Circuit breaker mechanical fault pre-diagnosis method based on time series analysis

Abstract

The application relates to the technical field of fault diagnosis, in particular to a circuit breaker mechanical fault pre-diagnosis method based on time series analysis. First, the operation data of the circuit breaker mechanical are collected in real time through the installation of sensors, and the data are preprocessed; then, the data are divided into independent time series sequences based on standard operation data by adopting a time series segmentation strategy, time domain features and frequency domain features are extracted for each independent time series sequence, and time series dependent patterns are extracted in combination with autocorrelation functions and cross-correlation functions; the time domain features, the frequency domain features and the time series dependent patterns are fused through a time-frequency-time series collaborative modeling framework; finally, based on the time series features, a federal Bayesian unit is constructed for fault diagnosis, wherein a local unit provides fault type and confidence prediction, and a global unit integrates the models of multiple local units through a federal learning framework. The application realizes circuit breaker mechanical fault pre-diagnosis through time series analysis.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, in particular to a circuit breaker mechanical fault pre-diagnosis method based on time series analysis. BACKGROUND

[0002] Traditional circuit breaker mechanical fault diagnosis methods mostly rely on artificial experience or simple statistical models, which are difficult to process and analyze complex multi-dimensional time series data, resulting in limitations in recognition. Existing algorithms often fail to effectively capture long-term dependencies in device operation when processing time series data, making it difficult to identify potential fault patterns. In addition, traditional methods often rely too much on static feature extraction methods in the extraction process of time series features, failing to fully exploit the time series dependencies and complex periodic fluctuations contained in time series data. Existing methods mainly rely on some simple statistics (such as mean, standard deviation) in time domain feature extraction, but lack in-depth mining of potential nonlinear and time-varying features in complex fault signals. Although frequency domain feature extraction can capture frequency components, traditional frequency domain analysis methods often fail to effectively reflect the frequency variation of circuit breaker mechanics under different working conditions.

[0003] Therefore, a circuit breaker mechanical fault pre-diagnosis method based on time series analysis is proposed. SUMMARY

[0004] The present application aims to provide a circuit breaker mechanical fault pre-diagnosis method based on time series analysis, which realizes circuit breaker mechanical fault pre-diagnosis through time series analysis.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] The circuit breaker mechanical fault pre-diagnosis method based on time series analysis comprises:

[0007] Install a sensor to collect real-time operation data of the circuit breaker mechanics for preprocessing and generate standard operation data;

[0008] Based on the standard operation data of the circuit breaker mechanics, a time series segmentation strategy is used for time series segmentation to generate independent time series sequences. Time domain features and frequency domain features are extracted for each independent time series sequence. Time series dependency patterns are extracted using autocorrelation functions and cross-correlation functions. The autocorrelation function measures the correlation of the same independent time series sequence at different time lags, and the cross-correlation function measures the mutual dependence of two different independent time series sequences at different time lags. Through a time-frequency-time series collaborative modeling framework, a cross-attention mechanism is used to establish relationships between different feature layers. The time domain features, frequency domain features and time series dependency patterns are fused to generate circuit breaker mechanical time series features.

[0009] Based on the mechanical timing characteristics of the circuit breaker, a federated Bayesian unit is constructed to perform circuit breaker mechanical fault diagnosis. The federated Bayesian unit includes a local unit and a global unit. The local unit provides fault type and confidence prediction. The local unit is subjected to fluctuation detection through a graph structure variational encoder. The learning rate of the local unit is dynamically adjusted according to the severity of the fluctuation, and the parameters of the locally trained Bayesian model are encrypted. The global unit periodically integrates the encrypted Bayesian model parameters of all local units to perform pre-diagnosis of circuit breaker mechanical faults.

[0010] The real-time operation data includes vibration data, temperature data, pressure data, displacement data and current data of the circuit breaker mechanical parts during operation;

[0011] The preprocessing includes using median filtering to remove noise, using a generative adversarial network to perform data filling on real-time operation data, and performing standardization processing through Z-score standardization to generate standard operation data.

[0012] The specific process of using the time series segmentation strategy to perform time series segmentation is as follows:

[0013] Obtain standard operating data of circuit breaker machinery;

[0014] Performing time series segmentation on the standard operating data using a time series segmentation strategy, wherein the time series segmentation strategy is combined with abnormality-sensitive segmentation of an autoencoder, and the autoencoder is trained using the standard operating data to learn and reconstruct a time series data pattern under a normal operating state, thereby generating a reconstruction error;

[0015] A reconstruction error distribution is calculated based on the reconstruction error, and a reconstruction error threshold is set according to the reconstruction error distribution. When the reconstruction error of a continuous time window continuously exceeds the reconstruction error threshold, the time point at which the deviation begins is marked as a new time series segmentation point to generate an independent time series sequence.

