A switch cabinet fault detection method based on multi-modal data fusion

Abstract

This application relates to power system fault detection technology, specifically to a switchgear fault detection method based on multimodal data fusion. It aims to address the issue that in actual operation, switchgear faults can arise from various factors such as electrical overload, insulation degradation, mechanical component wear, or abnormal temperature. Failure to identify and address these faults in a timely manner can easily lead to power outages or even equipment damage. Traditional methods, lacking collaborative analysis of multi-source information, often fail to comprehensively reflect the overall health status of the switchgear, resulting in frequent false alarms. For example, normal load fluctuations may be misjudged as faults, or weak signals may cause missed detections in the early stages of a real fault. This invention effectively solves the problems of false alarms and missed detections caused by the reliance on single parameters in traditional methods by real-time acquisition of multi-source operating data, preprocessing, feature extraction, multimodal fusion, and deep learning diagnostics, combined with early warning and location steps. It offers the advantages of comprehensive, accurate, and timely detection of switchgear faults.

Description

Technical Field

[0001] This application relates to power system fault detection technology, and more specifically, to a switchgear fault detection method based on multimodal data fusion. Background Technology

[0002] As a critical power distribution device in the power system, the operating status of switchgear directly determines the safety, stability, and continuity of power supply. In actual operation, switchgear may malfunction due to various factors such as electrical overload, insulation degradation, mechanical component wear, or abnormal temperature. If these malfunctions are not identified and addressed in a timely manner, they can easily lead to power outages or even equipment damage. Traditional fault detection methods have long relied on monitoring a single parameter, such as collecting only three-phase current and voltage data to analyze electrical characteristics, relying solely on partial discharge signals to assess insulation conditions, or monitoring hotspot areas only through temperature sensors.

[0003] This single-dimensional data acquisition method has fundamental flaws: focusing solely on electrical parameters fails to capture potential risks arising from abnormal mechanical characteristics; relying solely on temperature data may overlook the development trend of early partial discharge; and monitoring mechanical parameters alone makes it difficult to correlate with the dynamic process of electrical changes. Due to the lack of collaborative analysis of multi-source information, traditional methods often fail to comprehensively reflect the overall health status of the switchgear, leading to frequent false alarms, such as misjudging normal load fluctuations as faults or missing early-stage faults due to weak signals. More seriously, single-parameter detection struggles to provide early warning of faults, typically triggering alarms only when the fault has caused significant damage, delaying optimal handling and increasing the risk of system collapse.

[0004] Existing technologies attempt to improve detection capabilities by introducing intelligent algorithms. Some solutions employ graph neural networks to construct signal node association models to enhance the temporal consistency of multi-source data and utilize long short-term memory networks to extract sequence features. However, their design focus is biased towards inspection path planning and abnormal signal identification, failing to effectively integrate switchgear structural characteristics and sensor layout information, resulting in insufficient fault location accuracy and limited response capability to early, weak fault signals. Other solutions focus on high-voltage switchgear scenarios, using conditional reversible likelihood modeling combined with dual-baseline alarm mechanisms for risk classification. However, such methods have poor adaptability to medium and low-voltage switchgear and cannot cover the common characteristics of equipment at different voltage levels. Furthermore, these technologies generally suffer from excessively high algorithmic complexity, requiring substantial computational resources, resulting in high implementation costs and hindering widespread application in resource-constrained field environments.

[0005] Overall, the current technology system has significant shortcomings in terms of data fusion depth, early warning timeliness, positioning accuracy, and universality, and cannot meet the comprehensive needs of modern power systems for real-time, accurate, and economical fault detection.

[0006] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0007] (a) Technical problems to be solved The purpose of this application is to provide a switchgear fault detection method based on multimodal data fusion, which has the advantages of comprehensive, accurate and timely detection of switchgear faults.

[0008] (II) Technical Solution The present invention provides a switchgear fault detection method based on multimodal data fusion, comprising the following steps: Data acquisition: Real-time acquisition of multi-source operational data from the switchgear via sensors; Data preprocessing: Preprocessing the collected multi-source data; Feature extraction: Extracting key features from preprocessed data; Multimodal data fusion: The features of various parameters are fused and a deep learning model is used for fault diagnosis to obtain the fault type and severity of the switchgear; Fault warning and location: Based on the fault diagnosis results, timely fault warning information is issued and the fault location is accurately located.

[0009] Furthermore, this application also proposes a switchgear fault detection system based on multimodal data fusion, comprising: Data acquisition module: used to acquire multi-source operating data of the switchgear in real time; Data preprocessing module: used to preprocess the acquired data; Feature extraction module: used to extract key features from preprocessed data; Multimodal data fusion module: used to fuse the features of various parameters and use a deep learning model for fault diagnosis; Fault warning and location module: Used to issue fault warning information based on fault diagnosis results and accurately locate the fault location.

[0010] Furthermore, this application also proposes an electronic device comprising: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the aforementioned switchgear fault detection method based on multimodal data fusion; or, a non-transitory machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the aforementioned switchgear fault detection method based on multimodal data fusion.

[0011] (III) Beneficial Effects Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively solves the problems of false alarms and missed detections caused by the reliance on a single parameter in traditional methods by real-time acquisition of multi-source operating data, preprocessing, feature extraction, multimodal fusion and deep learning diagnosis, combined with early warning and localization steps. It has the advantages of comprehensive, accurate and timely detection of switchgear faults. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the logic structure of a switchgear fault detection method; Figure 2 This is a schematic diagram of the framework structure of the switchgear fault detection method. Detailed Implementation

[0014] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0015] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0016] Traditional switchgear fault detection methods primarily rely on single-parameter monitoring, resulting in low detection accuracy, high false alarm rates, and an inability to provide early warnings. Furthermore, existing technologies lack the capability for multi-source data fusion and analysis, failing to fully utilize the correlation information between various parameters, leading to reduced accuracy and reliability in fault diagnosis. In addition, some methods have limited applicability or high algorithmic complexity, hindering large-scale application.

[0017] In this regard, such as Figures 1-2This application proposes a switchgear fault detection method based on multimodal data fusion, comprising the following steps: S100, Data Acquisition: Real-time acquisition of multi-source operating data of the switchgear through sensors; S200, Data Preprocessing: Preprocess the collected multi-source data; S300, Feature Extraction: Extracting key features from preprocessed data; S400, Multimodal Data Fusion: The features of various parameters are fused, and a deep learning model is used for fault diagnosis to obtain the fault type and severity of the switchgear; S500, Fault Warning and Location: Based on the fault diagnosis results, timely fault warning information is issued, and the fault location is accurately located.

[0018] For ease of understanding, the following explains some key terms in this embodiment: Switchgear refers to power distribution equipment used in power systems for controlling, protecting, and monitoring circuits. It typically consists of circuit breakers, disconnectors, load switches, and instrument transformers. Its operating status directly affects the safety and stability of the power system.

[0019] Multi-source operational data refers to various types of data reflecting the operating status of the switchgear, acquired from different monitoring points or through different types of sensors. These data include electrical parameters, temperature, partial discharge signals, and mechanical characteristic parameters. These data describe the equipment's operating status from different dimensions.

[0020] A sensor is a device that can sense and detect physical quantities and convert them into measurable electrical signals. In this application, the sensor is used to acquire real-time operating data of the switchgear.

