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

Abstract

The present invention provides a method and system for analyzing device telemetry data faults based on machine learning, which relates to the field of fault analysis technology. The method includes standardizing multidimensional telemetry data, extracting time series features and statistical features, screening core feature sets, constructing lightweight feature extraction rules, calculating feature component contribution scores and assigning adaptive weight coefficients, generating fault feature representations and matching them with a fault diagnosis rule base, thereby determining the fault type and probability, achieving lightweight edge data processing and timely generation of fault warnings, and improving the accuracy and efficiency of fault diagnosis.

Description

Technical Field

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

[0002] With the rapid development of industrial IoT technology, real-time monitoring and fault diagnosis of equipment telemetry data have become critical for ensuring the safe and stable operation of industrial production. Industrial equipment generates a large amount of multidimensional telemetry data during operation, which contains information about the equipment's operating status and potential faults. Traditional equipment fault diagnosis relies primarily on manual judgment. With the development of machine learning technology, data-driven intelligent fault analysis methods have gradually become a research hotspot. Currently, the industrial sector has begun to experiment with applying machine learning algorithms to equipment fault prediction and diagnosis. By learning and analyzing historical equipment operating data, fault warning models are established to achieve early identification of equipment abnormalities and accurate determination of fault types.

[0003] Existing technologies often use full-feature analysis when processing multi-dimensional telemetry data, lacking an effective feature screening mechanism. This results in high model calculation complexity, making it difficult to run efficiently on edge computing devices and unable to meet the needs of real-time fault diagnosis in industrial sites.

[0004] Traditional fault diagnosis methods do not adequately consider the importance of different features. They usually use fixed weights or empirical settings to process each feature component. They are unable to dynamically adjust feature weights according to the equipment operating status, making it difficult to accurately reflect the contribution of each feature to fault judgment under different working conditions, thereby reducing the accuracy of fault diagnosis.

[0005] The existing fault diagnosis system that collaborates between the edge and the cloud has a low degree of integration, and the edge computing capabilities are not fully utilized. 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. It cannot meet the actual needs of industrial sites for rapid fault diagnosis and timely warning. Summary of the Invention

[0006] The embodiments of the present invention provide a method and system for analyzing device telemetry data faults based on machine learning, which can solve the problems in the prior art.

[0007] A first aspect of an embodiment of the present invention provides a method for analyzing device telemetry data faults based on machine learning, comprising:

[0008] receiving multidimensional telemetry data collected by a telemetry data collection module of a device, performing standardization processing on the multidimensional telemetry data to generate standardized data;

[0009] Extracting the temporal features and statistical features of the standardized data and establishing feature evaluation indicators; screening a core feature set based on the feature evaluation indicators; constructing lightweight feature extraction rules based on the core feature set to achieve lightweight edge data processing;

[0010] 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;

[0011] 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 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;

[0012] The edge fault warning information, the fault type and the current fault probability are transmitted to the cloud management platform for storage and analysis.

[0013] 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:

[0014] Receive device telemetry data and generate a multidimensional time series feature sequence; calculate the autocorrelation coefficient of each feature in the multidimensional time series feature sequence and the mutual correlation coefficient between the features;

[0015] 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 a weighted combination of the autocorrelation coefficient, the cross-correlation coefficient, the information gain value, and the inter-class dispersion to obtain a feature importance score;

[0016] Sorting the multidimensional time series feature sequence in descending order based on the feature importance score, selecting the top-ranked features as a candidate feature set; calculating the mutual information between the features in the candidate feature set, constructing a feature correlation matrix, and screening a core feature subset from the candidate feature set based on a maximum correlation and minimum redundancy criterion;

[0017] 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;

[0018] 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.

[0019] 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:

[0020] receiving input feature data, inputting the input feature data into a feature neuron network, calculating an activation state value of each feature neuron, and calculating a corresponding neurotransmitter concentration change rate based on the activation state value, wherein the neurotransmitter concentration change rate is determined by a neurotransmitter release rate and a decay rate;

[0021] Calculating a dynamic activity threshold of a characteristic neuron based on the rate of change of the neurotransmitter concentration, wherein the dynamic activity threshold is adaptively adjusted with the cumulative change of the neurotransmitter concentration; and calculating a synaptic strength modulation factor using the difference between the dynamic activity threshold and the neurotransmitter concentration;

[0022] multiplying the input feature data, the synaptic weight coefficient and the neurotransmitter concentration to calculate the local field response strength, and calculating the feature importance score based on the local field response strength and the synaptic strength modulation factor;

[0023] According to the feature importance score, the Hebbian learning rule is executed to calculate the enhancement amount of the synaptic weight; based on the enhancement amount of the synaptic weight, the feedback inhibition mechanism is executed to calculate the inhibition amount of the synaptic weight;

[0024] The enhancement amount of the synaptic weight is superimposed with the inhibition amount of the synaptic weight, the synaptic weight coefficient is updated, and the updated synaptic weight coefficient is normalized to obtain a final adaptive weight coefficient.

[0025] According to the feature importance score, executing the Hebbian learning rule to calculate the enhancement amount of the synaptic weight; based on the enhancement amount of the synaptic weight, executing the feedback inhibition mechanism to calculate the inhibition amount of the synaptic weight includes:

[0026] Input feature data is input into a neural network, a sigmoid activation function is used to perform a nonlinear transformation on the input feature data, the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron are calculated, the neuron with the largest activity value is selected as the winning neuron based on a competitive learning algorithm, and the weight enhancement parameter is dynamically adjusted using the activity state of the winning neuron;

[0027] According to the feature importance score, the Hebbian learning rule is executed to determine the basic enhancement coefficient through the feature importance score, and the weight adjustment coefficient is calculated using the dynamic threshold algorithm in combination with the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron and the current weight saturation, and the enhancement amount of the synaptic weight is obtained according to the weight adjustment coefficient;

[0028] Performing a square sum operation on the enhancement amount of the synaptic weight to obtain an enhancement energy value, and determining whether the enhancement energy value exceeds a preset energy upper limit threshold;

[0029] Calculating the rate of change of the enhancement amount of the synaptic weight in the time dimension, substituting the rate of change and the mean parameter and variance parameter of the Gaussian function into the Gaussian function to construct a Gaussian inhibition function;

[0030] Based on the enhancement amount of the synaptic weight, a feedback inhibition mechanism is executed, and the inhibition strength coefficient is calculated using the output value of the Gaussian inhibition function. The inhibition amount of the synaptic weight is obtained by combining the inhibition learning rate, the current synaptic weight and the degree of neuronal inactivation.

[0031] Generating a fault feature representation based on the adaptive weight coefficient and the device state feature vector; building a fault diagnosis rule base, and matching the fault feature representation with the fault diagnosis rule base includes:

[0032] Multiplying the device state feature vector by the adaptive weight coefficient to generate an initial fault feature representation; constructing a device fault knowledge graph, processing the device fault knowledge graph using a graph neural network, learning the feature representation of each node, and extracting a main diagnostic path based on the feature representation;

[0033] Building an initial fault diagnosis rule base according to the main diagnostic path, matching the initial fault feature representation with the rules in the initial fault diagnosis rule base to obtain a rule matching effect;

[0034] Calculating a confidence score for each rule based on the rule matching effect, dynamically adjusting the rule weight according to the confidence score, and deleting rules whose confidence scores are lower than a preset confidence threshold;

[0035] New rule candidates are generated using the equipment fault knowledge graph, and the rules with confidence scores higher than a preset confidence threshold in the rule candidates are added to the initial fault diagnosis rule base to obtain an optimized fault diagnosis rule base; the initial fault feature representation is finally matched with the optimized fault diagnosis rule base, and the fault diagnosis result is output.

