Classification method for fault data of low-voltage primary electrical equipment

Abstract

The electrical low-voltage primary equipment fault data classification method disclosed herein includes: triggering a corresponding data processing strategy based on electromagnetic interference intensity data to obtain a standardized dataset; filtering the standardized dataset using a pre-built correlation model to obtain a core dataset; assessing the equipment fault stage of the target equipment based on the deviation of the core dataset; adjusting the weights of various operational data in the core dataset based on the equipment fault stage to obtain a weighted fusion feature set; and using a pre-trained fault classification model to perform fault classification and identification to determine the fault classification result of the target equipment. This disclosure can dynamically adjust the data processing strategy according to the electromagnetic interference intensity, improving data quality; using a correlation model for filtering to retain core operational data; and adjusting weights based on the assessed specific fault evolution stage, effectively adapting to the fault evolution cycle and improving the sensitivity to early, subtle fault identification.

Description

Technical Field

[0001] This disclosure relates to the field of electrical equipment fault diagnosis technology, and in particular to a method for classifying fault data of low-voltage primary electrical equipment. Background Technology

[0002] Low-voltage primary electrical equipment is responsible for the transmission and control of electrical energy in power distribution systems. During long-term operation, it is susceptible to faults such as insulation aging, poor contact, and overheating due to electrical stress, mechanical wear, and environmental factors. These faults initially present with weak signals but deteriorate rapidly over time. Most unplanned power outages in power distribution systems are caused by these faults. Therefore, classifying and identifying fault data early is crucial for improving troubleshooting efficiency and reducing maintenance costs.

[0003] Existing fault data classification methods have the following technical shortcomings. First, fixed noise reduction thresholds cannot adapt to dynamic electromagnetic interference such as pulses and harmonics in the field, resulting in poor data quality. Furthermore, during the data removal process, redundant data removal relies solely on purely mathematical statistical correlations without considering the physical fault mechanism, easily leading to the accidental deletion of key complementary features. Second, the signal characteristics of equipment differ significantly between the early and late stages of a fault. Existing methods use fixed feature extraction sets, which cannot dynamically adapt to the entire fault evolution cycle, causing missed detection of early, weak faults and misjudgment of faults in the development stage.

[0004] In summary, there is an urgent need for a fault data classification method for low-voltage primary electrical equipment that has strong anti-interference capabilities and can identify the characteristics of the entire fault cycle, in order to overcome the shortcomings of related technologies. Summary of the Invention

[0005] This disclosure provides a method for classifying fault data of low-voltage primary electrical equipment, which addresses the shortcomings of existing fault data classification methods, such as weak anti-interference capability and poor ability to identify weak signals.

[0006] In view of the above problems, firstly, embodiments of this disclosure provide a method for classifying fault data of low-voltage primary electrical equipment, including: Acquire electromagnetic interference intensity data at the power distribution site and descriptive data of various operating data of the target equipment; Based on the electromagnetic interference intensity data, the corresponding data processing strategy is triggered to process the descriptive data of each operational data to obtain a standardized dataset; By using a pre-built correlation model, the standardized dataset is filtered based on the changes in description data of different operational data in the fault scenario, to obtain the core dataset. Based on the deviation between the current value and the normal state value of the descriptive data of each operational data in the core dataset, the equipment failure stage of the target device is assessed. Based on the equipment failure stage, the weights of each operational data item in the core dataset are adjusted to obtain a weighted fusion feature set; A pre-trained fault classification model is used to identify faults based on a weighted fusion feature set, and the fault classification result of the target device is determined.

[0007] In conjunction with the first aspect, in one possible implementation, the step of triggering a corresponding data processing strategy based on the electromagnetic interference intensity data to process the descriptive data of each operational data item to obtain a standardized dataset includes: The effective value of the electromagnetic interference intensity data is calculated based on the electromagnetic interference intensity data, and the electromagnetic interference level at the power distribution site is determined based on the effective value; the electromagnetic interference level includes: weak interference level, medium interference level, and strong interference level. When the power distribution site is under a low interference level, the moving average filtering noise reduction strategy is triggered to process the noise reduction of various operating data. When the power distribution site is under medium interference level, the Kalman filter noise reduction strategy is triggered to reduce noise in various operating data; In the case of strong interference at the power distribution site, noise reduction is performed on various operating data based on wavelet threshold noise reduction and Kalman filter noise reduction strategies respectively. Data cleaning is performed on the various operational data after noise reduction to fill in missing values ​​and remove outliers. After data cleaning, all operational data are standardized according to a preset standardization formula to obtain a standardized dataset.

[0008] In conjunction with the first aspect, in one possible implementation, the standardized dataset is filtered based on a pre-built correlation model, using pre-determined descriptions of changes in different operational data within a fault scenario, to obtain a core dataset, including: The correlation coefficient formula is used to calculate the statistical correlation between various operational data. Based on predetermined fault mechanism correlation factors, and combined with operational data, the fault mechanism correlation degree is calculated using the fault mechanism correlation degree calculation formula; wherein, the fault mechanism correlation factors are determined according to the synchronicity of changes in operational data in the fault scenario; Based on the statistical correlation and fault mechanism correlation of the data, a comprehensive correlation is obtained through weighted fusion. Based on the calculated comprehensive correlation of each piece of operational data, each piece of operational data is scored according to a preset correlation threshold. Based on the scoring results, operational data in the standardized dataset that are lower than the preset score are removed to obtain the core dataset.

[0009] In conjunction with the first aspect, in one possible implementation, assessing the equipment failure stage of the target device based on the deviation between the current value and the normal state value of the descriptive data of each operational data item in the core dataset includes: Calculate the deviation between the current value and the normal state value for each piece of running data; Based on the number of operational data points that deviate from the normal state value and the degree of deviation of each operational data point, the equipment failure stage of the target equipment is determined according to the preset evaluation criteria; wherein, the equipment failure stage includes: the nascent stage, the development stage, and the outbreak stage.

[0010] In conjunction with the first aspect, in one possible implementation, the step of adjusting the weights of various operational data items in the core dataset based on the equipment failure stage to obtain a weighted fusion feature set includes: Based on the equipment failure stage of the target device, extract the operational data corresponding to that equipment failure stage from the core dataset; Based on the ratio of the descriptive data of each extracted operational data to the total descriptive data of all extracted operational data, calculate the information entropy corresponding to each operational data. Based on the information entropy, the weight of each extracted running data item is determined, and the descriptive data of each running data item is weighted and fused with the weight to obtain a weighted fused feature set.

[0011] In conjunction with the first aspect, in one possible implementation, the step of employing a pre-trained fault classification model to perform fault classification and identification based on a weighted fusion feature set, and determining the fault classification result of the target device, includes: The weighted fusion feature set is input into the fault classification model; the fault classification model is a convolutional neural network architecture model trained with fault category as the true value and corresponding operating data as the input. The feature extraction layer in the fault classification model is used to perform feature extraction on the weighted fused feature set; The extracted features are input into the feature fusion layer of the fault classification model for feature fusion operation; The probability of the fused features is calculated by the classification function in the fault classification model, and the output includes the fault category of the target device and the corresponding confidence level.

