Switch equipment mechanical fault prediction system and method based on deep learning

Abstract

The invention discloses a switch equipment mechanical fault prediction system and method based on deep learning, and relates to the technical field of fault prediction. The system comprises a signal acquisition and preprocessing module, a multi-fault decoupling module, a layered fault identification module and a causal verification and positioning output module. According to the method, vibration and current signals are collected and preprocessed to remove interference, independent components of causal constraints are combined to analyze and decompose multi-fault mixed features, fault scenes are judged in a layered mode, fault categories are matched, fault reasonability is verified based on a causal relation graph, and fault severity is quantified and sorted. According to the invention, the problems of difficult multi-fault coupling identification and inaccurate positioning in the prior art are solved, the maintenance pertinence is improved, the processing time is shortened, and the stable operation of the switchgear is ensured.

Description

Technical Field

[0001] This invention relates to the field of fault prediction technology, specifically to a deep learning-based mechanical fault prediction system and method for switchgear. Background Technology

[0002] Switchgear is a core and critical piece of equipment in a power system, and the stable operation of its mechanical structure directly determines the reliability of power supply. In actual operation and maintenance scenarios, mechanical faults in switchgear often exhibit multi-fault coupling characteristics. For example, spring fatigue and shaft jamming, contact wear and operating mechanism jamming are prone to occur simultaneously, and the vibration, current and other characteristic signals of different faults superimpose and interfere with each other.

[0003] Existing deep learning-based mechanical fault prediction technologies for switchgear often employ a scheme where hybrid fault features are directly input into the model for identification, without designing a feature decoupling mechanism for multi-fault coupled scenarios. This approach has significant drawbacks: it cannot effectively separate individual fault feature components from the hybrid features, making it difficult for the model to distinguish specific fault types within coupled faults, and even more difficult to accurately locate multiple fault sources.

[0004] The aforementioned defects directly prevent maintenance personnel from developing targeted maintenance strategies based on the prediction results, which can easily lead to omissions or mis-repairs, prolonging fault handling time, increasing the risk of power outages, and severely restricting the engineering application of deep learning fault prediction technology in complex fault scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based system and method for predicting mechanical faults in switching equipment, in order to solve the problems mentioned in the background art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A deep learning-based method for predicting mechanical faults in switching equipment includes the following steps:

[0008] S1. Collect vibration and current signals during the operation of the switching equipment, and remove environmental vibration interference and electromagnetic interference noise through filtering to obtain the signal to be analyzed containing fault characteristics;

[0009] S2. By combining independent component analysis with causal decoupling logic for mechanical faults in switching equipment, the signal to be analyzed is decoupled. By setting causal constraints between fault features, the superimposed multi-fault mixed features are decomposed into independent feature components that correspond one-to-one with a single fault.

[0010] S3. Perform layered identification actions sequentially. The first layer is to count the number of independent feature components and determine whether the current fault is a multi-fault coupling scenario. The second layer is to extract the key characterization parameters of each independent feature component and match them with a preset single fault feature library to determine the specific fault category corresponding to each independent component.

[0011] S4. Call the pre-built causal relationship map of the mechanical system of the switching equipment, perform correlation verification on each fault category obtained in S3, verify the logical rationality of the fault evolution, eliminate false fault features caused by environmental interference, and finally output the multi-fault source localization result; the multi-fault source localization result includes each fault type, corresponding fault source component and fault severity ranking.

[0012] Furthermore, S1 includes the following:

[0013] Vibration sensors are deployed at key mechanical components of the switchgear, and current sensors are connected in series in the power supply circuit of the opening and closing coils. The sensors are connected to the data acquisition terminal through shielded cables. Sampling parameters are set according to the Nyquist sampling theorem and the frequency range of mechanical fault characteristics of the switchgear. The sampling duration covers the complete opening and closing cycle. Simultaneously, the sensor analog signals are collected and converted into digital signals for storage to obtain the original vibration signal sequence and the original current signal sequence.

[0014] The min-max normalization algorithm is used to map the amplitudes of the two types of original signals to the [0,1] interval, eliminating the amplitude imbalance caused by the difference in sensor range, and obtaining normalized vibration signal and normalized current signal.

[0015] Interference is removed by combining wavelet filtering and notch filtering, specifically as follows:

[0016] After decomposing the normalized vibration signal using wavelets, thresholding and inverse transformation are performed to remove environmental vibration interference. Spectral analysis is then conducted on the processed vibration signal and normalized current signal to locate the dominant interference frequency. A notch filter is used to specifically remove electromagnetic interference noise, and the resulting signal is then integrated to obtain the signal to be analyzed.