[0036] An initial fault diagnosis rule base is constructed according to the main diagnostic path, and the initial fault feature representation is matched with the rules in the initial fault diagnosis rule base to obtain a rule matching effect, including:

[0037] Using a multi-head attention mechanism to extract features from the main diagnostic path, generating a query matrix, a key matrix, and a value matrix, obtaining initial attention weights by performing an inner product operation on the query matrix and the key matrix, and multiplying the initial attention weights by the value matrix to obtain an enhanced feature representation;

[0038] Constructing a generative adversarial network model, wherein a generator of the generative adversarial network model generates candidate rules based on the enhanced feature representation, wherein the candidate rules include fault feature vectors, temporal dependencies, fault category codes, and rule confidences, and a discriminator of the generative adversarial network model calculates the authenticity scores of the candidate rules;

[0039] screening the candidate rules based on the authenticity scores, retaining candidate rules with authenticity scores higher than a preset authenticity threshold as valid candidate rules; combining the main diagnostic path with the valid candidate rules to form an initial diagnostic rule, and calculating the validity score of the initial diagnostic rule based on the authenticity scores;

[0040] Constructing a time-varying fault feature map, extracting features from the time-varying fault feature map using spatiotemporal convolution, generating a new hidden layer state by combining the historical hidden layer state with the current input features; and calculating the degree of matching between the new hidden layer state and the initial diagnostic rule based on an attention mechanism.

[0041] The validity score and the matching degree are weighted and fused to obtain a rule matching score, and the number of rules whose rule matching scores exceed a preset matching threshold is counted to obtain a rule matching effect.

[0042] 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 the preset probability threshold, generate edge fault warning information including:

[0043] Acquiring fault characteristic data, and performing standardization processing on the fault characteristic data to obtain standardized characteristic data;

[0044] calculating a membrane potential change rate of the standardized characteristic data, the membrane potential change rate being obtained by a weighted combination of a ratio of the standardized characteristic data to a membrane time constant, a synaptic input current, and an external input current, and calculating an average membrane potential of a neuronal group based on the membrane potential change rate;

[0045] Multiplying the action potential of the preceding neuron by the action potential of the succeeding neuron to obtain a synaptic activity level, multiplying the difference between the synaptic activity level and the plasticity threshold by the learning rate to obtain a change in the synaptic weight, and updating the synaptic weight according to the change in the synaptic weight;

[0046] Calculating the firing probability of the neuron group based on the difference between the average membrane potential and the firing threshold, and performing a weighted combination of the firing probability and the prior probability obtained from historical fault data statistics through the updated synaptic weight to obtain a fault matching probability;

[0047] performing normalization processing on the fault matching probability to obtain a normalized matching probability, selecting a maximum value of the normalized matching probability as a current fault probability, and determining a corresponding fault type according to the current fault probability;

[0048] The current fault probability is compared with a preset probability threshold, and when the current fault probability exceeds the preset probability threshold, edge fault warning information is generated according to the fault type, the current fault probability, and the current timestamp.

[0049] A second aspect of an embodiment of the present invention provides a device telemetry data fault analysis system based on machine learning, including:

[0050] The first unit is configured to receive multidimensional telemetry data collected by a device telemetry data collection module, perform standardization processing on the multidimensional telemetry data, and generate standardized data;

[0051] The second unit is used to extract the time series features and statistical features of the standardized data and establish feature evaluation indicators; filter the core feature set according to the feature evaluation indicators; and construct lightweight feature extraction rules based on the core feature set to achieve lightweight edge data processing;

[0052] a third unit, configured to process the standardized data using the lightweight feature extraction rule to generate a device state feature vector, calculate a 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;

[0053] A fourth unit is 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, 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; and generate edge fault warning information when the current fault probability exceeds a preset probability threshold;

[0054] The fifth unit is used to transmit the edge fault warning information, the fault type and the current fault probability to the cloud management platform for storage and analysis.

[0055] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0056] processor;

[0057] a memory for storing processor-executable instructions;

[0058] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0059] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0060] The beneficial effects of this application are as follows:

[0061] The present invention standardizes multi-dimensional telemetry data, extracts time series features and statistical features, establishes feature evaluation indicators to screen the core feature set, and constructs lightweight feature extraction rules, thereby achieving lightweight edge data processing, reducing computational complexity and resource consumption, and improving the processing efficiency of edge devices.

[0062] The present invention calculates the contribution score of each characteristic component in the equipment state characteristic vector, assigns an adaptive weight coefficient accordingly, and generates a fault feature representation, thereby achieving accurate expression of equipment state characteristics under different working conditions, improving the accuracy and reliability of fault diagnosis, and reducing the false alarm rate.

[0063] The present invention builds a fault diagnosis rule library, calculates the matching probability of various types of faults, generates edge-end fault warning information when the current fault probability exceeds a preset threshold, and transmits relevant information to the cloud platform, realizing the collaborative work of the edge and the cloud. It can not only timely discover and warn of abnormal equipment status, but also conduct in-depth analysis and optimization through the cloud platform, thereby improving the reliability and intelligence level of the entire system. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 Schematic diagram of a process for analyzing device telemetry data faults based on machine learning according to an embodiment of the present invention;

[0065] Figure 2 This is a flow chart of a lightweight feature extraction method for edge terminals based on multi-dimensional temporal features according to an embodiment of the present invention;

[0066] Figure 3A histogram comparing the performance of neural networks based on the Hebbian learning rule and feedback inhibition mechanism according to an embodiment of the present invention;

[0067] Figure 4 This is a flow chart of a fault diagnosis rule matching method based on multi-head attention and GAN in an embodiment of the present invention;

[0068] Figure 5 This is a bar chart comparing the performance of edge fault warning systems based on the neuron model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0070] The technical solution of the present invention is described in detail below with reference to specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0071] Figure 1 FIG. 1 is a flow chart of a method for analyzing device telemetry data failure based on machine learning according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0072] 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;

[0073] Extracting the temporal features and statistical features of the standardized data and establishing feature evaluation indicators; screening a core feature set based on the feature evaluation indicators; constructing lightweight feature extraction rules based on the core feature set to achieve lightweight edge data processing;

[0074] 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;

[0075] 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 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;

[0076] The edge fault warning information, the fault type and the current fault probability are transmitted to the cloud management platform for storage and analysis.

[0077] In an optional embodiment, 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:

[0078] Receive device telemetry data and generate a multidimensional time series feature sequence; calculate the autocorrelation coefficient of each feature in the multidimensional time series feature sequence and the mutual correlation coefficient between the features;

[0079] 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 a weighted combination of the autocorrelation coefficient, the cross-correlation coefficient, the information gain value, and the inter-class dispersion to obtain a feature importance score;

[0080] Sorting the multidimensional time series feature sequence in descending order based on the feature importance score, selecting the top-ranked features as a candidate feature set; calculating the mutual information between the features in the candidate feature set, constructing a feature correlation matrix, and screening a core feature subset from the candidate feature set based on a maximum correlation and minimum redundancy criterion;

[0081] 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;

[0082] 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.

