meanwhile, the problem data is input into a pre-trained one-dimensional deep convolutional neural network model, which outputs a multi-class probability vector; where the multi-class probability vector represents the data-driven probability of belonging to each fault type. The physical confidence score and data-driven probability are weighted and fused to generate fault diagnosis results, and the comprehensive confidence score is labeled. Based on the fault diagnosis results, fault features are extracted, the risk exceedance multiple is calculated based on the fault features and the baseline threshold, the growth trend of fault features within the time window is analyzed, and the risk level is obtained by locating fault features in a two-dimensional risk matrix by combining the risk exceedance multiple and the growth trend. Fault features with a risk level greater than the risk threshold are taken as target features. Vibration signals in multi-channel target features are decomposed into intrinsic mode functions (IMFs). Fault characteristic frequencies are identified by fusing a hybrid multi-fault decomposition diagnostic model based on physical information. IMF components are selected as effective components. The effective components are combined and blind source separation technology is used to obtain mutually independent source signals. The source signals are jointly decomposed and output as risk dimensionality reduction features. The source signals contain fault characteristic information. Based on the risk characteristics of the samples, a risk identification model for industrial intelligent monitoring products based on graph neural networks is constructed. The data to be identified is input into the risk identification model for industrial intelligent monitoring products, and the risk identification results are output.

2. The method for identifying intelligent product risks based on big data according to claim 1, characterized in that, A method for extracting key monitoring information from the monitoring feature set and the associated feature level includes: Extracting differences in monitoring feature sets: For the same sample and the same feature data in the time series, the sliding window dynamic time warping algorithm is used to align the historical and current time series curves, quantify the self-offset, and obtain the degradation deviation rate through the current feature value, rolling baseline value and running time. The degradation rate is output as key monitoring information; when the deviation is less than 3%, it is marked as environmental interference, and when the deviation is greater than or equal to 8%, it is marked as equipment degradation signal. For the same sample with different features in a time series, the joint difference between different data is calculated by combining the Claus-Leibler divergence and Jaccard similarity distance. A temperature and humidity correction factor is introduced to correct the cross-data table bias rate of the same data at different times. The weighted K-means clustering algorithm is used to cluster the cross-data, and the feature contribution weights of the cross-data are calculated. ； ； in Let be the sum of the within-class variances of the i-th dimension feature. For hyperparameters, The number of feature dimensions. Let be the sum of the within-class variances of the j-th dimension feature. Let x be the i-th feature value. The contribution weight for the i-th dimension feature. The number of clusters in the clustering. For cluster index, For the k-th cluster, It is the i-th eigenvalue of the k-th cluster center; Features contributing more than 61% of the joint differences are output as the core anomaly set, and features contributing less than 50% are output as the composite fault combination. Key monitoring information includes degradation rate, core anomaly set, and complex fault combinations.

3. The method for identifying intelligent product risks based on big data according to claim 1, characterized in that, A method for obtaining information contrast by performing consistency comparison based on the key monitoring information includes: Obtain the historical baseline of degradation rate, align the historical baseline and real-time degradation rate to calculate dynamic time warping offset, and use p-value hypothesis testing to test the dynamic time warping offset to obtain the tested time warping offset; the null hypothesis of the p-value hypothesis test is that the degradation rate is less than or equal to the fluctuation range of the historical baseline, and p is greater than 0.05; the alternative hypothesis is that degradation is significantly accelerated, and p is less than or equal to 0.

05. A multimodal association model is used to align the distribution of core anomaly sets and calculate the information similarity of core anomaly information: ； ； in Core abnormal element and Information contrast between the two As the core abnormal element, As another core abnormal element, Core abnormal element and Average distribution between To extract from the core abnormal element The relative entropy divergence to the average distribution H, To extract from the core abnormal element The relative entropy divergence to the average distribution H; Construct a fault contribution radar chart based on the composite fault combination elements to obtain the contribution of the composite fault combination elements; Information contrast is calculated based on time warping offset, information similarity, and contribution: ； in For information contrast, This is the time warping offset. The contribution of the composite fault combination elements. The degradation rate weighting coefficient, The core anomaly weighting coefficient, For composite fault weighting coefficients.

