A power distribution cabinet fault operation and maintenance method and system based on big data

Abstract

This invention discloses a method and system for fault operation and maintenance of power distribution cabinets based on big data, including a multi-dimensional sensing terminal deployed in the power distribution cabinet and its surrounding environment. This invention collects full data through the multi-dimensional sensing terminal, performs cleaning, missing value imputation, and timestamp alignment, effectively solving the problems of noise interference and asynchrony in multi-source heterogeneous data, generating a high-quality standardized time-series dataset. Utilizing sliding window technology to fuse time-domain statistics, frequency-domain waveforms, and temperature field distribution features, it can accurately capture subtle signs of early faults and thermo-electric coupling characteristics of the power distribution cabinet. Employing a hybrid deep learning architecture combining ensemble learning and LSTM, it balances feature selection and time-series dependency modeling, significantly improving the accuracy of fault probability calculation and model generalization ability. Furthermore, based on health status scores, geographical location, and load levels, it generates differentiated maintenance strategies, achieving precise allocation of operation and maintenance resources, avoiding "one-size-fits-all" inspections, and reducing operation and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of power distribution cabinet fault operation and maintenance technology, specifically to a power distribution cabinet fault operation and maintenance method and system based on big data. Background Technology

[0002] With the advancement of smart grid construction, distribution cabinets, as key nodes in power distribution, are directly related to power supply quality and grid security in terms of operational reliability. Traditional distribution cabinet operation and maintenance mainly rely on manual periodic inspections and simple threshold alarm mechanisms. However, facing increasingly complex grid environments and high load operating pressures, this passive operation and maintenance approach has exposed many technical bottlenecks.

[0003] First, in terms of fault feature extraction, traditional methods are mostly limited to simple statistical features in the time domain (such as RMS and average values), lacking in-depth analysis of frequency domain waveforms and harmonic components, and neglecting the spatial distribution characteristics of the temperature field inside the distribution cabinet. For weak nonlinear features that can characterize the early aging of equipment (such as the chaotic characteristics of partial discharge and thermo-electric coupling characteristics), existing technologies often lack effective mathematical quantification methods, resulting in insufficient extraction of fault symptoms and difficulty in detecting latent defects.

[0004] Furthermore, in constructing fault prediction models, existing diagnostic algorithms mostly employ a single machine learning model (such as support vector machines or basic neural networks), making it difficult to simultaneously ensure the efficiency of feature selection and the capture of time-series dependencies. A single model exhibits poor generalization ability when faced with extremely imbalanced sample classes (where faulty samples are far fewer than normal samples), and during model training, the cost of misjudging various fault types is typically treated equally, lacking a cost-sensitive learning mechanism for severe faults, resulting in a low recognition rate for high-risk faults.

[0005] Finally, in terms of generating operation and maintenance strategies, existing systems are mostly based on fixed maintenance cycles or simple state quantity threshold triggers, failing to comprehensively consider the dynamic changes in real-time equipment health scores, geographical accessibility, and load importance levels. This "one-size-fits-all" approach to strategy generation lacks a quantitative priority evaluation model, leading to unreasonable allocation of operation and maintenance resources. Furthermore, once deployed, the model remains fixed for a long period, lacking the ability to perform online incremental updates and adaptive adjustments based on new fault feedback data. Summary of the Invention

[0006] To address the aforementioned technical problems, this paper provides a method and system for fault operation and maintenance of power distribution cabinets based on big data. This technical solution solves the problems mentioned in the background section.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect of the present invention, a method for fault operation and maintenance of power distribution cabinets based on big data is provided, comprising: By deploying multi-dimensional sensing terminals in and around the power distribution cabinet, the operating status data, environmental parameter data and historical operation and maintenance record data of the power distribution cabinet are collected to form an original multi-source heterogeneous dataset. Data preprocessing is performed on the original multi-source heterogeneous dataset, including data cleaning, missing value imputation, timestamp alignment, and multi-source data fusion, to generate a standardized time-series feature dataset; Based on a standardized time-series feature dataset, a multi-dimensional state feature vector is constructed using the sliding time window technique, and key fault symptom features are extracted. These features include time-domain statistical features, frequency-domain waveform features, and temperature field distribution features. Key fault symptom features are input into a pre-trained power distribution cabinet fault prediction model for deep feature learning and fault probability calculation, outputting the health status score and potential fault type of the power distribution cabinet. The fault prediction model is trained based on a historical fault sample library using a hybrid deep learning architecture that combines ensemble learning algorithms and long short-term memory networks. Based on the health status score and potential fault types, combined with the geographical location information of the distribution cabinet and the current load importance level, differentiated operation and maintenance strategies are generated and pushed to the operation and maintenance terminal.