[0017] Furthermore, S2 includes the following:

[0018] The signal to be analyzed is reconstructed into a multi-channel signal matrix to meet the input requirements of the decoupling algorithm;

[0019] Based on the mechanical structure principle of switching equipment, the causal relationships between fault characteristics are sorted out, and causal direction constraints and feature uniqueness constraints are set, which are transformed into quantitative causal constraint conditions; the quantitative causal constraint conditions include causal correlation index threshold constraints and feature similarity threshold constraints.

[0020] An improved independent component analysis objective function incorporating the aforementioned causal constraints is constructed, and the separation matrix is ​​iteratively updated using the natural gradient descent method until the objective function converges.

[0021] Substituting the signal to be analyzed into the optimal separation matrix yields the initial decoupled components. Invalid components are eliminated, and the causal constraint satisfaction is verified. Finally, independent feature components corresponding one-to-one with a single fault are obtained.

[0022] Furthermore, S3 includes the following:

[0023] The number of independent feature components is counted. When the number of components is 1, it is determined to be a single fault scenario. When the number of components is not less than 2, it is determined to be a multi-fault coupled scenario.

[0024] For each independent feature component, extract key representation parameters in multiple dimensions in the time domain, frequency domain, and time-frequency domain, and construct the feature vector corresponding to each component;

[0025] A single fault feature library is preset, the fault discrimination of each parameter is calculated based on mutual information entropy and weights are assigned, and a weighted similarity algorithm is used to calculate the matching degree between the feature vector and the standard feature vector in the fault feature library;

[0026] Based on a preset matching threshold, the specific fault category corresponding to each independent component is determined, and a set of fault categories without repetition is obtained by combining the fault scenarios.

[0027] Furthermore, S4 includes the following:

[0028] A causal relationship graph of a pre-set switchgear mechanical system is provided. The graph includes fault type nodes, component nodes, and directed edges representing the causal evolution of faults. The directed edges are accompanied by causal association weights and evolution time lag attributes.

[0029] The set of fault categories obtained in S3 is associated and mapped with the fault nodes in the causal relationship graph;

[0030] The mapping fault categories are validated for correlation based on three dimensions: causal direction, correlation strength, and evolutionary time sequence, and a comprehensive correlation strength index is calculated. The comprehensive correlation strength index is calculated by combining the graph correlation weights and the fault feature matching similarity.

[0031] Set a correlation strength threshold, remove false fault features with a comprehensive correlation strength lower than the threshold, and obtain a valid fault set that passes the verification.

[0032] Construct a fault severity index that integrates feature deviation and associated impact range to quantitatively rank the effective fault set;

[0033] The system integrates the types of valid faults, their corresponding source components, and severity ranking results to output multi-fault source location results.

[0034] A deep learning-based mechanical fault prediction system for switching equipment includes: a signal acquisition and preprocessing module, a multi-fault decoupling module, a hierarchical fault identification module, and a causal verification and location output module.

[0035] The signal acquisition and preprocessing module acquires vibration and current signals during the operation of the switching equipment, and removes environmental vibration interference and electromagnetic interference noise through filtering to obtain the signal to be analyzed containing fault characteristics.

[0036] The multi-fault decoupling module adopts a combination of independent component analysis and causal decoupling logic for mechanical faults in switching equipment to perform decoupling processing on the signal to be analyzed. By setting causal constraints between fault features, the superimposed mixed features of multiple faults are decomposed into independent feature components that correspond one-to-one with a single fault.

[0037] The hierarchical fault identification module executes hierarchical identification actions in sequence. First, it counts the number of independent feature components to determine whether the current fault is a multi-fault coupled scenario. Then, it extracts the key characterization parameters of each independent feature component and matches them with the preset single fault feature library to determine the specific fault category corresponding to each independent component.

[0038] The causal verification and location output module calls the pre-built causal relationship map of the mechanical system of the switching equipment to perform correlation verification on each identified fault category to verify the logical rationality of the fault evolution, eliminate false fault features caused by environmental interference, and finally output the multi-fault source location results including each fault type, corresponding fault source component, and fault severity ranking.

[0039] Furthermore, the signal acquisition and preprocessing module includes a signal acquisition unit and a signal preprocessing unit;

[0040] The signal acquisition unit deploys vibration sensors at key mechanical components of the switchgear and connects current sensors in series in the power supply circuit of the opening and closing coils. It connects to the data acquisition terminal through shielded cables, sets sampling parameters based on the Nyquist sampling theorem and the frequency range of mechanical fault characteristics of the switchgear, synchronously acquires analog signals from the sensors and converts them into digital signals for storage, thereby obtaining the original vibration signal sequence and the original current signal sequence.

[0041] The signal preprocessing unit uses the min-max normalization algorithm to uniformly map the amplitudes of the two types of original signals to the [0,1] interval to eliminate amplitude imbalance. Then, it uses a combination of wavelet filtering and notch filtering to remove interference and integrate the signals to be analyzed.