[0016] The specific process of generating the mechanical timing characteristics of the circuit breaker is as follows:

[0017] The time domain features, the frequency domain features and the time sequence dependency pattern are integrated through a time-frequency-time sequence collaborative modeling framework to generate a mechanical time sequence feature of the circuit breaker;

[0018] The time-frequency-time series collaborative modeling framework includes a time domain feature layer, a frequency domain feature layer, a time series dependency pattern layer and a fusion layer. The time domain feature layer obtains the time domain features, the time domain feature layer obtains the frequency domain features, and the time series dependency pattern layer obtains the time series dependency pattern.

[0019] The fusion layer establishes relationships between different feature layers through a cross-attention mechanism, performs feature fusion, and generates mechanical timing features of the circuit breaker.

[0020] The time domain features are statistical quantities in the standard operating data of the circuit breaker mechanism, including mean, variance, standard deviation and peak value;

[0021] The frequency domain features are the main frequency and the spectral barycenter obtained by converting the standard operating data of the circuit breaker mechanism into the frequency domain through fast Fourier transform;

[0022] The time series dependent mode is extracted through the autocorrelation function and the cross-correlation function, and the autocorrelation function measures the correlation of the same independent time series sequence at different time lags;

[0023] The cross-correlation function measures the mutual dependence of two different independent time series sequences at different time lags.

[0024] The federated Bayesian unit includes a local unit and a global unit, the local unit locally extracts the circuit breaker mechanical time series features of each independent time series sequence to predict the fault type and confidence, and uploads the local model parameters to the global unit;

[0025] The global unit integrates the model parameters uploaded from multiple local units through a federated learning framework to build a global fault diagnosis model.

[0026] The specific process of fault type and confidence prediction provided by the local unit is:

[0027] The local unit receives the circuit breaker mechanical time series features generated by the time-frequency-time series collaborative modeling framework;

[0028] Using historical fault data and normal operating data, a fault classification model based on Bayesian inference is constructed to generate posterior probability;

[0029] According to the calculated posterior probability, the confidence of each fault type, i.e., the probability of belonging to the fault type, is output;

[0030] The parameters of the locally trained Bayesian model are encrypted and uploaded to the global unit;

[0031] The local unit is subjected to volatility detection by a graph structure variational encoder, the learning rate of the local unit is dynamically adjusted according to the severity of the volatility, and the parameters of the locally trained Bayesian model are encrypted.

[0032] The specific process of the fault classification model based on Bayesian inference is:

[0033] Define the fault types of the circuit breaker mechanism, including operation mechanism failure, transmission component failure, contact system failure, insulation component failure and auxiliary circuit failure;

[0034] A local Bayesian classification model is built based on historical data. By calculating the prior probability of the fault type and the conditional probability of the time series characteristics under a given fault type, the model is trained to identify the time series characteristic patterns corresponding to different fault types. During the Bayesian classification model training process, a Gaussian mixture model is combined with the conditional probability distribution to model the conditional probability distribution.

[0035] The classification layer is used to perform pre-diagnosis on the input circuit breaker mechanical timing characteristics.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. The present invention uses a multi-head cross-attention mechanism to parallel process and fuse time domain, frequency domain and timing-dependent features, effectively overcoming the limitations of the traditional cross-attention mechanism; each attention head independently focuses on the relationship between specific feature layers, ensuring that the model can capture richer feature information from multiple angles; at the same time, through the additive fusion splicing method, the mutual influence between different feature layers is strengthened, making the generated circuit breaker mechanical timing features more comprehensive and accurate, enhancing the expressive power of the model, and improving the accuracy of fault diagnosis, especially in complex timing data, and can better identify potential fault modes and abnormal behaviors.

[0038] 2. This invention uses comprehensive feature extraction across the time, frequency, and timing-dependent domains to more accurately capture the multidimensional information of circuit breaker mechanical operating data. In particular, it leverages autocorrelation functions to reveal the inherent patterns and periodic variations of internal motion within the device, while cross-correlation functions provide in-depth analysis of lead-lag relationships and fault propagation paths between different sensor signals (such as vibration and current). These detailed timing-dependent patterns, combined with time and frequency domain features, significantly enhance sensitivity and identification of early-stage potential circuit breaker faults, effectively reducing false alarm rates and providing maintenance personnel with more accurate fault location information, thereby significantly improving device reliability and predictive maintenance.