[0012] In conjunction with the first aspect, in one possible implementation, the fault classification models are deployed on the edge computing terminals corresponding to the target devices; The method further includes: The parameters of the fault classification models of each edge computing terminal are obtained. Based on the sample size of each fault classification model, the parameters of each fault classification model are aggregated. Based on the aggregated parameters, each fault classification model is fine-tuned.

[0013] In conjunction with the first aspect, in one possible implementation, it further includes: Based on the fault classification results, the accuracy of the fault classification model's inference is evaluated, and the fault classification model is adjusted based on the accuracy evaluation results, including: The fault classification results output by the fault classification model within a preset time period are compared with the actual fault situation at the power distribution site to determine the accuracy of the fault classification model's inference. If the accuracy is lower than a preset threshold, determine whether the fault category in the fault classification result is included in the pre-training data of the fault classification model; If the fault category is included in the pre-training data, adjust the preset correlation threshold and reconstruct the core dataset; If the fault category is not included in the pre-training data, a transfer learning process is performed to enable the fault classification model to learn new fault categories.

[0014] In conjunction with the first aspect, in one possible implementation, the execution of the transfer learning process includes: The sample data of the new fault category is mixed with the sample data in the pre-training data to construct the target domain dataset for transfer learning; The current fault classification model is used as the source domain model, and the network layer parameters used for feature extraction in the source domain model are frozen. Using the target domain dataset, the feature fusion layer and classification function in the source domain model are adjusted through an adaptive learning rate.

[0015] Secondly, this disclosure also provides a fault data classification device for low-voltage primary electrical equipment, including: The processing module is used to acquire electromagnetic interference intensity data and descriptive data of various operating data of the target equipment at the power distribution site; and to trigger the corresponding data processing strategy based on the electromagnetic interference intensity data to process the descriptive data of various operating data to obtain a standardized dataset. The filtering module is used to filter the standardized dataset based on the changes in the description data of different operating data in the fault scenario through a pre-built correlation model to obtain the core dataset. The evaluation module is used to evaluate the equipment failure stage of the target device based on the deviation between the current value and the normal state value of the descriptive data of each operational data in the core dataset. The first adjustment module is used to adjust the weights of various operational data in the core dataset based on the equipment failure stage to obtain a weighted fusion feature set; The identification module is used to identify faults based on a weighted fusion feature set using a pre-trained fault classification model, and to determine the fault classification result of the target device.

[0016] The beneficial effects of the embodiments disclosed herein include: This disclosure provides a method for classifying fault data of low-voltage primary electrical equipment. The method includes: acquiring electromagnetic interference intensity data and descriptive data of various operating data of the target equipment at the power distribution site; triggering a corresponding data processing strategy based on the electromagnetic interference intensity data to process the descriptive data of various operating data to obtain a standardized dataset; filtering the standardized dataset based on the changes in the descriptive data of different operating data in the fault scenario using a pre-constructed correlation model to obtain a core dataset; assessing the equipment fault stage of the target equipment based on the deviation between the current value and the normal state value of the descriptive data of various operating data in the core dataset; adjusting the weights of various operating data in the core dataset based on the equipment fault stage to obtain a weighted fusion feature set; and using a pre-trained fault classification model to perform fault classification and identification based on the weighted fusion feature set to determine the fault classification result of the target equipment. This disclosure can dynamically adjust the data processing strategy according to the electromagnetic interference intensity, improving the data quality under complex operating conditions; and using the correlation model for filtering to retain core operating data; and adjusting the weights based on the assessed specific fault evolution stage to effectively adapt to the fault evolution cycle and improve the sensitivity of identifying early weak faults. Attached Figure Description

[0017] Figure 1 A flowchart illustrating the method for classifying fault data of low-voltage primary electrical equipment provided in this embodiment of the disclosure; Figure 2 A schematic diagram of the structure of the electrical low-voltage primary equipment fault data classification device provided in the embodiments of this disclosure. Detailed Implementation

[0018] This disclosure provides a method for classifying fault data of low-voltage primary electrical equipment. Preferred embodiments of this disclosure are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit this disclosure. Furthermore, the embodiments and features described herein can be combined with each other unless otherwise specified.

[0019] This disclosure provides a method for classifying fault data of low-voltage primary electrical equipment, such as... Figure 1 As shown, it can be implemented as follows: S101. Obtain electromagnetic interference intensity data and descriptive data of various operating data of the target equipment at the power distribution site; S102. Trigger the corresponding data processing strategy according to the electromagnetic interference intensity data, process the descriptive data of each running data, and obtain a standardized dataset. S103. By using a pre-constructed correlation model, the standardized dataset is filtered based on the changes in the description data of different operational data in the fault scenario, to obtain the core dataset. S104. Based on the deviation between the current value and the normal state value of the description data of each running data in the core dataset, assess the equipment failure stage of the target device. S105. Based on the equipment failure stage, adjust the weights of each operational data item in the core dataset to obtain a weighted fusion feature set; S106. Using a pre-trained fault classification model, fault classification and identification are performed based on a weighted fusion feature set to determine the fault classification result of the target equipment.

[0020] In this embodiment of the disclosure, the power distribution site can refer to a low-voltage power distribution room; the target equipment can refer to the low-voltage primary equipment within the low-voltage power distribution room. Descriptive data of the various operational data of the target equipment can be collected by sensors deployed at key monitoring points of the low-voltage primary equipment. Specifically, the various operational data can be divided into multiple modalities, such as electrical quantity data, environmental quantity data, and status quantity data. These collected data can be integrated into a dataset, for example, a dataset can be defined as... .

[0021] In the above set, Electrical quantity data refers to basic physical quantities (such as voltage) used to describe the operating status and power transmission characteristics of target equipment. Current Active power (etc.), and the descriptive data can be the measured values ​​of the above physical quantities; these physical quantities can be obtained from the incoming end, outgoing end and main circuit of the target equipment, etc., and the sampling frequency can be 20kHz~50kHz. Environmental data refers to physical and chemical parameters (such as equipment surface temperature) used to describe the external operating conditions of the target equipment and the state of the equipment under environmental influences. Ambient humidity and dust concentration (etc.), and the descriptive data can be the measured values ​​of the above parameters; these parameters can be obtained from the heat-prone parts of the target device, the area within 1m around the device, etc., and the sampling frequency can be 5Hz~10Hz. State data refers to intrinsic characteristic parameters used to describe the target device (such as the number of switching actions). Insulation resistance Mechanical vibration amplitude These parameters can be collected from the mechanical transmission components, main circuit and housing of the target device, etc., and the sampling frequency can be 0.1Hz~20Hz. Extended modal data can be provided; this part of the data can be selected according to specific needs, such as partial discharge signals and electromagnetic interference intensity signals. wait.