4. The method for identifying intelligent product risks based on big data according to claim 1, characterized in that, A method for extracting and jointly decomposing temporal anomaly relationship features from the fuzzy data includes: By combining correlation differences with fuzzy computation to quantify anomaly sensitivity, fuzzy data is input, and an attention mechanism is used to calculate dynamic correlation weights. The association strength between fuzzy data and the global sequence is learned through a self-attention mechanism. If the association weights of outliers are concentrated in the neighborhood, the local focus is extracted. The Jensen-Shannon divergence between the current input attention distribution and the normal attention distribution is calculated. When the Jensen-Shannon divergence is greater than 0.25, it is marked as an outlier and the global deviation is extracted. The local focus and global deviation are used as temporal outlier relationship features. The causal weights of temporal anomaly relationship features are obtained. When the causal weight is greater than the causal threshold, a weighted fuzzy rule is used to enhance the fuzzy causal strength. When the Mahalanobis distance is greater than the distance threshold, fuzzy C-means cluster center distance and membership degree correction are used to enhance the fuzzy cluster outlier. When the path deviation is greater than the deviation threshold, fuzzy dynamic time warping is used to align historical cycles to enhance the seasonal gradation pattern. The fuzzy relationship features include fuzzy causal strength, fuzzy cluster outlier, and seasonal gradation pattern. Fuzzy trend decomposition, fuzzy seasonal decomposition, and fuzzy residual decomposition are performed on the fuzzy relation features to obtain the trend component, seasonal component, and residual component, as expressed in the following expression: ； ； ； in Let be the fuzzy relation feature at time t. Let be the membership degree at time t. The intercept is... The linear drift coefficient is... For the ridge regression regularization strength, To set an upper limit for monitoring time, Let c be the frequency value of the c-th frequency component. , For the c-th triangular fuzzy number to be estimated, Let t be the order of the Fourier series, and t be the current time. Let be the fuzzy residual component at time t. Let be the seasonal component at time t. Let be the high-dimensional residual at time t. For the fuzzy data at time t, Let be the estimated value of the trend component at time t. For model parameter vectors, ; The trend component, seasonal component, and residual component are output as fuzzy features, and the time series anomaly relationship features are labeled.

5. The method for identifying intelligent product risks based on big data according to claim 1, characterized in that, A method for coupling the fuzzy features and the risk dimensionality reduction features according to sample units includes: For each sample, the fuzzy feature vector and the risk dimensionality reduction feature vector are batch normalized and mapped to similar distribution intervals. The dimensions of the fuzzy feature vector and the risk dimensionality reduction feature vector are unified through a fully connected layer to obtain the first fuzzy feature and the first risk feature. Using the first fuzzy feature as the query and the first risk feature as the key and value, calculate the attention weight matrix to obtain the second fuzzy feature affected by the risk feature; using the first risk feature as the query and the first fuzzy feature as the key and value, calculate the attention weight to obtain the second risk feature affected by the fuzzy feature. The second fuzzy feature and the second risk feature are concatenated, and then passed through a fully connected layer and a sigmoid activation function to generate a gate value between 0 and 1, expressed as: ； in This is the gate value. This is the second fuzzy feature. This is the second risk characteristic. This is the weight matrix. For bias terms; The second fuzzy feature and the second risk feature are weighted and summed using a gating value to generate the sample risk feature, expressed as follows: ； in As a sample risk characteristic, This is for element-wise multiplication.

6. The method for identifying intelligent product risks based on big data according to claim 1, characterized in that, A method for constructing a risk identification model for industrial intelligent monitoring products based on graph neural networks according to the risk characteristics of the samples includes: The entire industrial monitoring system is constructed as a heterogeneous graph. The node set includes monitoring unit nodes and subsystem nodes; the edge set represents the edge relationships between nodes; and the meta-path is a semantically defined meta-path based on node and edge relationships. Among them, the monitoring unit nodes are the sample risk characteristics of the monitored smart products, and the subsystem nodes are functional units. The edge relationships include physical connection relationships, functional dependency relationships, and membership relationships. The risk identification model for industrial intelligent monitoring products includes a node internal and external feature enhancement module, a system dynamic representation module, a feature fusion and heterogeneous graph relationship aggregation module, and risk discrimination and model training. The node internal and external feature enhancement module learns attention vectors for each node type, maps nodes to the corresponding attention vectors according to node type, and adaptively weights the original features of the nodes. The system dynamic characterization module takes the sample risk characteristics of multiple time steps within each monitoring unit as input, uses the Transformer encoder to capture long-term dependencies in the sequence, introduces a temporal attention layer, automatically focuses on the key time step with the greatest risk evolution, and generates a global dynamic characterization. The feature fusion and heterogeneous graph relationship aggregation module includes feature fusion and heterogeneous graph relationship aggregation. The heterogeneous graph relationship aggregation includes node-level attention and semantic-level attention. Feature fusion concatenates static enhanced features and dynamic temporal features and passes them through a fully connected layer to generate the initial embedding of the node. Node-level attention uses a graph attention network to calculate the attention coefficients of the node on the neighbors of the metapath and aggregates them to obtain the metapath-specific embedding of the node. Semantic-level attention introduces a semantic-level attention vector, calculates the global importance weight of the metapath in the risk identification task, and weights and fuses all metapath-specific embeddings to obtain the final embedding. Risk discrimination and model training include a risk discriminator and a loss function; the risk discriminator takes the final embedding of the node as input to a multilayer perceptron, connects it to a sigmoid activation function, and outputs the risk probability of the monitoring unit; the loss function adopts focus loss.