[0039] 3. By constructing federated Bayesian units, this approach effectively combines the privacy protection of local data with the accuracy of the global fault diagnosis model. Local units independently process time series data and generate fault predictions, quantifying the uncertainty of the predictions through Bayesian inference, providing strong support for accurate determination of fault type and confidence level. Global units aggregate local model parameters through a federated learning framework, ensuring data privacy while also improving diagnostic accuracy without directly exchanging data. This approach effectively improves the accuracy and safety of circuit breaker mechanical fault diagnosis and exhibits good adaptability and scalability in multi-device environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1This is a flow chart of the circuit breaker mechanical fault pre-diagnosis method based on time series analysis of the present invention;

[0041] Figure 2 It is a pre-processing flow chart of the present invention;

[0042] Figure 3 This is a data flow chart of the circuit breaker mechanical fault pre-diagnosis method based on time series analysis of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 are within the scope of protection of the present invention.

[0044] Example 1:

[0045] See also Figure 1 The present invention provides a method for pre-diagnosis of mechanical faults of circuit breakers based on time series analysis. The technical solution is as follows:

[0046] Install sensors to collect real-time operating data of circuit breaker machinery for pre-processing and generate standard operating data;

[0047] Based on the standard operating data of the circuit breaker machine, a time series segmentation strategy is used to perform time series segmentation to generate independent time series sequences. Time domain features and frequency domain features are extracted for each independent time series sequence. The time series dependency pattern is extracted by combining the autocorrelation function and the cross-correlation function. The time domain features, the frequency domain features, and the time series dependency pattern are fused through a time-frequency-time series collaborative modeling framework to generate the circuit breaker machine time series features.

[0048] A federated Bayesian mechanism is constructed based on the mechanical timing characteristics of the circuit breaker to perform circuit breaker mechanical fault diagnosis. The federated Bayesian unit includes a local unit and a global unit. The local unit provides fault type and confidence prediction, and the global unit integrates the contents of multiple local units through a federated learning framework to perform pre-diagnosis of circuit breaker mechanical faults.

[0049] Specifically, the real-time operation data includes vibration data, temperature data, pressure data, displacement data and current data of the mechanical parts of the circuit breaker during operation;

[0050] The preprocessing includes using median filtering to remove noise, using generative adversarial networks to fill in real-time operation data, and performing standardization processing through Z-score standardization to generate standard operation data. Figure 2; Set the sampling frequency of real-time running data to 100 Hz, collect 100 data points per second, collect 6000 data points per minute, form a time series data sequence; Take the vibration data as an example, assuming that there is a sudden noise value in the original vibration data at a certain time, the value of the vibration data at a certain time point is 1200, but in fact it should be within the normal fluctuation range of 50-60; Using median filtering, assuming that the sliding window size is 3, the median filter will select one point on each side of the current point, and calculate the median of these points to replace the noise, such as the two data points around the current time point are 55 and 56, then the value after median filtering is 56, effectively removing the noise;

[0051] The data filling is missing due to sensor failure when the temperature data t=10s, the generator obtained by training through the generative adversarial network generates temperature values consistent with the time sequence rule based on the time points before and after it (the temperature is 30°C at t=9s and 32°C at t=11s), and the time sequence consistency loss ensures that the filled data is consistent with the data around the time point in trend and change mode, the time sequence consistency loss minimizes the difference between the filled data and the data around the time point through the gradient consistency loss, the similarity measure and the periodic constraint, ensuring that the generated temperature values not only fall within the mean range, but also conform to the change trend of the data; The gradient change rate between the filled value and the time points before and after it ensures that the generated data is consistent with the surrounding data in change rate; The similarity measure quantifies the difference between the generated data and the real data through L1 loss, L2 loss or time difference loss; The periodic constraint forces the generated data to follow a periodic trend, so that the generated filled data is consistent with the periodic characteristics.

[0052] The standardization process is to use Z-score standardization to unify all data on a standard scale, the specific formula is:

[0053]

[0054] Wherein is the original data, is the mean of the feature, is the standard deviation of the feature, is the generated standard running data.

[0055] Through the above preprocessing process, the quality and accuracy of the running data of the mechanical components of the circuit breaker can be effectively improved, the noise is removed by median filtering, the missing data is filled by the generative adversarial network and the time sequence consistency loss is combined to ensure the consistency of the data change trend, and the Z-score standardization process eliminates the dimensional difference between features. These steps effectively improve the accuracy, consistency and comparability of the data, providing more reliable input for fault diagnosis.