[0021] Data preprocessing involves cleaning, denoising, format conversion, and normalization of the raw collected data to eliminate redundancy, noise, and inconsistencies, improve data quality, and make it suitable for subsequent feature extraction and model training.

[0022] Feature extraction involves identifying and quantifying key information from preprocessed data that effectively characterizes the operating status or failure modes of switchgear. These features are high-level abstract representations of the data, helping to reduce data dimensionality and highlight failure modes.

[0023] Multimodal data fusion integrates features extracted from data of different modalities to form a more comprehensive and discriminative feature representation. Through fusion, the complementarity and redundancy between different modalities can be utilized to improve the accuracy of fault diagnosis.

[0024] Deep learning models are a class of machine learning models based on artificial neural networks. They possess multi-layered nonlinear transformation structures and are capable of automatically learning complex feature representations and patterns from large amounts of data. In this application, a deep learning model is used for fault diagnosis of fused multimodal data.

[0025] Fault diagnosis involves determining whether a switchgear has a fault based on collected and analyzed operational data, and identifying the type, location, and severity of the fault.

[0026] Fault early warning is to issue alarm information in advance based on the diagnosis results before the switch cabinet failure occurs or in the early stage of the failure, so that timely intervention measures can be taken to avoid the failure from escalating or causing serious accidents.

[0027] Fault location is the process of determining the specific physical location of a fault, whether inside or outside the switchgear, so that maintenance personnel can quickly and accurately carry out repairs.

[0028] The switchgear fault detection method of this application first involves data acquisition. Data acquisition can be achieved in various ways. For example, it can be done through manual inspection and recording, where staff periodically observe and record the operating parameters of the switchgear. Alternatively, basic analog sensors, such as thermistors or current transformers, can be deployed to periodically sample the temperature or current of the switchgear, and the acquired data can be stored in local storage media.

[0029] Subsequently, the collected multi-source data undergoes preprocessing. The raw data may contain noise, missing values, or inconsistent formats. Data preprocessing may include simple mean filtering to smooth the data, using linear interpolation to fill in missing values, or converting data of different formats into a standardized tabular form to facilitate subsequent processing.

[0030] Next, key features are extracted from the preprocessed data. Feature extraction involves selecting and calculating statistical characteristics from the data, such as the maximum, minimum, average, and variance. For data with time-series characteristics, periodic or trend features can be extracted to reflect the data's changing patterns over time.

[0031] Furthermore, multimodal data fusion is performed, fusing the features of various parameters, and a deep learning model is used for fault diagnosis to obtain the fault type and severity of the switchgear. The extracted features from different modalities can be easily concatenated to form a unified feature vector. This feature vector is then input into a deep learning model, such as a multilayer perceptron, which learns the mapping relationship between the features and the switchgear fault type through model training, thereby outputting the diagnostic results of the fault type and severity.

[0032] In the multimodal data fusion step, this application adopts a self-attention mechanism based on the query-key-value mechanism, which allows the model to automatically calculate weights based on the correlation between features when fusing features of different modalities. The deep learning model adopts a combination structure of 3-5 layers of convolutional neural network and 1-2 layers of long short-term memory (LSTM) network. The convolutional neural network is used to extract spatial distribution features such as amplitude, frequency and phase distribution of partial discharge signal, as well as local key features such as temperature data change trend and gradient features. The LSTM network is used to capture the dynamic evolution and trend of time series data such as electrical parameters (such as three-phase current, three-phase voltage and power factor) and mechanical characteristic parameters (such as opening and closing time, opening and closing speed and opening and closing stroke).

[0033] Finally, based on the fault diagnosis results, timely fault warning information is issued, and the fault location is accurately pinpointed. When the fault type or severity output by the fault diagnosis model reaches a preset threshold, an early warning mechanism can be triggered, such as issuing an alarm via audible and visual alarm devices, or notifying maintenance personnel via SMS, email, or other means. Fault location can be initially determined by reviewing the switchgear equipment ledger or combining it with the experience of maintenance personnel to identify the area where the fault may occur.

[0034] This application effectively solves the problems of low detection accuracy, high false alarm rate, insufficient early warning capability, and poor data fusion caused by single-parameter monitoring in traditional switchgear fault detection by integrating the entire process of multi-source data acquisition, preprocessing, feature extraction, multimodal fusion diagnosis, and fault early warning and location. As a result, it achieves comprehensive perception and in-depth analysis of the switchgear's operating status, improves the accuracy of fault type identification and severity determination, and can promptly issue early warning information and accurately locate fault positions, thereby enhancing the safety and reliability of switchgear operation.

[0035] In some of the embodiments described above in this application, multi-source operating data is proposed to comprehensively collect the operating status of the switchgear. However, in this process, since the specific data type is not clearly defined, the data collection may be incomplete or inconsistent, and it may not be able to cover all potential fault points of the switchgear, thereby affecting the accuracy of subsequent multimodal data fusion and the reliability of fault diagnosis, and exacerbating the limitations of traditional single parameter detection.

[0036] In this regard, this application further clarifies that the multi-source operating data includes electrical parameters, partial discharge signals, temperature data, and mechanical characteristic parameters.

[0037] Specifically, the electrical parameters refer to electrical quantities that reflect the operating status of the switchgear circuit, such as current, voltage, and power. Their function is to monitor the electrical load, insulation status, and the presence of abnormalities such as short circuits and overloads in real time. In practical applications, these parameters can be collected using sensors such as current transformers (CTs) and voltage transformers (PTs), or obtained through smart meters integrated within the switchgear.

[0038] The partial discharge signal refers to the electromagnetic or acoustic signal generated by partial discharge phenomena occurring in the insulating medium. Partial discharge is an important early sign of insulation degradation in switchgear, and monitoring its signal can effectively provide early warning of potential insulation faults. This signal can be acquired using ultra-high frequency (UHF) sensors, extra-high frequency (TEV) sensors, or acoustic sensors, which can capture discharge pulses or acoustic waves at different frequency bands.

[0039] The temperature data refers to the temperature values ​​of key components inside the switchgear (such as contacts, busbars, cable joints, etc.). Abnormal temperatures are often a direct manifestation of faults such as overload, poor contact, or inadequate heat dissipation. Continuous monitoring of temperature changes can promptly detect potential overheating issues and prevent the fault from escalating. Temperature data can be acquired through non-contact measurement using an infrared thermal imager, or through contact measurement using platinum resistance temperature sensors (RTDs), thermocouples, or fiber optic temperature sensors installed in critical locations.

[0040] The mechanical characteristic parameters refer to the mechanical motion characteristics exhibited by the operating mechanisms of circuit breakers, disconnectors, and other devices within the switchgear during the opening and closing process. These parameters include, but are not limited to, opening and closing time, opening and closing speed, overtravel, and bounce. Monitoring these mechanical characteristic parameters helps assess the health of the operating mechanisms and detect mechanical faults such as jamming, wear, and spring failure. Data acquisition is typically achieved through displacement sensors (such as linear displacement sensors), velocity sensors, or high-speed camera systems combined with dedicated analysis software.