[0083] like Figure 2 As shown, the method includes:

[0084] Receive telemetry data from equipment, including information collected by various sensors such as temperature, pressure, and vibration frequency. For a certain piece of industrial equipment, temperature, pressure, and vibration data are collected at 10 samples per second, forming a three-dimensional time series feature sequence. For example, during a one-hour monitoring period, three time series containing 36,000 data points each are generated.

[0085] For the generated multidimensional time series feature sequences, the autocorrelation coefficient of each feature and the cross-correlation coefficient between features are calculated. The autocorrelation coefficient reflects the correlation of a single feature at different time points, while the cross-correlation coefficient indicates the degree of association between different features. Taking temperature and pressure as examples, the calculated autocorrelation coefficient for 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 the two features.

[0086] Information entropy calculation and Fisher discriminant analysis were then performed on the multidimensional time series feature sequences. Information entropy calculation yields the information gain value for each feature, which measures the amount of information contained in the feature. Fisher discriminant analysis yields the inter-class dispersion, which is used to assess the feature's ability to distinguish between different classes. In practical applications, the information gain value for the temperature feature was calculated to be 0.75, with an inter-class dispersion of 0.68; the information gain value for the pressure feature was 0.62, with an inter-class dispersion of 0.55; and the information gain value for the vibration feature was 0.83, with an inter-class dispersion of 0.72.

[0087] The autocorrelation coefficient, cross-correlation coefficient, information gain, and inter-class dispersion are weighted together to produce a feature importance score. The weighting scheme can be adjusted based on the application scenario. For example, the weights for the four indicators can be set to 0.2, 0.2, 0.3, and 0.3, respectively. Based on this weighting scheme, the calculated importance scores for the three features, temperature, pressure, and vibration, are 0.735, 0.595, and 0.765, respectively.

[0088] Sort the multidimensional time series features in descending order based on their importance scores and select the top-ranked features as the candidate feature set. In the above example, vibration and temperature features are selected as the candidate feature set. If the system actually contains more features, the top 5 or 10 features are selected to form the candidate set.

[0089] 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, with higher values ​​indicating greater redundancy between the features. In the vibration and temperature example, assume the calculated mutual information value is 0.42, forming a 2×2 correlation matrix.

[0090] A core feature subset is selected from the candidate feature set based on the maximum relevance and minimum redundancy criterion. This criterion aims to select the feature combination that is most relevant to the target variable and has minimal redundancy. In practice, features that meet these criteria are added one by one through an iterative process. If there are five features in the candidate set, three core features that are non-redundant and rich in information are ultimately selected.

[0091] Perform principal component analysis on the core feature subset and calculate the contribution of each principal component. Principal component analysis converts the original features into mutually orthogonal principal components, each of which represents a certain proportion of the variance in the original data. Suppose a principal component analysis is performed on three core features, resulting in three principal components: PC1, PC2, and PC3, with contribution rates of 65%, 25%, and 10%, respectively.

[0092] Select the principal component combinations whose cumulative contribution reaches the preset contribution threshold. If the contribution threshold is set to 85%, PC1 and PC2 are selected because their cumulative contribution is 90%. Based on these principal components, construct feature extraction rules, specifically mapping the original features to the principal component space.

[0093] Applying the constructed feature extraction rules to real-time telemetry data yields lightweight feature representations. These feature representations have lower dimensionality while retaining the key information of the original data. For example, if the original data has 10 dimensions, the feature extraction rules can reduce it to 3, significantly reducing the burden of data processing and transmission.

[0094] Calculate the reconstruction error of the lightweight feature representation relative to the original feature. The reconstruction error measures the amount of information lost during feature extraction; smaller errors indicate more complete information retained. On a test dataset, the calculated reconstruction error was 0.072. If the preset error threshold is 0.1, the feature extraction rule is considered acceptable.

[0095] When the reconstruction error is less than a preset error threshold, feature extraction rules are deployed to the edge computing end, achieving lightweight edge data processing. By deploying these rules, edge devices can locally preprocess and reduce the dimensionality of raw sensor data, transmitting only lightweight features with high information content, significantly reducing communication bandwidth requirements and improving system response speed.

[0096] In an optional embodiment, 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 based on the contribution score includes:

[0097] receiving input feature data, inputting the input feature data into a feature neuron network, calculating an activation state value of each feature neuron, and calculating a corresponding neurotransmitter concentration change rate based on the activation state value, wherein the neurotransmitter concentration change rate is determined by a neurotransmitter release rate and a decay rate;

[0098] Calculating a dynamic activity threshold of a characteristic neuron based on the rate of change of the neurotransmitter concentration, wherein the dynamic activity threshold is adaptively adjusted with the cumulative change of the neurotransmitter concentration; and calculating a synaptic strength modulation factor using the difference between the dynamic activity threshold and the neurotransmitter concentration;

[0099] multiplying the input feature data, the synaptic weight coefficient and the neurotransmitter concentration to calculate the local field response strength, and calculating the feature importance score based on the local field response strength and the synaptic strength modulation factor;

[0100] According to the feature importance score, the Hebbian learning rule is executed to calculate the enhancement amount of the synaptic weight; based on the enhancement amount of the synaptic weight, the feedback inhibition mechanism is executed to calculate the inhibition amount of the synaptic weight;

[0101] The enhancement amount of the synaptic weight is superimposed with the inhibition amount of the synaptic weight, the synaptic weight coefficient is updated, and the updated synaptic weight coefficient is normalized to obtain a final adaptive weight coefficient.

[0102] Input feature data is received, which can be, for example, multi-dimensional sensor data of industrial equipment, including indicators such as temperature, pressure, vibration, and current, with dimensions of up to 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 established in the network accordingly. Each feature neuron calculates its activation state value based on the received input value. The activation state value is calculated using a nonlinear mapping method. For example, when the temperature feature input is 0.78, the corresponding activation state value is 0.83.

[0103] Based on the activation state value of each characteristic neuron, the system calculates the corresponding rate of change of neurotransmitter concentration. The rate of change of neurotransmitter concentration is determined by the neurotransmitter release rate and decay rate. The neurotransmitter release rate is proportional to the activation state value of the characteristic neuron. The higher the activation state value, the greater the release rate. The decay rate reflects the natural dissipation rate of the neurotransmitter in the synaptic cleft. For example, when the activation state value of the temperature characteristic neuron is 0.83, the corresponding neurotransmitter release rate is 0.05 and the decay rate is 0.02, resulting in a final calculated rate of change of neurotransmitter concentration of 0.03.

[0104] The dynamic activity threshold of the characteristic neuron is calculated based on the rate of change of neurotransmitter concentration. This threshold is not fixed but adaptively adjusts with the cumulative change in neurotransmitter concentration. As the neurotransmitter concentration continues to increase, the neuron's dynamic activity threshold increases accordingly, and vice versa. For example, if the neurotransmitter concentration corresponding to the temperature feature increases cumulatively from an initial value of 0.2 to 0.5, its dynamic activity threshold is raised from 0.3 to 0.45. The system uses the difference between the dynamic activity threshold and the neurotransmitter concentration to calculate the synaptic strength modulation factor. 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, corresponding to a synaptic strength modulation factor of 1.2.