[0056] An autoencoder model was trained based on standard operating data. The autoencoder, through an encoder and decoder, maps the input time series data into a low-dimensional space and is able to reconstruct the time series under normal operating conditions. The training process allows the autoencoder to learn the normal operating pattern. The autoencoder's architecture consists of an input layer (five features), two hidden layers (64 and 32 neurons), and a latent space layer with 16 neurons. The decoder's architecture is the opposite of the encoder, restoring the original input dimensions. The mean squared error (MSE) loss function was used, and the Adam algorithm was used as the optimizer. The initial learning rate was 0.001, and the learning rate was reduced by 50% every 10 epochs. The batch size was set to 64, and the training epochs were 50. Dropout (0.2) regularization was used to prevent overfitting. During training, the model learned the patterns of the standard operating data and optimized data reconstruction capabilities by minimizing the reconstruction error.

[0057] After the model training is completed, the standard operating data is input into the autoencoder to calculate the reconstruction error of each time series. The reconstruction error represents the difference between the autoencoder reconstructed data and the original data. It can be measured by the mean square error. The reconstruction errors of all time points are counted to obtain the distribution of the reconstruction error. According to the characteristics of the reconstruction error distribution, a reconstruction error threshold is set using the maximum reconstruction error method. That is, during the training process, the reconstruction errors of all time points are calculated, and the maximum reconstruction error value is found. This maximum reconstruction error value is used as the threshold. When the reconstruction error of any time point exceeds this maximum value, it will be regarded as a potential abnormal behavior.

[0058] The reconstruction error of each window is calculated using a sliding time window. If the reconstruction error of consecutive time windows continuously exceeds the set threshold, it means that the data in that part has deviated from the normal pattern, which may indicate the onset of a fault. This time point is recorded as a new time series segmentation point. Based on the detected time series segmentation point, the data is segmented into multiple independent time series subsequences. Each subsequence represents an independent time period and contains complete time series data and potential fault information.

[0059] Through the reconstruction error method based on the autoencoder, the standard operating data of the circuit breaker machinery can be effectively modeled and anomalies detected. By setting the threshold by the maximum reconstruction error method, data that deviates from the normal operating mode can be accurately identified, thereby detecting potential faults at an early stage. At the same time, the sliding time window strategy ensures fine-grained detection of data, which can capture the starting point of the fault in a timely manner, greatly improving the accuracy and response speed of fault warning.

[0060] Extract the time series features from the independent time series, fuse the time domain features, the frequency domain features and the time series dependency pattern through the time-frequency-time series collaborative modeling framework to generate the circuit breaker mechanical time series features. Figure 3 ;

[0061] The time-frequency-time series collaborative modeling framework includes a time domain feature layer, a frequency domain feature layer, a time series dependency pattern layer and a fusion layer. The time domain feature layer obtains the time domain features, the time domain features obtain frequency domain features, and the time series dependency pattern layer obtains the time series dependency pattern.

[0062] The fusion layer uses a cross-attention mechanism to establish relationships between different feature layers, perform feature fusion, and generate circuit breaker mechanical timing features. Furthermore, multiple independent attention heads are used in the cross-attention mechanism to simultaneously focus on the relationships between different feature layers. Each head can learn the mutual influence between specific features in different subspaces. By computing multiple heads in parallel, the model can capture feature relationships from different perspectives and provide a more comprehensive feature understanding.

[0063] Multiple attention heads are created for each pair of feature layers in the time domain feature layer, the frequency domain feature layer, and the timing dependency pattern layer, and cross-calculations are performed on them. Specifically, cross-calculations are performed on the query vector of each feature layer with the key and value vectors of other layers. Each attention head processes these calculations independently, thereby extracting the potential relationships between each feature layer. Different attention heads process different relationships. For example, one head may focus on the relationship between time domain features and timing dependency, while another head may focus on the connection between frequency domain features and timing dependency. After parallel calculation by multiple heads, the outputs of each head are spliced ​​and merged through linear transformation to generate the mechanical timing features of the circuit breaker. The splicing adopts additive fusion to strengthen the mutual influence between feature layers. The mechanical timing features of the circuit breaker simultaneously contain information from the time domain, frequency domain, and timing dependency layers.

[0064] The time domain, frequency domain and timing-dependent features are processed and fused in parallel through a multi-head cross-attention mechanism, effectively overcoming the limitations of the traditional cross-attention mechanism; each attention head independently focuses on the relationship between specific feature layers, ensuring that the model can capture richer feature information from multiple angles; at the same time, through the additive fusion splicing method, the mutual influence between different feature layers is strengthened, making the generated circuit breaker mechanical timing features more comprehensive and accurate, enhancing the expressive power of the model and improving the accuracy of fault diagnosis, especially in complex timing data, and being able to better identify potential fault modes and abnormal behaviors.