[0008] Preferably, the specific process of cleaning and imputing missing values ​​in the original multi-source heterogeneous dataset is as follows: The 3σ principle based on statistical distribution or Hampel identifier is used to remove physically impossible values ​​from the operational status data and abnormal jump points in the environmental parameters. For discrete missing data caused by communication interruption, a K-nearest neighbor-based interpolation algorithm is used to fill in the missing data. The interpolation calculation formula is as follows: ; in, The fill value for the missing points. For the first Observations of neighboring nodes, The corresponding weighting coefficient is usually taken as... , This refers to the electrical or Euclidean distance between the missing point and its neighboring nodes. The preset number of neighboring nodes; For continuously missing data, forward predictions based on a time-series autoregressive moving average model are used for imputation. The prediction formula is as follows: ; in, For a moment The predicted value, For constant terms, These are the autoregressive coefficients. The moving average coefficient is... It is a white noise sequence. and These are the autoregression order and the moving average order, respectively, to ensure the temporal continuity of the data.

[0009] Preferably, the specific implementation of timestamp alignment and multi-source data fusion of the original multi-source heterogeneous dataset includes: Using the unique device identifier of the power distribution cabinet as the key, the collected structured operation data, semi-structured environmental logs, and unstructured maintenance text records are associated. A linear interpolation resampling method is employed to map sensor data from different sampling frequencies onto the same time axis, generating time-series sequences with equal time intervals. The interpolation formula is as follows: ; in, For the interpolation result, For the target timestamp, and The start and end times of the interpolation interval. and These are the original observations for the corresponding time period; The coupling relationship between multidimensional parameters is analyzed using the Pearson correlation coefficient matrix, and the calculation formula is as follows: ; in, For variables and The correlation coefficient, and For the first Observations of each sample point and These are the means of the two sets of data. To determine the sample size, highly collinear features with redundancy exceeding a preset threshold are removed, while feature dimensions with independent information contributions are retained.

[0010] Preferably, the step of constructing a multi-dimensional state feature vector using the sliding time window technique and extracting key fault symptom features specifically includes: Set a fixed-length sliding time window, with the window length L covering one complete thermal stability cycle of the power distribution cabinet, and the window sliding step size Δt being an integer multiple of the sampling period; In the time domain, the statistical characteristics of current, voltage, and vibration signals within the calculation window are determined, including the mean μ and variance. Skewness S and kurtosis The formula for calculating kurtosis is: ; in, For the first in the window Each sample value, This represents the average of the data within the window. Standard deviation The number of sample points within the window, and the peak factor is extracted simultaneously. ,and ; In the frequency domain, a fast Fourier transform is performed on the time-domain signal to extract the amplitude of the fundamental component. Total Harmonic Distortion (THD) and Energy Spectral Density in Specific High-Frequency Bands The formula for calculating THD is: ; in, For the first The amplitude of the second harmonic component, The highest harmonic order; In the spatial dimension, the temperature field matrix inside the cabinet is constructed based on infrared thermal imaging data, and the norm of the temperature gradient at each measuring point and the offset displacement of the hot spot center are calculated.

[0011] Preferably, the extraction of temperature field distribution features further includes: The interior of the distribution cabinet is divided into several three-dimensional mesh units, and the temperature rise rate of each unit is calculated. and the Euclidean distance from the historical average temperature rise curve ; The thermal diffusivity at key contact points inside the cabinet is calculated using the difference scheme of the heat conduction equation. The coefficient is then normalized to the square of the load current to obtain the thermal-electric coupling characteristic index η, which characterizes the abnormal change in contact resistance. The formula for calculating the thermal diffusivity is: ; in, The thermal conductivity of the material, For density, Specific heat capacity; The formula for calculating the thermal-electric coupling characteristic index η is: ; in, The thermal diffusivity is calculated at the current moment. The baseline thermal diffusivity under historical normal operating conditions. This is the current load current. This is the rated current.

[0012] Preferably, the hybrid deep learning architecture of the pre-trained power distribution cabinet fault prediction model comprises a two-level cascaded structure: The first stage employs an ensemble learning algorithm based on gradient boosting decision trees to perform preliminary screening and nonlinear transformation on the input key fault symptom features, outputting a high-dimensional abstract feature vector. The superposition process of the m-th tree is represented as follows: ; in, For the first The prediction function of a decision tree. The learning rate; The second level will use the abstract feature vector As input, it is fed into the input gate. Forgotten Gate and output gate The constructed multi-layer Long Short-Term Memory (LSTM) network optimizes the dependency weights on the time series using the backpropagation algorithm. The cell state update formula for LSTM units is as follows: ; ; in, for Cellular state at any given moment For the hidden layer output, ⊙ represents the Hadamard product. As candidate memory units, the probability distribution P(y|X) of various faults is finally output through a fully connected layer and a Softmax activation function.