[0105] The input feature data, synaptic weight coefficient and neurotransmitter concentration are multiplied to calculate the local field response strength. 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 strength is 0.78×0.65×0.5=0.2535. Based on the local field response strength and the synaptic strength modulation factor, the system calculates the feature importance score. The feature importance score reflects the contribution of a specific feature to the current device state judgment. Continuing with 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.

[0106] Based on the feature importance scores, the system applies the Hebbian learning rule to calculate the amount of synaptic weight enhancement. Hebbian learning embodies the principle that "simultaneously activated neurons strengthen connections." The higher the feature importance score, the greater the synaptic weight enhancement. For example, the temperature feature has an importance score of 0.3042, corresponding to a synaptic weight enhancement of 0.018. Furthermore, to prevent the excessive enhancement of certain feature weights, which can lead to the neglect of other features, the system implements a feedback inhibition mechanism based on the synaptic weight enhancement to calculate the amount of synaptic weight inhibition. This inhibition calculation takes into account the overall network balance. The synaptic weight inhibition corresponding to the temperature feature is 0.008.

[0107] The synaptic weight enhancement and inhibition are superimposed to 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 updated synaptic weight coefficients of all features to ensure that the sum of the weights is 1, resulting in the final adaptive weight coefficient. Assuming that the system processes temperature, pressure, and vibration features simultaneously, the updated weights are 0.66, 0.45, and 0.39, respectively. The final adaptive weight coefficients obtained after normalization are 0.44, 0.3, and 0.26, respectively. These weight coefficients reflect the relative importance of each feature in device status assessment and are dynamically adjusted as the device's operating status changes.

[0108] Through the above implementation, the present invention can dynamically adjust the weight coefficient of each characteristic component according to the real-time changes of the equipment operation data, thereby improving the accuracy and adaptability of anomaly detection.

[0109] In an optional embodiment, executing the Hebbian learning rule based on the feature importance score to calculate the enhancement amount of the synaptic weight; executing the feedback inhibition mechanism based on the enhancement amount of the synaptic weight to calculate the inhibition amount of the synaptic weight includes:

[0110] Input feature data is input into a neural network, a sigmoid activation function is used to perform a nonlinear transformation on the input feature data, the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron are calculated, the neuron with the largest activity value is selected as the winning neuron based on a competitive learning algorithm, and the weight enhancement parameter is dynamically adjusted using the activity state of the winning neuron;

[0111] According to the feature importance score, the Hebbian learning rule is executed to determine the basic enhancement coefficient through the feature importance score, and the weight adjustment coefficient is calculated using the dynamic threshold algorithm in combination with the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron and the current weight saturation, and the enhancement amount of the synaptic weight is obtained according to the weight adjustment coefficient;

[0112] Performing a square sum operation on the enhancement amount of the synaptic weight to obtain an enhancement energy value, and determining whether the enhancement energy value exceeds a preset energy upper limit threshold;

[0113] Calculating the rate of change of the enhancement amount of the synaptic weight in the time dimension, substituting the rate of change and the mean parameter and variance parameter of the Gaussian function into the Gaussian function to construct a Gaussian inhibition function;

[0114] Based on the enhancement amount of the synaptic weight, a feedback inhibition mechanism is executed, and the inhibition strength coefficient is calculated using the output value of the Gaussian inhibition function. The inhibition amount of the synaptic weight is obtained by combining the inhibition learning rate, the current synaptic weight and the degree of neuronal inactivation.

[0115] By evaluating the importance of features, and then combining the Hebbian learning rule to calculate the synaptic weight enhancement, and calculating the weight inhibition through the feedback inhibition mechanism, the dynamic adjustment of the neural network weights is achieved.

[0116] The system receives 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], representing the raw values ​​of different features. This feature data is then fed into a neural network layer consisting of 15 neurons. For each input feature, the system applies a nonlinear transformation using the sigmoid activation function, mapping the input value to the interval (0, 1). For example, an input value of 0.5 is converted to 0.622 by the sigmoid function. Thus, all input features, after transformation, form the activity state vector of the presynaptic neuron.

[0117] Calculate the activity state of the postsynaptic neuron. Assume the synaptic weight matrix of the current network is W, with dimensions of 10×15, representing the connection strengths between the 10 input features and the 15 postsynaptic neurons. The activity state of the postsynaptic neuron is calculated by multiplying the input features by the corresponding weights, summing the results, and then applying the sigmoid activation function. For example, the activity value of the first postsynaptic neuron is 0.75.

[0118] Based on the principle of competitive learning, the system identifies the neuron with the highest activity value as the winning neuron. Assuming that the activity value of the third neuron is 0.88, which is the highest among all neurons, it is selected as the winning neuron. The system dynamically adjusts the weight enhancement parameter according to the activity state of this winning neuron. Specifically, if the activity value of the winning neuron is higher than the preset threshold of 0.8, the weight enhancement parameter is set to a higher value of 0.05; if the activity 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 activity value of the winning neuron is 0.88, which exceeds the threshold of 0.8, the weight enhancement parameter is set to 0.05.

[0119] The Hebbian learning rule is applied based on the feature importance scores. Assume that the feature importance evaluation module provides a score vector of [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 base boost coefficients. For example, for the first feature with an importance score of 0.8, its base boost coefficient might be set to 0.04; for the second feature with a score of 0.4, its coefficient might be set to 0.02.

[0120] The dynamic threshold algorithm calculates the weight adjustment coefficient based on the activity status of the presynaptic and postsynaptic neurons, as well as the current weight saturation. Weight saturation refers to the ratio of the current weight to the maximum weight. For example, if the current value of a weight is 0.6 and the maximum value is 1, its saturation is 0.6. The dynamic threshold algorithm dynamically adjusts the learning rate based on weight saturation: the higher the saturation, the lower the learning rate, preventing excessive weight growth. For example, for a weight with a saturation of 0.6, the adjustment coefficient is 0.75.

[0121] By multiplying the base enhancement coefficient, the activity state of the pre- and post-synaptic neurons, and the weight adjustment coefficient, the system calculates the enhancement of the synaptic weight. For example, the enhancement of the connection from the first feature to the third neuron is 0.04 × 0.622 × 0.88 × 0.75 = 0.0164.

[0122] The squared sum of the enhancements for all synaptic weights is calculated to obtain the enhancement energy value. Assume the calculated enhancement energy value is 0.12. The system determines whether this value exceeds the preset energy upper threshold of 0.2. If so, the system scales down all enhancements to ensure that the total energy does not exceed the threshold. If not, the enhancement remains unchanged. In this example, the enhancement energy value of 0.12 is less than the threshold of 0.2, so no scaling is required.

[0123] Calculate the rate of change of the synaptic weight enhancement over time. Assuming the enhancement over three consecutive time steps is [0.0164, 0.0158, 0.0170], the rate of change can be calculated as the difference between the two most recent steps, 0.0012. This rate of change, along with the mean parameter of 0 and the variance parameter of 0.01, is substituted into the Gaussian function to construct the Gaussian inhibition function. The output of the Gaussian function indicates that inhibition is weaker when the rate of change is close to the mean and stronger when it deviates from the mean.