[0065] The time domain feature extraction process is to extract time domain statistics from the standard operating data of the circuit breaker machine, including the mean, variance, standard deviation and peak value of the standard operating data of the circuit breaker machine; the mean is the average value of the calculated data, which represents the central trend of the signal; the variance is the degree of dispersion of the calculated data, which reflects the amplitude of the signal fluctuation; the standard deviation is the square root of the variance, which is used to measure the dispersion of the data; the peak value is the maximum value in the calculated data, which indicates the extreme situation of the signal.

[0066] The frequency domain feature extraction is to apply fast Fourier transform to the standard operating data of the circuit breaker machine to convert the data from the time domain to the frequency domain to generate frequency domain features; the fast Fourier transform can decompose time series data into different frequency components, revealing the frequency characteristics of the signal. The frequency domain features include the main frequency and the center of gravity of the spectrum. The main frequency is the frequency position with the maximum energy in the spectrum, that is, the most significant frequency component in the signal; the main frequency reflects the main vibration frequency of the circuit breaker machine and can indicate normal or abnormal operating status; the center of gravity of the spectrum is the weighted average frequency of each frequency component on the spectrum, which measures the overall distribution of the frequency components and reflects the concentration of the signal frequency characteristics.

[0067] The time series dependency pattern extraction extracts the time series dependency pattern through the autocorrelation function and the cross-correlation function. The autocorrelation function calculates the correlation of the same independent time series at different time lags and evaluates the time dependency of the data, that is, checks the correlation of the signal itself in the past and future moments, following the formula:

[0068]

[0069] in Indicates the circuit breaker mechanism at time point The instantaneous value of a certain standard operating data at the current moment; represents the lag, i.e. the time interval; Indicates the circuit breaker mechanism at time point The same standard operating data; Indicates the long-term average value of the standard operating data timing signal of the circuit breaker machinery. represents the expected value operator, Operational data for circuit breaker at a point in time and The covariance between Running data timing signal for circuit breaker The variance of is the autocorrelation coefficient;

[0070] The cross-correlation function measures the mutual dependence between two independent time series at different time lags, revealing the temporal correlation between the two signals, and follows the formula:

[0071]

[0072] in Indicates the first time point of the circuit breaker mechanism Standard operating data of Indicates the second time of the circuit breaker mechanism Standard operating data of Representing time series At the time point The value of , where the hysteresis refers to Relative to The lag or lead, when When comparing and the future , The changes occurred in Lag, that is lag behind ; When Ahead of ; and Represents the circuit breaker operation data timing signal and the long-term average of and Represents the circuit breaker operation data timing signal and The standard deviation of represents the expected value operator, Running data for circuit breakers At the time point The deviation and At the time point The covariance between the deviations of is the mutual correlation coefficient;

[0073] By calculating the autocorrelation coefficient sequence of each independent time series and the mutual correlation coefficient sequence between different types of time series sequences, and then extracting key features from the values, peak positions and curve shapes of these coefficient sequences as quantitative representations of the time series dependency pattern, a time series dependency pattern is generated.

[0074] The comprehensive feature extraction of these time-domain, frequency-domain, and timing-dependent patterns enables more accurate capture of multi-dimensional information about circuit breaker mechanical operating data. In particular, the autocorrelation function reveals the inherent patterns and periodic variations in the device's internal motion, while the cross-correlation function provides in-depth analysis of the lead-lag relationships and fault propagation paths between different sensor signals (such as vibration and current). These detailed timing-dependent patterns, combined with time-domain and frequency-domain features, significantly enhance sensitivity and identification of potential circuit breaker faults in their early stages, effectively reducing false alarm rates and providing maintenance personnel with more accurate fault location information, thereby significantly improving equipment reliability and predictive maintenance.

[0075] Based on the mechanical time series characteristics of the circuit breaker, a federated Bayesian unit is constructed to perform circuit breaker mechanical fault and diagnosis, such as Figure 3 As shown, the federated Bayesian unit consists of a local unit and a global unit. The local unit locally predicts the fault type and confidence level for each independent time series extracted from the circuit breaker mechanical timing features and uploads the local model parameters to the global unit. Specifically, the local unit receives each independent time series from the circuit breaker mechanical unit and preprocesses the data to extract time domain features, frequency domain features, and timing dependency patterns to generate circuit breaker mechanical timing features. The extracted circuit breaker mechanical timing features are then analyzed and the fault type is predicted using the trained local model. This process outputs a fault prediction result and a corresponding confidence level based on the input time series data. For example, the local model may be trained based on temperature, pressure, vibration data, etc. to predict whether an overload, short circuit, or other type of fault has occurred and output the corresponding probability value. The local unit then uploads the prediction result and model parameters (such as weights and biases) acquired during the training process to the global unit.