[0042] Furthermore, the multi-fault decoupling module includes a constraint construction and objective function optimization unit and a decoupling operation and component selection unit;

[0043] The constraint construction and objective function optimization unit reconstructs the signal to be analyzed into a multi-channel signal matrix, sorts out the causal relationship between fault features based on the mechanical structure principle of switching equipment, sets causal direction constraints and feature uniqueness constraints and transforms them into quantitative constraints, constructs an improved independent component analysis objective function that incorporates causal constraints, and iteratively updates the separation matrix through the natural gradient descent method until the objective function converges.

[0044] The decoupling operation and component screening unit substitutes the signal to be analyzed into the optimal separation matrix to obtain the initial decoupling components, eliminates invalid components and verifies the causal constraint satisfaction, and finally obtains the independent feature components that correspond one-to-one with a single fault.

[0045] Furthermore, the hierarchical fault identification module includes a fault scenario determination unit and a feature extraction and matching unit;

[0046] The fault scenario determination unit counts the number of independent feature components. When the number of components is 1, it is determined to be a single fault scenario. When the number of components is not less than 2, it is determined to be a multi-fault coupled scenario.

[0047] The feature extraction and matching unit extracts key characterization parameters in the time domain, frequency domain, and time-frequency domain for each independent feature component and constructs a feature vector. It calculates the fault discrimination of each parameter based on mutual information entropy and assigns weights. It uses a weighted similarity algorithm to calculate the matching degree between the feature vector and the standard feature vector in the preset single fault feature library. It combines the preset matching threshold to determine the fault category and integrates them to obtain a set of fault categories without repetition.

[0048] Furthermore, the causality verification and location output module includes a multi-dimensional causality verification unit and a fault location and sorting output unit;

[0049] The multi-dimensional causal verification unit pre-sets a causal relationship graph of a mechanical system for switching equipment containing fault type nodes, component nodes, and directed edges with attributes. It associates and maps the set of fault categories with the fault nodes in the graph, and performs correlation verification from three dimensions: causal direction, correlation strength, and evolution time sequence. It calculates a comprehensive correlation strength index that combines the correlation weight of the graph with the similarity of fault feature matching, and sets a threshold to remove false fault features to obtain an effective fault set.

[0050] The fault location and ranking output unit constructs a fault severity index that integrates feature deviation and associated influence range, quantifies and ranks the effective fault set, integrates the type of effective fault, corresponding source component and severity ranking results, and outputs multi-fault source location results.

[0051] Compared with existing technologies, the beneficial effects of this invention are as follows: By combining independent component analysis with the causal decoupling logic of mechanical faults in switching equipment, this invention introduces dual quantitative constraints of causal direction and feature uniqueness. It accurately decomposes the superimposed mixed features of multiple faults into independent feature components corresponding one-to-one with a single fault, specifically solving the problem of mixed feature interference caused by the lack of a multi-fault decoupling mechanism in existing technologies. This fundamentally achieves effective separation of single fault features. Furthermore, by employing a hierarchical identification strategy, this invention first determines whether the current fault is a single fault scenario or a multi-fault coupled scenario based on the number of independent feature components. Then, it extracts multi-dimensional key characterization parameters in the time domain, frequency domain, and time-frequency domain of each independent feature component to construct a feature vector. Combined with a mutual information entropy weighted similarity matching algorithm, it determines the specific fault category, overcoming the deficiency of existing technologies in distinguishing specific fault types within coupled faults. This improves the accuracy of fault identification and the adaptability to different fault scenarios. This invention uses a pre-defined causal relationship map of a switchgear mechanical system containing fault type nodes, component nodes, and quantified attributes. It verifies the correlation of identified fault categories from three dimensions: causal direction, correlation strength, and evolution sequence to eliminate false fault features. At the same time, it constructs a fault severity index that integrates feature deviation and correlation influence range to quantify and rank valid faults. This invention solves the problem that existing technologies cannot accurately locate multiple fault sources and achieves precise identification of fault source components and clear division of fault handling priorities. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a schematic diagram of a module of a deep learning-based mechanical fault prediction system for switching equipment according to the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Please see Figure 1 The present invention provides the following technical solution:

[0056] A deep learning-based mechanical fault prediction system for switching equipment includes: a signal acquisition and preprocessing module, a multi-fault decoupling module, a hierarchical fault identification module, and a causal verification and location output module.

[0057] The signal acquisition and preprocessing module acquires vibration and current signals during the operation of the switching equipment, and removes environmental vibration interference and electromagnetic interference noise through filtering to obtain the signal to be analyzed containing fault characteristics.