[0022] In one possible implementation, electromagnetic interference intensity data can be obtained by acquiring signals in a specific space through a radio frequency antenna, and then refining them through analysis algorithms to obtain a quantitative index characterizing the electromagnetic energy radiation level of the space where the target device is located.

[0023] Since electromagnetic interference can cause abnormalities in the operational data collected by sensors, this disclosure introduces a dynamic anti-interference mechanism. Based on the electromagnetic interference intensity data of the space where the target device is located, different noise reduction algorithms and filtering strategies are matched to process the collected descriptive data. Combined with data cleaning and other operations, a standardized dataset with high data quality is obtained.

[0024] This allows for the removal of redundant data from the standardized dataset that is not highly correlated with the failure mechanism of the target device. Furthermore, it avoids the shortcomings of redundancy removal methods in related technologies that are detached from physical mechanism constraints. This disclosure uses a correlation model to calculate the synchronicity of changes in different operational data under failure scenarios to obtain the correlation between operational data. This allows for the filtering of the standardized dataset, retaining only the operational data with high correlation to the failure scenario after removing redundant data, thus obtaining the core dataset.

[0025] Based on the obtained core dataset, various operational data can be extracted from it, and the deviation between the current value of the descriptive data for each operational data point and the baseline normal state value can be calculated. Then, based on the degree of deviation, the current stage of the target device's fault evolution can be assessed.

[0026] Based on the failure stage of the target device, corresponding weights can be assigned to each operational data item in the core dataset to obtain a weighted fusion feature set. This feature set can be used to change the attention of the fault classification model to different operational data, enabling the fault classification model to better capture the essence of the fault represented by the descriptive data of the operational data.

[0027] Then, the weighted fusion feature set can be input into the fault classification model, and the fault classification model can infer based on the data in the weighted fusion feature set to finally determine the fault of the target device, thereby realizing the classification of faults based on the fault data of the device (i.e., the data in the operation data that is strongly correlated with the fault of the device).

[0028] In summary, this disclosure can dynamically adjust the data processing strategy according to the intensity of electromagnetic interference, thereby improving the data quality under complex operating conditions; and it uses a correlation model for filtering to retain core operational data; at the same time, it adjusts the weights based on the assessed specific fault evolution stage, effectively adapting to the fault evolution cycle and improving the sensitivity of identifying early and weak faults.

[0029] In another embodiment provided in this disclosure, the step S102 above, "triggering the corresponding data processing strategy according to the electromagnetic interference intensity data, processing the descriptive data of each operational data to obtain a standardized dataset," can be implemented as follows: Step 1: Calculate the effective value of the electromagnetic interference intensity data based on the electromagnetic interference intensity data, and determine the electromagnetic interference level of the power distribution site based on the effective value; The electromagnetic interference levels include: weak interference level, medium interference level and strong interference level. Step 2: When the power distribution site is under weak interference level, trigger the moving average filtering noise reduction strategy to perform noise reduction processing on various operating data; Step 3: When the power distribution site is under medium interference level, trigger the Kalman filter noise reduction strategy to perform noise reduction processing on various operating data; Step 4: When the power distribution site is under strong interference, noise reduction is performed on various operating data based on wavelet threshold noise reduction and Kalman filter noise reduction strategies respectively. Step 5: Perform data cleaning on the various operational data after noise reduction to fill in missing values ​​and remove outliers. Step 6: Perform standardization operations on the cleaned data according to the preset standardization formula to obtain a standardized dataset.

[0030] In this embodiment of the disclosure, the effective value of the electromagnetic interference intensity data can be calculated based on the root mean square of the electromagnetic interference intensity data over a specific time period. It can be divided into three levels. Specifically, under the weak interference level... ; Medium interference level Under strong interference level .

[0031] Under weak interference levels, the interference at the power distribution site is mainly stable noise, thus triggering a moving average filtering strategy with low computational complexity. By averaging the data points within the sliding window, slight data fluctuations can be smoothed with minimal computing power. In one possible implementation, the size of the sliding window... The formula for the moving average filtering noise reduction strategy is: ;in, These are the filtered data points (i.e., the descriptive data of the filtered running data). These are the original data points (i.e., the descriptive data of the collected operational data). The size of the sliding window can be selected. Alternatively, other odd values ​​can be selected as needed.

[0032] Under medium interference levels, target equipment in power distribution sites will be subject to certain electromagnetic interference, thus a Kalman filter noise reduction strategy can be used. Utilizing the state-space prediction and update mechanism of the Kalman filter, the state at the previous time step is combined with the predicted state at the current time step, thereby filtering out random noise in the operating data. Specifically, the state equation of the Kalman filter noise reduction strategy is as follows: The observation equation for the Kalman filtering noise reduction strategy is: .in, This is the state transition matrix; The system noise matrix; The observation matrix; , These are system noise and observation noise (variance dynamically adjusted based on interference intensity), respectively. For the current moment The predicted state (i.e., the predicted current running data); For the previous moment The state (i.e., the running data of the previous moment). for The observed values ​​at a given time (i.e., the measured operational data).

[0033] Under high interference levels, strong electromagnetic interference exists in power distribution sites. A combination of wavelet thresholding and Kalman filtering denoising strategies can be used. First, wavelet thresholding is used to decompose the waveform of continuous operating data, specifically using the dB4 wavelet with a decomposition level of up to three. Then, adaptive thresholding is applied to the high-frequency system to remove high-frequency impulse noise. Subsequently, the signal reconstructed from the wavelet is fed into a Kalman filter for secondary filtering and smoothing. This combined strategy can eliminate complex and strong interference while protecting the real operating data from being filtered out.

[0034] This allows for data cleaning of the noise-reduced operational data to fill in missing values ​​and remove outliers. Specifically, linear interpolation can be used to fill in missing values ​​in continuous operational data; and... The criteria combine equipment failure thresholds to eliminate outliers. If an outlier occurs... If the value exceeds the equipment fault warning threshold, the descriptive data of the operating data can be determined to be an outlier, and the median of the operating data should be used to replace the value. For the first One set of running data; The average of the data over a period of time; The standard deviation of the data over a period of time.

[0035] After data cleaning, the various operational data can be standardized according to a preset standardization formula to obtain a standardized dataset. This addresses the issue of differences in physical dimensions among the various operational data points, mapping them to a unified scale range, thus obtaining a standardized dataset. The preset standardization formula can be the Z-score standardization formula, specifically... In this formula, The fault offset coefficient (its value range can be...) (This can be adjusted according to the type of target equipment). The fault critical offset for the mode corresponding to the operating data (e.g., the voltage fault critical offset is ±10% of the rated voltage). This refers to the cleaned runtime data. This is the running data before data cleaning. After completing the above operations, a standardized dataset can be obtained. .