[0076] Through the collaboration of local and global units, the accuracy and efficiency of circuit breaker mechanical fault diagnosis are improved while ensuring data privacy. Local units independently process and analyze their own time series data, extract key features, and perform fault predictions. This allows for real-time identification of each circuit breaker's status and potential fault types, while also quantifying the confidence level of the predictions. The global unit, through a federated learning framework, aggregates model parameters from different local units, integrating multi-source information to generate a global diagnostic model.

[0077] The local model is a fault classification model based on Bayesian inference. It employs Bayesian inference to handle different types of circuit breaker mechanical faults. It infers the circuit breaker fault type by combining historical data with real-time time series data for joint modeling. The Bayesian model uses prior probabilities (historical fault data or expert knowledge) and likelihood functions (generated probabilities based on feature data) to calculate posterior probabilities and predict the probability of each fault type. Based on these predictions, the local model outputs the fault type and the corresponding confidence level. The confidence level indicates the reliability of the diagnosis, quantifies the uncertainty of the prediction, and helps determine the occurrence and severity of the fault.

[0078] The global unit receives model parameters uploaded by multiple local units. These parameters may vary in different local units because the timing data and fault types processed by each local unit have local differences. The global unit aggregates and fuses the model parameters uploaded by each local unit through the federated learning framework, and finally constructs a global fault diagnosis model. Under the entire federated learning framework, the data of each local unit is not directly transmitted to the global unit. Instead, only the model parameters are uploaded. This mechanism effectively protects the privacy of each local unit and prevents the leakage of sensitive data.

[0079] In each local unit, the time series data is modeled through a graph-structured variational encoder. Each time point in the time series data is regarded as a node in the graph, and edges are established between the nodes through time series dependencies to form the graph structure of the time series data. The graph-structured variational encoder extracts volatility features from the time series data of the local unit through graph convolution operations. The output of each node is the volatility feature at that time point, and it is mapped to the latent space through the encoder to obtain a low-dimensional representation of the data. For each time series window, the variational encoder performs volatility analysis based on statistics such as the standard deviation and mean change of the volatility feature. If the volatility is severe (for example, the volatility index exceeds the set threshold), it means that the device state changes drastically, and the volatility of the local unit is determined to be large. The learning rate of the local unit is dynamically adjusted according to the severity of the volatility.

[0080] By constructing federated Bayesian units, this approach effectively combines the privacy protection of local data with the accuracy of the global fault diagnosis model. Local units independently process time series data and generate fault predictions, quantifying the uncertainty of these predictions through Bayesian inference, providing strong support for accurate determination of fault type and confidence level. Global units aggregate local model parameters through a federated learning framework, ensuring data privacy while also improving diagnostic accuracy without directly exchanging data. This approach effectively improves the accuracy and safety of circuit breaker mechanical fault diagnosis and exhibits good adaptability and scalability in multi-device environments.

[0081] The fault types are defined according to the working principle of the circuit breaker and common fault types, including operating mechanism failure, transmission component failure, contact system failure, insulation component failure and auxiliary circuit failure.

[0082] Using historical fault data, a Bayesian classification model for each fault type is constructed. During this process, the prior probability of each fault type, that is, the probability of the fault occurring, is calculated. Using this historical data, combined with specific time series features, the model is trained to identify the time series feature patterns of different fault types. At the same time, a Gaussian mixture model is used to model the conditional probability distribution of each fault type. The Gaussian mixture model is applied to model the characteristic data of each fault type, and an independent Gaussian mixture model is trained for each fault type. The expectation maximization algorithm is used to estimate the parameters of each model, including the mean, covariance matrix, and weights. The expectation maximization algorithm iteratively adjusts these parameters until the model converges, thereby accurately representing the conditional probability distribution of each fault type. The trained Gaussian mixture model is embedded in the Bayesian classification model to more accurately capture the conditional probability distribution of each fault type, thereby improving the accuracy and robustness of fault diagnosis and effectively enhancing the model's adaptability to complex time series data.

