Switch cabinet state intelligent detection method and system based on multi-sensor fusion

Abstract

The invention relates to the technical field of power equipment state monitoring, and provides a switch cabinet state intelligent detection method and system based on multi-sensor fusion, and the method comprises the steps: obtaining state data corresponding to a switch cabinet through a temperature sensor, a vibration sensor, a partial discharge sensor, a current sensor and a voltage sensor which are disposed in the switch cabinet; inputting the obtained state data into a preset data preprocessing module, performing denoising processing and normalization processing on the state data, obtaining preprocessed multi-source data, and generating a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a preset hierarchical fusion model, carrying out feature layer fusion on the multi-dimensional feature vector to obtain an intermediate fusion feature, combining historical operation data and historical fault data of the switch cabinet, carrying out decision-making layer weighted fusion on the intermediate fusion feature to generate a comprehensive state feature vector, and outputting the comprehensive state feature vector. And triggering a corresponding early warning system or a fault alarm system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power equipment state monitoring, in particular to a switch cabinet state intelligent detection method and system based on multi-sensor fusion. BACKGROUND

[0002] As the core equipment for realizing power distribution, control and protection in the power system, the reliability of the operation state of the switch cabinet directly affects the safety and stability of the power system. The traditional switch cabinet state detection method usually relies on a single type of sensor (such as a temperature sensor, a partial discharge sensor, etc.) to independently monitor, which has the problems of single detection dimension and incomplete abnormal feature recognition, and it is difficult to accurately capture the equipment hidden danger under the coupling of multiple factors. With the development of multi-sensor fusion technology, some existing schemes try to improve the detection effect through multi-source data fusion, but generally have the following defects: 1. Single data fusion level: only simple fusion at the feature layer or the decision layer, without forming a hierarchical and progressive fusion mechanism, and it is difficult to fully exploit the complementary information of different sensor data. For example, traditional feature layer fusion mostly uses simple splicing or average weighting, without considering the historical reliability difference of sensors; the decision layer fusion lacks deep use of historical fault data, resulting in insufficient reliability of the fusion result.

[0003] 2. Insufficient state evaluation model precision: relying on artificial design of feature engineering and traditional machine learning algorithms, the adaptability to complex fault modes is poor, it is difficult to automatically learn the nonlinear correlation between multi-dimensional features, resulting in low state classification accuracy.

[0004] 3. Coarse abnormal response mechanism: only triggering early warning according to a single feature threshold, without comprehensive judgment combining the abnormal degree of multi-dimensional features, which is prone to false positives or false negatives, and lacks precise positioning of fault types, making it difficult to guide operation and maintenance decisions.

[0005] Therefore, a method is needed to solve at least one of the above problems. SUMMARY

[0006] The present application provides a switch cabinet state intelligent detection method and system based on multi-sensor fusion, aiming to solve the problem in the prior art that no scheme has organically combined a hierarchical fusion model (combining the feature layer and the decision layer), a deep learning state evaluation model, and an association analysis of historical operation data and fault data, resulting in that the comprehensiveness, accuracy and intelligence level of switch cabinet state detection cannot meet the actual needs.

[0007] In a first aspect, the present application provides a switch cabinet state intelligent detection method based on multi-sensor fusion, comprising: The status data of the switch cabinet is obtained by temperature sensors, vibration sensors, partial discharge sensors, current sensors and voltage sensors arranged in the switch cabinet. The status data includes at least temperature data, vibration data, partial discharge data, current data and voltage data. The acquired status data is input into a preset data preprocessing module to perform noise reduction and normalization processing on the status data, thereby obtaining preprocessed multi-source data and generating multi-dimensional feature vectors. The multi-dimensional feature vectors are then input into a preset hierarchical fusion model to perform feature layer fusion on the multi-dimensional feature vectors, thereby obtaining intermediate fusion features. Combined with the historical operation data and historical fault data of the switchgear, the intermediate fusion features are then subjected to decision-level weighted fusion to generate a comprehensive status feature vector. The comprehensive state feature vector is input into the pre-trained state assessment model, which is built based on a deep learning algorithm and trained with historical sample data. The comprehensive state feature vector is processed by the state assessment model to output the real-time state level of the switchgear. The real-time state level includes normal state, warning state and fault state. When the real-time status level is a warning state or a fault state, the corresponding warning information or fault type judgment result is generated based on the degree of abnormality of each dimension of the comprehensive status feature vector. The real-time detection data, status assessment results and abnormal information are stored in the preset database, triggering the corresponding warning system or fault alarm system.

[0008] In some embodiments, acquiring preprocessed multi-source data includes: for preprocessed temperature data, using a preset temperature feature extraction algorithm to extract temperature change trend features and abnormal temperature fluctuation features; for preprocessed vibration data, using a preset vibration feature extraction algorithm to extract time-domain features, frequency-domain features, and time-frequency-domain features of vibration signals; for preprocessed partial discharge data, using a preset discharge feature extraction algorithm to extract amplitude features, phase features, and discharge frequency features of discharge pulses; and for preprocessed current and voltage data, extracting amplitude features, waveform distortion features, and power factor features of current and voltage.

[0009] In some embodiments, the denoising and normalization of the state data to obtain preprocessed multi-source data and generate multi-dimensional feature vectors includes: denoising the temperature data using a moving average filtering algorithm to remove random noise and abnormal jump points; denoising the vibration data using a wavelet denoising algorithm to remove high-frequency noise through multi-layer wavelet decomposition and thresholding; denoising the partial discharge data using a pulse recognition algorithm to eliminate interference signals by combining the time-domain waveform characteristics of the discharge pulse; denoising the current and voltage data using a median filtering algorithm to remove periodic interference and spike noise; and normalizing the denoised data by normalizing the temperature data to a preset temperature range, and scaling the vibration data, partial discharge data, current data, and voltage data to a preset uniform numerical range according to their data types to eliminate dimensional differences between different sensor data.

[0010] In some embodiments, the multi-dimensional feature vector includes temperature features, vibration features, partial discharge features, and current-voltage features corresponding to temperature data, vibration data, partial discharge data, current data, and voltage data, respectively. The step of inputting the multi-dimensional feature vector into a preset hierarchical fusion model and performing feature layer fusion to obtain intermediate fusion features includes: concatenating temperature features, vibration features, partial discharge features, and current-voltage features according to feature dimensions to form an initial fusion vector containing each type of feature; performing dimensionality reduction on the initial fusion vector using principal component analysis, retaining principal components whose cumulative variance contribution rate exceeds a preset threshold, removing redundant features, and generating intermediate fusion features containing key information; or, using a weighted fusion method, assigning initial weights to temperature features, vibration features, partial discharge features, and current-voltage features based on historical reliability data of each type of sensor, and generating intermediate fusion features through weighted summation.

[0011] In some embodiments, the step of combining historical operation data and historical fault data of the switchgear to perform decision-level weighted fusion of intermediate fusion features to generate a comprehensive state feature vector includes: extracting historical operation data of the target switchgear and historical fault data of similar switchgear from a preset database; cleaning and extracting features from the historical data to construct a historical feature dataset; using an association rule algorithm to analyze the correlation between each dimension feature and the fault type in the historical feature data; calculating the weight coefficient of each dimension feature in the decision-level fusion based on the correlation; performing a weighted operation on the intermediate fusion features and the weight coefficients; and correcting the weighted result by combining the probability distribution of fault occurrence in the historical data to generate a comprehensive state feature vector containing historical experience information.

[0012] In some embodiments, the state assessment model is constructed based on a deep learning algorithm and trained using historical sample data, including: constructing a state assessment model using a multi-layer neural network structure, wherein the neural network includes at least an input layer, several hidden layers, and an output layer, the number of neurons in the input layer matches the dimension of the comprehensive state feature vector, and the number of neurons in the output layer matches the number of categories of real-time state levels; dividing the historical sample data into a training set and a validation set, wherein the training set includes sample data of normal state, warning state, and fault state and their corresponding labels; and training the state assessment model using a backpropagation algorithm, with the state level prediction accuracy of the sample data as the optimization objective, and adjusting the weights and bias parameters of the neural network to make the prediction error of the model on the validation set less than a preset threshold.

[0013] In some embodiments, processing the comprehensive state feature vector through the state evaluation model to output the real-time state level of the switchgear includes: inputting the comprehensive state feature vector into the input layer of the state evaluation model, performing feature transformation sequentially through the activation functions of each hidden layer, generating probability values ​​for each state level in the output layer; setting a preset state level judgment threshold; outputting a normal state when the probability value of a normal state is greater than a first threshold; outputting a warning state when the probability value of a warning state is greater than a second threshold and higher than the probability values ​​of other states; and outputting a fault state when the probability value of a fault state is greater than a third threshold and higher than the probability values ​​of other states, wherein the first threshold is greater than the second threshold, and the second threshold is greater than the third threshold.