[0036] In another embodiment provided in this disclosure, step S103 above, "filtering the standardized dataset to obtain the core dataset based on the changes in the description data of different operational data in the fault scenario using a pre-constructed correlation model," can be implemented as follows: Step 1: Calculate the statistical correlation between the various operational data using the correlation coefficient formula; Step 2: Based on the predetermined fault mechanism correlation factors and combined with the operational data, calculate the fault mechanism correlation degree using the fault mechanism correlation degree calculation formula; wherein, the fault mechanism correlation factors are determined according to the synchronicity of changes in operational data in the fault scenario; Step 3: Based on the statistical correlation and fault mechanism correlation of the data, obtain the comprehensive correlation through weighted fusion; Step 4: Based on the comprehensive correlation of each piece of running data calculated, score each piece of running data according to the preset correlation threshold, and remove running data in the standardized dataset that are lower than the preset score based on the score results to obtain the core dataset.

[0037] In this embodiment, two different sets of operational data can be extracted from the core dataset. These two sets of operational data can belong to two different dimensions. The correlation coefficient can be calculated using the Pearson correlation coefficient, as shown in the following formula: ;in, To determine the correlation between data; and These are the two extracted runtime data for calculating the correlation between the two samples. and The description data at the i-th sampling point, here and It is the running data that exists in the form of feature vectors in a standardized dataset; and , which are the mean values ​​of the descriptive data for the corresponding running data.

[0038] Furthermore, the correlation degree of the fault mechanism can be calculated using the formula for calculating the correlation degree of the fault mechanism. The formula for calculating the correlation degree of the fault mechanism is: ;in The correlation degree between the fault mechanism and the fault mechanism.

[0039] This is a fault mechanism correlation factor, which is determined based on the correlation between different operating data in fault scenarios. For example, the mechanism correlation factor between insulation resistance and temperature. Mechanistic correlation factors between humidity and vibration That is, the change in insulation resistance has a strong synchronous change characteristic with the change in temperature of the target equipment, and the change in humidity has a strong synchronous change characteristic with the vibration of the target equipment.

[0040] and These are the differences between the fault state and normal state of the description data for the two operational data points, respectively. and These are the rated value ranges for the description data of the two operating data points.

[0041] Furthermore, a weighted fusion method can be used to obtain the comprehensive correlation between the two sets of operational data. ;in The weighting coefficient can take the following values: Values ​​that can be obtained in fault scenarios In normal scenarios, the values ​​that can be obtained are... 6. This ensures that the data association mechanism takes priority in the event of a failure, and the data retained under normal circumstances takes priority.

[0042] This can further improve the overall correlation. correlation threshold For comparison, the correlation threshold here can be set between 0.75 and 0.85. In one possible implementation... .like If a data point is deemed redundant, it can be scored, and the data point with the higher overall score will be retained. The scoring process can utilize two dimensions: one is the fault sensitivity score. The second is the sampling frequency normalization score. The scoring formula is: .like Then, two types of data are retained, ultimately resulting in the core dataset after removing redundancy. The sensitivity score quantifies the sensitivity of operational data to fault response; the sampling frequency normalization score quantifies the sampling density of operational data, indicating that higher sampling density means more complete sampled operational data. In one possible implementation, the sensitivity score can be a standardized score between 0 and 1, with higher scores indicating greater sensitivity of operational data to fault response; the sampling frequency normalization score can be a standardized score between 0 and 1 assigned based on the actual sampling frequency of the operational data, with higher sampling frequencies and better information completeness resulting in higher scores.

[0043] In another embodiment provided in this disclosure, step S104 above, "assessing the equipment failure stage of the target device based on the deviation between the current value and the normal state value of the descriptive data of each operational data in the core dataset," can be implemented as follows: Step 1: Calculate the deviation between the current value and the normal state value for each running data item; Step 2: Based on the number of operational data points that deviate from the normal state value and the degree of deviation of each operational data point, determine the equipment failure stage of the target equipment according to the preset evaluation criteria; wherein, the equipment failure stage includes: the nascent stage, the development stage, and the outbreak stage.

[0044] In this embodiment, the current value of the descriptive data for each operational data item can be extracted from the core dataset in real time, and the difference between each value and its corresponding normal state value can be calculated. This allows for the assessment of the target device's fault stage based on the number and degree of deviation in the operational data.

[0045] The above process can be represented by the following formula: ;in, The current value of the data describes the data being processed. The description data refers to the normal state values ​​of the running data; The number of items in the running data; These are characteristic indicators of the fault stage. Specific preset evaluation criteria can be used for the nascent stage. (i.e., only 1 to 2 data points deviate slightly from the normal range); development phase (i.e., multiple operational data deviate, but none are fatally exceeded); outbreak period (That is, critical operational data are severely out of control and approaching the critical fault threshold). In the nascent stage, the fault is relatively hidden, and the operational data is in a state of slight and gradual change; in the development stage, the fault is evolving, and multiple operational characteristics may be coupled; in the outbreak stage, the target equipment will be on the verge of failure or has already experienced a serious fault, and the operational data will show fatal sudden changes.

[0046] In another embodiment provided in this disclosure, step S105 above, "adjusting the weights of various operational data in the core dataset based on the equipment failure stage to obtain a weighted fusion feature set," can be implemented as follows: Step 1: Based on the equipment failure stage of the target device, extract the operational data corresponding to that failure stage from the core dataset; Step 2: Calculate the information entropy corresponding to each piece of operational data based on the ratio of the descriptive data of each extracted operational data to the total descriptive data of all extracted operational data. Step 3: Based on the information entropy, determine the weight of each extracted running data item, and perform a weighted fusion operation on the description data of each running data item and the weight to obtain a weighted fusion feature set.

[0047] In this embodiment of the disclosure, if the fault of the target device is in its early stages, it is necessary to extract the insulation resistance change rate from the core dataset. Vibration amplitude standard deviation Operational data, such as these, is primarily used to describe highly sensitive and subtle characteristics of the target device during operation. This is the reference value for insulation resistance; Cumulative running time The insulation resistance value was measured later.

[0048] If the target equipment is being assessed as being in the development stage of a fault, then it is necessary to extract the voltage harmonic distortion rate (THD) and the temperature-resistance coupling coefficient. These operational data, including runtime data, exhibit mutual coupling during the fault development phase. This is the actual measured temperature value of the equipment. This is the measured insulation resistance value.

[0049] If the target device is being assessed as being in the flare-up phase of a fault, then the peak current needs to be extracted. Temperature rise rate Operational data, such as these, exhibits abrupt changes during the period of failure outbreak.

[0050] Then, the sequence of descriptive data for each extracted operational data item can be normalized, and the ratio of the descriptive data value of a specific operational data item at a particular moment to the total sequence of that data item can be calculated. Subsequently, according to the information entropy calculation formula, the information entropy value corresponding to each operational data item is calculated sequentially. The specific information entropy calculation formula is as follows: ; The value of the information entropy of the i-th running data; The normalized probability of the eigenvalues; This refers to the j-th data point describing the i-th running data.