[0083] During the training process, Bayesian inference uses prior and conditional probabilities to calculate posterior probabilities. Posterior probabilities reflect the likelihood of a fault type given observed data (i.e., input time series features). This process not only identifies the fault type but also quantifies the confidence level of the prediction, enhancing diagnostic reliability. The trained Bayesian classification model uses the classification layer to perform a preliminary diagnosis of the input circuit breaker mechanical time series features. At this layer, the input time series feature data is compared with the model learned during training. The posterior probability is calculated, and the predicted fault type and its corresponding confidence level are output. This diagnoses potential circuit breaker fault types and provides decision support.

[0084] Example 2:

[0085] A power company employed the present invention's method for predicting circuit breaker failures based on time series analysis for its operational circuit breaker equipment. The company installed multiple sensors on each circuit breaker component, including vibration, temperature, pressure, displacement, and current sensors, to collect real-time operational data. The data collection frequency was set to 100 Hz, with 100 data points collected per second and 6,000 data points collected per minute, forming a time series data set.

[0086] During the data preprocessing stage, median filtering technology is first used to remove noise from the collected vibration data. In particular, when sudden noise occurs in the vibration data (for example, the value of a vibration data point is 1200, while the normal fluctuation range is 50-60), the noise point is replaced with the median of the two surrounding data points through a sliding window method, thereby effectively removing anomalies in the data.

[0087] Furthermore, in the event of a sensor failure (e.g., temperature sensor data loss), a generative adversarial network (GAN) is used to fill in the missing data. Specifically, when temperature data is missing at t=10s, the trained generator network generates a fill-in value that conforms to the temporal order based on the preceding and following time points (e.g., the temperature is 30°C at t=9s and 32°C at t=11s). This ensures data consistency and accuracy, and ensures that the fill-in data follows the changing pattern of the surrounding time points.

[0088] Next, all data were normalized using Z-score normalization so that data of different features were on the same scale for subsequent analysis.

[0089] Through the time series segmentation strategy, the autoencoder is used to learn and reconstruct the standard operating data, and the abnormal behavior is determined by calculating the reconstruction error. When the reconstruction error exceeds the preset threshold, it is marked as the time point of potential failure and an independent time series subsequence is generated.

[0090] When constructing time series features, a time-frequency-time series collaborative modeling framework is used to fuse time-domain features, frequency-domain features, and time-series dependency patterns to generate circuit breaker mechanical time series features. This feature fusion is further optimized through a multi-head cross-attention mechanism, ensuring that the model can extract key feature relationships from different perspectives, thereby more accurately capturing circuit breaker mechanical failure modes.

[0091] Using a Bayesian classification model combined with local Bayesian inference, fault types are classified. Each local unit outputs the fault type and confidence level based on the extracted time series features. The model parameters of multiple local units are uploaded to the global unit via a federated learning framework to construct a fault diagnosis model, ultimately achieving accurate fault pre-diagnosis for circuit breaker mechanisms.

[0092] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A circuit breaker mechanical fault pre-diagnosis method based on time series analysis, characterized in that: include: Install sensors to collect real-time operating data of circuit breaker machinery for pre-processing and generate standard operating data; Based on the standard operating data of the circuit breaker machinery, a time series segmentation strategy is used to perform time series segmentation to generate independent time series sequences. Time domain features and frequency domain features are extracted for each independent time series sequence. The time series dependency pattern is extracted by combining the autocorrelation function and the cross-correlation function. The autocorrelation function measures the correlation of the same independent time series sequence at different time lags to generate an autocorrelation coefficient sequence; the cross-correlation function measures the mutual dependence of two different independent time series sequences at different time lags to generate a cross-correlation coefficient sequence. The values, peak positions and curve shapes are extracted based on the autocorrelation coefficient sequence and the cross-correlation coefficient sequence to generate a time series dependency pattern; the time-frequency-time series collaborative modeling framework is used to establish relationships between different feature layers through a cross-attention mechanism, and the time domain features, the frequency domain features and the time series dependency pattern are fused to generate the circuit breaker machinery time series features; A federated Bayesian unit is constructed based on the mechanical timing characteristics of the circuit breaker to perform pre-diagnosis of the circuit breaker mechanical fault; the federated Bayesian unit includes a local unit and a global unit. The local unit locally predicts the fault type and confidence of the circuit breaker mechanical timing characteristics extracted from each independent timing sequence, and uploads the local model parameters to the global unit; the global unit integrates the model parameters uploaded from multiple local units through a federated learning framework to construct a global fault diagnosis model; the local unit provides fault type and confidence prediction, and the local unit is subjected to fluctuation detection through a graph structure variational encoder. The learning rate of the local unit is dynamically adjusted according to the severity of the fluctuation, and the parameters of the locally trained Bayesian model are encrypted. The global unit periodically integrates the encrypted Bayesian model parameters of all local units to perform pre-diagnosis of the circuit breaker mechanical fault.