[0124] Based on the amount of enhancement in the synaptic weight, the system implements a feedback inhibition mechanism. The system uses the output of the Gaussian inhibition function to calculate the inhibition strength coefficient. For example, for a rate of change of 0.0012, the Gaussian function output is 0.9, corresponding to an inhibition strength coefficient of 0.1. The system combines the inhibitory learning rate of 0.02, the current synaptic weight of 0.6, and the neuron inactivation level of 0.3 to calculate the amount of inhibition in the synaptic weight. For example, the inhibition is 0.1 × 0.02 × 0.6 × 0.3 = 0.00036.

[0125] The enhancement and inhibition are combined to update the synaptic weights. For example, for an initial weight of 0.6, the enhancement is 0.0164, and the inhibition is 0.00036, the updated weight is 0.6 + 0.0164 - 0.00036 = 0.61604. In this way, the system optimizes neural network weights based on feature importance, increasing the network's sensitivity to important features. At the same time, a feedback inhibition mechanism prevents excessive weight growth, maintaining network stability and generalization capabilities.

[0126] Figure 3 This is a bar chart comparing the performance of neural networks based on the Hebbian learning rule and feedback inhibition mechanism according to an embodiment of the present invention:

[0127] The figure compares the performance of three mechanisms, traditional Hebbian learning, improved feedback inhibition, and dynamic threshold adjustment, across five key performance metrics. In terms of weight convergence speed, the dynamic threshold adjustment mechanism performs best, reaching 93.8%, a 17.3 percentage point improvement over the traditional method. In terms of feature extraction accuracy, improved feedback inhibition achieves the best result, 91.7%, a 9.4 percentage point improvement over the traditional method. In terms of overfitting resistance, the dynamic threshold adjustment mechanism leads significantly, reaching 91.2%, a 22.5 percentage point improvement over the traditional method. In terms of computational efficiency, traditional Hebbian learning performs well, reaching 85.4%, but is not much different from the other two improved methods. In terms of energy consumption optimization, the dynamic threshold adjustment mechanism performs the best, achieving an optimization result of 94.3%, a 31.5 percentage point improvement over the traditional method. Overall, the dynamic threshold adjustment mechanism demonstrates significant advantages across multiple performance metrics, with particularly significant improvements in convergence speed, overfitting resistance, and energy optimization, demonstrating that this mechanism can better balance learning effectiveness and computational resource consumption.

[0128] In an optional embodiment, generating a fault feature representation based on the adaptive weight coefficient and the device state feature vector; building a fault diagnosis rule base, and matching the fault feature representation with the fault diagnosis rule base includes:

[0129] Multiplying the device state feature vector by the adaptive weight coefficient to generate an initial fault feature representation; constructing a device fault knowledge graph, processing the device fault knowledge graph using a graph neural network, learning the feature representation of each node, and extracting a main diagnostic path based on the feature representation;

[0130] Building an initial fault diagnosis rule base according to the main diagnostic path, matching the initial fault feature representation with the rules in the initial fault diagnosis rule base to obtain a rule matching effect;

[0131] Calculating a confidence score for each rule based on the rule matching effect, dynamically adjusting the rule weight according to the confidence score, and deleting rules whose confidence scores are lower than a preset confidence threshold;

[0132] New rule candidates are generated using the equipment fault knowledge graph, and the rules with confidence scores higher than a preset confidence threshold in the rule candidates are added to the initial fault diagnosis rule base to obtain an optimized fault diagnosis rule base; the initial fault feature representation is finally matched with the optimized fault diagnosis rule base, and the fault diagnosis result is output.

[0133] After obtaining the device state feature vector and adaptive weight coefficient, multiply them together to generate the initial fault signature. For example, the device state feature vector includes multidimensional features such as temperature, vibration, noise, and current. For example, suppose the feature vector of a 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 signature value is 52.07.

[0134] When building a knowledge graph for equipment failures, entities such as equipment components, fault types, symptoms, and environmental factors are considered nodes, and relationships between entities are considered edges. For example, for compressor equipment, nodes include "bearing," "bearing overheating," "abnormal vibration," and "inadequate lubrication," while edges include relationships such as "cause," "manifested by," and "caused by." A complete knowledge graph can contain hundreds of nodes and thousands of edges, forming a complex network of fault associations.

[0135] A graph neural network is used to process the equipment fault knowledge graph and learn the feature representation of each node. During implementation, the nodes in the knowledge graph are first randomly initialized, and each node is assigned a 64-dimensional feature vector. Through a message passing mechanism, each node aggregates information from neighboring nodes and, after multiple layers of nonlinear transformations, updates its own feature representation. After three rounds of iteration, each node acquires a feature representation that incorporates the graph's structural information.

[0136] Based on the learned node feature representations, the primary diagnostic path is extracted. This method calculates the similarity between the initial fault feature representation and the features of each fault node, selects the top five fault nodes with the highest similarity, and traces the paths from these nodes to the fault cause node to form the primary diagnostic path. For example, for a "bearing overheating" fault, the diagnostic paths are "bearing overheating → abnormal vibration → insufficient lubrication" and "bearing overheating → abnormal temperature → cooling system failure."

[0137] An initial fault diagnosis rule base is constructed based on the extracted main diagnostic paths. Each rule consists of a precondition, conclusion, and initial weight. For example, the rule "IF temperature > 80°C AND vibration > 10 THEN bearing overheat with a weight of 0.8" or the rule "IF current > 15°C AND temperature > 75°C THEN motor overload with a weight of 0.75" is considered. The initial rule base typically contains 50-100 rules, covering the main fault types.

[0138] The initial fault feature representation is matched against the rules in the initial fault diagnosis rule base to obtain the rule matching results. The matching process calculates the degree to which the feature representation satisfies the rule prerequisites and generates a matching score. For example, for the motor feature vector above, the matching score for the first rule is 0.92, and the matching score for the second rule is 0.78.

[0139] A confidence score is calculated for each rule based on its matching performance. The confidence score takes into account the rule's 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, its confidence score is calculated as 0.9 × 0.85 × 0.7 = 0.5355.

[0140] Dynamically adjust rule weights based on confidence scores, deleting rules with confidence scores below a preset confidence threshold. For example, if the preset confidence threshold is 0.4, rules with confidence scores below 0.4 will be deleted. For retained rules, their weights are updated based on their confidence scores, for example, new weight = old weight × (1 + confidence score × 0.2).

[0141] Generate new rule candidates using the equipment failure knowledge graph. By analyzing important paths in the knowledge graph that are not covered by the current rule base, supplementary rules are generated. For example, if a strong correlation between "current fluctuation" and "bearing wear" is found but not covered by the current rule base, the rule "IF current fluctuation > 20% AND operating time > 5000h THEN bearing wear is high WITH weight 0.65" is generated.

[0142] The newly generated rule candidates are evaluated and those with confidence scores above a preset confidence threshold are added to the initial fault diagnosis rule base, resulting in an optimized fault diagnosis rule base. If a new rule has a confidence score of 0.58, which is above the threshold of 0.4, it is added to the rule base. The optimized rule base typically contains 20% fewer inefficient rules than the initial rule base, while also adding 15% more high-quality new rules.

[0143] The initial fault signature is finally matched against the optimized fault diagnosis rule base, and the fault diagnosis results are output. The matching process takes rule weights into account, calculates a comprehensive matching score, selects the top three fault types with the highest scores as the diagnosis results, and provides a confidence percentage. For example, the final diagnosis results of "bearing overheating (87%)," "inadequate lubrication (76%)," and "cooling system failure (62%)" provide maintenance personnel with clear fault location and cause.