[0051] In the formula for calculating information entropy The value can be determined based on the amount of operational data extracted at different stages of the fault. For example, if the fault is in its early stages and 10 operational data items were extracted, then the value can be set to... Assuming the fault is in its development phase and 14 operational data points have been extracted, the following values ​​can be selected: Assuming the fault is in its outbreak phase and 12 operational data items have been extracted, the following values ​​can be selected: .

[0052] Then, the weight of each piece of running data can be calculated separately. The specific weight calculation formula is as follows: ; Let be the weight of the i-th operational data item. In one possible implementation, the weights of insulation resistance, vibration characteristics, etc., during the nascent stage can be 0.3~0.4; the weights of temperature, harmonic characteristics, etc., during the development stage can be 0.25~0.35; and the weights of current, power characteristics, etc., during the outbreak stage can be 0.35~0.45. Then, a weighted fusion operation can be performed on the operational data extracted from the fault stage to obtain the corresponding weighted fusion feature set. , The value is determined based on the operational data extracted at different fault stages.

[0053] In another embodiment provided in this disclosure, step S106 above, "using a pre-trained fault classification model, performing fault classification and identification based on a weighted fusion feature set, and determining the fault classification result of the target device," can be implemented as follows: Step 1: Input the weighted fusion feature set into the fault classification model; the fault classification model is a convolutional neural network architecture model trained with fault category as the ground truth and corresponding operating data as the input. Step 2: Use the feature extraction layer in the fault classification model to perform feature extraction on the weighted fusion feature set; Step 3: Input the extracted features into the feature fusion layer of the fault classification model for feature fusion operation; Step 4: Calculate the probability of the fused features using the classification function in the fault classification model, and output the fault classification result containing the fault category of the target device and the corresponding confidence level.

[0054] In this embodiment of the disclosure, the fault classification model can be a model based on a convolutional neural network architecture. This model can be trained using pre-acquired historical fault samples. During the training process, the specific fault category is determined as the ground truth label, and the corresponding historical operating data is used as the input quantity as the corresponding training data.

[0055] In one possible implementation, the fault classification model may include an input layer, a feature extraction layer, a feature fusion layer, a classification layer, and an output layer. The input layer receives a weighted fused feature set F with dimension p.

[0056] The feature extraction layer consists of two convolutional layers (Conv1 with 16 3×1 convolutional kernels and Conv2 with 32 3×1 convolutional kernels) and one max pooling layer (with a 2×1 pooling window), using ReLU activation. Since the input weighted fused feature set has undergone a fault-phase weighting operation, redundant interference has been removed and the core running data has been amplified, enabling the convolutional kernels to capture the dependencies between running data sequences and spatial topological features.

[0057] The feature fusion layer is a single fully connected layer (with 2p neurons) activated by LeakyReLU. The feature tensors output from the convolutional and pooling layers are input to the feature fusion layer of the fault classification model. This feature fusion layer can reduce the dimensionality and map multi-channel local spatiotemporal features, aggregating them into a one-dimensional feature vector.

[0058] The classification layer employs a classification function (specifically, the Softmax function) to output five preset fault categories (normal, poor contact, insulation aging, mechanical jamming, and overheating erosion). The classification layer calculates the probability distribution of the feature vectors using the classification function, mapping it to probability values ​​for each preset fault category. It then outputs the option with the highest probability value as the specific fault category for the target device, simultaneously outputting the maximum probability value corresponding to this determination as the confidence level. The output layer outputs the fault category and confidence level as the fault classification result.

[0059] In another embodiment provided in this disclosure, the fault classification model is deployed on the edge computing terminal corresponding to the target device; Equipment fault data classification methods also include: The parameters of the fault classification models of each edge computing terminal are obtained. Based on the sample size of each fault classification model, the parameters of each fault classification model are aggregated. Based on the aggregated parameters, each fault classification model is fine-tuned.

[0060] In this embodiment of the disclosure, to adapt to the computing power and data privacy protection requirements of edge terminals, a federated learning architecture combined with a lightweight CNN can be adopted to achieve cross-site data collaborative training and local real-time inference. Each power distribution site (e.g., a low-voltage distribution room) can be equipped with an edge computing terminal. This edge computing terminal can collect operational data of the target equipment in the power distribution site, process the operational data, and input it into the edge computing terminal. The fault classification model deployed in the edge computing terminal then performs inference based on the operational data to determine whether the target equipment has malfunctioned.

[0061] Furthermore, a corresponding cloud server can be set up in the control center of the power distribution system where the power distribution site is located. This cloud server can periodically send parameter retrieval requests to each edge computing terminal. Based on the amount of local fault samples accumulated by each edge computing terminal during operation, the internal parameters of each fault classification model are aggregated to obtain global parameters. Then, the cloud server can distribute the global parameters to each edge computing terminal, allowing for fine-tuning of the fault classification models deployed on each edge computing terminal.

[0062] During this process, the cloud server can obtain the local parameters of each fault classification model. And aggregate these local parameters into global parameters. The specific formula for this process can be: ;in, The number of edge computing terminals communicating with the cloud server; The sample size of local faults accumulated by the k-th edge computing terminal; This represents the sample size of all edge computing terminals. The purpose of this fine-tuning process is to iteratively refine the loss function of the fault classification model, which can be expressed using the cross-entropy loss function formula: ; For real-world labeling. Fault classification models deployed on edge computing terminals can employ model quantization (e.g., INT8 quantization) and layer fusion techniques to compress model size and adapt to the lightweight chips of edge computing terminals.

[0063] In yet another embodiment provided in this disclosure, the equipment fault data classification method further includes the following steps: Based on the fault classification results, the accuracy of the fault classification model's inference is evaluated, and the fault classification model is adjusted based on the accuracy evaluation results. This can be implemented through the following steps: Step 1: Compare the fault classification results output by the fault classification model within the preset time period with the actual fault situation at the power distribution site to determine the accuracy of the fault classification model's inference. Step 2: If the accuracy is lower than a preset threshold, determine whether the fault category in the fault classification result is included in the pre-training data of the fault classification model. Step 3: If the fault category is included in the pre-training data, adjust the preset correlation threshold and reconstruct the core dataset; Step 4: If the fault category is not included in the pre-training data, perform a transfer learning process to enable the fault classification model to learn new fault categories.

[0064] In this embodiment, the fault classification results output by the edge computing terminal can be compared and verified one by one with the corresponding actual fault situations. The ratio of the number of correctly judged samples to the total number of fault samples in that time period is calculated, thereby determining the true accuracy of the current fault classification model in reasoning at the actual power distribution site. .

[0065] If the accuracy rate falls below a preset threshold, it indicates that the current fault classification model's performance is no longer sufficient for operational requirements, necessitating a corresponding model diagnostic process. This preset threshold could be 93%. Then it is necessary to determine the reason for the decrease in accuracy. First, it is necessary to check whether the true fault categories corresponding to these incorrectly identified faults exist in the pre-training data used to train the model, in order to distinguish whether the decrease in accuracy is due to incorrect identification of existing faults or the appearance of unknown faults.