[0013] Preferably, the training process of the fault prediction model includes: Construct a historical fault sample library with four levels of tags, including normal state, general defects, serious faults and critical faults; The SMOTE oversampling algorithm is used to synthesize and augment minority class fault samples, balancing the sample class distribution. The formula for generating it is: ; in, For the original minority class samples, Let it be a random sample from its k nearest neighbors. A random number between [0, 1]; Introduce penalty factors for different fault severity levels into the loss function. This makes the cost of misclassifying high-risk faults higher than that of low-risk faults, resulting in a higher cost for the weighted cross-entropy loss function. Defined as: ; in, For the sample size, For the number of categories, For real labels, To predict probabilities, For the first Penalty weights for different types of failures; Dynamically adjust the learning rate using the Adam optimizer The formula is adjusted as follows: ; in, The initial learning rate, , The decay rate is used, and an early stopping method is employed on the validation set to prevent overfitting of the model until the F1-Score on the validation set converges.

[0014] Preferably, the specific logic for generating differentiated operation and maintenance strategies is as follows: A mapping function between health status scores and failure probabilities is established, classifying health status into five levels: excellent, normal, attentive, abnormal, and severe, and assigning a corresponding risk weight to each level. ; Based on geographical location information, calculate the estimated travel time for maintenance personnel to reach the distribution cabinet from their work site. and road condition complexity coefficient ; Set the load loss weighting factor based on the current load importance level. ; The maintenance priority index is calculated using a weighted summation formula. The calculation formula is: ; in, , , Here are the normalized weight coefficients of each influencing factor, and ; When the index exceeds the preset threshold At that time, a structural chemical work order containing specific maintenance time windows, a list of required spare parts, and risk prevention and control measures will be automatically generated.

[0015] Preferably, the method also includes the step of incrementally updating the fault prediction model: Real-time collection of fault feedback data verified on-site by maintenance personnel, and addition of it to the historical sample library as new training samples; Set model performance monitoring indicators. When the model's prediction accuracy on the new dataset drops beyond the preset range, or when the operating conditions of the power distribution cabinet undergo structural changes, trigger the model retraining mechanism. The formula for calculating prediction accuracy is: ; in, It is a true positive. It is a true negative. It was a false positive. It is a false negative; The weight parameters W and bias b of the LSTM network are fine-tuned and updated using the latest time-series dataset. The update formula is as follows: ; ; in, For learning rate, For gradient operators, The loss function is denoted by , and k-fold cross-validation is used to evaluate the generalization ability of the updated model.

[0016] In a second aspect of the invention, a big data-based power distribution cabinet fault maintenance system is also provided, comprising: The data acquisition module is used to collect the operating status data, environmental parameter data and historical operation and maintenance record data of the power distribution cabinet through multi-dimensional sensing terminals deployed in the power distribution cabinet and its surrounding environment, forming an original multi-source heterogeneous dataset. The preprocessing module is used to preprocess the original multi-source heterogeneous dataset, including data cleaning, missing value imputation, timestamp alignment and multi-source data fusion, to generate a standardized time-series feature dataset. The extraction module is used to construct a multi-dimensional state feature vector based on a standardized time-series feature dataset using a sliding time window technique, and to extract key fault symptom features, including time-domain statistical features, frequency-domain waveform features, and temperature field distribution features. The output module is used to input key fault symptom features into a pre-trained power distribution cabinet fault prediction model, perform deep feature learning and fault probability calculation, and output the health status score and potential fault type of the power distribution cabinet. The fault prediction model is trained based on a historical fault sample library and using a hybrid deep learning architecture that combines ensemble learning algorithms and long short-term memory networks. The generation module is used to generate differentiated operation and maintenance strategies based on the health status score and potential fault types, combined with the geographical location information of the distribution cabinet and the current load importance level, and push them to the operation and maintenance terminal.