[0041] Through the aforementioned technical solution, this application clearly defines the specific types of multi-source operational data, ensuring the comprehensiveness and relevance of data acquisition. Electrical parameters, partial discharge signals, temperature data, and mechanical characteristic parameters comprehensively cover the operating status of the switchgear from four dimensions: electrical, insulation, thermal, and mechanical, effectively avoiding the limitations of traditional single-parameter detection. This multi-dimensional, high-precision raw data input provides a rich and reliable information source for subsequent multi-modal data fusion steps, greatly improving the accuracy and reliability of fault diagnosis. Given the guaranteed integrity and consistency of data acquisition, the deep learning model can more effectively learn and identify the characteristics of different fault modes, thereby achieving accurate judgment of the fault type and severity of the switchgear, significantly enhancing early warning capabilities, and reducing the risk of false alarms and missed alarms.

[0042] In some of the embodiments described above in this application, multi-source operational data is proposed to comprehensively collect the operating status of switchgear. However, in the implementation process, due to the lack of specific definition of data content, the collected data may be incomplete or inaccurate, affecting the accuracy of subsequent feature extraction and fusion, thereby reducing the reliability of fault diagnosis and early warning capability.

[0043] In this regard, this application further proposes that the multi-source operating data includes electrical parameters, partial discharge signals, temperature data, and mechanical characteristic parameters. Specifically, the electrical parameters include three-phase current, three-phase voltage, and power factor; the partial discharge signals include amplitude, frequency, and phase distribution; the temperature data includes temperature values ​​of key parts of the switchgear; and the mechanical characteristic parameters include opening and closing time, opening and closing speed, and opening and closing stroke.

[0044] Specifically, the three-phase current and three-phase voltage among the electrical parameters are core parameters reflecting the basic operating status of the switchgear electrical circuit. Three-phase current is used to monitor the load condition of each phase and determine whether there are abnormalities such as overload, short circuit, or three-phase imbalance; three-phase voltage is used to monitor power supply quality and determine whether there are problems such as undervoltage, overvoltage, or voltage imbalance. This can be achieved by real-time sampling using current transformers and voltage transformers, converting analog signals into digital signals for processing; or by using integrated intelligent sensors to directly output digital current and voltage values. Power factor is an important indicator for measuring the operating efficiency of electrical equipment, reflecting the relationship between active power and apparent power. An abnormal power factor may indicate problems such as excessive reactive power loss, insulation aging, or harmonic pollution. This can be achieved by directly measuring the power factor using a power factor meter installed in the circuit; or by calculating it from the collected three-phase current and three-phase voltage data combined with the phase difference.

[0045] The amplitude of the partial discharge signal represents the intensity of the discharge, usually measured in picocoulombs (pC). A larger amplitude indicates higher discharge energy and potentially more severe insulation defects. This can be achieved by capturing the electromagnetic wave signals generated by partial discharge using ultra-high frequency (UHF) or extra-high frequency (TEV) sensors, amplifying and quantizing them; or by detecting ultrasonic signals generated by partial discharge using acoustic sensors and converting them into electrical signals for amplitude analysis. Frequency represents the number of discharges per unit time; changes in frequency can reflect the development trend of insulation defects. This can be achieved by counting and statistically analyzing the acquired partial discharge pulse signals; or by setting a threshold and recording and analyzing the frequency of pulses exceeding the threshold. Phase distribution refers to the position of the partial discharge pulse within the power frequency voltage cycle. Different discharge types exhibit different characteristics in phase distribution, which helps distinguish fault types. This can be achieved by synchronously acquiring the partial discharge signal and power frequency voltage, plotting a phase-resolved partial discharge (PRPD) spectrum for analysis; or by using digital signal processing technology to calculate the phase angle of the discharge pulse relative to the voltage zero-crossing point.

[0046] The temperature data includes temperature values ​​for critical components of the switchgear. These critical components typically refer to areas prone to heat generation or sensitive to temperature, such as contacts, busbar connection points, circuit breaker bodies, cable joints, and transformer windings. Abnormal temperature increases in these areas are often early signs of overload, poor contact, or insulation aging. This can be achieved by installing resistance temperature detectors (RTDs) or thermocouples at these critical components for contact-based temperature measurement; or by using infrared thermal imagers for non-contact inspection to obtain the surface temperature distribution of these critical components.

[0047] The opening and closing time in the mechanical characteristic parameters refers to the time required for the circuit breaker in the switchgear to completely separate / close its contacts from receiving the opening / closing command. Too long or too short a time may indicate a fault in the operating mechanism. This can be achieved by installing a travel sensor and a time relay on the circuit breaker operating mechanism to accurately record the action time; or by monitoring the current or voltage waveform of the operating coil and combining it with the state changes of the auxiliary contact switch to estimate the time. The opening and closing speed refers to the speed at which the moving contact of the circuit breaker moves during the opening / closing process. Abnormal speed may indicate insufficient power in the operating mechanism, excessive friction, or failure of the buffer device. This can be achieved by installing a displacement sensor on the moving contact of the circuit breaker to measure the contact displacement in real time and performing differential calculations on the displacement data to obtain the speed; or by recording the contact movement process with a high-speed camera and then performing image analysis to obtain speed information. The opening and closing travel refers to the distance the moving contact of the circuit breaker moves during the opening / closing process. Insufficient or excessive travel may lead to poor contact, reduced arc extinguishing capability, or mechanism jamming. This can be achieved by installing a linear displacement sensor on the moving contact of the circuit breaker for precise measurement; or by using a mechanical limit switch in combination with auxiliary contact signals to indirectly determine whether the travel is complete.

[0048] Through the above technical solutions, this application provides a detailed and comprehensive definition of multi-source operating data, concretizing abstract data categories into directly measurable physical quantities. Specifically, electrical parameters are defined as three-phase current, three-phase voltage, and power factor, enabling precise capture of electrical fault characteristics such as overload, short circuit, three-phase imbalance, and abnormal operating efficiency in switchgear. Partial discharge signals are refined into amplitude, frequency, and phase distribution, allowing for a comprehensive characterization of the intensity, occurrence pattern, and type of insulation defects, significantly improving the ability to identify early insulation hazards. Temperature data is limited to temperature values ​​of key parts of the switchgear, making thermal fault monitoring more targeted and enabling timely detection of problems such as contact overheating and poor connections. Mechanical characteristic parameters are specified to opening and closing time, opening and closing speed, and opening and closing stroke, enabling accurate diagnosis of mechanical fault modes such as jamming, wear, and abnormal displacement of the operating mechanism. This refined definition of data sources fundamentally solves the problem of incomplete or inaccurate data acquisition, ensuring the comprehensiveness, accuracy, and directionality of input data, and providing high-quality basic data for subsequent data preprocessing, feature extraction, and multimodal data fusion. Therefore, this application can significantly improve the accuracy and reliability of fault diagnosis, enhance the ability to identify weak and early faults, thereby achieving more timely and effective fault warning and location, effectively avoiding false alarms and missed alarms caused by data ambiguity, and ensuring the safe and stable operation of the switchgear.

[0049] In some of the embodiments described above in this application, data preprocessing and feature extraction steps are proposed to prepare multi-source data for multimodal fusion and fault diagnosis. However, in this process, due to the lack of specific noise removal methods, the data may be distorted by noise interference. At the same time, feature extraction is not optimized for different types of data, and the extracted features may be insufficient or irrelevant, thereby affecting the accuracy of subsequent fusion and diagnosis and failing to accurately reflect the potential fault status of the switchgear.

