Equipment telemetry data fault analysis method and system based on machine learning

Abstract

The invention provides an equipment telemetry data fault analysis method and system based on machine learning, and relates to the technical field of fault analysis, and the method comprises the steps: carrying out the standardization processing of multi-dimensional telemetry data, extracting time sequence features and statistical features, screening a core feature set, and constructing a lightweight feature extraction rule; according to the method, feature component contribution degree scores are calculated, adaptive weight coefficients are distributed, fault feature representation is generated, and a fault diagnosis rule base is matched, so that the fault type and probability are determined, edge end data processing lightweight and timely generation of fault early warning are realized, and the accuracy and efficiency of fault diagnosis are improved.

Description

Technical Field

[0001] The present invention relates to fault analysis technology, and in particular to a method and system for fault analysis of device telemetry data based on machine learning. Background Art

[0002] With the rapid development of industrial Internet of Things technology, the real-time monitoring and fault diagnosis of device telemetry data have become key links to ensure the safe and stable operation of industrial production. A large amount of multi-dimensional telemetry data is generated during the operation of industrial devices, and this data contains device operation status and potential fault information. Traditional device fault diagnosis mainly relies on manual experience judgment. With the development of machine learning technology, data-driven intelligent fault analysis methods have gradually become a research hotspot. At present, the industrial field has begun to attempt to apply machine learning algorithms to device fault prediction and diagnosis. By learning and analyzing the historical operation data of devices, a fault warning model is established to achieve early identification of device abnormal states and accurate judgment of fault types.

[0003] Existing technologies often analyze using all features when dealing with multi-dimensional telemetry data, lacking an effective feature screening mechanism, resulting in high model calculation complexity, being difficult to operate efficiently on edge computing devices, and unable to meet the requirements of real-time fault diagnosis in industrial fields.

[0004] Traditional fault diagnosis methods do not adequately consider the importance of different features, usually using fixed weights or empirically set methods to process each feature component, unable to dynamically adjust feature weights according to the device operation status, being difficult to accurately reflect the contribution of each feature to fault judgment under different working conditions, and reducing the accuracy of fault diagnosis.

[0005] The existing fault diagnosis systems that collaborate between the edge and the cloud have low integration, the edge computing capabilities are not fully utilized, and a large amount of raw data needs to be transmitted to the cloud for processing, which not only increases the network transmission burden but also prolongs the fault response time, and cannot meet the actual requirements of rapid fault diagnosis and timely warning in industrial fields. Summary of the Invention

[0006] Embodiments of the present invention provide a method and system for fault analysis of device telemetry data based on machine learning, which can solve the problems in the existing technologies.

[0007] In a first aspect of embodiments of the present invention, a method for fault analysis of device telemetry data based on machine learning is provided, including: Receiving multi-dimensional telemetry data collected by a device telemetry data acquisition module, performing normalization processing on the multi-dimensional telemetry data to generate normalized data; Extracting the time series features and statistical features of the standardized data and establishing feature evaluation indicators; screening a core feature set according to the feature evaluation indicators; constructing a lightweight feature extraction rule based on the core feature set to achieve lightweight edge data processing; Processing the standardized data using the lightweight feature extraction rule to generate a device state feature vector, calculating a contribution score of each feature component in the device state feature vector, and assigning an adaptive weight coefficient to each feature component according to the contribution score; Based on the adaptive weight coefficient and the device state feature vector, a fault feature representation is generated; a fault diagnosis rule base is constructed, and the fault feature representation is matched with the fault diagnosis rule base; the matching probability of each type of fault is calculated, the maximum matching probability is selected as the current fault probability, and the corresponding fault type is determined; when the current fault probability exceeds a preset probability threshold, edge fault warning information is generated; The edge fault warning information, the fault type and the current fault probability are transmitted to the cloud management platform for storage and analysis.

[0008] Filtering a core feature set according to the feature evaluation index; constructing a lightweight feature extraction rule based on the core feature set to achieve lightweight edge data processing includes: Receiving equipment telemetry data to generate a multi-dimensional time series feature sequence; respectively calculating the autocorrelation coefficient of each feature in the multi-dimensional time series feature sequence and the mutual correlation coefficient between the features; Performing information entropy calculation and Fisher discriminant analysis on the multidimensional time series feature sequence to obtain the information gain value and inter-class dispersion of each feature, and performing weighted combination of the autocorrelation coefficient, the mutual correlation coefficient, the information gain value and the inter-class dispersion to obtain a feature importance score; Based on the feature importance score, the multidimensional time series feature sequence is sorted in descending order, and the features with the highest order are selected as the candidate feature set; the mutual information between the features in the candidate feature set is calculated, a feature correlation matrix is constructed, and a core feature subset is selected from the candidate feature set based on the maximum correlation and minimum redundancy criterion; Performing principal component analysis on the core feature subset, calculating the contribution rate of each principal component, and selecting a principal component combination whose cumulative contribution rate reaches a preset contribution threshold; constructing a feature extraction rule based on the principal component combination, and applying the feature extraction rule to real-time telemetry data to obtain a lightweight feature representation; The reconstruction error of the lightweight feature representation relative to the original feature is calculated. When the reconstruction error is less than a preset error threshold, the feature extraction rule is deployed to the edge computing end to achieve lightweight edge data processing.

[0009] Processing the standardized data using the lightweight feature extraction rules to generate a device state feature vector, calculating the contribution score of each feature component in the device state feature vector, and allocating an adaptive weight coefficient to each feature component according to the contribution score includes: Receiving input feature data, inputting the input feature data into a feature neural network, calculating the activation state value of each feature neuron, and calculating the corresponding neurotransmitter concentration change rate according to the activation state value, where the neurotransmitter concentration change rate is jointly determined by the neurotransmitter release rate and the decay rate; Calculating the dynamic activity threshold of the feature neuron based on the neurotransmitter concentration change rate, where the dynamic activity threshold is adaptively adjusted with the cumulative change of the neurotransmitter concentration; calculating the synaptic strength modulation factor using the difference between the dynamic activity threshold and the neurotransmitter concentration; Multiplying the input feature data, the synaptic weight coefficient, and the neurotransmitter concentration to calculate the local field response intensity, and calculating the feature importance score according to the local field response intensity and the synaptic strength modulation factor; According to the feature importance score, executing the Hebbian learning rule to calculate the increment of the synaptic weight; based on the increment of the synaptic weight, executing a feedback inhibition mechanism to calculate the inhibition amount of the synaptic weight; Superposing the increment of the synaptic weight and the inhibition amount of the synaptic weight, updating the synaptic weight coefficient, and normalizing the updated synaptic weight coefficient to obtain the final adaptive weight coefficient.

[0010] According to the feature importance score, executing the Hebbian learning rule to calculate the increment of the synaptic weight; based on the increment of the synaptic weight, executing a feedback inhibition mechanism to calculate the inhibition amount of the synaptic weight includes: Inputting the input feature data into a neural network, performing a non-linear transformation on the input feature data using a sigmoid activation function, calculating the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron, selecting the neuron with the largest activity value as the winning neuron based on a competitive learning algorithm, and dynamically adjusting the weight enhancement parameter using the activity state of the winning neuron; According to the feature importance score, executing the Hebbian learning rule, determining a basic enhancement coefficient through the feature importance score, combining the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron and the current weight saturation, calculating a weight adjustment coefficient using a dynamic threshold algorithm, and obtaining the increment of the synaptic weight according to the weight adjustment coefficient; Performing a square summation operation on the increment of the synaptic weight to obtain an enhancement energy value, and determining whether the enhancement energy value exceeds a preset energy upper limit threshold; Calculate the change rate of the increment of the synaptic weight in the time dimension, substitute the change rate, the mean parameter, and the variance parameter of the Gaussian function into the Gaussian function to construct a Gaussian inhibition function; Based on the increment of the synaptic weight, execute a feedback inhibition mechanism, calculate the inhibition intensity coefficient using the output value of the Gaussian inhibition function, and combine the inhibition learning rate, the current synaptic weight, and the neuron inactivation degree to obtain the inhibition amount of the synaptic weight.