[0066] If the judgment result indicates that the model has made an incorrect judgment on the existing fault, it means that the original preset correlation threshold is not suitable for the current situation, and that operational data with discriminative value under the current operating condition was removed in the previous process. At this time, the preset correlation threshold in step S103 can be adjusted, and the redundancy removal standard can be relaxed appropriately, so that the part of the operational data that was originally filtered out can be retained. On this basis, filtering can be carried out again to construct a core dataset containing more operational data, thereby correcting the judgment bias caused by the lack of input information.

[0067] If the determination result indicates that an unknown fault has occurred at the power distribution site, the transfer learning process can be initiated to extend the model and train the fault classification model to learn new fault categories.

[0068] In another embodiment provided in this disclosure, the execution of the transfer learning process can be implemented as follows: Step 1: Mix the sample data of the new fault category with the sample data in the pre-training data to construct the target domain dataset for transfer learning; Step 2: Use the current fault classification model as the source domain model and freeze the network layer parameters used for feature extraction in the source domain model; Step 3: Using the target domain dataset, adjust the feature fusion layer and classification function in the source domain model using an adaptive learning rate.

[0069] In this embodiment of the disclosure, sample data related to new fault categories can be acquired and mixed with sample data from existing pre-training data in a certain proportion. In one possible implementation, sample data from more than 30 new fault categories can be mixed with more than 100 existing sample data to obtain a target domain dataset.

[0070] Furthermore, the fault classification model of the edge computing terminal can be designated as the source domain model for transfer learning. In this process, the weight parameters of some network layers of the source domain model (e.g., convolutional layers in the feature extraction layer) can be frozen to preserve the feature extraction capability obtained from model training.

[0071] Furthermore, the constructed target domain dataset can be input into the aforementioned source domain model, and an adaptive learning rate optimization algorithm can be used for training (e.g., an initial learning rate of 0.001, decaying by 10% each round for 20 rounds). During training, the adaptive learning rate algorithm will only fine-tune the model's network layers (e.g., feature fusion layers) and classification function to achieve knowledge expansion and upgrading of the model.

[0072] It should be noted that, in addition to the fault classification results, the content output by the edge computing terminal may also include the target device number, collection timestamp, fault stage, core characteristics (i.e., the operational data with higher weight in the fault) and operation and maintenance suggestions (e.g., strengthening insulation monitoring in the nascent stage and immediate shutdown for maintenance in the outbreak stage).

[0073] In addition, a tiered storage and maintenance strategy can be adopted. For example, for fault classification results with a confidence level of ≥90%, a dual storage strategy of local (i.e., edge computing terminal) and cloud (i.e. cloud server) can be adopted, and regular maintenance work orders can be generated to instruct staff to handle the fault. For fault classification results with 85% ≤ confidence level < 90%, a strategy of local storage combined with cloud backup can be adopted to trigger operation and maintenance review reminders to prompt staff to re-examine the specific faults; Data with a confidence level of <85% is stored locally and marked as pending verification, with additional annotations to be added after on-site verification by staff.

[0074] This disclosure has the following significant advantages over related technologies: Firstly, the noise reduction strategy can be dynamically adjusted based on the intensity of electromagnetic interference, thereby improving the signal-to-noise ratio under different interference scenarios and solving the problem of data quality fluctuation caused by complex interference in low-voltage power distribution rooms.

[0075] Secondly, the fault mechanism is associated with the operating data, and the correlation degree calculation is used to indicate the accuracy of the identification of redundant operating data, so as to avoid the erroneous removal of key data in fault scenarios and reduce the computational burden caused by invalid redundancy.

[0076] Third, the weights are dynamically adjusted for different stages of the fault to achieve early warning of the fault.

[0077] Fourth, under the premise of protecting data privacy, it can be trained collaboratively on target equipment in multiple power distribution sites to improve the generalization ability of the model; and the quantized model is used to adapt to lightweight chip deployment.

[0078] Fifth, the model can quickly adapt to new fault samples, significantly improving its scalability and thus meeting the fault diagnosis needs after new loads are connected.

[0079] Sixth, the output can include fault stage, core characteristics, and maintenance recommendations. Furthermore, it can be used to link maintenance processes according to confidence levels, shortening fault diagnosis time, reducing maintenance costs, and significantly improving the reliability of low-voltage power distribution systems.

[0080] This document also provides a fault classification process based on the method of this disclosure as an example to illustrate the implementation process of the present invention in conjunction with specific data and actual conditions: Assume the power distribution site is a power distribution room in a thermal power plant; the target equipment for monitoring consists of three low-voltage circuit breakers (specific models could be: MCCB 630A, rated voltage 380V, rated current 630A); the core of the edge computing terminal could be the Huawei Ascend 310 Enhanced Edition; specific operating data can be collected using an anti-interference sensor array. Among the collected multimodal data, electrical quantities can include voltage U (25kHz), current I (25kHz), and power P (25kHz), environmental quantities can include contact temperature T (8Hz), humidity H (8Hz), state quantities can include the number of actions N (0.5Hz), insulation resistance R (0.3Hz), and vibration amplitude A (15Hz), and extended quantities can include electromagnetic interference intensity E (1kHz).

[0081] Obtained through monitoring This indicates that the power distribution site is at a medium interference level, and adaptive Kalman filtering can be used. The system noise variance is set to 0.01, and the observation noise variance is set to 0.02. During data cleaning, three missing voltage values ​​are filled, and four current outliers are removed (the 3σ range is 236A to 404A; outliers such as 420A and 435A can be replaced with the median of 318A). Standardization uses a modified Z-score method. The voltage fault critical offset is 38V (10% of the rated 380V), thus obtaining the standardized dataset. .

[0082] Setting up fault scenarios Calculate the overall correlation: where temperature T is related to insulation resistance R. , , , , , None of them are redundant and can be retained; temperature T and dust concentration C This indicates the existence of redundancy, with a temperature fault sensitivity score of 0.8 and a dust score of 0.3. The dust data is redundant; temperature data is retained. The final core dataset is... .

[0083] calculate This indicates that the current fault is in its development stage, and 14 features can be extracted and their weights adjusted. For example, temperature has a weight of 0.3, harmonic distortion rate has a weight of 0.28, insulation resistance has a weight of 0.25, and the weights of other operational data are ≤0.17, thus constructing a weighted fusion feature set with a dimension of 14. .

[0084] This hypothesis assumes that the power distribution site can be federated for training from three other power distribution sites (e.g., the power distribution room in a commercial complex). It is assumed that power distribution room 1 has 700 samples, power distribution room 2 has 650 samples, and power distribution room 3 has 600 samples. Each site undergoes local training for 20 rounds, and the loss function converges to 0.08.