[0017] Compared with existing technologies, this invention provides a method and system for fault operation and maintenance of power distribution cabinets based on big data, which has the following beneficial effects: This invention collects full data through a multi-dimensional sensing terminal, and performs cleaning, missing value imputation, and timestamp alignment, effectively solving the problems of noise interference and asynchrony of multi-source heterogeneous data, generating a high-quality standardized time-series dataset. Utilizing sliding window technology to fuse time-domain statistics, frequency-domain waveforms, and temperature field distribution features, it can accurately capture subtle signs of early faults in distribution cabinets and their thermo-electric coupling characteristics. Employing a hybrid deep learning architecture combining ensemble learning and LSTM, it balances feature selection and time-series dependency modeling, significantly improving the accuracy of fault probability calculation and model generalization ability. Furthermore, based on health status scores, geographical location, and load levels, it generates differentiated maintenance strategies, enabling precise allocation of maintenance resources, avoiding "one-size-fits-all" inspections, reducing maintenance costs, and improving the reliability of power grid operation. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the method flow for S101-S105 in this invention; Figure 2 This is a schematic diagram of the method flow for S201-S203 in this invention; Figure 3 This is a schematic diagram of the method flow for S301-S304 in this invention; Figure 4 This is a schematic diagram of the method flow for S401-S404 in this invention; Figure 5 This is a schematic diagram of the method flow for S501-S505 in this invention. Detailed Implementation

[0019] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0020] Example 1 Please refer to Figure 1 As shown, in a first aspect of the present invention, a method for fault operation and maintenance of power distribution cabinets based on big data is provided, comprising: S101. By deploying multi-dimensional sensing terminals in the distribution cabinet and its surrounding environment, the operating status data, environmental parameter data and historical operation and maintenance record data of the distribution cabinet are collected to form an original multi-source heterogeneous dataset. S102. Perform data preprocessing on the original multi-source heterogeneous dataset, including data cleaning, missing value imputation, timestamp alignment, and multi-source data fusion, to generate a standardized time-series feature dataset. S103. Based on a standardized time-series feature dataset, a multi-dimensional state feature vector is constructed using the sliding time window technique, and key fault symptom features are extracted. The features include time-domain statistical features, frequency-domain waveform features, and temperature field distribution features. S104. Input the key fault symptom features into the pre-trained distribution cabinet fault prediction model, perform deep feature learning and fault probability calculation, and output the health status score and potential fault type of the distribution cabinet. The fault prediction model is trained based on the historical fault sample library and uses a hybrid deep learning architecture that combines ensemble learning algorithm and long short-term memory network. S105. Based on the health status score and potential fault types, combined with the geographical location information of the distribution cabinet and the current load importance level, generate differentiated operation and maintenance strategies and push them to the operation and maintenance terminal.

[0021] As will be understood by those skilled in the art, this invention effectively solves the problems of noise interference and asynchrony in multi-source heterogeneous data by collecting full data through multi-dimensional sensing terminals, and performing cleaning, missing value imputation, and timestamp alignment, generating a high-quality standardized time-series dataset; by using sliding window technology to integrate time-domain statistics, frequency-domain waveforms, and temperature field distribution features, it can accurately capture subtle signs of early faults in distribution cabinets and their thermo-electric coupling characteristics; by adopting a hybrid deep learning architecture combining ensemble learning and LSTM, it balances feature selection and time-series dependency modeling, significantly improving the accuracy of fault probability calculation and model generalization ability; and by generating differentiated maintenance strategies based on health status scores, geographical location, and load levels, it achieves precise allocation of operation and maintenance resources, avoids "one-size-fits-all" inspections, reduces operation and maintenance costs, and improves the reliability of power grid operation.

[0022] Please refer to Figure 2 As shown, the specific process of data cleaning and missing value imputation for the original multi-source heterogeneous dataset is as follows: S201. Use the 3σ principle based on statistical distribution or Hampel identifier to remove physically impossible values ​​in the operating status data and abnormal jump points in the environmental parameters. S202. For discrete missing data caused by communication interruption, a K-nearest neighbor-based interpolation algorithm is used to fill in the missing data. The interpolation calculation formula is as follows: ; in, The fill value for the missing points. For the first Observations of neighboring nodes, The corresponding weighting coefficient is usually taken as... , This refers to the electrical or Euclidean distance between the missing point and its neighboring nodes. The preset number of neighboring nodes; S203. For continuously missing data, forward predictions based on a time-series autoregressive moving average model are used for imputation. The prediction formula is as follows: ; in, For a moment The predicted value, For constant terms, These are the autoregressive coefficients. The moving average coefficient is... It is a white noise sequence. and These are the autoregression order and the moving average order, respectively, to ensure the temporal continuity of the data.