[0050] In response, this application further proposes to use an adaptive filtering algorithm to remove noise interference and retain useful signals in the data preprocessing step; and to extract the amplitude, frequency and phase distribution features of the partial discharge signal and the trend and gradient features of the temperature data in the feature extraction step.

[0051] Specifically, adaptive filtering algorithms are digital signal processing techniques that automatically adjust filter parameters based on the statistical characteristics of the input signal. Their core lies in using a feedback mechanism to enable the filter to effectively suppress noise and retain useful signals even in unknown or time-varying noise environments. For example, Kalman filtering can be used, which estimates the signal state and filters out noise in real time by combining a system dynamic model and the statistical characteristics of measured noise through prediction and updating stages; or the Least Mean Square (LMS) algorithm can be used, which iteratively adjusts the filter coefficients to minimize the mean square error between its output and the desired signal, thereby achieving noise suppression. Furthermore, the Recursive Least Squares (RLS) algorithm is also an effective adaptive filtering method, introducing a forgetting factor on top of LMS, resulting in better tracking performance for time-varying systems. For partial discharge signals, extracting their amplitude, frequency, and phase distribution characteristics is crucial for comprehensively characterizing discharge activity. Amplitude characteristics can include the peak amplitude, average amplitude, or maximum amplitude of the discharge pulse; these parameters directly reflect the intensity and energy of the discharge. Frequency characteristics refer to the number of discharge pulses occurring per unit time, revealing the activity level and occurrence patterns of the discharge. Phase distribution characteristics are typically represented by phase-resolved partial discharge (PRPD) maps. These maps statistically analyze the distribution of discharge pulses within the power frequency cycle, revealing the type of discharge (e.g., internal discharge, surface discharge, corona discharge) and its development trend, providing crucial information for determining the nature of insulation defects. For switchgear temperature data, extracting its trend and gradient characteristics provides a deeper understanding of the equipment's thermal state. Trend characteristics can be described using methods such as moving averages, exponential smoothing, or linear regression slopes to depict the direction and pattern of temperature changes over time, such as continuous rise, stabilization, or decline. Gradient characteristics reflect the rate and acceleration of temperature change by calculating the first or second difference of temperature. For example, the first difference indicates the rate of temperature rise or fall, while the second difference reveals the variation in the rate of temperature change. These characteristics are more effective than simple instantaneous temperature values ​​in identifying potential overheating risks or cooling anomalies.

[0052] Through the above technical solutions, this application effectively solves the problems of feature distortion caused by noise interference and insufficient or irrelevant feature extraction during data preprocessing. Specifically, the adaptive filtering algorithm can dynamically adjust the filtering parameters according to the complex electromagnetic environment and signal characteristics of the switchgear, efficiently eliminating various noise interferences while completely preserving weak fault signals and abrupt change features. This ensures the cleanliness and effectiveness of the input data from the source, avoiding the useful signal attenuation or feature distortion that may be caused by traditional fixed filtering methods. In terms of feature extraction, amplitude, frequency, and phase distribution features are specifically extracted for partial discharge signals, which can comprehensively characterize the discharge intensity, occurrence pattern, and discharge type of insulation defects, providing highly identifiable features for the identification of early insulation faults. At the same time, trend and gradient features are specifically extracted for temperature data, which can accurately capture the rate of abnormal temperature changes and local overheating trends, reflecting potential thermal fault risks earlier than simple temperature values. This combined design of "adaptive denoising and differentiated feature extraction" makes the data preprocessing and feature extraction stages highly adaptable to the characteristics of multi-source heterogeneous data from switchgear, significantly improving the completeness and relevance of effective features. This provides high-quality and highly reliable input for subsequent multimodal data fusion and deep learning models for fault diagnosis, thereby significantly improving the accuracy of switchgear fault detection and early warning capabilities.

[0053] In some of the solutions mentioned above in this application, a multimodal data fusion step is proposed to fuse multi-source data and perform fault diagnosis. However, in this process, due to the lack of an effective feature weight allocation mechanism, the fusion effect may be unsatisfactory, and the importance of different parameter features cannot be fully captured, resulting in a reduction in the accuracy and reliability of fault diagnosis, thereby affecting the early warning and accurate positioning capabilities.

[0054] In response, this application further proposes to employ an attention mechanism in the multimodal data fusion step to automatically learn the importance weights of different parameter features and improve the fusion effect; the deep learning model includes convolutional neural networks and recurrent neural networks.

[0055] Attention mechanisms are techniques that simulate human cognitive attention. Their core idea is to enable models to dynamically focus on more important parts of input data and assign them higher weights. In implementation, various forms can be adopted. For example, self-attention based on a query-key-value mechanism allows the model to automatically calculate weights based on the correlation between features when fusing features from different modalities; or channel attention can be used to weight different feature channels based on their importance, highlighting feature channels that contribute more to fault diagnosis. By introducing attention mechanisms, this application can solve the problem of the lack of effective feature weight allocation mechanisms in multimodal data fusion, enabling the model to adaptively focus on features more critical to fault diagnosis.

[0056] Convolutional Neural Networks (CNNs) are deep learning models particularly adept at processing data with local correlations. They extract local features from input data through convolutional kernels and effectively reduce model complexity by utilizing weight sharing and pooling operations. Implementation methods can include one-dimensional convolutional neural networks (1D CNNs), suitable for processing time-series data, such as raw waveforms of electrical or mechanical parameters, to capture their local temporal patterns; or two-dimensional convolutional neural networks (2D CNNs), suitable for processing transformed two-dimensional data, such as using time-frequency maps of partial discharge signals or temperature field distribution maps as input to extract their spatial features. In this application, convolutional neural networks are primarily used to efficiently extract key features with locality and spatial distribution characteristics from multimodal data.

[0057] Recurrent Neural Networks (RNNs) are deep learning models specifically designed for processing sequential data. Their unique recurrent structure enables them to capture temporal dependencies and long-term patterns in the data. Common implementations include Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), which effectively address the gradient vanishing or exploding problems that traditional RNNs may encounter when processing long sequences by introducing gating mechanisms. In this application, RNNs are primarily used to process data with temporal characteristics, such as electrical parameters (three-phase current, three-phase voltage, and power factor) and mechanical characteristic parameters (opening and closing time, opening and closing speed, and opening and closing stroke), to capture the dynamic patterns and trends of these parameters over time.

[0058] Through the above technical solution, this application effectively solves the problems of poor fusion effect, low diagnostic accuracy, and weak early warning capability caused by unreasonable feature weight allocation, weakening of key fault features, and insufficient spatiotemporal feature mining capability of deep learning models during multimodal data fusion. Specifically, an attention mechanism is introduced into the multimodal data fusion step, enabling the model to automatically learn and adaptively allocate importance weights for various features such as electrical parameters, partial discharge signals, temperature data, and mechanical characteristic parameters. This allows for the highlighting of sensitive features highly correlated with switchgear faults during fusion, while suppressing redundant or irrelevant information. This fundamentally solves the defects of fixed weights and inability to focus on key fault information in traditional fusion methods, significantly improving the effectiveness and relevance of feature fusion.