[0014] In some embodiments, when the real-time status level is a warning state or a fault state, generating corresponding warning information or fault type judgment results based on the degree of abnormality of each dimension feature in the comprehensive status feature vector includes: establishing a preset mapping rule library between abnormal features and warning levels and fault types, wherein the mapping rule library contains the correspondence between the abnormal threshold, abnormal duration and abnormal type of each dimension feature; comparing each dimension feature in the comprehensive status feature vector with the threshold in the mapping rule library, calculating the degree of deviation and duration of the feature value from the normal range; matching the corresponding warning level or fault type according to the degree of deviation and duration; if the temperature feature exceeds the high temperature threshold multiple times consecutively, generating a temperature abnormality warning; if the partial discharge frequency feature exceeds the fault threshold, combining the phase feature to match the preset discharge fault type, generating a specific fault type judgment result.

[0015] In some embodiments, storing real-time detection data, status assessment results, and anomaly information in a preset database to trigger a corresponding early warning system or fault alarm system includes: structuring and storing real-time detected temperature data, vibration data, partial discharge data, current data, and voltage data according to timestamps and switchgear numbers, and storing corresponding multi-dimensional feature vectors, comprehensive status feature vectors, and status assessment results; when the real-time status level is an early warning state, calling the corresponding early warning interface according to preset early warning rules to send a message containing anomaly characteristics and an early warning level to the monitoring platform; when the real-time status level is a fault state, triggering a local audible and visual alarm device and a remote fault alarm system, sending emergency alarm information containing fault type judgment results and fault location, and recording the alarm time and processing status to the database.

[0016] Secondly, this application provides an intelligent switchgear status detection system based on multi-sensor fusion, comprising: The data acquisition unit is used to acquire the status data of the switch cabinet through temperature sensors, vibration sensors, partial discharge sensors, current sensors and voltage sensors arranged in the switch cabinet. The status data includes at least temperature data, vibration data, partial discharge data, current data and voltage data. The vector generation unit is used to input the acquired state data into a preset data preprocessing module, perform noise reduction and normalization on the state data, obtain preprocessed multi-source data, and generate multi-dimensional feature vectors; input the multi-dimensional feature vectors into a preset hierarchical fusion model, perform feature layer fusion on the multi-dimensional feature vectors, obtain intermediate fusion features, and combine the historical operation data and historical fault data of the switchgear to perform decision layer weighted fusion on the intermediate fusion features to generate a comprehensive state feature vector; The level output unit is used to input the comprehensive state feature vector into the pre-trained state assessment model. The state assessment model is built based on a deep learning algorithm and trained with historical sample data. The comprehensive state feature vector is processed by the state assessment model to output the real-time state level of the switchgear. The real-time state level includes normal state, warning state and fault state. The result generation unit is used to generate corresponding early warning information or fault type judgment results based on the degree of abnormality of each dimension feature in the comprehensive state feature vector when the real-time state level is an early warning state or a fault state. It also stores the real-time detection data, state assessment results and abnormal information in a preset database and triggers the corresponding early warning system or fault alarm system.

[0017] This application provides a method and system for intelligent detection of switchgear status based on multi-sensor fusion. The method utilizes multiple types of sensors, including temperature, vibration, partial discharge, current, and voltage, for collaborative monitoring, covering multi-dimensional potential hazards such as thermal faults, mechanical faults, and insulation faults in switchgear, avoiding the problem of missed detection by a single sensor. The hierarchical fusion model balances the complementary multi-source information at the feature layer with the guidance of historical experience at the decision layer. Through principal component analysis, weighted fusion, and association rule algorithms, it effectively removes redundant features and dynamically allocates feature weights, improving data utilization efficiency. The deep learning status assessment model automatically learns the complex associations of multi-dimensional features, achieving significantly higher accuracy in classifying normal, warning, and fault states than traditional algorithms, reducing errors from manual intervention. Based on the degree and duration of deviations in abnormal features, specific warning information or fault type judgments are generated, providing clear maintenance guidance for operation and maintenance personnel, reducing fault handling costs, and improving the reliability of power system operation.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart illustrating the steps of an intelligent detection method for switchgear status based on multi-sensor fusion provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a switchgear status intelligent detection system based on multi-sensor fusion provided in an embodiment of this application; Figure 3 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0024] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0025] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0026] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0027] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0028] Switchgear, as the core equipment for power distribution, control, and protection in a power system, directly impacts the safety and stability of the power system due to the reliability of its operating status. Traditional switchgear status monitoring methods typically rely on independent monitoring using a single type of sensor (such as temperature sensors or partial discharge sensors), which suffers from limitations in detection dimensions and incomplete identification of abnormal features, making it difficult to accurately capture potential equipment hazards under the combined effects of multiple factors. With the development of multi-sensor fusion technology, some existing solutions attempt to improve detection performance through multi-source data fusion, but they generally suffer from the following drawbacks: 1. Single-layered data fusion: The fusion process is limited to simple fusion at the feature layer or decision layer, failing to establish a hierarchical and progressive fusion mechanism, making it difficult to fully leverage the complementary information from different sensor data. For example, traditional feature layer fusion often uses simple splicing or average weighting without considering the historical reliability differences of sensors; decision layer fusion lacks in-depth utilization of historical fault data, resulting in insufficient reliability of the fusion results.

[0029] 2. Insufficient accuracy of state assessment models: Relying on manually designed feature engineering and traditional machine learning algorithms, they are poorly adaptable to complex fault modes and have difficulty automatically learning the nonlinear correlation between multi-dimensional features, resulting in low state classification accuracy.

[0030] 3. The abnormal response mechanism is crude: it triggers warnings based on only a single feature threshold, without combining the degree of abnormality of multiple features for comprehensive judgment. This can easily lead to false alarms or missed alarms, and it lacks accurate location of fault types, making it difficult to guide operation and maintenance decisions.

[0031] In the existing technology, there is no solution that organically combines the hierarchical fusion model (combining the feature layer and the decision layer), the deep learning state assessment model, and the correlation analysis of historical operating data and fault data, which makes it difficult to meet the actual needs in terms of the comprehensiveness, accuracy and intelligence level of switch cabinet state detection.

[0032] Therefore, there is an urgent need for a detection method that can integrate multi-source heterogeneous data, hierarchically mine feature value, and accurately assess equipment status.

[0033] To resolve the above issues, please refer to [link / reference]. Figure 1 This intelligent detection method for switchgear status based on multi-sensor fusion can be implemented using computer equipment, which can be deployed on a single server or server cluster. It can also be deployed on handheld terminals, laptops, wearable devices, or robots, etc.

[0034] Specifically, such as Figure 1 As shown, the provided intelligent detection method for switchgear status based on multi-sensor fusion includes steps S101 to S104, which are detailed below: Step S101. Obtain the corresponding status data of the switch cabinet by means of temperature sensors, vibration sensors, partial discharge sensors, current sensors and voltage sensors arranged in the switch cabinet. The status data includes at least temperature data, vibration data, partial discharge data, current data and voltage data.

[0035] Specifically, by deploying various types of sensors, key status parameters of the switchgear are collected in real time, covering multiple physical quantities such as temperature, vibration, partial discharge, current, and voltage, thus constructing a comprehensive monitoring system.

[0036] Sensor selection and placement include: Temperature sensors: Thermocouples or infrared sensors are installed at heat-prone locations such as contacts, busbar connections, and cable joints to collect temperature data in real time (accuracy ±0.5℃, sampling frequency 1Hz). Vibration sensors: Accelerometers are installed on the cabinet surface or mechanical components (such as circuit breaker operating mechanisms) to collect vibration signals (frequency range 0-10kHz, sampling frequency 20kHz) for monitoring mechanical faults (such as loosening or wear). Partial discharge sensors: Ultrasonic sensors (detecting surface discharge) or high-frequency current transformers (HFCTs, detecting internal discharge) are used to collect discharge pulse signals (discharge quantity, number of discharges, phase distribution), with a sampling frequency of 1MHz or higher. Current / voltage sensors: Three-phase current and voltage RMS values ​​and waveform data are collected through Rogowski coils or transformers (sampling frequency 100Hz) for analyzing load conditions and abnormal electrical parameters.

[0037] The data acquisition hardware uses distributed data acquisition modules (such as PLCs and edge computing nodes), supports protocols such as Modbus and TCP / IP, and enables synchronous sampling and data transmission from multiple sensors, ensuring timestamp accuracy ≤1ms.

[0038] Step S102. Input the acquired status data into a preset data preprocessing module to perform noise reduction and normalization on the status data, obtain preprocessed multi-source data, and generate multi-dimensional feature vectors; input the multi-dimensional feature vectors into a preset hierarchical fusion model to perform feature layer fusion on the multi-dimensional feature vectors, obtain intermediate fusion features, and combine the historical operation data and historical fault data of the switchgear to perform decision layer weighted fusion on the intermediate fusion features to generate a comprehensive status feature vector.