[0023] The specific implementation of timestamp alignment and multi-source data fusion for original multi-source heterogeneous datasets includes: Using the unique device identifier of the power distribution cabinet as the key, the collected structured operation data, semi-structured environmental logs, and unstructured maintenance text records are associated. A linear interpolation resampling method is employed to map sensor data from different sampling frequencies onto the same time axis, generating time-series sequences with equal time intervals. The interpolation formula is as follows: ; in, For the interpolation result, For the target timestamp, and The start and end times of the interpolation interval. and These are the original observations for the corresponding time period; The coupling relationship between multidimensional parameters is analyzed using the Pearson correlation coefficient matrix, and the calculation formula is as follows: ; in, For variables and The correlation coefficient, and For the first Observations of each sample point and These are the means of the two sets of data. To determine the sample size, highly collinear features with redundancy exceeding a preset threshold are removed, while feature dimensions with independent information contributions are retained.

[0024] Please refer to Figure 3 As shown, a multi-dimensional state feature vector is constructed using the sliding time window technique, and key fault symptom features are extracted, specifically including: S301. Set a fixed-length sliding time window, the window length L covers one complete thermal stability cycle of the power distribution cabinet, and the window sliding step Δt is an integer multiple of the sampling period; S302. In the time domain, calculate the statistical characteristics of the current, voltage, and vibration signals within the window, including the mean μ and variance. Skewness S and kurtosis The formula for calculating kurtosis is: ; in, For the first in the window Each sample value, This represents the average of the data within the window. Standard deviation The number of sample points within the window, and the peak factor is extracted simultaneously. ,and ; S303. In the frequency domain, perform a Fast Fourier Transform on the time-domain signal to extract the amplitude of the fundamental component. Total Harmonic Distortion (THD) and Energy Spectral Density in Specific High-Frequency Bands The formula for calculating THD is: ; in, For the first The amplitude of the second harmonic component, The highest harmonic order; S304. In the spatial dimension, construct the temperature field matrix inside the cabinet based on infrared thermal imaging data, and calculate the norm of the temperature gradient at each measuring point and the offset displacement of the hot spot center.

[0025] Extraction of temperature field distribution characteristics also includes: The interior of the distribution cabinet is divided into several three-dimensional mesh units, and the temperature rise rate of each unit is calculated. and the Euclidean distance from the historical average temperature rise curve ; The thermal diffusivity at key contact points inside the cabinet is calculated using the difference scheme of the heat conduction equation. The coefficient is then normalized to the square of the load current to obtain the thermal-electric coupling characteristic index η, which characterizes the abnormal change in contact resistance. The formula for calculating the thermal diffusivity is: ; in, The thermal conductivity of the material, For density, Specific heat capacity; The formula for calculating the thermal-electric coupling characteristic index η is: ; in, The thermal diffusivity is calculated at the current moment. The baseline thermal diffusivity under historical normal operating conditions. This is the current load current. This is the rated current.

[0026] The hybrid deep learning architecture of the pre-trained power distribution cabinet fault prediction model consists of a two-level cascaded structure: The first stage employs an ensemble learning algorithm based on gradient boosting decision trees to perform preliminary screening and nonlinear transformation on the input key fault symptom features, outputting a high-dimensional abstract feature vector. The superposition process of the m-th tree is represented as follows: ; in, For the first The prediction function of a decision tree. The learning rate; The second level will abstract feature vectors. As input, it is fed into the input gate. Forgotten Gate and output gate The constructed multi-layer Long Short-Term Memory (LSTM) network optimizes the dependency weights on the time series using the backpropagation algorithm. The cell state update formula for LSTM units is as follows: ; ; in, for Cellular state at any given moment For the hidden layer output, ⊙ represents the Hadamard product. As candidate memory units, the probability distribution P(y|X) of various faults is finally output through a fully connected layer and a Softmax activation function.

[0027] Please refer to Figure 4 As shown, the training process of the fault prediction model includes: S401. Construct a historical fault sample library with four levels of labels, including normal state, general defects, serious faults and critical faults. S402. The SMOTE oversampling algorithm is used to synthesize and expand minority class fault samples, balance the sample class distribution, and synthesize samples. The formula for generating it is: ; in, For the original minority class samples, Let it be a random sample from its k nearest neighbors. A random number between [0, 1]; S403. Introduce penalty factors for different fault severity levels into the loss function. This makes the cost of misclassifying high-risk faults higher than that of low-risk faults, resulting in a higher cost for the weighted cross-entropy loss function. Defined as: ; in, For the sample size, For the number of categories, For real labels, To predict probabilities, For the first Penalty weights for different types of failures; S404. Dynamically adjust the learning rate using the Adam optimizer. The formula is adjusted as follows: ; in, The initial learning rate, , The decay rate is used, and an early stopping method is employed on the validation set to prevent overfitting of the model until the F1-Score on the validation set converges.