[0058] The multi-fault decoupling module adopts a combination of independent component analysis and causal decoupling logic for mechanical faults in switching equipment to perform decoupling processing on the signal to be analyzed. By setting causal constraints between fault features, the superimposed mixed features of multiple faults are decomposed into independent feature components that correspond one-to-one with a single fault.

[0059] The hierarchical fault identification module executes hierarchical identification actions in sequence. First, it counts the number of independent feature components to determine whether the current fault is a multi-fault coupled scenario. Then, it extracts the key characterization parameters of each independent feature component and matches them with the preset single fault feature library to determine the specific fault category corresponding to each independent component.

[0060] The causal verification and location output module calls the pre-built causal relationship map of the mechanical system of the switching equipment to perform correlation verification on each identified fault category to verify the logical rationality of the fault evolution, eliminate false fault features caused by environmental interference, and finally output the multi-fault source location results including each fault type, corresponding fault source component, and fault severity ranking.

[0061] The signal acquisition and preprocessing module includes a signal acquisition unit and a signal preprocessing unit;

[0062] The signal acquisition unit deploys vibration sensors at key mechanical components of the switchgear and connects current sensors in series in the power supply circuit of the opening and closing coils. It connects to the data acquisition terminal through shielded cables, sets sampling parameters based on the Nyquist sampling theorem and the frequency range of mechanical fault characteristics of the switchgear, synchronously acquires analog signals from the sensors and converts them into digital signals for storage, thereby obtaining the original vibration signal sequence and the original current signal sequence.

[0063] The signal preprocessing unit uses the min-max normalization algorithm to uniformly map the amplitudes of the two types of original signals to the [0,1] interval to eliminate amplitude imbalance. Then, it uses a combination of wavelet filtering and notch filtering to remove interference and integrate the signals to be analyzed.

[0064] The multi-fault decoupling module includes a constraint construction and objective function optimization unit and a decoupling operation and component selection unit;

[0065] The constraint construction and objective function optimization unit reconstructs the signal to be analyzed into a multi-channel signal matrix, sorts out the causal relationship between fault features based on the mechanical structure principle of switching equipment, sets causal direction constraints and feature uniqueness constraints and transforms them into quantitative constraints, constructs an improved independent component analysis objective function that incorporates causal constraints, and iteratively updates the separation matrix through the natural gradient descent method until the objective function converges.

[0066] The decoupling operation and component screening unit substitutes the signal to be analyzed into the optimal separation matrix to obtain the initial decoupling components, eliminates invalid components and verifies the causal constraint satisfaction, and finally obtains the independent feature components that correspond one-to-one with a single fault.

[0067] The hierarchical fault identification module includes a fault scenario determination unit and a feature extraction and matching unit;

[0068] The fault scenario determination unit counts the number of independent feature components. When the number of components is 1, it is determined to be a single fault scenario. When the number of components is not less than 2, it is determined to be a multi-fault coupled scenario.

[0069] The feature extraction and matching unit extracts key characterization parameters in the time domain, frequency domain, and time-frequency domain for each independent feature component and constructs a feature vector. It calculates the fault discrimination of each parameter based on mutual information entropy and assigns weights. It uses a weighted similarity algorithm to calculate the matching degree between the feature vector and the standard feature vector in the preset single fault feature library. It combines the preset matching threshold to determine the fault category and integrates them to obtain a set of fault categories without repetition.

[0070] The causality verification and location output module includes a multi-dimensional causality verification unit and a fault location and sorting output unit;

[0071] The multi-dimensional causal verification unit pre-sets a causal relationship graph of a mechanical system for switching equipment containing fault type nodes, component nodes, and directed edges with attributes. It associates and maps the set of fault categories with the fault nodes in the graph, and performs correlation verification from three dimensions: causal direction, correlation strength, and evolution time sequence. It calculates a comprehensive correlation strength index that combines the correlation weight of the graph with the similarity of fault feature matching, and sets a threshold to remove false fault features to obtain an effective fault set.

[0072] The fault location and ranking output unit constructs a fault severity index that integrates feature deviation and associated influence range, quantifies and ranks the effective fault set, integrates the type of effective fault, corresponding source component and severity ranking results, and outputs multi-fault source location results.

[0073] A deep learning-based method for predicting mechanical faults in switching equipment includes the following steps:

[0074] S1. Collect vibration and current signals during the operation of the switching equipment, and remove environmental vibration interference and electromagnetic interference noise through filtering to obtain the signal to be analyzed containing fault characteristics;

[0075] S2. By combining independent component analysis with causal decoupling logic for mechanical faults in switching equipment, the signal to be analyzed is decoupled. By setting causal constraints between fault features, the superimposed multi-fault mixed features are decomposed into independent feature components that correspond one-to-one with a single fault.