[0039] Specifically, data quality is improved through preprocessing such as denoising and normalization, and a two-layer fusion mechanism of "feature layer-decision layer" is constructed. By combining historical reliability and fault data of sensors, deep fusion of multi-source data is achieved.

[0040] The noise reduction process removes random noise by using Kalman filtering or moving average filtering on slowly varying signals such as temperature and current; and retains characteristic waveforms by using wavelet denoising or morphological filtering on transient signals such as vibration and partial discharge.

[0041] Using min-max normalization (formula: x′= (x xmin) / xmax X-score (xmin) or Z-score normalization can unify data of different dimensions to the interval [-1, 1] or [0, 1], thus eliminating the influence of scale differences.

[0042] Feature engineering extracts multi-dimensional feature vectors, including: time-domain features (mean, variance, peak value, kurtosis), frequency-domain features (FFT spectrum energy distribution), time-frequency domain features (wavelet packet energy entropy); statistical features specific to partial discharge (such as NQ-φ spectrum parameters, discharge repetition rate); and harmonic features of current and voltage (total harmonic distortion rate THD, amplitude of each harmonic).

[0043] The hierarchical fusion model includes: feature layer fusion employs a weighted feature concatenation method, dynamically allocating weights based on the sensor's historical reliability. The reliability index is defined as: wi = 1 / (1 + σi) 2 ),σi 2 =The standard deviation of historical data of sensor i is used to assign higher weights to sensors with high stability (such as current sensors) to avoid interference from noisy sensors. After fusion, an intermediate feature vector Fmid=[w1f1,w2f2…,wnfn] is generated.

[0044] The decision-making layer uses a weighted fusion approach, combining historical operating data (normal / abnormal operating condition samples) and historical fault data (fault types, pre-fault feature patterns) to construct a Bayesian weighted model or attention mechanism module. This allows for secondary fusion of the feature layer output: using historical fault data to train the contribution of each feature to different fault types (e.g., temperature anomalies have a higher weight for overheating faults). The weight matrix W={wij} is dynamically adjusted, where wij represents the importance of feature i to fault type j, and finally the comprehensive state feature vector Ffinal=W*Fmid is generated.

[0045] Step S103. Input the comprehensive state feature vector into the pre-trained state assessment model. The state assessment model is constructed based on a deep learning algorithm and trained using historical sample data. The comprehensive state feature vector is processed by the state assessment model to output the real-time state level of the switchgear. The real-time state level includes normal state, warning state and fault state.

[0046] Specifically, an end-to-end state assessment model is built based on deep learning, which automatically learns the nonlinear correlation between multi-dimensional features to improve the accuracy of complex fault mode identification.

[0047] The model architecture employs a multimodal fusion neural network, supporting different types of data input: for vibration and partial discharge waveform data, 1D-CNN is used to extract local features; for time-series data such as temperature and current, LSTM or Transformer is used to capture time dependencies; finally, multi-branch features are fused through fully connected layers to output state classification probabilities.

[0048] The training process includes: Sample construction: Collecting historical data and labeling it into three categories: normal, warning (early defects), and fault (clear fault type), and dividing it into training set, validation set, and test set in a 7:2:1 ratio; using the cross-entropy loss function and introducing focal loss to solve the sample imbalance problem; Optimization strategies include: Adam optimizer, dynamic adjustment of the learning rate, and decay of the learning rate if the accuracy of the validation set does not improve after every 50 rounds; Model output is expressed through real-time status level (normal / warning / fault) and confidence score, with the confidence score threshold set to 0.8 (which can be adjusted according to operation and maintenance needs).

[0049] Step S104. When the real-time status level is a warning state or a fault state, generate corresponding warning information or fault type judgment results based on the degree of abnormality of each dimension of the comprehensive status feature vector, and store the real-time detection data, status assessment results and abnormal information in the preset database to trigger the corresponding warning system or fault alarm system.

[0050] Specifically, the system quantifies the degree of anomalies based on comprehensive feature vectors to achieve accurate early warning and fault location, while storing all data for model iteration and operation and maintenance analysis.

[0051] Anomaly detection and response calculate the anomaly degree of each dimension's features when the status level is warning or fault: Anomaly degree i = |fi μi∣ / σi, (μi,σi are the mean / standard deviation of the normal state characteristics).

[0052] By combining preset rules (such as temperature anomaly > 3, partial discharge quantity suddenly increases by 2 times) with model feature attribution (such as gradient weighted activation map Grad-CAM), the dominant abnormal features are located, and specific early warning information (such as "contact temperature rises abnormally, infrared retest recommended") or fault type judgment (such as "abnormal mechanical vibration, suspected circuit breaker contact loosening").

[0053] Data storage and system linkage include: Database design: storing real-time detection data (timestamps, raw sensor values, feature vectors), status assessment results (level, confidence level), and anomaly information (anomaly ranking, fault type), supporting time-series databases (such as InfluxDB) or relational databases (MySQL); Early warning mechanism: triggering SMS and email alerts through API interface, and in the event of a fault, linking the switch cabinet intelligent terminal to cut off the power supply and pushing a work order (including fault location information) to the operation and maintenance platform to achieve a closed loop of "detection-assessment-response".

[0054] In some embodiments, acquiring preprocessed multi-source data includes: for preprocessed temperature data, using a preset temperature feature extraction algorithm to extract temperature change trend features and abnormal temperature fluctuation features; for preprocessed vibration data, using a preset vibration feature extraction algorithm to extract time-domain features, frequency-domain features, and time-frequency-domain features of vibration signals; for preprocessed partial discharge data, using a preset discharge feature extraction algorithm to extract amplitude features, phase features, and discharge frequency features of discharge pulses; and for preprocessed current and voltage data, extracting amplitude features, waveform distortion features, and power factor features of current and voltage.

[0055] For different types of status data (temperature, vibration, partial discharge, current / voltage), a dedicated feature extraction algorithm is used to extract key features reflecting the equipment status from the raw data and construct a multi-dimensional feature vector.

[0056] Temperature feature extraction includes: Temperature change trend features: The first-order difference mean (reflecting the rate of temperature rise) and second-order difference variance (reflecting the stability of temperature fluctuations) of temperature data are calculated using a sliding window (e.g., a 1-hour window). Linear regression is used to fit the slope of the temperature trend to identify long-term temperature rise trends. Abnormal temperature fluctuation features: Sudden jumps in temperature data exceeding the historical mean ±3σ are detected. The number and amplitude of abnormal points per unit time are statistically analyzed, generating features such as "abnormal fluctuation frequency" and "maximum fluctuation amplitude."

[0057] Vibration feature extraction includes: Time-domain features: calculating the mean, variance, peak value, kurtosis (reflecting waveform sharpness), and margin factor (measuring impact signal characteristics) of the vibration signal. Frequency-domain features: performing FFT transform on the vibration signal to extract the energy proportion of each frequency component (e.g., the energy proportion of high frequencies above 100Hz reflects the degree of mechanical loosening), the dominant peak frequency, and the corresponding amplitude. Time-frequency-domain features: using wavelet packet decomposition to calculate the energy entropy of each frequency band (e.g., 0-5kHz, 5-10kHz), characterizing the complexity of the signal's time-frequency distribution.

[0058] Partial discharge feature extraction includes: Amplitude features: extracting the peak amplitude, average amplitude, and the difference between the maximum and minimum amplitude of a single discharge pulse. Phase features: statistically analyzing the phase distribution of discharge pulses within the power frequency cycle (e.g., the percentage of discharges in the interval φ∈[0°,180°]), generating N-φ spectrum features. Discharge frequency features: calculating the number of discharges per unit time (e.g., per second) and the discharge repetition rate (the reciprocal of adjacent discharge intervals).

[0059] Current and voltage feature extraction includes: Amplitude features: extracting the effective values, maximum values, minimum values, and three-phase unbalance of three-phase current / voltage. Waveform distortion features: calculating the total harmonic distortion (THD) and the amplitude proportion of each harmonic (e.g., 3rd and 5th harmonics) using FFT to identify the degree of waveform distortion. Power factor features: calculating active power, reactive power, apparent power, and power factor (cosφ) to reflect load characteristics and energy utilization efficiency.

[0060] In some embodiments, the denoising and normalization of the state data to obtain preprocessed multi-source data and generate multi-dimensional feature vectors includes: denoising the temperature data using a moving average filtering algorithm to remove random noise and abnormal jump points; denoising the vibration data using a wavelet denoising algorithm to remove high-frequency noise through multi-layer wavelet decomposition and thresholding; denoising the partial discharge data using a pulse recognition algorithm to eliminate interference signals by combining the time-domain waveform characteristics of the discharge pulse; denoising the current and voltage data using a median filtering algorithm to remove periodic interference and spike noise; and normalizing the denoised data by normalizing the temperature data to a preset temperature range, and scaling the vibration data, partial discharge data, current data, and voltage data to a preset uniform numerical range according to their data types to eliminate dimensional differences between different sensor data.