[0085] The cloud servers that manage the power distribution systems at these power distribution sites use weighted average aggregated parameters (e.g., , , The global model validation set accuracy is 94.2%; when deployed on the edge computing terminal, the model size after INT8 quantization is 480KB and the inference latency is 38ms.

[0086] When five novel "harmonic overload" fault samples appeared in the test set used to test the pattern, the confidence level of the fault classification model output was 78%, which is lower than the preset threshold. Therefore, a transfer learning process can be initiated. Using the current basic fault classification model as the source domain model, the target domain dataset is set to a combination of 30 harmonic overload samples and 100 known fault samples. After training and fine-tuning, the fault classification model's accuracy in classifying harmonic overload faults improved to 92.5%, and the overall fault classification accuracy was 93.8%.

[0087] The final output results are: Low-voltage circuit breaker #1 (normal, confidence level 96%), Low-voltage circuit breaker #2 (poor contact, confidence level 93%), core characteristic: THD=4.8%. The maintenance suggestion is to "check the contact oxidation"); low-voltage circuit breaker #3 (harmonic overload, confidence level 92%, core characteristics: THD=12%, current distortion rate 8%, maintenance suggestion is to "investigate the load harmonic source"); regular maintenance work orders are generated for equipment #2 and #3, and the data is updated synchronously on the cloud server. A verification time is arranged for maintenance personnel to confirm that the fault type is consistent with the classification result.

[0088] This disclosure also provides an electrical low-voltage primary equipment fault data classification device, such as... Figure 2 As shown, it includes: Processing module 201 is used to acquire electromagnetic interference intensity data at the power distribution site and descriptive data of various operating data of the target equipment; and to trigger the corresponding data processing strategy based on the electromagnetic interference intensity data to process the descriptive data of various operating data to obtain a standardized dataset. The filtering module 202 is used to filter the standardized dataset based on the changes in the description data of different operating data in the fault scenario through a pre-built correlation model to obtain the core dataset. Evaluation module 203 is used to evaluate the equipment failure stage of the target device based on the deviation between the current value and the normal state value of the description data of each operational data in the core dataset. The first adjustment module 204 is used to adjust the weights of various operational data in the core dataset based on the equipment failure stage to obtain a weighted fusion feature set; The identification module 205 is used to identify faults based on a weighted fusion feature set using a pre-trained fault classification model to determine the fault classification result of the target device.

[0089] In another embodiment provided in this disclosure, the processing module 201 is used to calculate the effective value of the electromagnetic interference intensity data based on the electromagnetic interference intensity data, and determine the electromagnetic interference level of the power distribution site based on the effective value; the electromagnetic interference level includes: weak interference level, medium interference level, and strong interference level; when the power distribution site is at a weak interference level, a moving average filtering noise reduction strategy is triggered to perform noise reduction processing on various operating data; when the power distribution site is at a medium interference level, a Kalman filtering noise reduction strategy is triggered to perform noise reduction processing on various operating data; when the power distribution site is at a strong interference level, noise reduction processing is performed on various operating data based on wavelet threshold noise reduction and Kalman filtering noise reduction strategies respectively; data cleaning operation is performed on various operating data after noise reduction processing to fill missing values ​​and remove outliers; and standardization operation is performed on various operating data after data cleaning according to a preset standardization formula to obtain a standardized dataset.

[0090] In another embodiment provided in this disclosure, the filtering module 202 is used to calculate the statistical correlation degree between various operational data using a correlation coefficient formula; calculate the fault mechanism correlation degree based on a predetermined fault mechanism correlation factor and the operational data using a fault mechanism correlation degree calculation formula; wherein the fault mechanism correlation factor is determined according to the synchronicity of changes in operational data in the fault scenario; obtain a comprehensive correlation degree by weighted fusion based on the statistical correlation degree and the fault mechanism correlation degree; score each operational data based on a preset correlation degree threshold based on the calculated comprehensive correlation degree of each operational data, and remove operational data in the standardized dataset that are lower than the preset score based on the score results to obtain the core dataset.

[0091] In another embodiment provided in this disclosure, the evaluation module 203 is used to calculate the deviation between the current value and the normal state value of each running data item; based on the number of running data items where the current value deviates from the normal state value and the degree of deviation of each running data item, the equipment failure stage of the target device is determined according to a preset evaluation standard; wherein, the equipment failure stage includes: the budding stage, the development stage, and the outbreak stage.

[0092] In another embodiment provided in this disclosure, the first adjustment module 204 is used to extract operational data corresponding to the equipment failure stage of the target device from the core dataset; calculate the information entropy corresponding to each operational data based on the ratio of the descriptive data of each extracted operational data to the descriptive data of all extracted operational data; determine the weight of each extracted operational data based on the information entropy; and perform a weighted fusion operation on the descriptive data of each operational data and the weight to obtain a weighted fusion feature set.

[0093] In another embodiment provided in this disclosure, the identification module 205 is used to input the weighted fused feature set into a fault classification model; the fault classification model is a convolutional neural network architecture model trained with fault category as the ground truth and corresponding operating data as input; the feature extraction layer in the fault classification model is used to perform feature extraction operation on the weighted fused feature set; the extracted features are input into the feature fusion layer in the fault classification model for feature fusion operation; the fused features are probability calculated by the classification function in the fault classification model, and a fault classification result containing the fault category of the target device and the corresponding confidence level is output.

[0094] In yet another embodiment provided in this disclosure, such as Figure 2 As shown, the equipment fault data classification device also includes: an aggregation module 206, which is used to obtain the parameters of the fault classification model of each edge computing terminal, aggregate the parameters of each fault classification model according to the sample size included in each fault classification model, and fine-tune each fault classification model based on the aggregated parameters.

[0095] In yet another embodiment provided in this disclosure, such as Figure 2 As shown, the equipment fault data classification device further includes: a second adjustment module 207, used to evaluate the accuracy of the fault classification model inference based on the fault classification results, and adjust the fault classification model based on the accuracy evaluation results; The second adjustment module 207 is used to compare the fault classification results output by the fault classification model within a preset time period with the actual fault situation at the power distribution site to determine the accuracy of the fault classification model's inference; if the accuracy is lower than a preset threshold, it determines whether the fault category in the fault classification result is included in the pre-training data of the fault classification model; if the fault category is included in the pre-training data, it adjusts the preset correlation threshold and reconstructs the core dataset; if the fault category is not included in the pre-training data, it executes a transfer learning process to enable the fault classification model to learn new fault categories.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that the embodiments of this disclosure can be implemented in hardware or by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this disclosure.

[0097] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes in the drawings are not necessarily essential for implementing this disclosure.

[0098] Those skilled in the art will understand that the modules in the apparatus of the embodiments can be distributed in the apparatus of the embodiments as described in the embodiments, or they can be located in one or more devices different from this embodiment with corresponding changes. The modules of the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0099] The sequence numbers of the embodiments disclosed above are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0100] Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from its spirit and scope. Therefore, if such modifications and variations fall within the scope of the claims of this disclosure and their equivalents, this disclosure is also intended to include such modifications and variations.