[0011] Based on the adaptive weight coefficient and the device state feature vector, generate a fault feature representation; construct a fault diagnosis rule base, and match the fault feature representation with the fault diagnosis rule base, including: Multiply the device state feature vector by the adaptive weight coefficient to generate an initial fault feature representation; construct a device fault knowledge graph, process the device fault knowledge graph using a graph neural network, learn the feature representations of each node, and extract the main diagnosis path based on the feature representations; Construct an initial fault diagnosis rule base according to the main diagnosis path, match the initial fault feature representation with the rules in the initial fault diagnosis rule base to obtain the rule matching effect; Calculate the confidence score of each rule based on the rule matching effect, dynamically adjust the rule weight according to the confidence score, and delete the rules with confidence scores lower than the preset confidence threshold; Generate new rule candidates using the device fault knowledge graph, supplement the rules with confidence scores higher than the preset confidence threshold in the rule candidates to the initial fault diagnosis rule base to obtain an optimized fault diagnosis rule base; perform a final match between the initial fault feature representation and the optimized fault diagnosis rule base, and output the fault diagnosis result.

[0012] Construct an initial fault diagnosis rule base according to the main diagnosis path, and match the initial fault feature representation with the rules in the initial fault diagnosis rule base to obtain the rule matching effect, including: Use the multi-head attention mechanism to extract features from the main diagnosis path to generate a query matrix, a key matrix, and a value matrix, obtain the initial attention weight through the inner product operation of the query matrix and the key matrix, and multiply the initial attention weight by the value matrix to obtain an enhanced feature representation; Construct a generative adversarial network model, the generator of the generative adversarial network model generates candidate rules based on the enhanced feature representation, the candidate rules include a fault feature vector, a temporal dependence relationship, a fault category encoding, and a rule confidence, and the discriminator of the generative adversarial network model calculates the authenticity score of the candidate rules; Screen the candidate rules based on the authenticity score, and retain the candidate rules with authenticity scores higher than the preset authenticity threshold as valid candidate rules; combine the main diagnostic path with the valid candidate rules to form an initial diagnostic rule, and calculate the effectiveness score of the initial diagnostic rule based on the authenticity score; Construct a time-varying fault feature map, extract features from the time-varying fault feature map using spatio-temporal convolution, and generate a new hidden layer state by combining the historical hidden layer state and the current input features; calculate the matching degree between the new hidden layer state and the initial diagnostic rule based on the attention mechanism; Perform weighted fusion on the effectiveness score and the matching degree to obtain a rule matching score, count the number of rules whose rule matching scores exceed the preset matching threshold, and obtain the rule matching effect.

[0013] Calculate the matching probabilities of various types of faults, select the maximum matching probability as the current fault probability, and determine the corresponding fault type; when the current fault probability exceeds the preset probability threshold, generate edge-side fault warning information including: Obtain fault feature data, and perform standardization processing on the fault feature data to obtain standardized feature data; Calculate the membrane potential change rate of the standardized feature data, which is obtained through the weighted combination of the ratio of the standardized feature data to the membrane time constant, the synaptic input current, and the external input current, and calculate the average membrane potential of the neuron group according to the membrane potential change rate; Multiply the action potential of the previous neuron by the action potential of the subsequent neuron to obtain the synaptic activity degree, multiply the difference between the synaptic activity degree and the plasticity threshold by the learning rate to obtain the change amount of the synaptic weight, and update the synaptic weight according to the change amount of the synaptic weight; Calculate the firing probability of the neuron group based on the difference between the average membrane potential and the firing threshold, and perform weighted combination on the firing probability and the prior probability statistically obtained from historical fault data through the updated synaptic weight to obtain the fault matching probability; Perform normalization processing on the fault matching probability to obtain a normalized matching probability, select the maximum value of the normalized matching probability as the current fault probability, and determine the corresponding fault type according to the current fault probability; Compare the current fault probability with the preset probability threshold, and when the current fault probability exceeds the preset probability threshold, generate edge-side fault warning information according to the fault type, the current fault probability, and the current timestamp.

[0014] In the second aspect of the embodiments of the present invention, a device telemetry data fault analysis system based on machine learning is provided, including: The first unit is used to receive the multi-dimensional telemetry data collected by the device telemetry data acquisition module, perform standardization processing on the multi-dimensional telemetry data, and generate standardized data; The second unit is used to extract the temporal characteristics and statistical characteristics of the standardized data, establish feature evaluation indicators; screen the core feature set according to the feature evaluation indicators; construct lightweight feature extraction rules based on the core feature set to realize lightweight edge-side data processing; The third unit is used to process the standardized data using the lightweight feature extraction rules, generate a device status feature vector, calculate the contribution score of each feature component in the device status feature vector, and assign an adaptive weight coefficient to each feature component according to the contribution score; The fourth unit is used to generate a fault feature representation based on the adaptive weight coefficient and the device status feature vector; construct a fault diagnosis rule library, and match the fault feature representation with the fault diagnosis rule library; calculate the matching probabilities of various faults, select the maximum matching probability as the current fault probability, and determine the corresponding fault type; when the current fault probability exceeds a preset probability threshold, generate an edge-side fault warning message; The fifth unit is used to transmit the edge-side fault warning message, the fault type, and the current fault probability to the cloud management platform for storage and analysis.

[0015] In the third aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0016] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0017] The beneficial effects of this application are as follows: By performing standardization processing on multi-dimensional telemetry data, extracting temporal characteristics and statistical characteristics, establishing feature evaluation indicators to screen the core feature set, and constructing lightweight feature extraction rules, the present invention realizes lightweight edge-side data processing, reduces computational complexity and resource consumption, and improves the processing efficiency of edge-side devices.

[0018] The present invention calculates the contribution score of each feature component in the device status feature vector, assigns an adaptive weight coefficient accordingly, and generates a fault feature representation, realizing an accurate expression of the device status features under different working conditions, improving the accuracy and reliability of fault diagnosis, and reducing the false alarm rate.

[0019] The present invention constructs a fault diagnosis rule base, calculates the matching probabilities of various faults, generates edge - side fault warning information when the current fault probability exceeds a preset threshold, and transmits relevant information to the cloud platform, realizing the collaborative work between the edge side and the cloud. It can not only timely detect and warn of abnormal device states, but also perform in - depth analysis and optimization through the cloud platform, improving the reliability and intelligence level of the entire system. Brief Description of the Drawings

[0020] Figure 1 It is a schematic flowchart of the method for fault analysis of device telemetry data based on machine learning in an embodiment of the present invention; Figure 2 It is a flowchart of the edge - side lightweight feature extraction method based on multi - dimensional time - series features in an embodiment of the present invention; Figure 3 It is a bar chart comparing the performance of neural networks based on Hebbian learning rules and feedback inhibition mechanisms in an embodiment of the present invention; Figure 4 It is a flowchart of the fault diagnosis rule matching method based on multi - head attention and GAN in an embodiment of the present invention; Figure 5 It is a bar chart comparing the performance of the edge - side fault warning system based on neuron models in an embodiment of the present invention. Detailed Embodiments

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0022] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0023] Figure 1 It is a schematic flowchart of the method for fault analysis of device telemetry data based on machine learning in an embodiment of the present invention, as Figure 1 shown, the method includes: Receiving multi - dimensional telemetry data collected by a device telemetry data acquisition module, performing normalization processing on the multi - dimensional telemetry data to generate normalized data; Extracting the time series features and statistical features of the standardized data and establishing feature evaluation indicators; screening a core feature set according to the feature evaluation indicators; constructing a lightweight feature extraction rule based on the core feature set to achieve lightweight edge data processing; Processing the standardized data using the lightweight feature extraction rule to generate a device state feature vector, calculating a contribution score of each feature component in the device state feature vector, and assigning an adaptive weight coefficient to each feature component according to the contribution score; Based on the adaptive weight coefficient and the device state feature vector, a fault feature representation is generated; a fault diagnosis rule base is constructed, and the fault feature representation is matched with the fault diagnosis rule base; the matching probability of each type of fault is calculated, the maximum matching probability is selected as the current fault probability, and the corresponding fault type is determined; when the current fault probability exceeds a preset probability threshold, edge fault warning information is generated; The edge fault warning information, the fault type and the current fault probability are transmitted to the cloud management platform for storage and analysis.