[0028] Please refer to Figure 5 As shown, the specific logic for generating differentiated operation and maintenance strategies is as follows: S501. Establish a mapping function between health status scores and failure probabilities, classifying health status into five levels: excellent, normal, attentive, abnormal, and severe, and assigning a corresponding risk weight to each level. ; S502. Based on geographical location information, calculate the estimated travel time for maintenance personnel to reach the distribution cabinet from their work station. and road condition complexity coefficient ; S503. Set the load loss weighting coefficient according to the current load importance level. ; S504. Calculate the maintenance priority index using the weighted summation formula. The calculation formula is: ; in, , , Here are the normalized weight coefficients of each influencing factor, and ; S505. When the index exceeds the preset threshold At that time, a structural chemical work order containing specific maintenance time windows, a list of required spare parts, and risk prevention and control measures will be automatically generated.

[0029] It also includes the step of incrementally updating the fault prediction model: Real-time collection of fault feedback data verified on-site by maintenance personnel, and addition of it to the historical sample library as new training samples; Set model performance monitoring indicators. When the model's prediction accuracy on the new dataset drops beyond the preset range, or when the operating conditions of the power distribution cabinet undergo structural changes, trigger the model retraining mechanism. The formula for calculating prediction accuracy is: ; in, It is a true positive. It is a true negative. It was a false positive. It is a false negative; The weight parameters W and bias b of the LSTM network are fine-tuned and updated using the latest time-series dataset. The update formula is as follows: ; ; in, For learning rate, For gradient operators, The loss function is denoted by , and k-fold cross-validation is used to evaluate the generalization ability of the updated model.

[0030] In a second aspect of the invention, a big data-based power distribution cabinet fault maintenance system is also provided, comprising: The data acquisition module is used to collect the operating status data, environmental parameter data and historical operation and maintenance record data of the power distribution cabinet through multi-dimensional sensing terminals deployed in the power distribution cabinet and its surrounding environment, forming a raw multi-source heterogeneous dataset. The preprocessing module is used to preprocess the original multi-source heterogeneous dataset, including data cleaning, missing value imputation, timestamp alignment, and multi-source data fusion, to generate a standardized time-series feature dataset. The extraction module is used to construct a multi-dimensional state feature vector based on a standardized time-series feature dataset using a sliding time window technique, and to extract key fault symptom features, including time-domain statistical features, frequency-domain waveform features, and temperature field distribution features. The output module is used to input key fault symptom features into the pre-trained power distribution cabinet fault prediction model, perform deep feature learning and fault probability calculation, and output the health status score and potential fault type of the power distribution cabinet. The fault prediction model is trained based on a historical fault sample library and uses a hybrid deep learning architecture that combines ensemble learning algorithm and long short-term memory network. The generation module is used to generate differentiated operation and maintenance strategies based on the health status score and potential fault types, combined with the geographical location information of the distribution cabinet and the current load importance level, and push them to the operation and maintenance terminal.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for fault operation and maintenance of power distribution cabinets based on big data, characterized in that, include: By deploying multi-dimensional sensing terminals in and around the power distribution cabinet, the operating status data, environmental parameter data and historical operation and maintenance record data of the power distribution cabinet are collected to form an original multi-source heterogeneous dataset. Data preprocessing is performed on the original multi-source heterogeneous dataset, including data cleaning, missing value imputation, timestamp alignment, and multi-source data fusion, to generate a standardized time-series feature dataset; Based on a standardized time-series feature dataset, a multi-dimensional state feature vector is constructed using the sliding time window technique, and key fault symptom features are extracted. These features include time-domain statistical features, frequency-domain waveform features, and temperature field distribution features. Key fault symptom features are input into a pre-trained power distribution cabinet fault prediction model for deep feature learning and fault probability calculation, outputting the health status score and potential fault type of the power distribution cabinet. The fault prediction model is trained based on a historical fault sample library using a hybrid deep learning architecture that combines ensemble learning algorithms and long short-term memory networks. Based on the health status score and potential fault types, combined with the geographical location information of the distribution cabinet and the current load importance level, differentiated operation and maintenance strategies are generated and pushed to the operation and maintenance terminal.