[0059] Meanwhile, the deep learning model is explicitly defined as a combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs), fully leveraging the advantages of both. CNNs can efficiently mine spatial distribution features of partial discharge signals, such as amplitude, frequency, and phase distribution, as well as key local features like temperature data trends and gradient characteristics. RNNs, on the other hand, can accurately capture the dynamic evolution and trend patterns of time-series data, including electrical parameters (such as three-phase current, three-phase voltage, and power factor) and mechanical characteristic parameters (such as opening and closing time, opening and closing speed, and opening and closing stroke). Working together, these two technologies achieve comprehensive extraction and deep correlation modeling of the spatiotemporal features of multi-source heterogeneous data, significantly enhancing the accuracy and robustness of fault type identification and severity determination.

[0060] This integrated fault diagnosis mechanism, combining "adaptive weight fusion and deep spatiotemporal feature mining," is highly compatible with the multimodal data characteristics of switchgear, effectively improving the ability to identify minor and early-stage faults, and enhancing the reliability of fault diagnosis and early warning effects. By more accurately identifying fault types and severity, this application provides more reliable input for subsequent fault warning and location steps, thereby improving the overall performance of the switchgear fault detection method.

[0061] In some of the embodiments described above in this application, a fault warning and location step is proposed to issue a fault warning message and locate the fault location based on the fault diagnosis result. However, in its implementation, the location accuracy may be insufficient, and it is impossible to effectively combine the structural characteristics of the switch cabinet and the sensor layout to achieve accurate location, which leads to untimely fault handling or increased risk of misjudgment.

[0062] In response, this application further proposes to use a model-based fault location algorithm in the fault early warning and location steps, which combines the structural characteristics of the switchgear and the sensor layout to achieve accurate fault location.

[0063] Specifically, the model-based fault location algorithm refers to simulating the signal propagation, energy dissipation, or state change patterns during a fault by establishing physical or data models of the components inside the switchgear. For example, electrical, thermal, or mechanical physical models of the switchgear can be constructed, and the impact of different fault types and locations on sensor data can be predicted through finite element analysis, circuit simulation, and other methods. Alternatively, historical fault data can be used to train machine learning models (such as support vector machines, neural networks, decision trees, etc.) to learn the mapping relationship between fault characteristics and fault locations. These models provide a theoretical basis and computational framework for fault location, avoiding the limitations of traditional methods that rely solely on empirical judgment or single-parameter threshold determination, thereby improving the accuracy and reliability of fault location.

[0064] Simultaneously, during fault location, the inherent characteristics of the switchgear, such as its actual physical structure, internal component layout, connection methods, material properties, and fault propagation paths, are fully considered—that is, the structural features of the switchgear are taken into account. For example, a 3D CAD model or detailed structural drawings of the switchgear can be imported into the location system as input parameters for the algorithm, enabling the algorithm to understand the specific components where a fault may occur (such as circuit breakers, disconnectors, busbars, cable joints, etc.) and their spatial locations. Alternatively, a physical connection diagram or fault propagation network between the various components within the switchgear can be established to analyze the propagation path and attenuation characteristics of fault signals. By integrating this structural information, the location algorithm can more accurately correlate detected abnormal signals with specific physical locations, enhancing the accuracy of the location.

[0065] Furthermore, during fault location, information such as the actual installation location, quantity, type, and monitoring range of sensors within the switchgear is fully utilized, i.e., sensor layout is considered. For example, the precise spatial coordinates, monitoring area, and physical quantities that each sensor can sense (such as temperature, partial discharge signals, vibration, etc.) can be accurately mapped to the structural model of the switchgear, forming a sensor network topology. The location algorithm can weight the sensor data based on the differences in sensitivity of different sensors to fault signals at different locations, or utilize the spatial distribution characteristics of multiple sensor data for triangulation or multi-point location. This combined approach allows the location algorithm to fully utilize the spatial information of sensor data, improving the accuracy and robustness of fault location.

[0066] Through the above technical solution, this application effectively solves the problems of insufficient fault location accuracy and inability to effectively combine the structural characteristics of switchgear and sensor layout in existing technologies. Specifically, by adopting a model-based fault location algorithm, a standardized location model can be established based on the fault transmission mechanism and multimodal characteristic correlation of switchgear, avoiding the errors and randomness caused by the reliance on single signals and empirical judgments in traditional location methods, and improving the rigor and stability of the location logic. At the same time, by combining the algorithm with the structural characteristics of the switchgear itself and the actual installation depth of the sensors, the spatial distribution and physical correspondence of the sensor data are fully utilized, making the fault location calculation more consistent with the actual layout of the equipment, and enabling rapid convergence to specific components and specific points, significantly narrowing the scope of investigation. This three-in-one precise location mechanism of "mechanism model + physical structure + sensor layout" effectively improves the fault point location accuracy and response speed, reduces the risk of misjudgment and missed location, and provides reliable support for rapid repair and shortening power outage time. It forms a synergistic effect with the overall multimodal fault detection solution, further enhancing the engineering practicality and field implementation capability of the solution.

[0067] The following example will provide a more detailed explanation of the above technical solution: In a power substation, a switchgear fault detection system based on multimodal data fusion has been deployed to monitor the operating status of critical switchgear. This system aims to overcome the limitations of traditional single-parameter detection, improve early warning capabilities, enhance data fusion effects, and achieve accurate fault location.

[0068] First, during the data acquisition phase, the system acquires multi-source operational data from the switchgear in real time using various sensors. For example, current transformers and voltage transformers continuously collect electrical parameters such as three-phase current, three-phase voltage, and power factor; partial discharge sensors monitor partial discharge signals inside the switchgear, including their amplitude, frequency, and phase distribution; temperature sensors acquire real-time temperature data at key locations in the switchgear, such as busbar connections and circuit breaker contacts; and mechanical characteristic sensors record mechanical characteristic parameters such as opening and closing time, opening and closing speed, and opening and closing stroke. This comprehensive acquisition of multi-source data effectively compensates for the inadequacy of a single parameter in fully reflecting the switchgear's operating status.

[0069] Next, the data preprocessing stage begins. Since the collected raw data may contain various noises and interferences, the system employs an adaptive filtering algorithm to process this multi-source data. For example, the adaptive filtering algorithm can effectively remove power frequency interference and random noise in partial discharge signals while retaining useful discharge pulse signals. The algorithm can also smooth out instantaneous fluctuations in temperature data, ensuring the data quality for subsequent analysis. Through data cleaning, denoising, and normalization steps, high-quality input is provided for subsequent feature extraction and fault diagnosis.

[0070] In the feature extraction stage, the system extracts key features from the preprocessed data. Specifically, for partial discharge signals, the system extracts their amplitude, frequency, and phase distribution features, which reflect the intensity, activity, and type of the discharge. For temperature data, the system analyzes its changing trends and gradient characteristics to identify abnormal temperature rise patterns. For electrical parameters, features such as harmonic content and imbalance can be extracted. For mechanical characteristic parameters, their deviations from normal operating values ​​are extracted. The extraction of these features transforms the raw time-series data into a more diagnostically meaningful representation.