[0144] 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.

[0145] In an optional embodiment, an initial fault diagnosis rule base is constructed based on the main diagnostic path, and the initial fault feature representation is matched with the rules in the initial fault diagnosis rule base to obtain a rule matching effect, including:

[0146] Using a multi-head attention mechanism to extract features from the main diagnostic path, generating a query matrix, a key matrix, and a value matrix, obtaining initial attention weights by performing an inner product operation on the query matrix and the key matrix, and multiplying the initial attention weights by the value matrix to obtain an enhanced feature representation;

[0147] Constructing a generative adversarial network model, wherein a generator of the generative adversarial network model generates candidate rules based on the enhanced feature representation, wherein the candidate rules include fault feature vectors, temporal dependencies, fault category codes, and rule confidences, and a discriminator of the generative adversarial network model calculates the authenticity scores of the candidate rules;

[0148] screening the candidate rules based on the authenticity scores, retaining candidate rules with authenticity scores higher than a preset authenticity threshold as valid candidate rules; combining the main diagnostic path with the valid candidate rules to form an initial diagnostic rule, and calculating the validity score of the initial diagnostic rule based on the authenticity scores;

[0149] Constructing a time-varying fault feature map, extracting features from the time-varying fault feature map using spatiotemporal convolution, generating a new hidden layer state by combining the historical hidden layer state with the current input features; and calculating the degree of matching between the new hidden layer state and the initial diagnostic rule based on an attention mechanism.

[0150] The validity score and the matching degree are weightedly fused to obtain a rule matching score, and the number of rules whose rule matching scores exceed a preset matching threshold is counted to obtain a rule matching effect.

[0151] like Figure 4 As shown, the method includes:

[0152] A multi-head attention mechanism is used to extract features from the main diagnostic path. The system represents the main diagnostic path as a vector sequence D = {d1, d2, ..., dn}, where each vector di represents a feature of a step in the diagnostic path. The multi-head attention mechanism generates a query matrix Q, a key matrix K, and a value matrix V by setting eight attention heads, each with a dimension of 64.

[0153] Taking the fault diagnosis of a piece of power equipment as an example, the input main diagnostic path consists of the steps "voltage fluctuation-temperature anomaly-vibration anomaly-fault alarm." After linear projection, the diagnostic path vector is projected into the query space, key space, and value space. By calculating the dot product of the query matrix and the key matrix, an initial attention weight matrix is ​​obtained. The weight values ​​range from 0 to 1, reflecting the correlation between the diagnostic steps. After normalizing the initial weights, they are multiplied by the value matrix V to obtain the enhanced feature representation E = {e1, e2, ..., en}. Each feature vector has a dimension of 512, capturing the key information and inter-step dependencies in the diagnostic path.

[0154] A generative adversarial network model was constructed to generate diagnostic rules. The generator G employs a four-layer fully connected network structure. The input layer is the enhanced feature representation E. The number of neurons in the hidden layers is 256, 128, and 64, respectively. The output layer has the dimension of the rule representation. Based on the enhanced feature representation, the generator generates candidate rules R = {r1, r2, ..., rm}. Each rule ri consists of four components: a fault feature vector F (of dimension 128), a temporal dependency T (represented by a 20-dimensional vector), a fault category code C (one-hot encoding, with the dimension being the number of fault categories), and a rule confidence S (a scalar value ranging from 0 to 1).

[0155] In practical applications, such as motor bearing failure scenarios, the generated candidate rules include the feature vectors of "vibration frequency exceeds 200 Hz and lasts longer than 30 seconds, and the temperature rise rate exceeds 5°C / minute." The temporal relationship indicates that temperature anomalies precede vibration anomalies. The fault category code points to the "bearing wear" category, and the rule confidence is 0.92.

[0156] The discriminator D employs a three-layer convolutional network architecture, with candidate rules as input and a plausibility score (a scalar value between 0 and 1) as output. The discriminator calculates the plausibility score for each candidate rule by comparing the feature distribution differences between the candidate rule and the true expert rule. During training, the generator and discriminator perform adversarial learning with 5000 iterations, a learning rate of 0.0001, and a batch size of 32.

[0157] Candidate rules are screened based on the authenticity scores output by the discriminator. A authenticity threshold of 0.75 is set, and candidate rules with authenticity scores above the threshold are retained as valid candidate rules. The main diagnostic paths are combined with the valid candidate rules to form an initial diagnostic rule set I = {i1, i2, ..., ik}. For each initial diagnostic rule, its effectiveness score V is calculated based on the authenticity score. This is calculated as a weighted combination of the authenticity score and the rule complexity. The rule complexity is determined by the number of features and conditions contained in the rule.

[0158] A time-varying fault feature graph G is constructed, organizing the real-time fault data into a three-dimensional tensor with the dimensions [time step, number of sensors, feature dimension]. A spatiotemporal convolutional network is used to extract time-varying features. The network consists of three spatiotemporal convolutional layers with a kernel size of 3×3, a stride of 1, padding of 1, and output channels of 32, 64, and 128, respectively. The spatiotemporal convolution considers both the spatial relationship between sensors and the changing trends in the time series. A long short-term memory network is used to process sequential information, with a hidden state dimension of 256. A gating mechanism combines the historical hidden state h(t-1) with the current input feature x(t) to generate a new hidden state h(t).

[0159] Based on the attention mechanism, the matching degree M between the new hidden state h(t) and the initial diagnostic rule set I is calculated. For each rule i, its similarity score with the current state is calculated, with higher scores indicating a higher matching degree. The validity score V and the matching degree M are weighted and fused, with weights of 0.4 and 0.6, respectively, to obtain the rule matching score S. The matching threshold is set to 0.8, and the number of rules with matching scores exceeding the threshold is counted to obtain the rule matching effect. In a substation equipment fault diagnosis test, 65 valid candidate rules were screened from the initially generated 100 candidate rules. Ultimately, 42 rules had matching scores exceeding the threshold, resulting in a matching effect of 42%, which improves matching efficiency by approximately 23% compared to traditional rule matching methods.

[0160] In an optional embodiment, the matching probabilities of various types of faults are 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, generating edge fault warning information includes:

[0161] Acquiring fault characteristic data, and performing standardization processing on the fault characteristic data to obtain standardized characteristic data;

[0162] calculating a membrane potential change rate of the standardized characteristic data, the membrane potential change rate being obtained by a weighted combination of a ratio of the standardized characteristic data to a membrane time constant, a synaptic input current, and an external input current, and calculating an average membrane potential of a neuronal group based on the membrane potential change rate;

[0163] Multiplying the action potential of the preceding neuron by the action potential of the succeeding neuron to obtain a synaptic activity level, multiplying the difference between the synaptic activity level and the plasticity threshold by the learning rate to obtain a change in the synaptic weight, and updating the synaptic weight according to the change in the synaptic weight;

[0164] Calculating the firing probability of the neuron group based on the difference between the average membrane potential and the firing threshold, and performing a weighted combination of the firing probability and the prior probability obtained from historical fault data statistics through the updated synaptic weight to obtain a fault matching probability;

[0165] performing normalization processing on the fault matching probability to obtain a normalized matching probability, selecting a maximum value of the normalized matching probability as a current fault probability, and determining a corresponding fault type according to the current fault probability;

[0166] The current fault probability is compared with a preset probability threshold, and when the current fault probability exceeds the preset probability threshold, edge fault warning information is generated according to the fault type, the current fault probability, and the current timestamp.