7. A big data-based intelligent product risk identification system, used to perform the method described in any one of claims 1-6, characterized in that, include: The data acquisition and processing module is used to acquire monitoring data and related data of a preset wind turbine gearbox in the target industrial scenario, preprocess the monitoring data and related data, and establish a synchronization time mechanism. The monitoring data includes vibration data, process data, and online oil data. The related data includes maintenance history data, environmental data, offline report data, inspection data, and image data. Risk classification module: used to perform multi-source feature extraction on the monitoring data and the associated data to obtain monitoring feature set and associated feature level, extract key information from the monitoring feature set and the associated feature level to obtain key monitoring information, and perform consistency comparison based on the key monitoring information to obtain information contrast. If the information contrast is higher than the consistency threshold or lower than the consistency threshold and is obviously abnormal, the output is problem data; otherwise, the output is fuzzy data. Dimensionality Reduction Coupling Module: Used to extract and jointly decompose temporal anomaly relationship features from the fuzzy data to obtain fuzzy features, to perform risk identification and joint decomposition on the problematic data to obtain risk dimensionality reduction features, and to couple the fuzzy features and the risk dimensionality reduction features on a sample unit basis to obtain sample risk features; including: Problem data is input into the physical mechanism and hybrid integral model. A knowledge graph is constructed based on the physical structure of the gearbox, component information, and corresponding topological relationships. The meshing frequency of each gear stage and the fault characteristic frequency of the bearing are calculated based on the mechanical model. The meshing frequency, fault characteristic frequency, and sideband theoretical calculation formulas are stored in the database. The spectrum of vibration signals in the problem data is analyzed. The physical mechanism and hybrid integral model use the knowledge graph to quickly locate the frequency components with abnormally high energy. The fault characteristic frequencies include the inner ring, outer ring, rolling elements, and cage. Guided by the knowledge graph, the amplitude, kurtosis factor, and sideband energy related to the theoretical fault characteristic frequencies are extracted from the spectrum of the problem data to generate a preliminary list of fault hypotheses and calculate the physical confidence score of the hypotheses. Meanwhile, the problem data is input into a pre-trained one-dimensional deep convolutional neural network model, which outputs a multi-class probability vector; where the multi-class probability vector represents the data-driven probability of belonging to each fault type. The physical confidence score and data-driven probability are weighted and fused to generate fault diagnosis results, and the comprehensive confidence score is labeled. Based on the fault diagnosis results, fault features are extracted, the risk exceedance multiple is calculated based on the fault features and the baseline threshold, the growth trend of fault features within the time window is analyzed, and the risk level is obtained by locating fault features in a two-dimensional risk matrix by combining the risk exceedance multiple and the growth trend. Fault features with a risk level greater than the risk threshold are taken as target features. Vibration signals in multi-channel target features are decomposed into intrinsic mode functions (IMFs). Fault characteristic frequencies are identified by fusing a hybrid multi-fault decomposition diagnostic model based on physical information. IMF components are selected as effective components. The effective components are combined and blind source separation technology is used to obtain mutually independent source signals. The source signals are jointly decomposed and output as risk dimensionality reduction features. The source signals contain fault characteristic information. Modeling output module: used to construct a risk identification model for industrial intelligent monitoring products based on graph neural networks according to the risk characteristics of the samples, input the data to be identified into the risk identification model for industrial intelligent monitoring products, and output the risk identification results.

Images