[0024] In an optional implementation, screening a core feature set according to the feature evaluation index; constructing a lightweight feature extraction rule based on the core feature set to achieve lightweight edge data processing includes: Receiving equipment telemetry data to generate a multi-dimensional time series feature sequence; respectively calculating the autocorrelation coefficient of each feature in the multi-dimensional time series feature sequence and the mutual correlation coefficient between the features; Performing information entropy calculation and Fisher discriminant analysis on the multidimensional time series feature sequence to obtain the information gain value and inter-class dispersion of each feature, and performing weighted combination of the autocorrelation coefficient, the mutual correlation coefficient, the information gain value and the inter-class dispersion to obtain a feature importance score; Based on the feature importance score, the multidimensional time series feature sequence is sorted in descending order, and the features with the highest order are selected as the candidate feature set; the mutual information between the features in the candidate feature set is calculated, a feature correlation matrix is constructed, and a core feature subset is selected from the candidate feature set based on the maximum correlation and minimum redundancy criterion; Performing principal component analysis on the core feature subset, calculating the contribution rate of each principal component, and selecting a principal component combination whose cumulative contribution rate reaches a preset contribution threshold; constructing a feature extraction rule based on the principal component combination, and applying the feature extraction rule to real-time telemetry data to obtain a lightweight feature representation; The reconstruction error of the lightweight feature representation relative to the original feature is calculated. When the reconstruction error is less than a preset error threshold, the feature extraction rule is deployed to the edge computing end to achieve lightweight edge data processing.

[0025] likeFigure 2 As shown, the method includes: Receiving telemetry data from devices, which includes information collected by various sensors such as temperature, pressure, vibration frequency, etc. For a certain industrial device, temperature data, pressure data, and vibration data sampled 10 times per second are collected to form a three-dimensional time-series feature sequence. For example, during one hour of monitoring, three time series each containing 36,000 data points will be generated.

[0026] For the generated multi-dimensional time-series feature sequence, calculate the autocorrelation coefficient of each feature and the cross-correlation coefficient between features. The autocorrelation coefficient reflects the correlation of a single feature at different time points, and the cross-correlation coefficient indicates the degree of association between different features. Taking the two features of temperature and pressure as an example, the calculated autocorrelation coefficient of temperature is 0.85, indicating that the temperature data has strong temporal continuity; the cross-correlation coefficient between temperature and pressure is 0.62, indicating a moderate degree of correlation between these two features.

[0027] Subsequently, perform information entropy calculation and Fisher discriminant analysis on the multi-dimensional time-series feature sequence. Through information entropy calculation, the information gain value of each feature can be obtained to measure the amount of information contained in the feature. Through Fisher discriminant analysis, the between-class scatter can be obtained, which is used to evaluate the discrimination ability of the feature between different classes. In practical applications, the calculated information gain value of the temperature feature is 0.75, and the between-class scatter is 0.68; the information gain value of the pressure feature is 0.62, and the between-class scatter is 0.55; the information gain value of the vibration feature is 0.83, and the between-class scatter is 0.72.

[0028] Combine the autocorrelation coefficient, cross-correlation coefficient, information gain value, and between-class scatter through weighted combination to obtain the feature importance score. The weighting method can be adjusted according to the actual application scenario. For example, the weights of the four indicators can be set to 0.2, 0.2, 0.3, and 0.3 respectively. According to this weight configuration, the calculated importance scores of the three features of temperature, pressure, and vibration are 0.735, 0.595, and 0.765 respectively.

[0029] Based on the feature importance score, sort the multi-dimensional time-series feature sequence in descending order, and select the features with higher rankings as the candidate feature set. In the above example, the vibration and temperature features will be selected into the candidate feature set. If there are more features in the system, select the top 5 or top 10 features with the highest scores to form the candidate set.

[0030] Calculate the mutual information between the features in the candidate feature set and construct a feature correlation matrix. Mutual information measures the amount of information shared between two features, and the higher the value, the greater the redundancy between the features. In the example of vibration and temperature, assume that the calculated mutual information value is 0.42, forming a 2×2 correlation matrix.

[0031] The core feature subset is screened from the candidate feature set based on the maximum correlation and minimum redundancy criterion. This criterion aims to select the feature combination that is most relevant to the target variable and has the minimum redundancy among themselves. In actual operation, the features that meet the conditions are added one by one in an iterative manner. If there are 5 features in the candidate set, 3 core features that are non-redundant and rich in information are finally screened out.

[0032] Principal component analysis is performed on the core feature subset to calculate the contribution rate of each principal component. Principal component analysis transforms the original features into mutually orthogonal principal components, and each principal component represents a certain proportion of the variance of the original data. Suppose principal component analysis is performed on three core features, and three principal components PC1, PC2, and PC3 are obtained, and their contribution rates are 65%, 25%, and 10% respectively.

[0033] Select the combination of principal components whose cumulative contribution rate reaches the preset contribution threshold. If the contribution threshold is set to 85%, then PC1 and PC2 are selected because their cumulative contribution rate is 90%. Based on these principal components, a feature extraction rule is constructed, and the specific form is the mapping relationship from the original features to the principal component space.

[0034] The constructed feature extraction rule is applied to the real-time telemetry data to obtain a lightweight feature representation. These feature representations have lower dimensions but retain the main information of the original data. Suppose the original data has 10 dimensions and is reduced to 3 dimensions through the feature extraction rule, significantly reducing the data processing and transmission burden.

[0035] Calculate the reconstruction error of the lightweight feature representation relative to the original features. The reconstruction error measures the information loss during the feature extraction process, and the smaller the error, the more complete the retained information. On a certain test data set, the calculated reconstruction error is 0.072. If the preset error threshold is 0.1, then this feature extraction rule is considered acceptable.

[0036] When the reconstruction error is less than the preset error threshold, the feature extraction rule is deployed to the edge computing end to realize lightweight data processing at the edge. By deploying these rules, edge devices can preprocess and reduce the dimension of the original sensor data locally, and only transmit the lightweight features with high information content, significantly reducing the communication bandwidth requirement and improving the system response speed.

[0037] In an alternative embodiment, the standardized data is processed using the lightweight feature extraction rule to generate a device state feature vector, the contribution score of each feature component in the device state feature vector is calculated, and an adaptive weight coefficient is assigned to each feature component according to the contribution score, including: Receive input feature data, input the input feature data into a feature neural network, calculate the activation state value of each feature neuron, and calculate the corresponding neurotransmitter concentration change rate according to the activation state value. The neurotransmitter concentration change rate is jointly determined by the neurotransmitter release rate and the decay rate; Calculate the dynamic activity threshold of the feature neuron based on the neurotransmitter concentration change rate. The dynamic activity threshold is adaptively adjusted with the cumulative change of the neurotransmitter concentration; calculate the synaptic strength modulation factor using the difference between the dynamic activity threshold and the neurotransmitter concentration; Multiply the input feature data, the synaptic weight coefficient, and the neurotransmitter concentration to calculate the local field response intensity, and calculate the feature importance score according to the local field response intensity and the synaptic strength modulation factor; Execute the Hebbian learning rule based on the feature importance score to calculate the increment of the synaptic weight; based on the increment of the synaptic weight, execute the feedback inhibition mechanism to calculate the inhibition amount of the synaptic weight; Superimpose the increment of the synaptic weight and the inhibition amount of the synaptic weight to update the synaptic weight coefficient, and perform normalization processing on the updated synaptic weight coefficient to obtain the final adaptive weight coefficient.

[0038] Receive input feature data, which can be, for example, multi-dimensional sensor data of industrial equipment, including indicators such as temperature, pressure, vibration, current, etc., and its dimension can reach 50 - 100 dimensions. These input feature data are input into the feature neural network for processing. Taking the temperature sensor as an example, if the original input value is 85 °C, it is converted to a normalized value of 0.78 after standardization. The feature neural network contains neurons corresponding to each input feature. For a 50-dimensional input feature vector, 50 feature neurons will be correspondingly established in the network. Each feature neuron calculates its activation state value according to the received input value. The activation state value is calculated using a non-linear mapping method. For example, when the temperature feature input is 0.78, the corresponding activation state value is 0.83.

[0039] Based on the activation state value of each feature neuron, the system calculates the corresponding neurotransmitter concentration change rate. The neurotransmitter concentration change rate is jointly determined by the neurotransmitter release rate and the decay rate. The neurotransmitter release rate is proportional to the activation state value of the feature neuron. The higher the activation state value, the greater the release rate; while the decay rate reflects the natural dissipation speed of neurotransmitters in the synaptic cleft. For example, when the activation state value of the temperature feature neuron is 0.83, the corresponding neurotransmitter release rate is 0.05, the decay rate is 0.02, and the finally calculated neurotransmitter concentration change rate is 0.03.