[0071] The next stage is multimodal data fusion and fault diagnosis. The system fuses features extracted from different modalities. During this process, the system employs an attention mechanism to automatically learn and assign importance weights to features with different parameters. For example, when a local discharge signal becomes abnormal, the attention mechanism assigns higher weights to the local discharge feature; when the temperature continues to rise, the weight of the temperature feature increases accordingly. This dynamic weight allocation mechanism effectively improves the data fusion effect and overcomes the problem of poor data fusion capabilities in traditional methods. The fused features are then input into a deep learning model for fault diagnosis. This model combines the advantages of convolutional neural networks (CNNs) and recurrent neural networks (RNNs), where CNNs are used to capture the local spatial correlation of features, and RNNs are used to handle the temporal dependencies of features. Through learning from a large amount of historical fault data, the model can identify complex fault modes and output the fault type (e.g., insulation degradation, contact overheating, mechanical jamming, etc.) and severity of the switchgear. For example, when the model detects a continuous increase in the amplitude and frequency of local discharge, accompanied by a slow increase in local temperature, it can diagnose an early insulation degradation fault.

[0072] Finally, in the fault warning and location phase, the system promptly issues fault warning information based on the diagnostic results of the deep learning model. For example, when insulation degradation is detected to be in its initial stage, the system will issue a "Level 1 warning" to alert maintenance personnel. If the degradation worsens, it may be upgraded to a "Level 2 warning," recommending maintenance. Simultaneously, the system employs a model-based fault location algorithm, combining the structural characteristics of the switchgear and sensor layout information to accurately locate the fault. For instance, by analyzing the response differences of different temperature sensors and partial discharge sensors, the system can precisely pinpoint the possible location of the fault, such as the "circuit breaker arc-extinguishing chamber" or "bus connection point A." This early warning and precise location capability significantly outperforms the limitations of traditional methods that only detect faults after they occur, and it also solves the problem of insufficient fault location accuracy in existing technologies, providing maintenance personnel with timely and accurate decision-making support, thereby effectively preventing potential power accidents.

[0073] Example 1

[0074] Traditional switchgear fault detection methods primarily rely on single-parameter monitoring, resulting in low detection accuracy, high false alarm rates, and an inability to provide early warnings. Furthermore, existing technologies lack the capability for multi-source data fusion and analysis, failing to fully utilize the correlation information between various parameters, leading to reduced accuracy and reliability in fault diagnosis. In addition, some methods have limited applicability or high algorithmic complexity, hindering large-scale application.

[0075] To address this issue, this application proposes a switchgear fault detection system based on multimodal data fusion. This system, through modular design, achieves comprehensive acquisition, efficient processing, feature extraction, intelligent fusion, and accurate early warning of multi-source data, effectively overcoming the shortcomings of traditional methods such as reliance on single parameters, delayed early warning, data isolation, and limited applicability. Specifically, the data acquisition module is used to collect multi-source operating data of the switchgear in real time. By deploying devices such as temperature sensors, current transformers, and partial discharge detectors, it simultaneously acquires multi-dimensional data such as electrical parameters, temperature distribution, and mechanical vibration, avoiding the one-sidedness of single-parameter detection and providing a data foundation for comprehensive fault analysis. The data preprocessing module is used to preprocess the acquired data, performing data cleaning, noise reduction, and normalization operations. For example, it uses moving average filtering to eliminate signal noise, linear interpolation to fill missing data points, and converts data of different dimensions into a unified standard range, ensuring that subsequent processing is based on high-quality input and improving the stability and reliability of the system.

[0076] The feature extraction module extracts key features from preprocessed data, calculating statistical features such as mean, variance, and peak factor for different data types, or extracting time-frequency domain features such as wavelet coefficients to quantitatively characterize the switchgear's operating status and avoid interference from invalid features in diagnosis. The multimodal data fusion module fuses various parameter features and uses a deep learning model for fault diagnosis. First, the extracted feature vectors are concatenated or weighted to form a comprehensive feature representation. Then, the deep learning model automatically learns the complex relationships between features and outputs diagnostic results indicating the fault type and severity, solving the problem of poor data fusion capabilities. The fault warning and location module issues fault warning information based on the fault diagnosis results and accurately locates the fault. When the diagnostic result indicates that the fault probability exceeds a preset threshold, a multi-level warning mechanism is triggered, such as audible and visual alarms or remote notifications. Simultaneously, by combining the switchgear's topology and historical data, a spatial mapping algorithm is used to determine the specific location of the fault, achieving early warning and precise location.

[0077] Through the above technical solution, this application effectively integrates the entire process of multi-source data acquisition, preprocessing, feature extraction, intelligent fusion, and early warning positioning, achieving comprehensive perception and in-depth analysis of the switchgear's operating status. This significantly improves the accuracy of fault type identification and severity determination, enabling timely issuance of early warning information and precise fault location, thereby enhancing the safety and reliability of switchgear operation, meeting the applicability requirements of switchgear at different voltage levels, and reducing algorithm complexity, facilitating large-scale application.

[0078] In some of the solutions mentioned above in this application, a data acquisition module is proposed to collect multi-source operating data, and a data preprocessing module is proposed to preprocess the data. However, in this process, data acquisition may not be able to fully cover all the key parameter types required for switchgear operation, such as electrical parameters, partial discharge signals, temperature data, and mechanical characteristic parameters, resulting in incomplete data. At the same time, the preprocessing process may lack targeted processing units and cannot effectively remove noise interference, clean invalid data, or standardize data formats, thereby affecting the accuracy and reliability of subsequent feature extraction and fault diagnosis.

[0079] In this regard, this application further proposes that the data acquisition module includes a current transformer, a voltage transformer, a partial discharge sensor, a temperature sensor, and a mechanical characteristic sensor; and the data preprocessing module includes a data cleaning unit, a noise reduction unit, and a normalization unit.

[0080] Specifically, current transformers and voltage transformers are key devices for accurately measuring the electrical parameters of switchgear. Current transformers proportionally convert high current to low current for use in measurement and protection circuits; their implementation can include, but is not limited to, clamp-on current transformers, through-type current transformers, or open-type current transformers. Voltage transformers proportionally convert high voltage to low voltage for use by measuring instruments and relay protection devices; their implementation can include, but is not limited to, electromagnetic voltage transformers, capacitive voltage transformers, or photoelectric voltage transformers. Through these transformers, the system can obtain electrical parameters such as three-phase current, three-phase voltage, and power factor during switchgear operation in real time and accurately. Partial discharge sensors are used to detect partial discharge signals generated by insulation defects inside the switchgear. Partial discharge is an important indicator of insulation degradation, and its signal characteristics are crucial for early detection of potential faults. Partial discharge sensors can be implemented in ways including, but not limited to, ultrasonic sensors that detect ultrasonic signals generated by partial discharge; or ultra-high frequency (UHF) sensors that detect electromagnetic wave signals generated by partial discharge. These sensors can capture key information such as the amplitude, frequency, and phase distribution of partial discharges. Temperature sensors are used to monitor the temperature of critical parts of the switchgear in real time. Abnormal temperatures are a direct manifestation of faults such as overheating and poor contact in the switchgear. Temperature sensors can be implemented in various ways, including but not limited to resistance temperature detectors (RTDs), such as Pt100, which measure resistance as a function of temperature; or thermocouples, which measure temperature difference electromotive force; or infrared temperature sensors for non-contact temperature measurement. These sensors can provide accurate temperature data for various points inside the switchgear, used to analyze temperature change trends and gradient characteristics. Mechanical characteristic sensors are used to monitor the mechanical motion state of the switchgear operating mechanism. The mechanical characteristic parameters of the switchgear, such as opening and closing time, speed, and stroke, are important indicators for evaluating its mechanical performance and reliability. Mechanical characteristic sensors can be implemented in various ways, including but not limited to displacement sensors, such as linear grating rulers or magnetostrictive displacement sensors, for measuring opening and closing strokes; or speed sensors, such as encoders or laser tachometers, for measuring opening and closing speeds.