[0076] S3. Perform layered identification actions sequentially. The first layer is to count the number of independent feature components and determine whether the current fault is a multi-fault coupling scenario. The second layer is to extract the key characterization parameters of each independent feature component and match them with a preset single fault feature library to determine the specific fault category corresponding to each independent component.

[0077] S4. Call the pre-built causal relationship map of the mechanical system of the switching equipment, perform correlation verification on each fault category obtained in S3, verify the logical rationality of the fault evolution, eliminate false fault features caused by environmental interference, and finally output the multi-fault source localization result; the multi-fault source localization result includes each fault type, corresponding fault source component and fault severity ranking.

[0078] S1 includes the following:

[0079] Vibration sensors are deployed at key mechanical components of the switchgear, and current sensors are connected in series in the power supply circuit of the opening and closing coils. The sensors are connected to the data acquisition terminal through shielded cables. Sampling parameters are set according to the Nyquist sampling theorem and the frequency range of mechanical fault characteristics of the switchgear. The sampling duration covers the complete opening and closing cycle. Simultaneously, the sensor analog signals are collected and converted into digital signals for storage to obtain the original vibration signal sequence and the original current signal sequence.

[0080] The min-max normalization algorithm is used to map the amplitudes of the two types of original signals to the [0,1] interval, eliminating the amplitude imbalance caused by the difference in sensor range, and obtaining normalized vibration signal and normalized current signal.

[0081] Interference is removed by combining wavelet filtering and notch filtering, specifically as follows:

[0082] After decomposing the normalized vibration signal using wavelets, thresholding and inverse transformation are performed to remove environmental vibration interference. Spectral analysis is then conducted on the processed vibration signal and normalized current signal to locate the dominant interference frequency. A notch filter is used to specifically remove electromagnetic interference noise, and the resulting signal is then integrated to obtain the signal to be analyzed.

[0083] In this embodiment, a piezoelectric vibration sensor is fixedly deployed at a key mechanical component of the switchgear, and a Hall effect current sensor is connected in series in the power supply circuit of the opening and closing coil. The sensor is connected to a data acquisition terminal via a shielded cable. Based on the Nyquist sampling theorem and the characteristic frequency range of mechanical faults in the switchgear, it is assumed that the sampling frequency of the vibration signal is set to 10kHz-50kHz and the sampling frequency of the current signal is set to 1kHz-5kHz, with the sampling duration covering the complete opening and closing cycle. The analog signals of the two types of sensors are synchronously acquired by the data acquisition terminal, converted into digital signals, and stored in a standard format to obtain the original vibration signal sequence xv(t) and the original current signal sequence xc(t), where t is the sampling time, which is a positive integer between 1 and N, and N is the number of sampling points.

[0084] To eliminate the amplitude imbalance caused by the range difference between the vibration sensor and the current sensor, the min-max normalization algorithm is used to map the amplitudes of the two types of original signals to the [0,1] interval, thereby obtaining the normalized vibration signal xv'(t) and the normalized current signal xc'(t) respectively.

[0085] Two types of interference are eliminated by combining wavelet filtering and notch filtering, specifically:

[0086] Wavelet filtering removes environmental vibration interference. The db4 wavelet is used to decompose the normalized vibration signal into multiple layers. The noise standard deviation σ is estimated by the median method. The filtering threshold λ is set, and λ=σ·(2lnN)^(1 / 2). After soft thresholding, the signal is inversely transformed with the low-frequency approximate component to obtain the vibration signal after removing environmental vibration interference.

[0087] Notch filtering removes electromagnetic interference noise. Fast Fourier Transform (FFT) is performed on both the vibration signal and the normalized current signal to pinpoint the dominant interference frequency f0. A second-order IIR notch filter is used, with a center frequency fc = f0, a 3dB bandwidth of 1Hz-3Hz, and an attenuation coefficient of 0.95-0.99. The two types of signals are then input into their respective notch filters to obtain the denoised vibration and current signals, which are then integrated into the signal to be analyzed, X(t) = [Xv(t), Xc(t)]. T Where Xv(t) is the denoised vibration signal and Xc(t) is the denoised current signal.

[0088] S2 includes the following:

[0089] The signal to be analyzed is reconstructed into a multi-channel signal matrix to meet the input requirements of the decoupling algorithm;

[0090] Based on the mechanical structure principle of switching equipment, the causal relationships between fault characteristics are sorted out, and causal direction constraints and feature uniqueness constraints are set, which are transformed into quantitative causal constraint conditions; the quantitative causal constraint conditions include causal correlation index threshold constraints and feature similarity threshold constraints.

[0091] An improved independent component analysis objective function incorporating the aforementioned causal constraints is constructed, and the separation matrix is ​​iteratively updated using the natural gradient descent method until the objective function converges.