[0040] Calculate the dynamic activity threshold of characteristic neurons according to the change rate of neurotransmitter concentration. This threshold is not fixed, but is adaptively adjusted with the cumulative change of neurotransmitter concentration. When the neurotransmitter concentration continuously increases, the dynamic activity threshold of neurons will also increase accordingly, and vice versa. For example, if the neurotransmitter concentration corresponding to the temperature feature increases from the initial value of 0.2 to 0.5 cumulatively, its dynamic activity threshold will be up-regulated from 0.3 to 0.45. The system calculates the synaptic strength modulation factor using the difference between the dynamic activity threshold and the neurotransmitter concentration. When the neurotransmitter concentration exceeds the dynamic activity threshold, the synaptic strength is positively modulated; otherwise, it is negatively modulated. For the temperature feature, if the neurotransmitter concentration is 0.5 and the dynamic activity threshold is 0.45, the difference is 0.05, and the corresponding synaptic strength modulation factor is 1.2.

[0041] Multiply the input feature data, synaptic weight coefficient, and neurotransmitter concentration to calculate the local field response intensity. For example, for the temperature feature, if the input value is 0.78, the current synaptic weight is 0.65, and the neurotransmitter concentration is 0.5, the calculated local field response intensity is 0.78×0.65×0.5 = 0.2535. According to the local field response intensity and the synaptic strength modulation factor, the system calculates the feature importance score. The feature importance score reflects the contribution degree of a specific feature to the judgment of the current device state. Continuing the above example, if the synaptic strength modulation factor is 1.2, the importance score of the temperature feature is 0.2535×1.2 = 0.3042.

[0042] Based on the feature importance score, the system executes the Hebbian learning rule to calculate the increment of the synaptic weight. Hebbian learning embodies the principle of "the connection of neurons that are activated simultaneously is enhanced". The higher the feature importance score, the greater the increment of the synaptic weight. For example, the importance score of the temperature feature is 0.3042, and the corresponding increment of the synaptic weight is 0.018. At the same time, to prevent the weights of some features from being over-enhanced and causing other features to be ignored, the system executes a feedback inhibition mechanism based on the increment of the synaptic weight to calculate the inhibition amount of the synaptic weight. The calculation of the inhibition amount takes into account the balance of the overall network, and the inhibition amount of the synaptic weight corresponding to the temperature feature is 0.008.

[0043] Superimpose the increment and decrement of the synaptic weight, and update the synaptic weight coefficient. For the temperature feature, the new synaptic weight is 0.65 + 0.018 - 0.008 = 0.66. Finally, the system normalizes the synaptic weight coefficients of all updated features to ensure that the sum of the weights is 1, obtaining the final adaptive weight coefficients. Suppose the system processes three features: temperature, pressure, and vibration simultaneously. The updated weights are 0.66, 0.45, and 0.39 respectively. After normalization, the final adaptive weight coefficients obtained are 0.44, 0.3, and 0.26 respectively. These weight coefficients reflect the relative importance of each feature in the device state assessment and will be dynamically adjusted as the device operating state changes.

[0044] Through the above implementation manner, the present invention can dynamically adjust the weight coefficients of each feature component according to the real-time changes of the device operation data, improving the accuracy and adaptability of anomaly detection.

[0045] In an alternative implementation manner, according to the feature importance score, execute the Hebbian learning rule to calculate the increment of the synaptic weight; based on the increment of the synaptic weight, execute the feedback inhibition mechanism, and calculating the decrement of the synaptic weight includes: Input the input feature data into the neural network, perform a non-linear transformation on the input feature data using the sigmoid activation function, calculate the active states of the presynaptic neuron and the postsynaptic neuron, select the neuron with the largest activation value as the winning neuron based on the competitive learning algorithm, and dynamically adjust the weight enhancement parameter using the active state of the winning neuron; According to the feature importance score, execute the Hebbian learning rule, determine the basic enhancement coefficient through the feature importance score, combine the active states of the presynaptic neuron and the postsynaptic neuron and the current weight saturation, calculate the weight adjustment coefficient using the dynamic threshold algorithm, and obtain the increment of the synaptic weight according to the weight adjustment coefficient; Perform a square summation operation on the increment of the synaptic weight to obtain an enhancement energy value, and determine whether the enhancement energy value exceeds a preset energy upper limit threshold; Calculate the change rate of the increment of the synaptic weight in the time dimension, substitute the change rate, the mean parameter, and the variance parameter of the Gaussian function into the Gaussian function to construct a Gaussian inhibition function; Based on the increment of the synaptic weight, execute the feedback inhibition mechanism, calculate the inhibition intensity coefficient using the output value of the Gaussian inhibition function, and combine the inhibition learning rate, the current synaptic weight, and the neuron inactivation degree to obtain the decrement of the synaptic weight.

[0046] Through feature importance evaluation, then combining with the Hebbian learning rule to calculate the synaptic weight increment, and calculating the weight inhibition amount through the feedback inhibition mechanism, the dynamic adjustment of the neural network weights is realized.

[0047] Receive input feature data, such as a 10-dimensional feature vector [0.5, 0.3, 0.8, 0.2, 0.6, 0.4, 0.7, 0.1, 0.9, 0.35], which represent the original values of different features. These feature data are input into a neural network layer containing 15 neurons. For each input feature, the system uses the sigmoid activation function for non-linear transformation to map the input value to the interval (0, 1). For example, for the input value 0.5, the value after transformation by the sigmoid function is 0.622. In this way, all input features form the active state vector of the presynaptic neurons after transformation.

[0048] Calculate the active state of the postsynaptic neurons. Assume that the current synaptic weight matrix of the network is W, with dimensions 10×15, representing the connection strength between 10 input features and 15 postsynaptic neurons. The active state of the postsynaptic neurons is calculated by multiplying the input features with the corresponding weights and summing them, and then applying the sigmoid activation function. For example, for the first postsynaptic neuron, its active value is 0.75.

[0049] Based on the principle of competitive learning, the system identifies the neuron with the highest active value as the winning neuron. Assume that the active value of the third neuron is 0.88, which is the highest among all neurons, so it is selected as the winning neuron. The system dynamically adjusts the weight increment parameter according to the active state of this winning neuron. Specifically, if the active value of the winning neuron is higher than the preset threshold of 0.8, the weight increment parameter is set to a higher value of 0.05; if the active value is between 0.5 and 0.8, it is set to a medium value of 0.03; if it is lower than 0.5, it is set to a lower value of 0.01. In this example, since the active value of the winning neuron is 0.88, exceeding the threshold of 0.8, the weight increment parameter is set to 0.05.

[0050] Execute the Hebbian learning rule according to the feature importance scores. Assume that the score vector provided by the feature importance evaluation module is [0.8, 0.4, 0.9, 0.3, 0.7, 0.5, 0.6, 0.2, 0.85, 0.45], representing the importance of 10 features. The system uses these scores to determine the basic increment coefficients. For example, for the first feature with an importance score of 0.8, its basic increment coefficient can be set to 0.04; for the second feature with a score of 0.4, its coefficient can be set to 0.02.

[0051] The weight adjustment coefficient is calculated using a dynamic threshold algorithm by combining the presynaptic neuron activity state, the postsynaptic neuron activity state, and the current weight saturation. Weight saturation refers to the ratio of the current weight to the maximum weight. Suppose the current value of a weight is 0.6 and the maximum value is 1, then its saturation is 0.6. The dynamic threshold algorithm dynamically adjusts the learning rate according to the weight saturation: the higher the saturation, the lower the learning rate, to prevent the weight from growing excessively. For example, for a weight with a saturation of 0.6, the adjustment coefficient is 0.75.

[0052] The system calculates the increment of the synaptic weight by multiplying the base enhancement coefficient, the presynaptic and postsynaptic neuron activity states, and the weight adjustment coefficient. For example, for the connection from the first feature to the third neuron, the increment is 0.04×0.622×0.88×0.75 = 0.0164.

[0053] The squared sum operation is performed on the increments of all synaptic weights to obtain the enhancement energy value. Suppose the calculated enhancement energy value is 0.12. The system determines whether this value exceeds the preset energy upper limit threshold of 0.2. If it exceeds, the system will scale down all increments proportionally to ensure that the total energy does not exceed the threshold; if it does not exceed, the original increments remain unchanged. In this example, the enhancement energy value of 0.12 is less than the threshold of 0.2, so no scaling is required.

[0054] The change rate of the synaptic weight increment in the time dimension is calculated. Suppose the increments in three consecutive time steps are [0.0164, 0.0158, 0.0170], then the change rate can be the difference between the last two steps, which is 0.0012. The system substitutes this change rate and the mean parameter 0 and variance parameter 0.01 of the Gaussian function into the Gaussian function to construct a Gaussian inhibition function. The output value of the Gaussian function indicates that the inhibition is weaker when the change rate is close to the mean and stronger when it deviates from the mean.