[0081] The data cleaning unit in the data preprocessing module is used to identify and process missing values, outliers, or inconsistent data in the collected data. Its role is to ensure the integrity and accuracy of the data, preventing dirty data from interfering with subsequent analysis. Data cleaning can be implemented using methods including, but not limited to, statistical methods such as mean imputation, median imputation, or regression imputation of missing values; or machine learning methods such as using Isolation Forest or Local Outlier Factor (LOF) algorithms to detect and process outliers. The denoising unit removes various noise interferences from the collected data, retaining useful signal components. Noise can mask the true signal, reducing the accuracy of feature extraction and fault diagnosis. Denoising can be implemented using methods including, but not limited to, wavelet denoising, which decomposes the signal into different frequency components and removes noise through wavelet transform; or Kalman filtering, which estimates the true state of the signal through prediction and correction; or adaptive filtering algorithms that dynamically adjust filtering parameters according to signal characteristics. The normalization unit unifies data of different dimensions and ranges into a standardized interval. This helps eliminate the dimensional differences between different parameters, making them comparable in subsequent feature fusion and deep learning models, and preventing certain features with large numerical ranges from dominating model training. Normalization can be implemented in ways including but not limited to Min-Max normalization, which linearly maps the data to the [0,1] interval; or Z-score standardization, which transforms the data into a distribution with a mean of 0 and a standard deviation of 1.

[0082] Through the above technical solutions, the fault detection system of this application, in the data acquisition stage, can comprehensively and multidimensionally acquire the electrical parameters, partial discharge signals, temperature data, and mechanical characteristic parameters required for switchgear operation in real time by configuring current transformers, voltage transformers, partial discharge sensors, temperature sensors, and mechanical characteristic sensors. This comprehensive coverage of multi-source data effectively overcomes the limitations of traditional single-parameter detection, providing a rich and complete data foundation for subsequent fault diagnosis. In the data preprocessing stage, the data cleaning unit can effectively identify and process missing and abnormal data, ensuring data integrity and reliability; the denoising unit can accurately remove various noise interferences, allowing useful signals to be clearly presented; the normalization unit eliminates the dimensional differences between different types of data, making all data comparable in feature extraction and multimodal fusion, avoiding data bias. These refined data acquisition and preprocessing measures significantly improve the quality of input data, thus laying a solid foundation for the subsequent feature extraction module to accurately extract key features, the multimodal data fusion module to efficiently fuse various parameter features, and the use of deep learning models for fault diagnosis, ultimately improving the accuracy, reliability, and early warning capability of switchgear fault detection.

[0083] In some of the embodiments described above in this application, a feature extraction module, a multimodal data fusion module, and a fault early warning and location module are proposed to realize switchgear fault detection. However, in the implementation process, due to the lack of clear internal unit division, it may lead to low data processing efficiency, incomplete feature extraction, poor multi-source data fusion effect, and insufficient fault location accuracy, thereby affecting the accuracy and timeliness of overall fault diagnosis.

[0084] In this regard, this application further proposes that the feature extraction module includes a signal processing unit and a machine learning unit; the multimodal data fusion module includes a feature fusion unit and a deep learning model unit; and the fault warning and location module includes a warning unit and a fault location unit.

[0085] Specifically, the signal processing unit is the core component of the feature extraction module. It is primarily responsible for the initial processing and optimization of the raw, multi-source operational data to remove noise, enhance useful signals, and provide a high-quality data foundation for subsequent feature extraction. For example, digital filters can be used to perform frequency domain filtering to remove high-frequency or low-frequency noise; alternatively, time-frequency analysis methods such as wavelet transform and empirical mode decomposition can be used to decompose and reconstruct the signal to separate noise components from the effective signal. The machine learning unit is another key component of the feature extraction module. Its main function is to automatically learn and extract key features related to the switchgear fault modes from the signal-processed data. These features are the direct basis for fault diagnosis. For example, dimensionality reduction algorithms such as principal component analysis and linear discriminant analysis can be used to extract the most representative features from high-dimensional data; alternatively, traditional machine learning models such as support vector machines and decision trees can be used to learn the inherent patterns in the data through training, thereby extracting classification or regression features.

[0086] Furthermore, the feature fusion unit is the core of the multimodal data fusion module. Its responsibility is to effectively integrate heterogeneous features from different modalities after feature extraction, forming a unified and more discriminative feature vector. For example, cascade fusion can be used, directly concatenating feature vectors from different modalities to form a longer feature vector; alternatively, parallel fusion can be used, combining features from different modalities through weighted averaging, multiplicative fusion, or other methods. The deep learning model unit is another key component of the multimodal data fusion module. It receives the unified feature vector output by the feature fusion unit and utilizes the powerful capabilities of deep learning to identify and classify fault modes, thereby outputting the fault type and severity of the switchgear. For example, convolutional neural networks can be used, which are particularly suitable for processing features with local correlations; or recurrent neural networks can be used, which are adept at processing time-dependent sequence data.

[0087] Furthermore, the early warning unit is a crucial component of the fault early warning and location module. Its main function is to determine the existence of potential faults or the occurrence of faults based on the fault diagnosis results output by the deep learning model unit, and to promptly generate and issue early warning information. For example, a fault threshold can be set, triggering an early warning when the diagnostic results exceed the preset threshold; alternatively, a trend prediction algorithm can be used, combining historical data and current diagnostic results to predict the fault development trend and issue an early warning before the fault reaches a critical state. The fault location unit is another key component of the fault early warning and location module. Its goal is to accurately pinpoint the specific location of the fault after an early warning occurs, combining the physical structure of the switchgear and sensor layout information. For example, a model-based fault location algorithm can be used, establishing a physical model of the switchgear and inverting the fault source location using sensor data; alternatively, a machine learning-based location method can be used, training a model to learn the mapping relationship between fault characteristics and fault location.

[0088] Through the above technical solutions, the signal processing unit in the feature extraction module can perform preliminary processing on the raw data, remove noise, and enhance useful signals, providing high-quality input for the machine learning unit. The machine learning unit then automatically learns and extracts key features from the processed data, ensuring the comprehensiveness and adaptability of feature extraction and avoiding feature omissions or inaccuracies caused by a single method. The feature fusion unit in the multimodal data fusion module effectively integrates features from different modalities to form a unified and more discriminative feature vector, while the deep learning model unit uses these fused features for advanced diagnosis, thereby improving the accuracy and robustness of data correlation analysis. The early warning unit in the fault warning and location module can issue alarms in a timely manner based on diagnostic results, while the fault location unit combines the switch cabinet structure and sensor layout to accurately determine the location, which significantly enhances early warning capabilities and location accuracy. Overall, by meticulously dividing each functional module into units, this application optimizes the system execution process, improves the reliability and efficiency of fault detection, and thus effectively improves the accuracy and timeliness of overall fault diagnosis.