Claims

1. A method for classifying fault data of low-voltage primary electrical equipment, characterized in that, include: Acquire electromagnetic interference intensity data at the power distribution site and descriptive data of various operating data of the target equipment; Based on the electromagnetic interference intensity data, the corresponding data processing strategy is triggered to process the descriptive data of each operational data to obtain a standardized dataset; By using a pre-built correlation model, the standardized dataset is filtered based on the changes in description data of different operational data in the fault scenario, to obtain the core dataset. Based on the deviation between the current value and the normal state value of the descriptive data of each operational data in the core dataset, the equipment failure stage of the target device is assessed. Based on the equipment failure stage, the weights of each operational data item in the core dataset are adjusted to obtain a weighted fusion feature set; A pre-trained fault classification model is used to identify faults based on a weighted fusion feature set, and the fault classification result of the target device is determined.

2. The method as described in claim 1, characterized in that, The process involves triggering a corresponding data processing strategy based on the electromagnetic interference intensity data to process the descriptive data of each operational data item, resulting in a standardized dataset, including: The effective value of the electromagnetic interference intensity data is calculated based on the electromagnetic interference intensity data, and the electromagnetic interference level at the power distribution site is determined based on the effective value; the electromagnetic interference level includes: weak interference level, medium interference level, and strong interference level. When the power distribution site is under a low interference level, the moving average filtering noise reduction strategy is triggered to process the noise reduction of various operating data. When the power distribution site is under medium interference level, the Kalman filter noise reduction strategy is triggered to reduce noise in various operating data; In the case of strong interference at the power distribution site, noise reduction is performed on various operating data based on wavelet threshold noise reduction and Kalman filter noise reduction strategies respectively. Data cleaning is performed on the various operational data after noise reduction to fill in missing values ​​and remove outliers. After data cleaning, all operational data are standardized according to a preset standardization formula to obtain a standardized dataset.

3. The method as described in claim 1, characterized in that, The standardized dataset is filtered using a pre-built correlation model based on the changes in description data of different operational data in a fault scenario, to obtain a core dataset, including: The correlation coefficient formula is used to calculate the statistical correlation between various operational data. Based on predetermined fault mechanism correlation factors, and combined with operational data, the fault mechanism correlation degree is calculated using the fault mechanism correlation degree calculation formula; wherein, the fault mechanism correlation factors are determined according to the synchronicity of changes in operational data in the fault scenario; Based on the statistical correlation and fault mechanism correlation of the data, a comprehensive correlation is obtained through weighted fusion. Based on the calculated comprehensive correlation of each piece of operational data, each piece of operational data is scored according to a preset correlation threshold. Based on the scoring results, operational data in the standardized dataset that are lower than the preset score are removed to obtain the core dataset.

4. The method as described in claim 1, characterized in that, The step of assessing the equipment failure stage of the target device based on the deviation between the current value and the normal state value of the descriptive data of each operational data item in the core dataset includes: Calculate the deviation between the current value and the normal state value for each piece of running data; Based on the number of operational data points that deviate from the normal state value and the degree of deviation of each operational data point, the equipment failure stage of the target equipment is determined according to the preset evaluation criteria; wherein, the equipment failure stage includes: the nascent stage, the development stage, and the outbreak stage.

5. The method as described in claim 1, characterized in that, Based on the equipment failure stage, the weights of each operational data item in the core dataset are adjusted to obtain a weighted fusion feature set, including: Based on the equipment failure stage of the target device, extract the operational data corresponding to that equipment failure stage from the core dataset; Based on the ratio of the descriptive data of each extracted operational data to the total descriptive data of all extracted operational data, calculate the information entropy corresponding to each operational data. Based on the information entropy, the weight of each extracted running data item is determined, and the descriptive data of each running data item is weighted and fused with the weight to obtain a weighted fused feature set.

6. The method as described in claim 1, characterized in that, The method employs a pre-trained fault classification model, which performs fault classification and identification based on a weighted fusion feature set, and determines the fault classification result of the target device, including: The weighted fusion feature set is input into the fault classification model; the fault classification model is a convolutional neural network architecture model trained with fault category as the true value and corresponding operating data as the input. The feature extraction layer in the fault classification model is used to perform feature extraction on the weighted fused feature set; The extracted features are input into the feature fusion layer of the fault classification model for feature fusion operation; The probability of the fused features is calculated by the classification function in the fault classification model, and the output includes the fault category of the target device and the corresponding confidence level.

7. The method as described in claim 1, characterized in that, The fault classification models are deployed on the edge computing terminals corresponding to the target devices. The method further includes: The parameters of the fault classification models of each edge computing terminal are obtained. Based on the sample size of each fault classification model, the parameters of each fault classification model are aggregated. Based on the aggregated parameters, each fault classification model is fine-tuned.

8. The method as described in claim 3, characterized in that, Also includes: Based on the fault classification results, the accuracy of the fault classification model's inference is evaluated, and the fault classification model is adjusted based on the accuracy evaluation results, including: The fault classification results output by the fault classification model within a preset time period are compared with the actual fault situation at the power distribution site to determine the accuracy of the fault classification model's inference. If the accuracy is lower than a preset threshold, determine whether the fault category in the fault classification result is included in the pre-training data of the fault classification model; If the fault category is included in the pre-training data, adjust the preset correlation threshold and reconstruct the core dataset; If the fault category is not included in the pre-training data, a transfer learning process is performed to enable the fault classification model to learn new fault categories.

9. The method as described in claim 8, characterized in that, The transfer learning process includes: The sample data of the new fault category is mixed with the sample data in the pre-training data to construct the target domain dataset for transfer learning; The current fault classification model is used as the source domain model, and the network layer parameters used for feature extraction in the source domain model are frozen. Using the target domain dataset, the feature fusion layer and classification function in the source domain model are adjusted through an adaptive learning rate.

10. A fault data classification device for low-voltage primary electrical equipment, characterized in that, include: The processing module is used to acquire electromagnetic interference intensity data at the power distribution site and descriptive data of various operating data of the target equipment; Based on the electromagnetic interference intensity data, the corresponding data processing strategy is triggered to process the descriptive data of each operational data to obtain a standardized dataset; The filtering module is used to filter the standardized dataset based on the changes in the description data of different operating data in the fault scenario through a pre-built correlation model to obtain the core dataset. The evaluation module is used to evaluate the equipment failure stage of the target device based on the deviation between the current value and the normal state value of the descriptive data of each operational data in the core dataset. The first adjustment module is used to adjust the weights of various operational data in the core dataset based on the equipment failure stage to obtain a weighted fusion feature set; The identification module is used to identify faults based on a weighted fusion feature set using a pre-trained fault classification model, and to determine the fault classification result of the target device.

Images