2. The method for fault operation and maintenance of power distribution cabinets based on big data according to claim 1, characterized in that, The specific process of data cleaning and missing value imputation for the original multi-source heterogeneous dataset is as follows: The 3σ principle based on statistical distribution or Hampel identifier is used to remove physically impossible values ​​from the operational status data and abnormal jump points in the environmental parameters. For discrete missing data caused by communication interruption, a K-nearest neighbor-based interpolation algorithm is used to fill in the missing data. The interpolation calculation formula is as follows: ； in, The fill value for the missing points. For the first Observations of neighboring nodes, The corresponding weighting coefficient is usually taken as... , This refers to the electrical or Euclidean distance between the missing point and its neighboring nodes. The preset number of neighboring nodes; For continuously missing data, forward predictions based on a time-series autoregressive moving average model are used for imputation. The prediction formula is as follows: ； in, For a moment The predicted value, For constant terms, These are the autoregressive coefficients. The moving average coefficient is... It is a white noise sequence. and These are the autoregression order and the moving average order, respectively, to ensure the temporal continuity of the data.

3. The method for fault operation and maintenance of power distribution cabinets based on big data according to claim 2, characterized in that, The specific implementation of timestamp alignment and multi-source data fusion of the original multi-source heterogeneous dataset includes: Using the unique device identifier of the power distribution cabinet as the key, the collected structured operation data, semi-structured environmental logs, and unstructured maintenance text records are associated. A linear interpolation resampling method is employed to map sensor data from different sampling frequencies onto the same time axis, generating time-series sequences with equal time intervals. The interpolation formula is as follows: ； in, For the interpolation result, For the target timestamp, and The start and end times of the interpolation interval. and These are the original observations for the corresponding time period; The coupling relationship between multidimensional parameters is analyzed using the Pearson correlation coefficient matrix, and the calculation formula is as follows: ； in, For variables and The correlation coefficient, and For the first Observations of each sample point and These are the means of the two sets of data. To determine the sample size, highly collinear features with redundancy exceeding a preset threshold are removed, while feature dimensions with independent information contributions are retained.

4. The method for fault operation and maintenance of power distribution cabinets based on big data according to claim 3, characterized in that, The method of constructing a multi-dimensional state feature vector using the sliding time window technique and extracting key fault symptom features specifically includes: Set a fixed-length sliding time window, with the window length L covering one complete thermal stability cycle of the power distribution cabinet, and the window sliding step size Δt being an integer multiple of the sampling period; In the time domain, the statistical characteristics of current, voltage, and vibration signals within the calculation window are determined, including the mean μ and variance. Skewness S and kurtosis The formula for calculating kurtosis is: ； in, For the first in the window Each sample value, This represents the average of the data within the window. Standard deviation The number of sample points within the window, and the peak factor is extracted simultaneously. ,and ; In the frequency domain, a fast Fourier transform is performed on the time-domain signal to extract the amplitude of the fundamental component. Total Harmonic Distortion (THD) and Energy Spectral Density in Specific High-Frequency Bands The formula for calculating THD is: ； in, For the first The amplitude of the second harmonic component, The highest harmonic order; In the spatial dimension, the temperature field matrix inside the cabinet is constructed based on infrared thermal imaging data, and the norm of the temperature gradient at each measuring point and the offset displacement of the hot spot center are calculated.

5. The method for fault operation and maintenance of power distribution cabinets based on big data according to claim 4, characterized in that, The extraction of temperature field distribution features also includes: The interior of the distribution cabinet is divided into several three-dimensional mesh units, and the temperature rise rate of each unit is calculated. and the Euclidean distance from the historical average temperature rise curve ; The thermal diffusivity at key contact points inside the cabinet is calculated using the difference scheme of the heat conduction equation. The coefficient is then normalized to the square of the load current to obtain the thermal-electric coupling characteristic index η, which characterizes the abnormal change in contact resistance. The formula for calculating the thermal diffusivity is: ； in, The thermal conductivity of the material, For density, Specific heat capacity; The formula for calculating the thermal-electric coupling characteristic index η is: ； in, The thermal diffusivity is calculated at the current moment. The baseline thermal diffusivity under historical normal operating conditions. This is the current load current. This is the rated current.

6. The method for fault operation and maintenance of power distribution cabinets based on big data according to claim 5, characterized in that, The hybrid deep learning architecture of the pre-trained power distribution cabinet fault prediction model comprises a two-level cascaded structure: The first stage employs an ensemble learning algorithm based on gradient boosting decision trees to perform preliminary screening and nonlinear transformation on the input key fault symptom features, outputting a high-dimensional abstract feature vector. The superposition process of the m-th tree is represented as follows: ； in, For the first The prediction function of a decision tree. The learning rate; The second level will use the abstract feature vector As input, it is fed into the input gate. Forgotten Gate and output gate The constructed multi-layer Long Short-Term Memory (LSTM) network optimizes the dependency weights on the time series using the backpropagation algorithm. The cell state update formula for LSTM units is as follows: ； ； in, for Cellular state at any given moment For the hidden layer output, ⊙ represents the Hadamard product. As candidate memory units, the probability distribution P(y|X) of various faults is finally output through a fully connected layer and a Softmax activation function.