[0061] To address the noise characteristics of different types of data, a dedicated denoising algorithm is used to improve data quality. Normalization is used to eliminate dimensional differences, providing standardized input for subsequent fusion processing.

[0062] Denoising processing includes: Temperature data: using a 5-point moving average filter (window size is configurable), the formula is as follows: Remove random noise and short-term abnormal fluctuations (such as sensor communication jitter).

[0063] Vibration data were decomposed into three levels using db4 wavelets, and soft thresholding was applied to the high-frequency coefficients (the main noise concentration area) (threshold = σ(2lnN)). 0.5 (where σ is the noise standard deviation and N is the signal length), the signal is reconstructed to preserve the mechanical vibration characteristics.

[0064] Partial discharge data are processed by designing a pulse recognition algorithm, which uses time-domain characteristics such as the rise time (<1μs) and pulse width (1-10μs) of the discharge pulse to remove interference signals (such as electromagnetic compatibility noise) with excessive duration.

[0065] The current and voltage data are filtered using a 3-point median filter. The median of the adjacent points before and after each sampling point is taken to eliminate periodic interference (such as power frequency harmonic aliasing) and spike noise.

[0066] Normalization processing includes: Temperature data: normalized to the [0, 1] interval, the formula is T′= (T Tmin) / (Tmax) Tmin), where Tmax and Tmin are the upper and lower limits of the temperature when the equipment is operating normally (e.g., 80℃ and 30℃ respectively).

[0067] Vibration / partial discharge / current / voltage data: Select the normalization method according to the data type: Vibration acceleration (m / s²) 2 The partial discharge quantity (pC) is normalized using Z-score to eliminate the influence of dimensions; the effective values ​​of current / voltage are scaled proportionally to the range of [-1, 1], as shown in the formula: X′=2(X ) / (X max) Xmin).

[0068] In some embodiments, the multi-dimensional feature vector includes temperature features, vibration features, partial discharge features, and current-voltage features corresponding to temperature data, vibration data, partial discharge data, current data, and voltage data, respectively. The step of inputting the multi-dimensional feature vector into a preset hierarchical fusion model and performing feature layer fusion to obtain intermediate fusion features includes: concatenating temperature features, vibration features, partial discharge features, and current-voltage features according to feature dimensions to form an initial fusion vector containing each type of feature; performing dimensionality reduction on the initial fusion vector using principal component analysis, retaining principal components whose cumulative variance contribution rate exceeds a preset threshold, removing redundant features, and generating intermediate fusion features containing key information; or, using a weighted fusion method, assigning initial weights to temperature features, vibration features, partial discharge features, and current-voltage features based on historical reliability data of each type of sensor, and generating intermediate fusion features through weighted summation.

[0069] In the feature layer fusion stage, multi-dimensional feature vectors are concatenated and spliced ​​together, and principal component analysis (PCA) or weighted fusion methods are combined to generate intermediate fused features containing key information, reducing redundancy and highlighting effective features.

[0070] The initial fusion vector is constructed by concatenating temperature features (such as trend slope, abnormal frequency), vibration features (time domain, frequency domain, time-frequency domain features), partial discharge features (amplitude, phase, frequency), and current and voltage features (amplitude, distortion, power factor) in a dimensional manner to form an initial vector of length D=D1+D2+D3+D4 (D1-number of temperature features, and so on).

[0071] PCA dimensionality reduction fusion extracts eigenvalues ​​and eigenvectors by calculating the covariance matrix of the initial vectors and sorting them by eigenvalues ​​from largest to smallest. Retain principal components with a cumulative variance contribution rate ≥ 95% (such as the first k principal components), generate intermediate fusion features of dimension k, and remove noise and redundant information (such as highly correlated duplicate features).

[0072] The weighted fusion method includes: historical reliability data: statistical analysis of the false alarm rate and data missing rate of each sensor over the past year, defining a reliability index Ri = 1. (False positive rate + Missing rate), the higher the reliability, the greater the weight; Weight allocation: normalized weight wi=Ri / ∑Ri, weighted summation of each type of feature, the formula is F mid=∑wi*Fi, where Fi is a single type feature vector.

[0073] In some embodiments, the step of combining historical operation data and historical fault data of the switchgear to perform decision-level weighted fusion of intermediate fusion features to generate a comprehensive state feature vector includes: extracting historical operation data of the target switchgear and historical fault data of similar switchgear from a preset database; cleaning and extracting features from the historical data to construct a historical feature dataset; using an association rule algorithm to analyze the correlation between each dimension feature and the fault type in the historical feature data; calculating the weight coefficient of each dimension feature in the decision-level fusion based on the correlation; performing a weighted operation on the intermediate fusion features and the weight coefficients; and correcting the weighted result by combining the probability distribution of fault occurrence in the historical data to generate a comprehensive state feature vector containing historical experience information.

[0074] During the decision-making fusion phase, historical operational data and fault data are used to analyze the correlation between features and faults through association rules, dynamically adjust the fusion weights, and generate comprehensive state features that include historical experience.

[0075] Historical data preparation involves extracting the target switchgear's operational data (feature vectors under normal / warning / fault conditions) from the database for the past three years, as well as historical fault cases of similar switchgear (labeled with fault types, such as mechanical faults, discharge faults, and overheating faults). The historical data is then cleaned (samples with a missing rate >30% are deleted) and standardized (using the same normalization method as in Example 2) to construct a historical feature dataset H={h1,h2,…,hm}, where each sample contains a feature vector and a fault type label.

[0076] Association rule analysis uses the Apriori algorithm to mine association rules between "feature anomaly → fault type", calculates support (the proportion of samples containing the rule) and confidence (the probability that the rule is correct), and filters rules with confidence ≥ 80%. The association degree Sij between feature fi and fault type Cj is defined as confidence × support, which serves as the core basis for the decision layer weight coefficient.

[0077] Weighted fusion and correction are achieved by constructing a weight matrix W=[wij], where wij=Sij / ∑j Sij, weight the intermediate fusion feature Fmid: Fweight=W*Fmid; combine the historical fault probability distribution (e.g., mechanical faults account for 30%, discharge faults account for 50%), and perform probability correction on Fweight: Ffinal=Fweight×P(Cj) to strengthen the feature influence corresponding to high-frequency faults.

[0078] In some embodiments, the state assessment model is constructed based on a deep learning algorithm and trained using historical sample data, including: constructing a state assessment model using a multi-layer neural network structure, wherein the neural network includes at least an input layer, several hidden layers, and an output layer, the number of neurons in the input layer matches the dimension of the comprehensive state feature vector, and the number of neurons in the output layer matches the number of categories of real-time state levels; dividing the historical sample data into a training set and a validation set, wherein the training set includes sample data of normal state, warning state, and fault state and their corresponding labels; and training the state assessment model using a backpropagation algorithm, with the state level prediction accuracy of the sample data as the optimization objective, and adjusting the weights and bias parameters of the neural network to make the prediction error of the model on the validation set less than a preset threshold.

[0079] A state assessment model is built based on deep learning. It automatically learns the nonlinear relationships between features through multi-layer neural networks and trains the model using historical samples to improve the classification accuracy of complex fault modes.

[0080] The model architecture design includes: Input layer: the number of neurons equals the dimension of the comprehensive state feature vector (e.g., 100 dimensions); Hidden layer: 2-3 fully connected layers are used, with the number of neurons in the first hidden layer set to 200 (an empirical value, which can be optimized through grid search), and ReLU is used as the activation function (to solve the gradient vanishing problem); Output layer: the number of neurons equals the number of state levels (3 categories: normal, warning, fault), and Softmax is used as the activation function, outputting the probability values ​​of each state.

[0081] Sample splitting and training: The training set (70%), validation set (20%), and test set (10%) are split in a 7:2:1 ratio to balance the number of samples in each set (imbalanced data is handled through oversampling / undersampling); Loss function: Cross-entropy loss function. , where yi is the real label (one-hot encoding). To predict probabilities, the Adam optimizer is used in the optimization process (with an initial learning rate of 0.001). The accuracy is evaluated on the validation set after every 10 rounds of training. If there is no improvement after 5 consecutive rounds, the training is terminated early to prevent overfitting.

[0082] Model optimization is achieved by adding a Dropout layer (ratio 0.2) to prevent overfitting and using a Batch Normalization layer to accelerate convergence. The final model's classification accuracy on the test set must be ≥95%; otherwise, the number of hidden layers or the learning rate should be adjusted and the model retrained.