[0167] Fault signature data is collected from sensors on multiple edge devices, including multi-dimensional data such as temperature, pressure, vibration, and current. After acquiring this data, it is normalized to eliminate dimensional differences. For example, raw temperature data of 95°C, current data of 8.7A, and vibration frequency data of 62Hz collected by an edge computing node can be converted into dimensionless, standardized signature data using the Z-score normalization method: 0.82, 1.35, and -0.42, respectively.

[0168] After normalization, the membrane potential change rate is calculated. In this embodiment, the membrane time constant is set to 20ms, the synaptic input current weight is 0.7, and the external input current weight is 0.3. Taking the above-mentioned normalized 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. Based on this membrane potential change rate, according to the neuronal membrane potential update rule, the average membrane potential of the neuronal population can be calculated. Assuming that the initial average membrane potential is -70mV, after 10ms of integration, the average membrane potential is updated to -68.023mV.

[0169] The calculation of synaptic activity involves the interaction between the action potentials of preceding and succeeding neurons. In this example, assume that the action potentials of preceding neurons A, B, and C are 0.8, 0.6, and 0.7, respectively, and that of succeeding neurons D and E are 0.5 and 0.9, respectively. Through matrix multiplication, the synaptic activity matrix can be obtained. For example, the synaptic activity from A to D is 0.8 × 0.5 = 0.4. Assuming a plasticity threshold of 0.35 and a learning rate of 0.05, the change in synaptic weight from A to D is (0.4 - 0.35) × 0.05 = 0.0025. Similar calculations are performed on all synaptic weights and updated. For example, the initial weight of 0.6 is updated to 0.6025.

[0170] The firing probability of a neuron group is calculated based on the difference between the mean membrane potential and the firing threshold. If the firing threshold is set to -65mV, the difference is -68.023-(-65)=-3.023mV. Using a sigmoid function, the firing probability is 0.427. This is then combined with prior probabilities derived from historical fault data, such as a prior probability of 0.3 for overheating, 0.25 for overload, and 0.2 for vibration anomalies, and weighted using the updated synaptic weights.

[0171] If the synaptic weight associated with the overheating fault is 0.68, the matching probability of the overheating fault is calculated to be 0.427 × 0.3 × 0.68 = 0.0871. Similarly, the matching probabilities of other fault types are calculated, such as 0.0762 for the overload fault and 0.0513 for the vibration abnormality.

[0172] Normalize the fault matching probabilities so that the sum of the probabilities of all fault types equals 1. After normalization, the probability of an overheating fault is 0.0871 / (0.0871 + 0.0762 + 0.0513 + ...) = 0.407. The probability of an overload fault is 0.356, and the probability of an abnormal vibration is 0.237. Select the maximum value of 0.407 as the current fault probability, determining that the current fault type is an overheating fault.

[0173] The current failure probability of 0.407 is compared with a preset probability threshold (for example, 0.4). Since 0.407 is greater than 0.4, the system generates an edge fault warning. The warning includes the fault type (overheating), the failure probability (0.407), and the current timestamp (2023-05-15 14:28:36). This warning can be sent to the operation and maintenance management platform via a message queue, triggering the corresponding alarm mechanism.

[0174] This method can run in real time on edge computing devices to provide early predictions for various types of faults. The system continuously learns new fault patterns and improves prediction accuracy through adaptive adjustments to synaptic weights. For example, after deployment on an industrial production line, this method successfully predicted an overheating failure in critical equipment three hours in advance, allowing maintenance personnel ample time to intervene and avoiding losses caused by production interruptions. Field tests have shown that this method has an 87% accuracy rate for predicting common faults, a 23% improvement over traditional threshold monitoring methods, and a 35% reduction in false alarms.

[0175] This method has low computational complexity. On an edge computing device equipped with a quad-core processor and 4GB of memory, it can process 200 sets of feature data per second, meeting real-time monitoring requirements. The system also supports adaptive parameter adjustment, such as automatically adjusting membrane time constants and learning rates based on different device characteristics, enhancing the method's versatility and robustness.

[0176] Figure 5 This is a bar chart comparing the performance of the edge fault warning system based on the neuron model according to an embodiment of the present invention:

[0177] The figure compares and analyzes the performance differences among traditional threshold detection methods, machine learning models, and the neural dynamic membrane potential model across five key performance indicators. In terms of hardware fault detection accuracy, the neural dynamic membrane potential model achieved the best performance, reaching 92.8%, 14.5 percentage points higher than traditional methods. In terms of software anomaly recognition rate, the neural dynamic membrane potential model also led the pack, reaching 91.5%, an 18 percentage point improvement over traditional methods. In terms of network communication fault prediction, the neural dynamic membrane potential model achieved an accuracy of 88.7%, a 23.5 percentage point improvement over traditional methods. In terms of data anomaly detection sensitivity, the neural dynamic membrane potential model performed the best, reaching 93.2%, an 11.8 percentage point improvement over traditional methods. In terms of system performance degradation warning, the neural dynamic membrane potential model achieved a warning accuracy of 90.6%, a 20.9 percentage point improvement over traditional methods. Overall, the neural dynamic membrane potential model significantly outperformed traditional threshold detection methods and machine learning models across all evaluation indicators, particularly in fault prediction and anomaly detection, demonstrating the model's efficiency and reliability in system fault diagnosis and warning.

[0178] A second aspect of an embodiment of the present invention provides a device telemetry data fault analysis system based on machine learning, including:

[0179] The first unit is configured to receive multidimensional telemetry data collected by a device telemetry data collection module, perform standardization processing on the multidimensional telemetry data, and generate standardized data;

[0180] The second unit is used to extract the time series features and statistical features of the standardized data and establish feature evaluation indicators; filter the core feature set according to the feature evaluation indicators; and construct lightweight feature extraction rules based on the core feature set to achieve lightweight edge data processing;

[0181] a third unit, configured to process the standardized data using the lightweight feature extraction rule to generate a device state feature vector, calculate a 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;

[0182] A fourth unit is 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, 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; and generate edge fault warning information when the current fault probability exceeds a preset probability threshold;

[0183] The fifth unit is used to transmit the edge fault warning information, the fault type and the current fault probability to the cloud management platform for storage and analysis.