[0055] Based on the increment of the synaptic weight, the system executes a feedback inhibition mechanism. The system calculates the inhibition intensity coefficient using the output value of the Gaussian inhibition function. For example, for a change rate of 0.0012, the output value of the Gaussian function is 0.9, and the corresponding inhibition intensity coefficient is 0.1. The system combines the inhibition learning rate of 0.02, the current synaptic weight of 0.6, and the neuron inactivation degree of 0.3 to calculate the inhibition amount of the synaptic weight. For example, the inhibition amount is 0.1×0.02×0.6×0.3 = 0.00036.

[0056] Combine the enhancement amount and the inhibition amount to update the synaptic weight. For example, for an initial weight of 0.6, an enhancement amount of 0.0164, and an inhibition amount of 0.00036, the updated weight is 0.6 + 0.0164 - 0.00036 = 0.61604. In this way, the system realizes the optimization of the neural network weights based on the feature importance, improves the sensitivity of the network to important features, and at the same time prevents the weights from growing excessively through the feedback inhibition mechanism, maintaining the stability and generalization ability of the network.

[0057] Figure 3 The following is a bar chart comparing the performance of the neural network based on the Hebbian learning rule and the feedback inhibition mechanism in the embodiments of the present invention: The figure compares the performance differences of three mechanisms, namely traditional Hebbian learning, improved feedback inhibition, and dynamic threshold adjustment, in five key performance indicators. In terms of the weight convergence speed, the dynamic threshold adjustment mechanism performs best, reaching 93.8%, which is 17.3 percentage points higher than the traditional method; in terms of the feature extraction accuracy, the improved feedback inhibition reaches the best effect of 91.7%, which is 9.4 percentage points higher than the traditional method; in terms of the anti-overfitting ability, the dynamic threshold adjustment mechanism is significantly ahead, reaching 91.2%, which is 22.5 percentage points higher than the traditional method; in terms of the computing efficiency, traditional Hebbian learning performs better, reaching 85.4%, but the difference from the other two improved methods is not significant; in terms of the energy consumption optimization, the dynamic threshold adjustment mechanism performs the most prominently, reaching an optimization effect of 94.3%, which is 31.5 percentage points higher than the traditional method. Generally speaking, the dynamic threshold adjustment mechanism shows significant advantages in multiple performance indicators, especially in the convergence speed, anti-overfitting ability, and energy optimization, indicating that this mechanism can better balance the learning effect and the consumption of computing resources.

[0058] In an alternative implementation manner, based on the adaptive weight coefficient and the device state feature vector, generate a fault feature representation; construct a fault diagnosis rule library, and the matching of the fault feature representation with the fault diagnosis rule library includes: Multiply the device state feature vector by the adaptive weight coefficient to generate an initial fault feature representation; construct a device fault knowledge graph, use a graph neural network to process the device fault knowledge graph, learn the feature representations of each node, and extract the main diagnosis path based on the feature representations; Construct an initial fault diagnosis rule library according to the main diagnosis path, match the initial fault feature representation with the rules in the initial fault diagnosis rule library, and obtain the rule matching effect; Calculate the confidence score of each rule based on the rule matching effect, dynamically adjust the rule weights according to the confidence score, and delete the rules with confidence scores lower than the preset confidence threshold. Generate new rule candidates using the device fault knowledge graph, and supplement the rules in the rule candidates with confidence scores higher than the preset confidence threshold to the initial fault diagnosis rule base to obtain an optimized fault diagnosis rule base; finally match the initial fault feature representation with the optimized fault diagnosis rule base, and output the fault diagnosis result.

[0059] After obtaining the device state feature vector and the adaptive weight coefficient, multiply the two to generate the initial fault feature representation. For example, the device state feature vector includes multi-dimensional features such as temperature, vibration, noise, and current. Suppose the feature vector of a certain motor device is [85, 12.3, 68, 15.2], and the corresponding adaptive weight coefficient is [0.3, 0.25, 0.15, 0.3]. By multiplying and summing the two, the initial fault feature representation value is 52.07.

[0060] When constructing the device fault knowledge graph, entities such as device components, fault types, fault symptoms, and environmental factors are used as nodes, and the relationships between entities are used as edges. For example, for a compressor device, the nodes include "bearing", "bearing overheating", "abnormal vibration", "insufficient lubrication", etc., and the edges include relationships such as "causes", "manifested as", "reason is", etc. The complete knowledge graph can contain hundreds of nodes and thousands of edges, forming a complex fault association network.

[0061] Use a graph neural network to process the device fault knowledge graph and learn the feature representations of each node. In the implementation process, first randomly initialize the nodes in the knowledge graph and assign each node a 64-dimensional feature vector. Through the message passing mechanism, each node aggregates the information from adjacent nodes and updates its own feature representation after multiple layers of non-linear transformation. After 3 rounds of iteration, each node obtains a feature representation containing graph structure information.

[0062] Based on the learned node feature representations, extract the main diagnosis paths. The specific method is to calculate the similarity between the initial fault feature representation and the feature of each fault node, select the top 5 fault nodes with the highest similarity, and trace the paths from these nodes to the fault cause nodes to form the main diagnosis paths. For example, for the "bearing overheating" fault, the diagnosis paths are "bearing overheating → abnormal vibration → insufficient lubrication" and "bearing overheating → temperature anomaly → cooling system failure".

[0063] Construct an initial fault diagnosis rule base according to the extracted main diagnosis paths. Each rule contains a precondition, a conclusion, and an initial weight. For example, the rule "IF temperature > 80 AND vibration > 10 THEN bearing overheat WITH weight 0.8", and the rule "IF current > 15 AND temperature > 75 THEN motor overload WITH weight 0.75". The initial rule base usually contains 50 - 100 rules, covering the main fault types.

[0064] Match the initial fault feature representation with the rules in the initial fault diagnosis rule base to obtain the rule matching effect. The matching process calculates the degree to which the feature representation satisfies the rule precondition and generates a matching score. For example, for the above motor feature vector, the matching score of the first rule is 0.92, and the matching score of the second rule is 0.78.

[0065] Calculate the confidence score for each rule based on the rule matching effect. The confidence score comprehensively considers the rule matching score, historical matching accuracy, and rule coverage. For example, if a rule has a matching score of 0.9, a historical accuracy of 0.85, and a coverage of 0.7, then its confidence score is calculated as 0.9×0.85×0.7 = 0.5355.

[0066] Dynamically adjust the rule weights according to the confidence scores, and delete the rules with confidence scores lower than the preset confidence threshold. If the preset confidence threshold is 0.4, then the rules with confidence scores lower than 0.4 will be deleted. For the remaining rules, update the rule weights according to the confidence scores. For example, the new weight = old weight × (1 + confidence score × 0.2).

[0067] Generate new rule candidates using the equipment fault knowledge graph. By analyzing the important paths not covered by the current rule base in the knowledge graph, supplementary rules are generated. For example, if it is found that there is a strong correlation between "current fluctuation" and "bearing wear" but not covered by the current rule base, then the rule "IF current fluctuation > 20% AND running time > 5000h THEN high bearing wearability WITH weight 0.65" is generated.

[0068] Evaluate the newly generated rule candidates, and supplement the rules with confidence scores higher than the preset confidence threshold to the initial fault diagnosis rule base to obtain an optimized fault diagnosis rule base. If the confidence score of the new rule is 0.58, which is higher than the threshold of 0.4, then it is added to the rule base. The optimized rule base usually reduces 20% of the inefficient rules compared to the initial rule base, while increasing 15% of high-quality new rules.

[0069] Finally, match the initial fault feature representation with the optimized fault diagnosis rule base to output the fault diagnosis result. The matching process takes into account the rule weights, calculates the comprehensive matching score, selects the top-3 fault types with the highest scores as the diagnosis results, and gives the confidence percentage. For example, the final diagnosis results are "Bearing overheating (87%)", "Insufficient lubrication (76%)", "Cooling system failure (62%)", providing clear fault location and causes for maintenance personnel.

[0070] Through the above technical implementation, the present invention can effectively improve the accuracy and efficiency of equipment fault diagnosis, reduce the misdiagnosis rate, and provide reliable technical support for industrial equipment maintenance.