[0083] In some embodiments, processing the comprehensive state feature vector through the state evaluation model to output the real-time state level of the switchgear includes: inputting the comprehensive state feature vector into the input layer of the state evaluation model, performing feature transformation sequentially through the activation functions of each hidden layer, generating probability values ​​for each state level in the output layer; setting a preset state level judgment threshold; outputting a normal state when the probability value of a normal state is greater than a first threshold; outputting a warning state when the probability value of a warning state is greater than a second threshold and higher than the probability values ​​of other states; and outputting a fault state when the probability value of a fault state is greater than a third threshold and higher than the probability values ​​of other states, wherein the first threshold is greater than the second threshold, and the second threshold is greater than the third threshold.

[0084] The probability value output by the state assessment model is combined with a preset threshold to determine the real-time state level, ensuring that the determination of different states has a clear confidence standard.

[0085] The probability values ​​are generated by inputting the comprehensive state feature vector into the model and then outputting a three-dimensional probability vector [Pnormal, Pwarning, Pfault] through the Softmax layer, which satisfies Pnormal + Pwarning + Pfault = 1.

[0086] The threshold setting and judgment logic includes: First threshold (normal state): set to 0.9 (high confidence requirement). If Pnormal > 0.9 and is the maximum value, it is judged as a normal state; Second threshold (warning state): set to 0.7. If Pwarning > 0.7 and Pwarning > Pnormal, Pwarning > Pfault, it is judged as a warning state. The third threshold (fault status): is set to 0.8 (higher than the warning threshold, reflecting the high confidence requirement of the fault). If Pfault>0.8 and is the maximum value, it is judged as a fault status. If the probability value does not meet the above conditions (e.g., the highest probability < 0.7), it is marked as "state uncertain", triggering data resampling and secondary evaluation.

[0087] In some embodiments, when the real-time status level is a warning state or a fault state, generating corresponding warning information or fault type judgment results based on the degree of abnormality of each dimension feature in the comprehensive status feature vector includes: establishing a preset mapping rule library between abnormal features and warning levels and fault types, wherein the mapping rule library contains the correspondence between the abnormal threshold, abnormal duration and abnormal type of each dimension feature; comparing each dimension feature in the comprehensive status feature vector with the threshold in the mapping rule library, calculating the degree of deviation and duration of the feature value from the normal range; matching the corresponding warning level or fault type according to the degree of deviation and duration; if the temperature feature exceeds the high temperature threshold multiple times consecutively, generating a temperature abnormality warning; if the partial discharge frequency feature exceeds the fault threshold, combining the phase feature to match the preset discharge fault type, generating a specific fault type judgment result.

[0088] By establishing a mapping rule base between abnormal features and warning / fault types, and by comparing feature values ​​with thresholds, combined with the degree of deviation and duration, accurate abnormal information can be generated.

[0089] The mapping rule base is constructed by defining anomaly thresholds for features in each dimension (e.g., temperature > 80℃ is the high temperature threshold, vibration acceleration > 5m / s² is the high temperature threshold). 2 The threshold for severe vibration and the duration of abnormality (e.g., exceeding the threshold for three consecutive samples) are used to establish a mapping relationship between "characteristic abnormal combination → fault type". For example: abnormal temperature + current imbalance → poor contact of contacts; high frequency vibration energy increase + partial discharge phase concentration → loosening of mechanical parts.

[0090] Anomaly Calculation: Deviation Di = |fi μi∣ / σi (μi / σi is the mean / standard deviation of the normal state characteristics), Di>2 is defined as significant abnormality; Duration: count the number of abnormal points in the most recent N sampling points (e.g. N=10), if it exceeds 50%, it is judged as persistent abnormality.

[0091] Early warning and fault location include: Temperature anomaly warning: If the temperature characteristic Di>3 and exceeds the standard for 5 consecutive samplings, generate "Contact temperature abnormal, infrared temperature measurement verification recommended"; Discharge fault judgment: If the frequency of partial discharge is >100 times / second and the phase is concentrated in 90°±30°, match the "internal corona discharge" type in the rule library, and output the fault type and severity (e.g., "Surface discharge of phase B insulator, power outage maintenance recommended").

[0092] In some embodiments, storing real-time detection data, status assessment results, and anomaly information in a preset database to trigger a corresponding early warning system or fault alarm system includes: structuring and storing real-time detected temperature data, vibration data, partial discharge data, current data, and voltage data according to timestamps and switchgear numbers, and storing corresponding multi-dimensional feature vectors, comprehensive status feature vectors, and status assessment results; when the real-time status level is an early warning state, calling the corresponding early warning interface according to preset early warning rules to send a message containing anomaly characteristics and an early warning level to the monitoring platform; when the real-time status level is a fault state, triggering a local audible and visual alarm device and a remote fault alarm system, sending emergency alarm information containing fault type judgment results and fault location, and recording the alarm time and processing status to the database.

[0093] By storing detection data, evaluation results, and anomaly information in a structured manner, and triggering different alarm mechanisms according to the status level, the traceability of monitoring data and the automation of anomaly response are achieved.

[0094] The data storage design includes: a structured table structure comprising: a real-time detection table containing timestamps (accurate to milliseconds), switch cabinet number, sensor type, and raw data values; a feature storage table storing multi-dimensional feature vectors, intermediate fusion features, and comprehensive status features; and an evaluation result table recording status levels, probability values ​​for each status, and anomaly information (such as the name and value of the feature with the highest anomaly score). Storage technologies: time-series data uses InfluxDB (supporting high-concurrency writes), and structured metadata uses MySQL, supporting fast queries by time range and device number.

[0095] Warning and alarm triggering include: Warning status: Sending a JSON format message to the monitoring platform via HTTP interface, including device number, warning time, abnormal characteristics (such as "vibration time domain kurtosis value exceeds the standard by 1.5 times"), and suggested measures (such as "arrange vibration spectrum analysis"). Fault states include: locally triggered audible and visual alarm devices (flashing red indicator light + buzzer alarm); remotely sending emergency alarm information (including fault type and location, such as "Partial discharge fault in phase B of switch cabinet No. 3, power outage recommended immediately") via API call to SMS / email interface; recording alarm time and processing status (unprocessed / processing / processed) in the database, and automatically generating maintenance tasks in conjunction with the maintenance work order system. Data traceability and model iteration: stored historical data is used for periodic model updates (such as monthly retraining of the state assessment model), and abnormal cases are used to expand the mapping rule base and improve the system's self-optimization capabilities.

[0096] In some embodiments, to address the scarcity of historical fault data for new switchgear models, a federated transfer learning framework is adopted to jointly train the system using data from similar devices at multiple sites, while simultaneously adapting the target device's personalized characteristics through transfer learning.

[0097] The federated learning architecture includes: Participants: Each substation acts as a federated node, storing the operation data of its switchgear (including a small number of fault samples) locally. Raw data is not uploaded; only model parameters are uploaded. Global model: The Transformer architecture is used to process time-series features, including a multi-head self-attention layer to capture long-distance dependencies, and the output layer is a state classifier.

[0098] Transfer learning adaptation includes: Pre-training stage: Using normal / warning state data from all federated nodes, a global base model is trained using the federated averaging algorithm (FedAvg); Fine-tuning stage: For the target new model of switchgear, a small number of fault samples (e.g., 10 fault instances) are collected, the self-attention layer of the base model is frozen, and only the output layer parameters are fine-tuned. The difference in feature distribution between the new and old equipment is reduced by using domain adaptation loss (e.g., the maximum mean difference of MMD).

[0099] Privacy protection and weight aggregation are achieved by performing differential privacy processing (adding Laplacian noise and privacy budgeting) on ​​each node before uploading model parameters. =0.5); a weighted average is used during aggregation, and the weights are dynamically adjusted based on the amount of data in each node and the richness of fault samples (the weight of nodes with a high proportion of fault samples is increased).

[0100] In some embodiments, by overcoming the rigidity of fixed threshold warnings, a reinforcement learning agent is introduced to dynamically adjust the warning strategy based on real-time features and historical operation and maintenance feedback, thereby reducing false alarm rates and improving fault response efficiency.

[0101] The state space and action definition include: State: current comprehensive feature vector, historical 24-hour early warning records, and operation and maintenance personnel response time (e.g., the average processing time of the last 3 early warnings); Action: adjusting the early warning threshold (e.g., dynamically adjusting the temperature early warning threshold from 80℃ to 75℃~85℃), triggering an urgent alarm (early alarm when the probability of failure increases); Reward function: R=0.8*fault false negative rate -1 -0.5×false alarm rate -0.3×response delay, balancing missed alarms, false alarms and response efficiency, and calibrating the reward coefficient through historical operation and maintenance data.

[0102] The Deep Q-Network (DQN) architecture includes: Input layer: concatenating a comprehensive feature vector (100-dimensional) and historical early warning metadata (5-dimensional), totaling 105 dimensions; Hidden layer: two fully connected layers (200 and 100 neurons respectively), with ReLU activation function; Output layer: Q-value function, outputting the expected reward value for each action, with an action space of 10 discrete threshold adjustment levels (e.g., step size of 1°C).