[0089] Example 2

[0090] In some of the solutions mentioned above in this application, a switchgear fault detection method based on multimodal data fusion is proposed to solve the problems of limitations of traditional single parameter detection, insufficient early warning capability and poor data fusion capability. However, when implementing this method, there is a lack of efficient and dedicated execution equipment or storage media, which leads to a complex and inefficient implementation process, making it difficult to quickly deploy and reliably operate in practical applications, and increasing system cost and maintenance difficulty.

[0091] In response, this application proposes an electronic device comprising: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the aforementioned switchgear fault detection method based on multimodal data fusion; or, a non-transitory machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the aforementioned switchgear fault detection method based on multimodal data fusion.

[0092] This electronic device is a hardware platform for carrying and executing the fault detection method. It can be an industrial control computer (IPC) with high reliability and anti-interference capabilities, suitable for industrial environments; it can also be an embedded system, small in size and low in power consumption, directly integrated into switchgear or monitoring units; or it can be a high-performance server for centrally processing fault detection tasks of multiple switchgears. At least one processor is the core computing unit of the electronic device, responsible for parsing and executing instructions, performing data operations, and logic control. This processor can be a general-purpose central processing unit (CPU), such as a processor based on x86 or ARM architecture, providing powerful general-purpose computing capabilities; it can also be a dedicated digital signal processor (DSP), adept at handling complex algorithms in signal acquisition and feature extraction; or it can be a microcontroller (MCU), suitable for control tasks with high real-time requirements and relatively low computational complexity. The memory is used to store the instruction code of the fault detection method and various data generated during the execution of the method. The memory may include random access memory (RAM) for temporary storage of variables and intermediate results during program execution, enabling fast read and write operations; or it may include non-volatile memory (such as flash memory, solid-state drives, or EEPROM) for persistent storage of the operating system, application code, and trained deep learning models, ensuring data integrity even after power failure. Instructions are sequences of binary code that a processor can understand and execute, collectively constituting the software implementation of the switchgear fault detection method based on multimodal data fusion. These instructions can be compiled and linked machine code, executed directly by the processor; script code in an interpreted language (such as Python), running on the processor via an interpreter; or firmware programs burned into the embedded system's memory to implement low-level hardware control and method logic. Non-transitory machine-readable storage media is a physical carrier capable of permanently storing executable instructions without data loss after power failure. It can be a hard disk drive (HDD) or a solid-state drive (SSD), providing large-capacity, high-speed data storage and retrieval capabilities, suitable for storing complex deep learning models and large amounts of historical data; it can also be a USB flash drive or SD card, facilitating program deployment, updates, and migration; or it can be an optical disc (such as a CD-ROM or DVD-ROM) for software distribution and backup. Executable instructions are program code stored on non-transitory machine-readable storage media that can be directly loaded and run by a machine (such as a processor in an electronic device). These instructions contain all the logic for implementing the aforementioned switchgear fault detection method based on multimodal data fusion, including steps such as data acquisition, preprocessing, feature extraction, multimodal data fusion, and fault early warning and location. When the machine executes these instructions, the entire fault detection process can be completed automatically.

[0093] Through the above technical solutions, this application provides a dedicated hardware and software carrier, enabling the aforementioned switchgear fault detection method based on multimodal data fusion to operate efficiently and stably. At least one processor in the electronic device can quickly respond to and execute complex fault detection algorithms, including inference from deep learning models, thereby shortening fault diagnosis time. Instructions stored in the memory ensure the integrity and consistency of the method, avoiding errors that may be introduced by manual operation and improving the degree of automation. Executable instructions stored in a non-transitory machine-readable storage medium provide a convenient and reliable way for the deployment and distribution of the method, reducing the difficulty of system integration and maintenance. This integrated hardware and software solution enables the switchgear fault detection method to be effectively implemented from the theoretical level to practical applications, significantly improving the real-time performance, accuracy, and reliability of fault detection, and providing a solid technical guarantee for the safe and stable operation of the power system.

[0094] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A switch cabinet fault detection method based on multi-modal data fusion, characterized in that, Includes the following steps: Real-time acquisition of multi-source operational data from the switchgear using sensors; Preprocess the collected multi-source data; Extract key features from the preprocessed data; The characteristics of each parameter are fused, and a deep learning model is used for fault diagnosis to obtain the fault type and severity of the switchgear. Based on the fault diagnosis results, a fault warning message is issued, and the fault location is accurately pinpointed.

2. The switchgear fault detection method based on multi-modal data fusion according to claim 1, characterized in that, The multi-source operating data includes electrical parameters, partial discharge signals, temperature data, and mechanical characteristic parameters.

3. The switchgear fault detection method based on multi-modal data fusion according to claim 2, characterized in that, The electrical parameters include: Three-phase current, three-phase voltage, and power factor; The partial discharge signal includes amplitude, frequency, and phase distribution; The temperature data includes the temperature values ​​of key parts of the switchgear; the mechanical characteristic parameters include opening and closing time, opening and closing speed, and opening and closing stroke.

4. The method for switchgear fault detection based on multi-modal data fusion according to claim 1, characterized in that, The data preprocessing step employs an adaptive filtering algorithm to remove noise interference and retain useful signals. In the feature extraction step, the amplitude, frequency and phase distribution features of the partial discharge signal are extracted, and the trend and gradient features of the temperature data are extracted.

5. The method for switchgear fault detection based on multi-modal data fusion according to claim 1, characterized in that, The multimodal data fusion step employs an attention mechanism to automatically learn the importance weights of different parameter features, thereby improving the fusion effect. The deep learning models include convolutional neural networks and recurrent neural networks.

6. The method for switchgear fault detection based on multi-modal data fusion according to claim 1, characterized in that, The fault warning and location step employs a model-based fault location algorithm, which, combined with the structural characteristics of the switchgear and the sensor layout, enables accurate location of the fault.

7. A switchgear fault detection system based on multi-modal data fusion, characterized in that, include: Data acquisition module: used to collect multi-source operating data of the switchgear in real time; Data preprocessing module: used to preprocess the collected data; Feature extraction module: used to extract key features from preprocessed data; Multimodal data fusion module: used to fuse the features of various parameters and use a deep learning model for fault diagnosis; Fault warning and location module: Used to issue fault warning information based on fault diagnosis results and accurately locate the fault location.

8. The switchgear fault detection system based on multi-modal data fusion as claimed in claim 7, wherein, The data acquisition module includes a current transformer, a voltage transformer, a partial discharge sensor, a temperature sensor, and a mechanical characteristic sensor. The data preprocessing module includes a data cleaning unit, a noise reduction unit, and a normalization unit.

9. The switchgear fault detection system based on multi-modal data fusion as claimed in claim 7, wherein, The feature extraction module includes a signal processing unit and a machine learning unit; The multimodal data fusion module includes a feature fusion unit and a deep learning model unit; The fault warning and location module includes a warning unit and a fault location unit.

10. A non-transitory machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the switchgear fault detection method based on multimodal data fusion as described in any one of claims 1-6.

Images