[0071] In an alternative embodiment, an initial fault diagnosis rule base is constructed according to the main diagnosis path, and the initial fault feature representation is matched with the rules in the initial fault diagnosis rule base. The rule matching effects include: Use the multi-head attention mechanism to extract features from the main diagnosis path, generate a query matrix, a key matrix, and a value matrix. Obtain the initial attention weights through the inner product operation of the query matrix and the key matrix, and multiply the initial attention weights by the value matrix to obtain an enhanced feature representation; Construct a generative adversarial network model. The generator of the generative adversarial network model generates candidate rules based on the enhanced feature representation. The candidate rules include fault feature vectors, temporal dependence relationships, fault category encodings, and rule confidence levels. The discriminator of the generative adversarial network model calculates the authenticity scores of the candidate rules; Screen the candidate rules based on the authenticity scores, and retain the candidate rules with authenticity scores higher than the preset authenticity threshold as valid candidate rules; combine the main diagnosis path with the valid candidate rules to form an initial diagnosis rule, and calculate the effectiveness score of the initial diagnosis rule based on the authenticity scores; Construct a time-varying fault feature map, use spatio-temporal convolution to extract features from the time-varying fault feature map, and generate a new hidden state by combining the historical hidden state and the current input features; calculate the matching degree between the new hidden state and the initial diagnosis rule based on the attention mechanism; Perform weighted fusion on the effectiveness score and the matching degree to obtain a rule matching score, count the number of rules whose rule matching scores exceed the preset matching threshold, and obtain the rule matching effect.

[0072] As Figure 4 shown, the method includes: The multi-head attention mechanism is used to extract features from the main diagnosis path. The system represents the main diagnosis path as a vector sequence D = {d1, d2, ..., dn}, where each vector di represents a step feature in the diagnosis path. The multi-head attention mechanism generates a query matrix Q, a key matrix K, and a value matrix V by setting 8 attention heads, each with a dimension of 64.

[0073] Taking the fault diagnosis of a certain power equipment as an example, the input main diagnosis path contains steps such as "voltage fluctuation - temperature anomaly - vibration anomaly - fault alarm". After linear projection, the diagnosis path vectors are projected into the query space, key space, and value space. By calculating the dot product of the query matrix and the key matrix, the initial attention weight matrix is obtained, and the weight values range from 0 to 1, reflecting the correlation between each diagnosis step. After normalizing the initial weights, multiplying them with the value matrix V, an enhanced feature representation E = {e1, e2, ..., en} is obtained, and each feature vector has a dimension of 512, capturing the key information and the dependencies between steps in the diagnosis path.

[0074] A generative adversarial network model is constructed to generate diagnosis rules. The generator G adopts a four-layer fully connected network structure. The input layer is the enhanced feature representation E, and the number of neurons in the hidden layers are 256, 128, and 64 respectively, and the output layer dimension is the rule representation dimension. The generator generates candidate rules R = {r1, r2, ..., rm} based on the enhanced feature representation. Each rule ri consists of four components: a fault feature vector F (with a dimension of 128), a temporal dependency relationship T (represented by a 20-dimensional vector), a fault category encoding C (one-hot encoding, with a dimension equal to the number of fault categories), and a rule confidence S (a scalar value ranging from 0 to 1).

[0075] In practical applications, such as in the scenario of motor bearing faults, the generated candidate rules contain a feature vector of "vibration frequency exceeding 200Hz and duration greater than 30 seconds, and the temperature rising rate exceeding 5℃ / minute". The temporal relationship indicates that the temperature anomaly precedes the vibration anomaly. The fault category encoding points to the "bearing wear" category, and the rule confidence is 0.92.

[0076] The discriminator D adopts a three-layer convolutional network structure. The input is the candidate rules, and the output is a authenticity score (a scalar value between 0 and 1). The discriminator calculates the authenticity score of each candidate rule by comparing the feature distribution differences between the candidate rules and the real expert rules. During the training process, the generator and the discriminator perform adversarial learning, with 5000 iterations, a learning rate set to 0.0001, and a batch size of 32.

[0077] Filter the candidate rules based on the authenticity score output by the discriminator. Set the authenticity threshold to 0.75, and retain the candidate rules with authenticity scores higher than the threshold as valid candidate rules. Combine the main diagnostic path with the valid candidate rules to form the initial diagnostic rule set I = {i1, i2, ..., ik}. For each initial diagnostic rule, calculate its validity score V based on the authenticity score. The calculation method is a weighted combination of the authenticity score and the rule complexity, where the rule complexity is determined by the number of features and conditions included in the rule.

[0078] Construct a time-varying fault feature map G. Organize the real-time collected fault data into a three-dimensional tensor with dimensions [time step, number of sensors, feature dimension]. Use a spatio-temporal convolutional network to extract time-varying features. The network contains 3 spatio-temporal convolutional layers with a convolutional kernel size of 3×3, a stride of 1, and a padding of 1, and the number of output channels is 32, 64, and 128 respectively. Spatio-temporal convolution takes into account both the spatial relationship between sensors and the change trend in the time series. Combine a long short-term memory network to process sequence information with a hidden state dimension of 256. Generate a new hidden state h(t) by combining the historical hidden layer state h(t - 1) and the current input feature x(t) through a gating mechanism.

[0079] Calculate the matching degree M between the new hidden state h(t) and the initial diagnostic rule set I based on the attention mechanism. For each rule i, calculate its similarity score with the current state. The higher the score, the higher the matching degree. Perform a weighted fusion of the validity score V and the matching degree M with weights of 0.4 and 0.6 respectively to obtain the rule matching score S. Set the matching threshold to 0.8, count the number of rules whose rule matching scores exceed the threshold, and obtain the rule matching effect. In the fault diagnosis test of a substation device, 65 valid candidate rules were screened out from the initially generated 100 candidate rules. Finally, 42 rules had rule matching scores exceeding the threshold, and the matching effect was 42%, which was about 23% higher than the traditional rule matching method in terms of matching efficiency.

[0080] In an alternative implementation, calculate the matching probability of various types of faults, select the maximum matching probability as the current fault probability, and determine the corresponding fault type; when the current fault probability exceeds the preset probability threshold, generate edge-side fault warning information including: Obtain fault feature data, and perform standardization processing on the fault feature data to obtain standardized feature data; Calculate the membrane potential change rate of the standardized feature data. The membrane potential change rate is obtained through a weighted combination of the ratio of the standardized feature data to the membrane time constant, the synaptic input current, and the external input current. Calculate the average membrane potential of the neuron group based on the membrane potential change rate. Multiply the action potential of the previous neuron by the action potential of the subsequent neuron to obtain the synaptic activity level, multiply the difference between the synaptic activity level and the plasticity threshold by the learning rate to obtain the change amount of the synaptic weight, and update the synaptic weight according to the change amount of the synaptic weight; Calculate the firing probability of the neuron population based on the difference between the average membrane potential and the firing threshold, and perform weighted combination on the firing probability and the prior probability statistically obtained from the historical fault data through the updated synaptic weight to obtain the fault matching probability; Perform normalization processing on the fault matching probability to obtain the normalized matching probability, select the maximum value of the normalized matching probability as the current fault probability, and determine the corresponding fault type according to the current fault probability; Compare the current fault probability with the preset probability threshold. When the current fault probability exceeds the preset probability threshold, generate edge-side fault warning information according to the fault type, the current fault probability, and the current timestamp.

[0081] The fault feature data is collected from sensors of multiple edge devices and includes multi-dimensional data such as temperature, pressure, vibration, and current. After obtaining these data, perform standardization processing on them to eliminate the dimension difference. For example, for the original temperature data of 95 °C, current data of 8.7 A, and vibration frequency data of 62 Hz collected by a certain edge computing node, through the Z-score standardization method, the data can be converted into dimensionless standardized feature data, which are 0.82, 1.35, and -0.42 respectively.

[0082] After standardization processing, calculate the membrane potential change rate. In this embodiment, the membrane time constant is set to 20 ms, the synaptic input current weight is 0.7, and the external input current weight is 0.3. Taking the above standardized data as an example, the calculated membrane potential change rate is: (0.82 / 20)×0.3 + (1.35×0.7)×0.3 + (-0.42×0.7)×0.3 = 0.1977. According to this membrane potential change rate, according to the neuron membrane potential update rule, the average membrane potential of the neuron population can be calculated. Assuming that the initial average membrane potential is -70 mV, after 10 ms of integration, the average membrane potential is updated to -68.023 mV.