2. The circuit breaker mechanical fault pre-diagnosis method based on time series analysis according to claim 1, characterized in that: The real-time operation data includes vibration data, temperature data, pressure data, displacement data and current data of the circuit breaker mechanical parts during operation; The preprocessing includes using median filtering to remove noise, using a generative adversarial network to perform data filling on real-time operation data, and performing standardization processing through Z-score standardization to generate standard operation data.

3. The circuit breaker mechanical fault pre-diagnosis method based on time series analysis according to claim 1, characterized in that: The specific process of using the time series segmentation strategy to perform time series segmentation is as follows: Obtain standard operating data of circuit breaker machinery; Performing time series segmentation on the standard operating data using a time series segmentation strategy, wherein the time series segmentation strategy is combined with abnormality-sensitive segmentation of an autoencoder, and the autoencoder is trained using the standard operating data to learn and reconstruct a time series data pattern under a normal operating state, thereby generating a reconstruction error; A reconstruction error distribution is calculated based on the reconstruction error, and a reconstruction error threshold is set according to the reconstruction error distribution. When the reconstruction error of a continuous time window continuously exceeds the reconstruction error threshold, the time point at which the deviation begins is marked as a new time series segmentation point to generate an independent time series sequence.

4. The circuit breaker mechanical fault pre-diagnosis method based on time series analysis according to claim 1, characterized in that: The specific process of generating the mechanical timing characteristics of the circuit breaker is as follows: The time domain features, the frequency domain features and the time sequence dependency pattern are integrated through a time-frequency-time sequence collaborative modeling framework to generate a mechanical time sequence feature of the circuit breaker; The time-frequency-time series collaborative modeling framework includes a time domain feature layer, a frequency domain feature layer, a time series dependency pattern layer and a fusion layer. The time domain feature layer obtains the time domain features, the time domain feature layer obtains the frequency domain features, and the time series dependency pattern layer obtains the time series dependency pattern. The fusion layer establishes relationships between different feature layers through a cross-attention mechanism, performs feature fusion, and generates mechanical timing features of the circuit breaker.

5. The circuit breaker mechanical fault pre-diagnosis method based on time series analysis according to claim 4, characterized in that: The time domain features are statistics in the standard operation data of the circuit breaker machinery, including mean, variance, standard deviation and peak value; The frequency domain features are obtained by converting the standard operating data of the circuit breaker mechanism into the frequency domain through fast Fourier transform, including the main frequency and the center of gravity of the spectrum; The time series dependency pattern is extracted by autocorrelation function and cross-correlation function, wherein the autocorrelation function measures the correlation of the same independent time series at different time lags; The cross-correlation function measures the mutual dependence between two different independent time series at different time lags.

6. The circuit breaker mechanical fault pre-diagnosis method based on time series analysis according to claim 1, characterized in that: The specific process of the local unit providing fault type and confidence prediction is as follows: The local unit receives the mechanical time series features of the circuit breaker generated by fusion of the time-frequency-time series collaborative modeling framework; Using historical fault data and normal operation data, a fault classification model based on Bayesian inference is constructed to generate posterior probabilities; According to the calculated posterior probability, the confidence level of each fault type is output, that is, the probability of belonging to the fault type; Encrypting and uploading the parameters of the locally trained Bayesian model to the global unit; The fluctuation of the local unit is detected by a graph structure variational encoder, the learning rate of the local unit is dynamically adjusted according to the severity of the fluctuation, and the parameters of the locally trained Bayesian model are encrypted.

7. The circuit breaker mechanical fault pre-diagnosis method based on time series analysis according to claim 6, characterized in that: The specific process of the fault classification model based on Bayesian inference is as follows: Defining mechanical fault types of circuit breaker, including operating mechanism fault, transmission component fault, contact system fault, insulation component fault and auxiliary circuit fault; A local Bayesian classification model is built based on historical data. By calculating the prior probability of the fault type and the conditional probability of the time series characteristics under a given fault type, the model is trained to identify the time series characteristic patterns corresponding to different fault types. During the Bayesian classification model training process, a Gaussian mixture model is combined with the conditional probability distribution to model the conditional probability distribution. The classification layer is used to perform pre-diagnosis on the input circuit breaker mechanical timing characteristics.

Images