[0092] Substituting the signal to be analyzed into the optimal separation matrix yields the initial decoupled components. Invalid components are eliminated, and the causal constraint satisfaction is verified. Finally, independent feature components corresponding one-to-one with a single fault are obtained.

[0093] In this embodiment, the signal to be analyzed obtained in S1 is X(t) = [Xv(t), Xc(t)]. T Reconstructed into a multi-channel signal matrix X, where X∈R M×T Where M=2 is the number of signal channels and T is the number of sampling points; based on the mechanical system structure of switching equipment, the characteristic causal relationships of typical faults are analyzed and transformed into two types of quantitative constraints, specifically:

[0094] Causal direction constraint: If fault A1 is the cause of fault B1, such as shaft jamming causing operating mechanism jamming, then the characteristic signal of fault A1 leads the characteristic signal of fault B1, which is mathematically expressed as: CI(ai,aj)>γ, where ai and aj are two characteristic components to be decoupled, CI( , ) is the causal correlation index, and γ is the causal correlation threshold.

[0095] Feature uniqueness constraint: The similarity between the fault feature corresponding to each independent component and other components is less than a preset threshold to avoid component redundancy. Mathematically, it is expressed as: SIM(ai,aj)<β, and i≠j, where SIM( , ) is the feature similarity and β is the similarity threshold.

[0096] Based on the objective function of "maximizing the non-Gaussianity of the output components" in standard independent component analysis (ICA), the aforementioned causal constraint term is introduced to construct a dual optimization objective function: J(W) = ∑ i∈[1,K] H(ai)+μ∑ i≠j max(0,CI(ai,aj)-γ)+v∑ i≠j max(0,SIM(ai,aj)-β), where W∈R K×M This is a separation matrix, where K is the maximum possible number of faults, set based on common fault types of switching equipment; ai=wi TX represents the decoupled independent components, and wi is the i-th row vector of W; H(ai) = -∫p(ai)lnp(ai)dai is the differential entropy of component ai, used to measure non-Gaussianity, and the larger the value, the stronger the non-Gaussianity, which conforms to the ICA separation criterion; aj = wj T X is the decoupled independent component, and wj is the j-th row vector of W; μ and v are constraint weights; the last two terms are causal constraint penalty terms, which generate penalty points when the constraint conditions are not met, guiding the optimization of the objective function;

[0097] The objective function J(W) is minimized using the natural gradient descent method, and the separation matrix W is updated iteratively. The iterative formula is: W(t+1)=W(t)+η[(IE{tanh(a(t))a(t)T})W(t)-μ CI Jv SIM J], where η is the learning rate, ranging from 0.001 to 0.01, to ensure stable convergence of the iteration; I is the identity matrix, E{} is the expectation operator, and tanh() is the activation function; CI J and SIM J represents the gradients of the causal correlation constraint and the feature similarity constraint, respectively, and is used to adjust the update direction of the separation matrix to ensure that the decoupling satisfies the causal constraint.

[0098] When the difference of the objective function between two iterations satisfies |J(W(t+1))-J(W(t))|<ε, and ε=10 -6 The iteration stops when the number of iterations reaches a preset maximum value, yielding the optimal separation matrix W. * ;

[0099] Substitute the signal matrix X to be analyzed into the optimal separation matrix W. * The initial decoupling component set A = [a1, a2, ..., aK] is calculated. T The process involves filtering and calculating the fault feature contribution rate ηi for each component, where ηi = Var(ai∩F) / Var(ai), and F is a preset set of fault feature frequency bands. Components with ηi < 0.5 are removed and classified as irrelevant noise components. The causal constraint condition is verified again for the remaining components, i.e., CI(ai,aj) > γ and SIM(ai,aj) < β. Finally, a set of independent feature components A corresponding one-to-one with a single fault is obtained. * =[a1 * a2 * ,...,aL * ] T , where L≤K is the actual number of faults.

[0100] S3 includes the following:

[0101] The number of independent feature components is counted. When the number of components is 1, it is determined to be a single fault scenario. When the number of components is not less than 2, it is determined to be a multi-fault coupled scenario.

[0102] For each independent feature component, extract key representation parameters in multiple dimensions in the time domain, frequency domain, and time-frequency domain, and construct the feature vector corresponding to each component;

[0103] A single fault feature library is preset, the fault discrimination of each parameter is calculated based on mutual information entropy and weights are assigned, and a weighted similarity algorithm is used to calculate the matching degree between the feature vector and the standard feature vector in the fault feature library;

[0104] Based on a preset matching threshold, the specific fault category corresponding to each independent component is determined, and a set of fault categories without repetition is obtained by combining the fault scenarios.