[0083] The calculation of the synaptic activity level involves the interaction of action potentials between the pre - synaptic neurons and the post - synaptic neurons. In this example, assume that the action potentials of the pre - synaptic neurons A, B, and C are 0.8, 0.6, and 0.7 respectively, and the action potentials of the post - synaptic neurons D and E are 0.5 and 0.9 respectively. Through matrix multiplication, the synaptic activity level matrix can be obtained. For example, the synaptic activity level from A to D is 0.8×0.5 = 0.4. Assume that the plasticity threshold is 0.35 and the learning rate is 0.05, then the change in synaptic weight from A to D is (0.4 - 0.35)×0.05 = 0.0025. Similar calculations and updates are performed for all synaptic weights. For example, the initial weight 0.6 is updated to 0.6025.

[0084] Calculate the firing probability of the neuron population based on the difference between the average membrane potential and the firing threshold. If the firing threshold is set to - 65mV, then the difference is - 68.023 - (- 65)= - 3.023mV. Through mapping by the sigmoid function, the firing probability is 0.427. Then, combined with the prior probabilities obtained from the statistical analysis of historical fault data, such as the prior probability of overheating fault 0.3, the prior probability of overload fault 0.25, the prior probability of abnormal vibration 0.2, etc., weighted combination is performed through the updated synaptic weights.

[0085] If the synaptic weight related to the overheating fault is 0.68, then the matching probability of the overheating fault is calculated as 0.427×0.3×0.68 = 0.0871. Similarly, calculate the matching probabilities of other fault types, such as the overload fault 0.0762 and the abnormal vibration 0.0513.

[0086] Perform normalization on the fault matching probabilities so that the sum of the probabilities of all fault types is 1. After normalization, the probability of the overheating fault is 0.0871 / (0.0871 + 0.0762 + 0.0513+...)=0.407. The probability of the overload fault is 0.356, and the probability of the abnormal vibration is 0.237. Select the maximum value 0.407 as the current fault probability and determine that the current fault type is the overheating fault.

[0087] Compare the current fault probability 0.407 with the preset probability threshold (for example, 0.4). Since 0.407>0.4, the system generates an edge - side fault warning message. The warning message includes the fault type (overheating fault), the fault probability (0.407), and the current timestamp (2023 - 05 - 15 14:28:36). This warning message can be sent to the operation and maintenance management platform through the message queue and trigger the corresponding alarm mechanism.

[0088] This method can run in real time on edge computing devices to predict various types of faults early. The system continuously learns new fault patterns and improves the prediction accuracy by adaptively adjusting the synaptic weights. For example, after being deployed on an industrial production line, this method successfully predicted an overheating fault of a key device 3 hours in advance, giving maintenance personnel sufficient time to intervene and avoiding losses caused by production interruptions. Field tests show that the prediction accuracy of this method for common faults can reach 87%, which is 23% higher than that of traditional threshold monitoring methods, and the false alarm rate is reduced by 35%.

[0089] The computational complexity of this method is relatively low. On an edge computing device configured with a 4-core processor and 4GB of memory, it can process 200 sets of feature data per second, meeting the requirements of real-time monitoring. The system also supports adaptive adjustment of parameters. For example, it can automatically adjust parameters such as the membrane time constant and learning rate according to different device characteristics, enhancing the versatility and robustness of the method.

[0090] Figure 5 The following is a bar chart comparing the performance of the edge-side fault warning system based on the neuron model in the embodiments of the present invention: This figure compares and analyzes the performance differences of traditional threshold detection methods, machine learning models, and neuron dynamic membrane potential models in five key performance indicators. In terms of the detection accuracy of hardware faults, the neuron dynamic membrane potential model performs the best, reaching 92.8%, which is 14.5 percentage points higher than the traditional method; in terms of the recognition rate of software anomalies, the neuron dynamic membrane potential model also leads, reaching 91.5%, which is 18 percentage points higher than the traditional method; in terms of the prediction ability of network communication faults, the neuron dynamic membrane potential model reaches an accuracy rate of 88.7%, which is 23.5 percentage points higher than the traditional method; in terms of the detection sensitivity of data anomalies, the neuron dynamic membrane potential model performs the best, reaching 93.2%, which is 11.8 percentage points higher than the traditional method; in terms of the early warning of system performance degradation, the neuron dynamic membrane potential model reaches an early warning accuracy rate of 90.6%, which is 20.9 percentage points higher than the traditional method. Overall, the neuron dynamic membrane potential model is significantly superior to traditional threshold detection methods and machine learning models in all evaluation indicators, especially showing obvious advantages in fault prediction and anomaly detection, reflecting the high efficiency and reliability of this model in system fault diagnosis and early warning.

[0091] In the second aspect of the embodiments of the present invention, a device telemetry data fault analysis system based on machine learning is provided, including: A first unit for receiving multi-dimensional telemetry data collected by a device telemetry data acquisition module, performing normalization processing on the multi-dimensional telemetry data, and generating normalized data; A second unit, configured to extract the temporal features and statistical features of the standardized data, establish feature evaluation metrics; screen a core feature set according to the feature evaluation metrics; construct a lightweight feature extraction rule based on the core feature set to achieve lightweight edge - end data processing; A third unit, configured to process the standardized data by using the lightweight feature extraction rule, generate a device state feature vector, calculate the contribution score of each feature component in the device state feature vector, and allocate an adaptive weight coefficient to each feature component according to the contribution score; A fourth unit, configured to generate a fault feature representation based on the adaptive weight coefficient and the device state feature vector; construct a fault diagnosis rule base, and match the fault feature representation with the fault diagnosis rule base; calculate the matching probability of various faults, select the maximum matching probability as the current fault probability, and determine the corresponding fault type; when the current fault probability exceeds a preset probability threshold, generate an edge - end fault warning message; A fifth unit, configured to transmit the edge - end fault warning message, the fault type, and the current fault probability to a cloud management platform for storage and analysis.

[0092] In a third aspect of the embodiments of the present invention, an electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0093] In a fourth aspect of the embodiments of the present invention, a computer - readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0094] The present invention may be a method, a device, a system, and / or a computer program product. The computer program product may include a computer - readable storage medium, on which computer - readable program instructions for executing various aspects of the present invention are loaded.

[0095] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for fault analysis of device telemetry data based on machine learning, characterized in that, include: receiving multidimensional telemetry data collected by a telemetry data collection module of a device, performing standardization processing on the multidimensional telemetry data, and generating standardized data; Extracting the time series features and statistical features of the standardized data and establishing feature evaluation indicators; screening a core feature set according to the feature evaluation indicators; constructing a lightweight feature extraction rule based on the core feature set to achieve lightweight edge data processing; Processing the standardized data using the lightweight feature extraction rule to generate a device state feature vector, calculating a contribution score of each feature component in the device state feature vector, and assigning an adaptive weight coefficient to each feature component according to the contribution score; generating a fault feature representation based on the adaptive weight coefficient and the device state feature vector; Construct a fault diagnosis rule base, match the fault feature representation with the fault diagnosis rule base; calculate the matching probability of each type of fault, select the maximum matching probability as the current fault probability, and determine the corresponding fault type; when the current fault probability exceeds a preset probability threshold, generate edge fault warning information; The edge fault warning information, the fault type and the current fault probability are transmitted to the cloud management platform for storage and analysis.

2. The method according to claim 1, characterized in that, Filtering a core feature set according to the feature evaluation index; constructing a lightweight feature extraction rule based on the core feature set to achieve lightweight edge data processing includes: Receiving equipment telemetry data to generate a multi-dimensional time series feature sequence; respectively calculating the autocorrelation coefficient of each feature in the multi-dimensional time series feature sequence and the mutual correlation coefficient between the features; Performing information entropy calculation and Fisher discriminant analysis on the multidimensional time series feature sequence to obtain the information gain value and inter-class dispersion of each feature, and performing weighted combination of the autocorrelation coefficient, the mutual correlation coefficient, the information gain value and the inter-class dispersion to obtain a feature importance score; Based on the feature importance score, the multidimensional time series feature sequence is sorted in descending order, and the features with the highest order are selected as the candidate feature set; the mutual information between the features in the candidate feature set is calculated, a feature correlation matrix is constructed, and a core feature subset is selected from the candidate feature set based on the maximum correlation and minimum redundancy criterion; Performing principal component analysis on the core feature subset, calculating the contribution rate of each principal component, and selecting a principal component combination whose cumulative contribution rate reaches a preset contribution threshold; constructing a feature extraction rule based on the principal component combination, and applying the feature extraction rule to real-time telemetry data to obtain a lightweight feature representation; The reconstruction error of the lightweight feature representation relative to the original feature is calculated. When the reconstruction error is less than a preset error threshold, the feature extraction rule is deployed to the edge computing end to achieve lightweight edge data processing.