[0103] Online learning and policy updates are achieved by the agent selecting actions based on real-time status (ε-greedy policy, initial ε=0.3, decaying to 0.1 during training), observing the feedback from the operation and maintenance system (such as whether there are missed or false alarms) and updating the Q value; after processing 100 warning records, batch training is performed using historical data in the experience replay pool (storing the most recent 1000 interaction records) to prevent overfitting.

[0104] This application provides a method and system for intelligent detection of switchgear status based on multi-sensor fusion. The method utilizes multiple types of sensors, including temperature, vibration, partial discharge, current, and voltage, for collaborative monitoring, covering multi-dimensional potential hazards such as thermal faults, mechanical faults, and insulation faults in switchgear, avoiding the problem of missed detection by a single sensor. The hierarchical fusion model balances the complementary multi-source information at the feature layer with the guidance of historical experience at the decision layer. Through principal component analysis, weighted fusion, and association rule algorithms, it effectively removes redundant features and dynamically allocates feature weights, improving data utilization efficiency. The deep learning status assessment model automatically learns the complex associations of multi-dimensional features, achieving significantly higher accuracy in classifying normal, warning, and fault states than traditional algorithms, reducing errors from manual intervention. Based on the degree and duration of deviations in abnormal features, specific warning information or fault type judgments are generated, providing clear maintenance guidance for operation and maintenance personnel, reducing fault handling costs, and improving the reliability of power system operation.

[0105] Please see Figure 2 As shown, Figure 2 This is a schematic diagram of the structure of a multi-sensor fusion-based intelligent switchgear status detection system 200 provided in this application embodiment. The multi-sensor fusion-based intelligent switchgear status detection system 200 is used to execute the steps of the multi-sensor fusion-based intelligent switchgear status detection method shown in the above embodiments. The multi-sensor fusion-based intelligent switchgear status detection system 200 can be a single server or a server cluster, or it can be a terminal, such as a handheld terminal, laptop computer, wearable device, or robot.

[0106] like Figure 2 As shown, the intelligent switchgear status detection system 200 based on multi-sensor fusion includes: The data acquisition unit 201 is used to acquire the status data of the switch cabinet through temperature sensors, vibration sensors, partial discharge sensors, current sensors and voltage sensors arranged in the switch cabinet. The status data includes at least temperature data, vibration data, partial discharge data, current data and voltage data. The vector generation unit 202 is used to input the acquired state data into a preset data preprocessing module, perform noise reduction and normalization on the state data, obtain preprocessed multi-source data, and generate multi-dimensional feature vectors; input the multi-dimensional feature vectors into a preset hierarchical fusion model, perform feature layer fusion on the multi-dimensional feature vectors, obtain intermediate fusion features, and combine the historical operation data and historical fault data of the switchgear to perform decision layer weighted fusion on the intermediate fusion features to generate a comprehensive state feature vector; The level output unit 203 is used to input the comprehensive state feature vector into the pre-trained state assessment model. The state assessment model is built based on a deep learning algorithm and trained with historical sample data. The comprehensive state feature vector is processed by the state assessment model to output the real-time state level of the switch cabinet. The real-time state level includes normal state, warning state and fault state. The result generation unit 204 is used to generate corresponding early warning information or fault type judgment results based on the degree of abnormality of each dimension feature in the comprehensive state feature vector when the real-time state level is an early warning state or a fault state, and to store the real-time detection data, state assessment results and abnormal information in a preset database, thereby triggering the corresponding early warning system or fault alarm system.

[0107] In some embodiments, acquiring preprocessed multi-source data includes: for preprocessed temperature data, using a preset temperature feature extraction algorithm to extract temperature change trend features and abnormal temperature fluctuation features; for preprocessed vibration data, using a preset vibration feature extraction algorithm to extract time-domain features, frequency-domain features, and time-frequency-domain features of vibration signals; for preprocessed partial discharge data, using a preset discharge feature extraction algorithm to extract amplitude features, phase features, and discharge frequency features of discharge pulses; and for preprocessed current and voltage data, extracting amplitude features, waveform distortion features, and power factor features of current and voltage.

[0108] In some embodiments, the denoising and normalization of the state data to obtain preprocessed multi-source data and generate multi-dimensional feature vectors includes: denoising the temperature data using a moving average filtering algorithm to remove random noise and abnormal jump points; denoising the vibration data using a wavelet denoising algorithm to remove high-frequency noise through multi-layer wavelet decomposition and thresholding; denoising the partial discharge data using a pulse recognition algorithm to eliminate interference signals by combining the time-domain waveform characteristics of the discharge pulse; denoising the current and voltage data using a median filtering algorithm to remove periodic interference and spike noise; and normalizing the denoised data by normalizing the temperature data to a preset temperature range, and scaling the vibration data, partial discharge data, current data, and voltage data to a preset uniform numerical range according to their data types to eliminate dimensional differences between different sensor data.

[0109] In some embodiments, the multi-dimensional feature vector includes temperature features, vibration features, partial discharge features, and current-voltage features corresponding to temperature data, vibration data, partial discharge data, current data, and voltage data, respectively. The step of inputting the multi-dimensional feature vector into a preset hierarchical fusion model and performing feature layer fusion to obtain intermediate fusion features includes: concatenating temperature features, vibration features, partial discharge features, and current-voltage features according to feature dimensions to form an initial fusion vector containing each type of feature; performing dimensionality reduction on the initial fusion vector using principal component analysis, retaining principal components whose cumulative variance contribution rate exceeds a preset threshold, removing redundant features, and generating intermediate fusion features containing key information; or, using a weighted fusion method, assigning initial weights to temperature features, vibration features, partial discharge features, and current-voltage features based on historical reliability data of each type of sensor, and generating intermediate fusion features through weighted summation.

[0110] In some embodiments, the step of combining historical operation data and historical fault data of the switchgear to perform decision-level weighted fusion of intermediate fusion features to generate a comprehensive state feature vector includes: extracting historical operation data of the target switchgear and historical fault data of similar switchgear from a preset database; cleaning and extracting features from the historical data to construct a historical feature dataset; using an association rule algorithm to analyze the correlation between each dimension feature and the fault type in the historical feature data; calculating the weight coefficient of each dimension feature in the decision-level fusion based on the correlation; performing a weighted operation on the intermediate fusion features and the weight coefficients; and correcting the weighted result by combining the probability distribution of fault occurrence in the historical data to generate a comprehensive state feature vector containing historical experience information.

[0111] In some embodiments, the state assessment model is constructed based on a deep learning algorithm and trained using historical sample data, including: constructing a state assessment model using a multi-layer neural network structure, wherein the neural network includes at least an input layer, several hidden layers, and an output layer, the number of neurons in the input layer matches the dimension of the comprehensive state feature vector, and the number of neurons in the output layer matches the number of categories of real-time state levels; dividing the historical sample data into a training set and a validation set, wherein the training set includes sample data of normal state, warning state, and fault state and their corresponding labels; and training the state assessment model using a backpropagation algorithm, with the state level prediction accuracy of the sample data as the optimization objective, and adjusting the weights and bias parameters of the neural network to make the prediction error of the model on the validation set less than a preset threshold.

[0112] In some embodiments, processing the comprehensive state feature vector through the state evaluation model to output the real-time state level of the switchgear includes: inputting the comprehensive state feature vector into the input layer of the state evaluation model, performing feature transformation sequentially through the activation functions of each hidden layer, generating probability values ​​for each state level in the output layer; setting a preset state level judgment threshold; outputting a normal state when the probability value of a normal state is greater than a first threshold; outputting a warning state when the probability value of a warning state is greater than a second threshold and higher than the probability values ​​of other states; and outputting a fault state when the probability value of a fault state is greater than a third threshold and higher than the probability values ​​of other states, wherein the first threshold is greater than the second threshold, and the second threshold is greater than the third threshold.

[0113] In some embodiments, when the real-time status level is a warning state or a fault state, generating corresponding warning information or fault type judgment results based on the degree of abnormality of each dimension feature in the comprehensive status feature vector includes: establishing a preset mapping rule library between abnormal features and warning levels and fault types, wherein the mapping rule library contains the correspondence between the abnormal threshold, abnormal duration and abnormal type of each dimension feature; comparing each dimension feature in the comprehensive status feature vector with the threshold in the mapping rule library, calculating the degree of deviation and duration of the feature value from the normal range; matching the corresponding warning level or fault type according to the degree of deviation and duration; if the temperature feature exceeds the high temperature threshold multiple times consecutively, generating a temperature abnormality warning; if the partial discharge frequency feature exceeds the fault threshold, combining the phase feature to match the preset discharge fault type, generating a specific fault type judgment result.