[0105] In this embodiment, the set of independent feature components A based on the output of S2 * =[a1 * a2 * ,...,aL * ] T Based on the failure evolution patterns of switching equipment, scenario judgment rules were formulated:

[0106] If L=1, it is determined to be a single fault scenario. This independent component contains only the feature information of one type of fault, and a single fault matching strategy is subsequently adopted.

[0107] If L≥2, it is determined to be a multi-fault coupling scenario, and each independent component corresponds to a non-repeating fault type. A multi-fault parallel matching strategy is then adopted.

[0108] For each independent feature component ag * Where g ranges from 1 to L, six core characterization parameters are extracted from three categories: time domain, frequency domain, and time-frequency domain, to construct a feature vector Fg=[fg1,fg2,...,fg6] that comprehensively reflects the fault characteristics. T The specific parameters are as follows: the time-domain parameters include peak value fg1, mean value fg2, and variance fg3. Peak value fg1 is the maximum amplitude of the component signal, which directly reflects the fault impact intensity; mean value fg2 is the average value of the component signal, which reflects the overall offset characteristics of the signal; variance fg3 is the fluctuation degree of the component signal, which reflects the change in signal stability caused by the fault.

[0109] Frequency domain parameters include characteristic frequency fg4 and peak power spectral density fg5; for independent characteristic components ag * Performing a Fast Fourier Transform (FFT) yields the frequency spectrum Ag(f). The frequency point with the largest amplitude is extracted as the characteristic frequency fg4, where fg4 = arg max.f |Ag(f)|; The peak value of the power spectral density fg5 is the maximum value of the power spectral density function, reflecting the degree of energy concentration at the fault characteristic frequency;

[0110] The time-frequency domain parameter is wavelet entropy fg6, and the db4 wavelet is used to adjust ag. * A three-level decomposition is performed, and the information entropy of the time-frequency components at each level is calculated and normalized to reflect the time-frequency distribution complexity of the fault characteristics.

[0111] Standard samples of common mechanical faults in switching equipment were collected through extensive experiments. For each fault sample, a single fault feature library G=[G1,G2,...,GK] was constructed based on the extraction method of core characterization parameters. T Where Gh = [gh1, gh2, ..., gh6] T Let Fg be the standard feature vector of the h-th type of fault, and K be the total number of fault types. The core representation parameters corresponding to the feature vector Fg and the standard feature vector in the single fault feature library G are all mapped to the interval [0,1] by the min-max algorithm.

[0112] The fault discrimination degree of each parameter is calculated based on mutual information entropy. The higher the discrimination degree, the greater the weight. The weight vector is Q=[q1,q2,...,q6]. T Satisfy ∑ m∈[1,6] qm=1; The weighted cosine similarity formula is used to calculate the matching degree between the feature vector Fg to be identified and the standard feature vector Gh in the fault feature database. The calculation formula is as follows:

[0113] S(g,h)=(Fg T ·diag(Q)·Gh) / [(Fg T ·diag(Q)·Fg)·(Gh T ·diag(Q)·Gh)]^(1 / 2),

[0114] Where S(g,h) is the weighted similarity between the g-th independent component and the h-th fault, and diag(Q) is the diagonal matrix constructed from the weight vectors. The closer the similarity is to 1, the higher the matching degree between the component and the fault.

[0115] Set a matching threshold S0, and for each independent component ag * Calculate the similarity S(g,h) between it and all faults in the feature library. If there exists a unique h... * If S(g,h)≥S0, then determine ag * The corresponding fault category is h * Fault type; if there are multiple h satisfying S(g,h)≥S0, take the h with the highest similarity. *As a judgment result; if all S(g,h) < S0, it is temporarily judged as an unidentified fault, pending further confirmation by S4 correlation verification; assuming the fault feature library K=3, where h=1 is spring fatigue, h=2 is shaft jamming, h=3 is contact wear; assuming for independent component a1 * The calculated values ​​are S(1,1)=0.6, S(1,2)=0.85, and S(1,3)=0.72, with a matching threshold of S0=0.7. The candidate faults satisfying S(g,h)≥S0 are h=2 and h=3, with h=2 having the highest similarity (shaft jamming). Therefore, h... * =2, that is, determine a1 * The corresponding fault category is "shaft jamming".

[0116] Based on the determined fault scenario, integrate the fault category determination results of all independent components; for a single fault scenario, output a uniquely determined fault category; for a multi-fault coupled scenario, output multiple sets of non-repeating fault categories.

[0117] S4 includes the following:

[0118] A causal relationship graph of a pre-set switchgear mechanical system is provided. The graph includes fault type nodes, component nodes, and directed edges representing the causal evolution of faults. The directed edges are accompanied by causal association weights and evolution time lag attributes.