3. The method according to claim 1, characterized in that Processing the standardized data using the lightweight feature extraction rule to generate a device state feature vector, calculating a contribution score of each feature component in the device state feature vector, and assigning an adaptive weight coefficient to each feature component according to the contribution score includes: Receive input feature data, input the input feature data into a feature neural network, calculate the activation state value of each feature neuron, and calculate the corresponding neurotransmitter concentration change rate according to the activation state value. The neurotransmitter concentration change rate is jointly determined by the neurotransmitter release rate and the decay rate; Calculate the dynamic activity threshold of the feature neuron based on the neurotransmitter concentration change rate. The dynamic activity threshold is adaptively adjusted with the cumulative change of the neurotransmitter concentration; calculate the synaptic strength modulation factor using the difference between the dynamic activity threshold and the neurotransmitter concentration; Multiply the input feature data, the synaptic weight coefficient, and the neurotransmitter concentration to calculate the local field response intensity, and calculate the feature importance score according to the local field response intensity and the synaptic strength modulation factor; According to the feature importance score, execute the Hebbian learning rule to calculate the increment of the synaptic weight; based on the increment of the synaptic weight, execute the feedback inhibition mechanism to calculate the inhibition amount of the synaptic weight; Superimpose the increment of the synaptic weight and the inhibition amount of the synaptic weight, update the synaptic weight coefficient, and perform normalization processing on the updated synaptic weight coefficient to obtain the final adaptive weight coefficient.

4. The method according to claim 3, wherein According to the feature importance score, execute the Hebbian learning rule to calculate the increment of the synaptic weight; Based on the increment of the synaptic weight, the feedback inhibition mechanism is executed, and calculating the inhibition amount of the synaptic weight includes: Input the input feature data into the neural network, perform nonlinear transformation on the input feature data using the sigmoid activation function, calculate the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron, select the neuron with the largest activity value as the winning neuron based on the competitive learning algorithm, and dynamically adjust the weight enhancement parameter using the activity state of the winning neuron; According to the feature importance score, execute the Hebbian learning rule, determine the basic enhancement coefficient through the feature importance score, combine the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron and the current weight saturation, calculate the weight adjustment coefficient using the dynamic threshold algorithm, and obtain the increment of the synaptic weight according to the weight adjustment coefficient; Perform a square summation operation on the increment of the synaptic weight to obtain an enhancement energy value, and determine whether the enhancement energy value exceeds a preset energy upper limit threshold; Calculate the change rate of the increment of the synaptic weight in the time dimension, substitute the change rate, the mean parameter, and the variance parameter of the Gaussian function into the Gaussian function to construct a Gaussian inhibition function; Based on the increment of the synaptic weight, execute the feedback inhibition mechanism, calculate the inhibition intensity coefficient using the output value of the Gaussian inhibition function, and combine the inhibition learning rate, the current synaptic weight, and the neuron inactivation degree to obtain the inhibition amount of the synaptic weight.

5. The method according to claim 1, characterized in that, Generate a fault feature representation based on the adaptive weight coefficient and the device state feature vector; Construct a fault diagnosis rule base, and match the fault feature representation with the fault diagnosis rule base, including: Multiply the device status feature vector by the adaptive weight coefficient to generate an initial fault feature representation; construct a device fault knowledge graph, use a graph neural network to process the device fault knowledge graph, learn the feature representations of each node, and extract the main diagnostic path based on the feature representations; Construct an initial fault diagnosis rule base according to the main diagnostic path, match the initial fault feature representation with the rules in the initial fault diagnosis rule base, and obtain the rule matching effect; Calculate the confidence score of each rule based on the rule matching effect, dynamically adjust the rule weight according to the confidence score, and delete the rules with confidence scores lower than the preset confidence threshold; Use the device fault knowledge graph to generate new rule candidates, supplement the rules with confidence scores higher than the preset confidence threshold in the rule candidates to the initial fault diagnosis rule base to obtain an optimized fault diagnosis rule base; finally match the initial fault feature representation with the optimized fault diagnosis rule base, and output the fault diagnosis result.

6. The method according to claim 5, wherein Construct an initial fault diagnosis rule base according to the main diagnostic path, match the initial fault feature representation with the rules in the initial fault diagnosis rule base, and the rule matching effect includes: Use the multi-head attention mechanism to extract features from the main diagnostic path to generate a query matrix, a key matrix, and a value matrix. Obtain the initial attention weight through the inner product operation of the query matrix and the key matrix, and multiply the initial attention weight by the value matrix to obtain an enhanced feature representation; Construct a generative adversarial network model. The generator of the generative adversarial network model generates candidate rules based on the enhanced feature representation. The candidate rules include a fault feature vector, a temporal dependence relationship, a fault category encoding, and a rule confidence. The discriminator of the generative adversarial network model calculates the authenticity score of the candidate rules; Filter the candidate rules based on the authenticity score, and retain the candidate rules with authenticity scores higher than the preset authenticity threshold as valid candidate rules; combine the main diagnostic path with the valid candidate rules to form an initial diagnostic rule, and calculate the effectiveness score of the initial diagnostic rule based on the authenticity score; Construct a time-varying fault feature map, use spatio-temporal convolution to extract features from the time-varying fault feature map, and generate a new hidden state by combining the historical hidden state and the current input feature; calculate the matching degree between the new hidden state and the initial diagnostic rule based on the attention mechanism; Perform weighted fusion on the effectiveness score and the matching degree to obtain a rule matching score, count the number of rules whose rule matching scores exceed the preset matching threshold, and obtain the rule matching effect.

7. The method according to claim 1, characterized in that, Calculate the matching probabilities of various faults, select the maximum matching probability as the current fault probability, and determine the corresponding fault type; when the current fault probability exceeds the preset probability threshold, generate edge-side fault warning information including: Obtain fault feature data, and perform standardization processing on the fault feature data to obtain standardized feature data; Calculate the membrane potential change rate of the standardized feature data, where the membrane potential change rate is obtained through a weighted combination of the ratio of the standardized feature data to the membrane time constant, the synaptic input current, and the external input current, and calculate the average membrane potential of the neuron population based on the membrane potential change rate; Multiply the action potential of the previous neuron by the action potential of the subsequent neuron to obtain the synaptic activity degree, multiply the difference between the synaptic activity degree and the plasticity threshold by the learning rate to obtain the change amount of the synaptic weight, and update the synaptic weight according to the change amount of the synaptic weight; Calculate the firing probability of the neuron population based on the difference between the average membrane potential and the firing threshold, and perform a weighted combination of the firing probability and the prior probability statistically obtained from the historical failure data through the updated synaptic weight to obtain the failure matching probability; Perform a normalization process on the failure matching probability to obtain a normalized matching probability, select the maximum value of the normalized matching probability as the current failure probability, and determine the corresponding failure type according to the current failure probability; Compare the current failure probability with a preset probability threshold, and when the current failure probability exceeds the preset probability threshold, generate edge - side failure warning information according to the failure type, the current failure probability, and the current timestamp.

8. A machine learning-based device telemetry data fault analysis system for implementing the method described in any one of the preceding claims 1-7, characterized in that, Comprising: A first unit, configured to receive the multi - dimensional telemetry data collected by the device telemetry data acquisition module, perform a normalization process on the multi - dimensional telemetry data, and generate standardized data; A second unit, configured to extract the timing features and statistical features of the standardized data, establish a feature evaluation index; screen a core feature set according to the feature evaluation index; construct a lightweight feature extraction rule based on the core feature set to achieve lightweight edge - side data processing; A third unit, configured to process the standardized data using the lightweight feature extraction rule, generate a device state feature vector, calculate the contribution score of each feature component in the device state feature vector, and assign an adaptive weight coefficient to each feature component according to the contribution score; A fourth unit, configured to generate a failure feature representation based on the adaptive weight coefficient and the device state feature vector; Construct a failure diagnosis rule library, match the failure feature representation with the failure diagnosis rule library; calculate the matching probabilities of various failures, select the maximum matching probability as the current failure probability, and determine the corresponding failure type; when the current failure probability exceeds the preset probability threshold, generate edge - side failure warning information; A fifth unit, configured to transmit the edge - side failure warning information, the failure type, and the current failure probability to the cloud management platform for storage and analysis.

9. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images