[0114] In some embodiments, storing real-time detection data, status assessment results, and anomaly information in a preset database to trigger a corresponding early warning system or fault alarm system includes: structuring and storing real-time detected temperature data, vibration data, partial discharge data, current data, and voltage data according to timestamps and switchgear numbers, and storing corresponding multi-dimensional feature vectors, comprehensive status feature vectors, and status assessment results; when the real-time status level is an early warning state, calling the corresponding early warning interface according to preset early warning rules to send a message containing anomaly characteristics and an early warning level to the monitoring platform; when the real-time status level is a fault state, triggering a local audible and visual alarm device and a remote fault alarm system, sending emergency alarm information containing fault type judgment results and fault location, and recording the alarm time and processing status to the database.

[0115] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the intelligent switchgear status detection system and its modules described above based on multi-sensor fusion can be referred to the corresponding processes in the embodiments of the intelligent switchgear status detection method based on multi-sensor fusion described above, and will not be repeated here.

[0116] The aforementioned intelligent detection method for switchgear status based on multi-sensor fusion can be implemented as a computer program, which can be used in, for example... Figure 2 It runs on the system shown.

[0117] Please see Figure 3 , Figure 3 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0118] The storage medium can store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any intelligent detection method for switchgear status based on multi-sensor fusion.

[0119] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0120] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any intelligent detection method for switch cabinet status based on multi-sensor fusion.

[0121] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 3The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0122] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0123] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: The status data of the switch cabinet is obtained by temperature sensors, vibration sensors, partial discharge sensors, current sensors and voltage sensors arranged in the switch cabinet. The status data includes at least temperature data, vibration data, partial discharge data, current data and voltage data. The acquired status data is input into a preset data preprocessing module to perform noise reduction and normalization processing on the status data, thereby obtaining preprocessed multi-source data and generating multi-dimensional feature vectors. The multi-dimensional feature vectors are then input into a preset hierarchical fusion model to perform feature layer fusion on the multi-dimensional feature vectors, thereby obtaining intermediate fusion features. Combined with the historical operation data and historical fault data of the switchgear, the intermediate fusion features are then subjected to decision-level weighted fusion to generate a comprehensive status feature vector. The comprehensive state feature vector is input into the pre-trained state assessment model, which is built based on a deep learning algorithm and trained with historical sample data. The comprehensive state feature vector is processed by the state assessment model to output the real-time state level of the switchgear. The real-time state level includes normal state, warning state and fault state. When the real-time status level is a warning state or a fault state, the corresponding warning information or fault type judgment result is generated based on the degree of abnormality of each dimension of the comprehensive status feature vector. The real-time detection data, status assessment results and abnormal information are stored in the preset database, triggering the corresponding warning system or fault alarm system.

[0124] In some embodiments, acquiring preprocessed multi-source data includes: for preprocessed temperature data, using a preset temperature feature extraction algorithm to extract temperature change trend features and abnormal temperature fluctuation features; for preprocessed vibration data, using a preset vibration feature extraction algorithm to extract time-domain features, frequency-domain features, and time-frequency-domain features of vibration signals; for preprocessed partial discharge data, using a preset discharge feature extraction algorithm to extract amplitude features, phase features, and discharge frequency features of discharge pulses; and for preprocessed current and voltage data, extracting amplitude features, waveform distortion features, and power factor features of current and voltage.

[0125] In some embodiments, the denoising and normalization of the state data to obtain preprocessed multi-source data and generate multi-dimensional feature vectors includes: denoising the temperature data using a moving average filtering algorithm to remove random noise and abnormal jump points; denoising the vibration data using a wavelet denoising algorithm to remove high-frequency noise through multi-layer wavelet decomposition and thresholding; denoising the partial discharge data using a pulse recognition algorithm to eliminate interference signals by combining the time-domain waveform characteristics of the discharge pulse; denoising the current and voltage data using a median filtering algorithm to remove periodic interference and spike noise; and normalizing the denoised data by normalizing the temperature data to a preset temperature range, and scaling the vibration data, partial discharge data, current data, and voltage data to a preset uniform numerical range according to their data types to eliminate dimensional differences between different sensor data.

[0126] In some embodiments, the multi-dimensional feature vector includes temperature features, vibration features, partial discharge features, and current-voltage features corresponding to temperature data, vibration data, partial discharge data, current data, and voltage data, respectively. The step of inputting the multi-dimensional feature vector into a preset hierarchical fusion model and performing feature layer fusion to obtain intermediate fusion features includes: concatenating temperature features, vibration features, partial discharge features, and current-voltage features according to feature dimensions to form an initial fusion vector containing each type of feature; performing dimensionality reduction on the initial fusion vector using principal component analysis, retaining principal components whose cumulative variance contribution rate exceeds a preset threshold, removing redundant features, and generating intermediate fusion features containing key information; or, using a weighted fusion method, assigning initial weights to temperature features, vibration features, partial discharge features, and current-voltage features based on historical reliability data of each type of sensor, and generating intermediate fusion features through weighted summation.

[0127] In some embodiments, the step of combining historical operation data and historical fault data of the switchgear to perform decision-level weighted fusion of intermediate fusion features to generate a comprehensive state feature vector includes: extracting historical operation data of the target switchgear and historical fault data of similar switchgear from a preset database; cleaning and extracting features from the historical data to construct a historical feature dataset; using an association rule algorithm to analyze the correlation between each dimension feature and the fault type in the historical feature data; calculating the weight coefficient of each dimension feature in the decision-level fusion based on the correlation; performing a weighted operation on the intermediate fusion features and the weight coefficients; and correcting the weighted result by combining the probability distribution of fault occurrence in the historical data to generate a comprehensive state feature vector containing historical experience information.

[0128] In some embodiments, the state assessment model is constructed based on a deep learning algorithm and trained using historical sample data, including: constructing a state assessment model using a multi-layer neural network structure, wherein the neural network includes at least an input layer, several hidden layers, and an output layer, the number of neurons in the input layer matches the dimension of the comprehensive state feature vector, and the number of neurons in the output layer matches the number of categories of real-time state levels; dividing the historical sample data into a training set and a validation set, wherein the training set includes sample data of normal state, warning state, and fault state and their corresponding labels; and training the state assessment model using a backpropagation algorithm, with the state level prediction accuracy of the sample data as the optimization objective, and adjusting the weights and bias parameters of the neural network to make the prediction error of the model on the validation set less than a preset threshold.

[0129] In some embodiments, processing the comprehensive state feature vector through the state evaluation model to output the real-time state level of the switchgear includes: inputting the comprehensive state feature vector into the input layer of the state evaluation model, performing feature transformation sequentially through the activation functions of each hidden layer, generating probability values ​​for each state level in the output layer; setting a preset state level judgment threshold; outputting a normal state when the probability value of a normal state is greater than a first threshold; outputting a warning state when the probability value of a warning state is greater than a second threshold and higher than the probability values ​​of other states; and outputting a fault state when the probability value of a fault state is greater than a third threshold and higher than the probability values ​​of other states, wherein the first threshold is greater than the second threshold, and the second threshold is greater than the third threshold.

[0130] In some embodiments, when the real-time status level is a warning state or a fault state, generating corresponding warning information or fault type judgment results based on the degree of abnormality of each dimension feature in the comprehensive status feature vector includes: establishing a preset mapping rule library between abnormal features and warning levels and fault types, wherein the mapping rule library contains the correspondence between the abnormal threshold, abnormal duration and abnormal type of each dimension feature; comparing each dimension feature in the comprehensive status feature vector with the threshold in the mapping rule library, calculating the degree of deviation and duration of the feature value from the normal range; matching the corresponding warning level or fault type according to the degree of deviation and duration; if the temperature feature exceeds the high temperature threshold multiple times consecutively, generating a temperature abnormality warning; if the partial discharge frequency feature exceeds the fault threshold, combining the phase feature to match the preset discharge fault type, generating a specific fault type judgment result.

[0131] In some embodiments, storing real-time detection data, status assessment results, and anomaly information in a preset database to trigger a corresponding early warning system or fault alarm system includes: structuring and storing real-time detected temperature data, vibration data, partial discharge data, current data, and voltage data according to timestamps and switchgear numbers, and storing corresponding multi-dimensional feature vectors, comprehensive status feature vectors, and status assessment results; when the real-time status level is an early warning state, calling the corresponding early warning interface according to preset early warning rules to send a message containing anomaly characteristics and an early warning level to the monitoring platform; when the real-time status level is a fault state, triggering a local audible and visual alarm device and a remote fault alarm system, sending emergency alarm information containing fault type judgment results and fault location, and recording the alarm time and processing status to the database.

[0132] The embodiments of this application also provide a computer-readable storage medium storing a computer program, the computer program including program instructions, and the processor executing the program instructions to implement the steps of the intelligent detection method for switch cabinet status based on multi-sensor fusion provided in the above embodiments of this application.