[0184] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including:

[0185] processor;

[0186] a memory for storing processor-executable instructions;

[0187] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0188] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0189] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0190] 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 it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for analyzing equipment telemetry data faults 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 temporal features and statistical features of the standardized data and establishing feature evaluation indicators; screening a core feature set based on the feature evaluation indicators; constructing lightweight feature extraction rules 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; Constructing a fault diagnosis rule base and matching the fault feature representation with the fault diagnosis rule base; calculating the matching probability of each type of fault, selecting the maximum matching probability as the current fault probability, and determining the corresponding fault type; generating edge fault warning information when the current fault probability exceeds a preset probability threshold, including: multiplying the device state feature vector by the adaptive weight coefficient to generate an initial fault feature representation; constructing a device fault knowledge graph, processing the device fault knowledge graph using a graph neural network, learning the feature representation of each node, and extracting a main diagnostic path based on the feature representation; Building an initial fault diagnosis rule base according to the main diagnostic path, matching the initial fault feature representation with the rules in the initial fault diagnosis rule base to obtain a rule matching effect; Calculating a confidence score for each rule based on the rule matching effect, dynamically adjusting the rule weight according to the confidence score, and deleting rules whose confidence scores are lower than a preset confidence threshold; Generate new rule candidates using the equipment fault knowledge graph, add rules with confidence scores higher than a preset confidence threshold among 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 a fault diagnosis result; 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: Receive device telemetry data and generate a multidimensional time series feature sequence; calculate the autocorrelation coefficient of each feature in the multidimensional 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 a weighted combination of the autocorrelation coefficient, the cross-correlation coefficient, the information gain value, and the inter-class dispersion to obtain a feature importance score; Sorting the multidimensional time series feature sequence in descending order based on the feature importance score, selecting the top-ranked features as a candidate feature set; calculating the mutual information between the features in the candidate feature set, constructing a feature correlation matrix, and screening a core feature subset from the candidate feature set based on a 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: receiving input feature data, inputting the input feature data into a feature neuron network, calculating an activation state value of each feature neuron, and calculating a corresponding neurotransmitter concentration change rate based on the activation state value, wherein the neurotransmitter concentration change rate is determined by a neurotransmitter release rate and a decay rate; Calculating a dynamic activity threshold of a characteristic neuron based on the rate of change of the neurotransmitter concentration, wherein the dynamic activity threshold is adaptively adjusted with the cumulative change of the neurotransmitter concentration; and calculating a 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 strength, and calculating the feature importance score based on the local field response strength and the synaptic strength modulation factor; According to the feature importance score, the Hebbian learning rule is executed to calculate the enhancement amount of the synaptic weight; based on the enhancement amount of the synaptic weight, the feedback inhibition mechanism is executed to calculate the inhibition amount of the synaptic weight; The enhancement amount of the synaptic weight is superimposed with the inhibition amount of the synaptic weight, the synaptic weight coefficient is updated, and the updated synaptic weight coefficient is normalized to obtain a final adaptive weight coefficient.

4. The method according to claim 3, characterized in that According to the feature importance score, the Hebbian learning rule is executed to calculate the enhancement amount of the synaptic weight; Based on the enhancement amount of the synaptic weight, executing a feedback inhibition mechanism, and calculating the inhibition amount of the synaptic weight includes: Input feature data is input into a neural network, a sigmoid activation function is used to perform a nonlinear transformation on the input feature data, the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron are calculated, the neuron with the largest activity value is selected as the winning neuron based on a competitive learning algorithm, and the weight enhancement parameter is dynamically adjusted using the activity state of the winning neuron; According to the feature importance score, the Hebbian learning rule is executed to determine the basic enhancement coefficient through the feature importance score, and the weight adjustment coefficient is calculated using the dynamic threshold algorithm in combination with the activity state of the presynaptic neuron and the activity state of the postsynaptic neuron and the current weight saturation, and the enhancement amount of the synaptic weight is obtained according to the weight adjustment coefficient; Performing a square sum operation on the enhancement amount of the synaptic weight to obtain an enhancement energy value, and determining whether the enhancement energy value exceeds a preset energy upper limit threshold; Calculating the rate of change of the enhancement amount of the synaptic weight in the time dimension, substituting the rate of change and the mean parameter and variance parameter of the Gaussian function into the Gaussian function to construct a Gaussian inhibition function; Based on the enhancement amount of the synaptic weight, a feedback inhibition mechanism is executed, and the inhibition strength coefficient is calculated using the output value of the Gaussian inhibition function. The inhibition amount of the synaptic weight is obtained by combining the inhibition learning rate, the current synaptic weight and the degree of neuronal inactivation.

5. The method according to claim 1, wherein An initial fault diagnosis rule base is constructed according to the main diagnostic path, and the initial fault feature representation is matched with the rules in the initial fault diagnosis rule base to obtain a rule matching effect, including: Using a multi-head attention mechanism to extract features from the main diagnostic path, generating a query matrix, a key matrix, and a value matrix, obtaining initial attention weights by performing an inner product operation on the query matrix and the key matrix, and multiplying the initial attention weights by the value matrix to obtain an enhanced feature representation; Constructing a generative adversarial network model, wherein a generator of the generative adversarial network model generates candidate rules based on the enhanced feature representation, wherein the candidate rules include fault feature vectors, temporal dependencies, fault category codes, and rule confidences, and a discriminator of the generative adversarial network model calculates the authenticity scores of the candidate rules; screening the candidate rules based on the authenticity scores, retaining candidate rules with authenticity scores higher than a preset authenticity threshold as valid candidate rules; combining the main diagnostic path with the valid candidate rules to form an initial diagnostic rule, and calculating the validity score of the initial diagnostic rule based on the authenticity scores; Constructing a time-varying fault feature map, extracting features from the time-varying fault feature map using spatiotemporal convolution, generating a new hidden layer state by combining the historical hidden layer state with the current input features; and calculating the degree of matching between the new hidden layer state and the initial diagnostic rule based on an attention mechanism. The validity score and the matching degree are weighted and fused to obtain a rule matching score, and the number of rules whose rule matching scores exceed a preset matching threshold is counted to obtain a rule matching effect.

6. The method according to claim 1, wherein 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 the preset probability threshold, generate edge fault warning information including: Acquiring fault characteristic data, and performing standardization processing on the fault characteristic data to obtain standardized characteristic data; calculating a membrane potential change rate of the standardized characteristic data, the membrane potential change rate being obtained by a weighted combination of a ratio of the standardized characteristic data to a membrane time constant, a synaptic input current, and an external input current, and calculating an average membrane potential of a neuronal group based on the membrane potential change rate; Multiplying the action potential of the preceding neuron by the action potential of the succeeding neuron to obtain a synaptic activity level, multiplying the difference between the synaptic activity level and the plasticity threshold by the learning rate to obtain a change in the synaptic weight, and updating the synaptic weight according to the change in the synaptic weight; Calculating the firing probability of the neuron group based on the difference between the average membrane potential and the firing threshold, and performing a weighted combination of the firing probability and the prior probability obtained from historical fault data statistics through the updated synaptic weight to obtain a fault matching probability; performing normalization processing on the fault matching probability to obtain a normalized matching probability, selecting a maximum value of the normalized matching probability as a current fault probability, and determining a corresponding fault type according to the current fault probability; The current fault probability is compared with a preset probability threshold, and when the current fault probability exceeds the preset probability threshold, edge fault warning information is generated according to the fault type, the current fault probability, and the current timestamp.

7. A machine learning-based equipment telemetry data fault analysis system, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is configured to receive multidimensional telemetry data collected by a device telemetry data collection module, perform standardization processing on the multidimensional telemetry data, and generate standardized data; The second unit is used to extract the temporal features and statistical features of the standardized data and establish feature evaluation indicators; filter the core feature set according to the feature evaluation indicators; and construct lightweight feature extraction rules based on the core feature set to achieve lightweight edge data processing; a third unit, configured to process the standardized data using the lightweight feature extraction rule to generate a device state feature vector, calculate a 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 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 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 fifth unit is used to transmit the edge fault warning information, the fault type and the current fault probability to the cloud management platform for storage and analysis.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images