7. A method for fault operation and maintenance of power distribution cabinets based on big data as described in claim 6, characterized in that, The training process of the fault prediction model includes: Construct a historical fault sample library with four levels of tags, including normal state, general defects, serious faults and critical faults; The SMOTE oversampling algorithm is used to synthesize and augment minority class fault samples, balancing the sample class distribution. The formula for generating it is: ； in, For the original minority class samples, Let it be a random sample from its k nearest neighbors. A random number between [0, 1]; Introduce penalty factors for different fault severity levels into the loss function. This makes the cost of misclassifying high-risk faults higher than that of low-risk faults, resulting in a higher cost for the weighted cross-entropy loss function. Defined as: ； in, For the sample size, For the number of categories, For real labels, To predict probabilities, For the first Penalty weights for different types of failures; Dynamically adjust the learning rate using the Adam optimizer The formula is adjusted as follows: ； in, The initial learning rate, , The decay rate is used, and an early stopping method is employed on the validation set to prevent overfitting of the model until the F1-Score on the validation set converges.

8. A method for fault operation and maintenance of power distribution cabinets based on big data as described in claim 7, characterized in that, The specific logic for generating differentiated operation and maintenance strategies is as follows: A mapping function between health status scores and failure probabilities is established, classifying health status into five levels: excellent, normal, attentive, abnormal, and severe, and assigning a corresponding risk weight to each level. ; Based on geographical location information, calculate the estimated travel time for maintenance personnel to reach the distribution cabinet from their work site. and road condition complexity coefficient ; Set the load loss weighting factor based on the current load importance level. ; The maintenance priority index is calculated using a weighted summation formula. The calculation formula is: ； in, , , Here are the normalized weight coefficients of each influencing factor, and ; When the index exceeds the preset threshold At that time, a structural chemical work order containing specific maintenance time windows, a list of required spare parts, and risk prevention and control measures will be automatically generated.

9. A method for fault operation and maintenance of power distribution cabinets based on big data as described in claim 8, characterized in that, It also includes the step of incrementally updating the fault prediction model: Real-time collection of fault feedback data verified on-site by maintenance personnel, and addition of it to the historical sample library as new training samples; Set model performance monitoring indicators. When the model's prediction accuracy on the new dataset drops beyond the preset range, or when the operating conditions of the power distribution cabinet undergo structural changes, trigger the model retraining mechanism. The formula for calculating prediction accuracy is: ； in, It is a true positive. It is a true negative. It was a false positive. It is a false negative; The weight parameters W and bias b of the LSTM network are fine-tuned and updated using the latest time-series dataset. The update formula is as follows: ； ； in, For learning rate, For gradient operators, The loss function is denoted by , and the k-fold cross-validation method is used to evaluate the generalization ability of the updated model.

10. A big data-based power distribution cabinet fault operation and maintenance system, used to implement the big data-based power distribution cabinet fault operation and maintenance method as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect the operating status data, environmental parameter data and historical operation and maintenance record data of the power distribution cabinet through multi-dimensional sensing terminals deployed in the power distribution cabinet and its surrounding environment, forming an original multi-source heterogeneous dataset. The preprocessing module is used to preprocess the original multi-source heterogeneous dataset, including data cleaning, missing value imputation, timestamp alignment and multi-source data fusion, to generate a standardized time-series feature dataset. The extraction module is used to construct a multi-dimensional state feature vector based on a standardized time-series feature dataset using a sliding time window technique, and to extract key fault symptom features, including time-domain statistical features, frequency-domain waveform features, and temperature field distribution features. The output module is used to input key fault symptom features into a pre-trained power distribution cabinet fault prediction model, perform deep feature learning and fault probability calculation, and output the health status score and potential fault type of the power distribution cabinet. The fault prediction model is trained based on a historical fault sample library and using a hybrid deep learning architecture that combines ensemble learning algorithms and long short-term memory networks. The generation module is used to generate differentiated operation and maintenance strategies based on the health status score and potential fault types, combined with the geographical location information of the distribution cabinet and the current load importance level, and push them to the operation and maintenance terminal.

Images