[0133] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0134] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for intelligent detection of switchgear status based on multi-sensor fusion, characterized in that, include: The status data of the switch cabinet is obtained by temperature sensors, vibration sensors, partial discharge sensors, current sensors and voltage sensors arranged in the switch cabinet. The status data includes at least temperature data, vibration data, partial discharge data, current data and voltage data. The acquired status data is input into a preset data preprocessing module to perform noise reduction and normalization processing on the status data, thereby obtaining preprocessed multi-source data and generating multi-dimensional feature vectors. The multi-dimensional feature vectors are then input into a preset hierarchical fusion model to perform feature layer fusion on the multi-dimensional feature vectors, thereby obtaining intermediate fusion features. Combined with the historical operation data and historical fault data of the switchgear, the intermediate fusion features are then subjected to decision layer weighted fusion to generate a comprehensive status feature vector. The comprehensive state feature vector is input into the pre-trained state assessment model, which is built based on a deep learning algorithm and trained with historical sample data. The comprehensive state feature vector is processed by the state assessment model to output the real-time state level of the switchgear. The real-time state level includes normal state, warning state and fault state. When the real-time status level is a warning state or a fault state, the corresponding warning information or fault type judgment result is generated based on the degree of abnormality of each dimension of the comprehensive status feature vector. The real-time detection data, status assessment results and abnormal information are stored in the preset database, triggering the corresponding warning system or fault alarm system.

2. The method according to claim 1, characterized in that, The acquisition of preprocessed multi-source data includes: For the preprocessed temperature data, a preset temperature feature extraction algorithm is used to extract temperature change trend features and abnormal temperature fluctuation features; For the preprocessed vibration data, a preset vibration feature extraction algorithm is used to extract the time-domain features, frequency-domain features, and time-frequency-domain features of the vibration signal; For the preprocessed partial discharge data, a preset discharge feature extraction algorithm is used to extract the amplitude features, phase features, and discharge frequency features of the discharge pulse; For the preprocessed current and voltage data, the amplitude characteristics, waveform distortion characteristics, and power factor characteristics of the current and voltage are extracted.

3. The method according to claim 1, characterized in that, The process of denoising and normalizing the state data to obtain preprocessed multi-source data and generating multi-dimensional feature vectors includes: The temperature data is denoised using a moving average filtering algorithm to remove random noise and anomalous jump points; the vibration data is denoised using a wavelet denoising algorithm, which removes high-frequency noise through multi-layer wavelet decomposition and thresholding. A pulse recognition algorithm is used to identify partial discharge data and eliminate interference signals by combining the time-domain waveform characteristics of the discharge pulse. Median filtering algorithm is used on current and voltage data to remove periodic interference and spike noise; After denoising, the various types of data are normalized. Temperature data is normalized to a preset temperature range, and vibration data, partial discharge data, current data, and voltage data are scaled to a preset uniform numerical range according to their data types to eliminate the dimensional differences between different sensor data.

4. The method according to claim 1, characterized in that, The multi-dimensional feature vector includes temperature features, vibration features, partial discharge features, and current / voltage features corresponding to temperature data, vibration data, partial discharge data, current data, and voltage data, respectively. The step of inputting the multi-dimensional feature vector into a preset hierarchical fusion model to perform feature layer fusion and obtain intermediate fused features includes: Temperature features, vibration features, partial discharge features, and current and voltage features are concatenated and spliced ​​according to feature dimensions to form an initial fusion vector containing each type of feature; Principal component analysis is used to reduce the dimensionality of the initial fusion vector. Principal components with a cumulative variance contribution rate exceeding a preset threshold are retained, redundant features are removed, and intermediate fusion features containing key information are generated. Alternatively, a weighted fusion method is used, and initial weights are assigned to temperature features, vibration features, partial discharge features, and current and voltage features based on historical reliability data of various types of sensors. Intermediate fusion features are generated by weighted summation.

5. The method according to claim 1, characterized in that, The process involves combining historical operating data and historical fault data of the switchgear, performing decision-level weighted fusion of intermediate fusion features, and generating a comprehensive state feature vector, including: Historical operation data of the target switchgear and historical fault data of similar switchgear are extracted from the preset database. The historical data is cleaned and features are extracted to construct a historical feature dataset. The association rule algorithm is used to analyze the correlation between the features of each dimension and the fault type in historical feature data, and the weight coefficient of each dimension feature in the decision-making layer fusion is calculated based on the correlation. The intermediate fusion features and weight coefficients are weighted and the weighted result is corrected by combining the probability distribution of fault occurrence in historical data, thus generating a comprehensive state feature vector containing historical experience information.

6. The method according to claim 1, characterized in that, The state assessment model is built based on a deep learning algorithm and trained using historical sample data, including: A state evaluation model is constructed using a multi-layer neural network structure. The neural network includes at least an input layer, several hidden layers, and an output layer. The number of neurons in the input layer matches the dimension of the comprehensive state feature vector, and the number of neurons in the output layer matches the number of categories of the real-time state level. The historical sample data is divided into a training set and a validation set. The training set contains sample data of normal state, warning state, and fault state, as well as their corresponding labels. The backpropagation algorithm is used to train the state assessment model, with the state level prediction accuracy of the sample data as the optimization objective. By adjusting the weights and bias parameters of the neural network, the prediction error of the model on the validation set is made less than a preset threshold.

7. The method according to claim 1, characterized in that, The process of processing the comprehensive state feature vector through the state assessment model to output the real-time state level of the switchgear includes: The comprehensive state feature vector is input into the input layer of the state evaluation model, and the features are transformed sequentially through the activation functions of each hidden layer. The probability values ​​of each state level are generated in the output layer. A preset threshold for determining the state level is used. When the probability value of the normal state is greater than the first threshold, the normal state is output. When the probability value of the warning state is greater than the second threshold and higher than the probability values ​​of other states, the warning state is output. When the probability value of a fault state is greater than the third threshold and higher than the probability values ​​of other states, the fault state is output, where the first threshold is greater than the second threshold, and the second threshold is greater than the third threshold.

8. The method according to claim 1, characterized in that, When the real-time status level is a warning state or a fault state, the corresponding warning information or fault type judgment result is generated based on the degree of anomaly of each dimension of the comprehensive status feature vector, including: Establish a pre-defined mapping rule base for abnormal features, warning levels, and fault types. The mapping rule base includes the correspondence between abnormal thresholds, abnormal durations, and abnormal types for each dimension of features. By comparing the features of each dimension in the comprehensive state feature vector with the thresholds in the mapping rule base, the degree and duration of deviation of the feature values ​​from the normal range are calculated. Based on the degree of deviation and duration, the corresponding warning level or fault type is matched. If the temperature feature exceeds the high temperature threshold multiple times in a row, a temperature abnormality warning is generated. If the partial discharge frequency feature exceeds the fault threshold, the preset discharge fault type is matched with the phase feature to generate a specific fault type judgment result.

9. The method according to claim 1, characterized in that, The step of storing real-time detection data, status assessment results, and anomaly information into a preset database to trigger the corresponding early warning system or fault alarm system includes: The real-time temperature data, vibration data, partial discharge data, current data, and voltage data are stored in a structured manner according to the timestamp and switch cabinet number, and the corresponding multi-dimensional feature vector, comprehensive state feature vector, and state assessment results are stored. When the real-time status level is in the warning state, the corresponding warning interface is called according to the preset warning rules to send a message containing abnormal characteristics and warning level to the monitoring platform; When the real-time status level is a fault state, the local audible and visual alarm device and the remote fault alarm system are triggered, sending emergency alarm information containing the fault type judgment result and the location of the fault, and recording the alarm time and processing status to the database.

10. A smart switchgear status detection system based on multi-sensor fusion, characterized in that, include: The data acquisition unit is used to acquire the status data of the switch cabinet through temperature sensors, vibration sensors, partial discharge sensors, current sensors and voltage sensors arranged in the switch cabinet. The status data includes at least temperature data, vibration data, partial discharge data, current data and voltage data. The vector generation unit is used to input the acquired state data into a preset data preprocessing module, perform noise reduction and normalization on the state data, obtain preprocessed multi-source data, and generate multi-dimensional feature vectors; input the multi-dimensional feature vectors into a preset hierarchical fusion model, perform feature layer fusion on the multi-dimensional feature vectors, obtain intermediate fusion features, and combine the historical operation data and historical fault data of the switchgear to perform decision layer weighted fusion on the intermediate fusion features to generate a comprehensive state feature vector; The level output unit is used to input the comprehensive state feature vector into the pre-trained state assessment model. The state assessment model is built based on a deep learning algorithm and trained with historical sample data. The comprehensive state feature vector is processed by the state assessment model to output the real-time state level of the switchgear. The real-time state level includes normal state, warning state and fault state. The result generation unit is used to generate corresponding early warning information or fault type judgment results based on the degree of abnormality of each dimension feature in the comprehensive state feature vector when the real-time state level is an early warning state or a fault state. It also stores the real-time detection data, state assessment results and abnormal information in a preset database and triggers the corresponding early warning system or fault alarm system.

Images