[0119] The set of fault categories obtained in S3 is associated and mapped with the fault nodes in the causal relationship graph;

[0120] The mapping fault categories are validated for correlation based on three dimensions: causal direction, correlation strength, and evolutionary time sequence, and a comprehensive correlation strength index is calculated. The comprehensive correlation strength index is calculated by combining the graph correlation weights and the fault feature matching similarity.

[0121] Set a correlation strength threshold, remove false fault features with a comprehensive correlation strength lower than the threshold, and obtain a valid fault set that passes the verification.

[0122] Construct a fault severity index that integrates feature deviation and associated impact range to quantitatively rank the effective fault set;

[0123] The system integrates the types of valid faults, their corresponding source components, and severity ranking results to output multi-fault source location results.

[0124] In this embodiment, the preset causal relationship diagram of the mechanical system of the switching equipment specifically includes:

[0125] A directed graph H=(N,L,B) is used, where the node set N=NF∪NP, and NF represents fault type nodes and NP represents component nodes. The directed edge set L is divided into two types of edges: fault-fault edges, which represent the causal evolution relationship between faults, and fault-component edges, which represent the source component corresponding to the fault. Each fault-fault edge in the attribute set B is accompanied by two core attributes, namely the causal association weight λij and the evolution time lag τij. The causal association weight λij represents the causal influence strength of Fi on Fj, and the evolution time lag τij represents the time difference between the occurrence of Fi and the occurrence of Fj. Through several sets of simulation experiments with different fault types, the causal association weight λij and the evolution time lag τij are corrected to ensure that the graph is consistent with the actual fault evolution law.

[0126] The fault category set F output by S3 is mapped one-to-one with the fault type nodes to obtain the mapping set FY, and FY={(Fl,NF,l)|l=1,2,...,L}, which clarifies the corresponding node and associated component of each identified fault in the graph;

[0127] For the mapped fault set, the logical rationality is verified from three dimensions: causal direction, correlation strength, and evolutionary sequence, as follows:

[0128] For causal direction verification, for any two faults Fi, Fj∈FY in the set, query whether there is a directed edge Fi→Fj or Fj→Fi in the graph; if there is an undirected edge associated fault pair, mark it as an association to be verified and proceed to the next step of strength verification.

[0129] For the quantitative calculation of association strength, a comprehensive association strength index Cij is constructed, which combines the map association weight and S3 feature similarity. The specific formula is: Cij=λij·[(S(i)+S(j)] / 2, where S(i) and S(j) are the maximum matching similarity of Fi and Fj in S3, respectively; and Cij∈[0,1], the larger the value, the stronger the association between the fault pairs.

[0130] For evolutionary timing verification, for fault pairs Fi→Fj with causal relationships, the corresponding independent components ai in S2 are extracted. * ,aj * Given the starting times ti and tj, verify that tj-ti≥τij; if not satisfied, mark it as a timing anomaly.

[0131] Set a correlation strength threshold C0. If a single fault Fl∈FY has Cil<C0 with all other faults in the set, and no associated component matches the fault node in the graph, it is judged as a false fault and is removed. If the fault pair (Fi,Fj) has Cij<C0 and the timing is abnormal, the fault with lower correlation strength is removed, such as S(i)<S(j) then Fi is removed. The valid fault set Fv that passes the verification is retained, and the number of elements in the valid fault set Fv is less than or equal to L.

[0132] Construct a fault severity index Z(Fv), and Z(Fv) = α·Δ(Fv) + β·Ω(Fv), where Δ(Fv) is the characteristic deviation, calculated by the formula: Δ(Fv) = ||Fg - Gh * || / max H ||FG||,||Fg-Gh * || represents the difference between the feature vector Fg to be identified and the standard feature vector Gh. * Euclidean distance, max H ||FG|| represents the maximum Euclidean distance between all fault feature vectors in the fault feature library and their corresponding standard vectors, used for normalization; Ω(Fv) is the range of influence, calculated using the formula: Ω(Fv) = count(Fv→Fk) / max NF count(F→Fk) and count(Fv→Fk) are the number of faults in the graph with Fv as the cause, and max(Fv→Fk) is the maximum number of faults in the graph. NF count(F→Fk) is the maximum number of impact faults of all fault nodes in the graph; α and β are weighting coefficients, and satisfy α+β=1;

[0133] The effective fault set Fv is sorted in descending order based on the severity index Z(Fv). The larger the Z(Fv) value, the higher the severity of the fault, and the higher the priority for maintenance. Finally, the effective fault types, corresponding source components, and severity ranking results are integrated to output the multi-fault source location results. The format example is: {Fault type: Shaft jamming, Source component: Shaft assembly, Severity: 1 (Z=0.82); Fault type: Operating mechanism jamming, Source component: Operating mechanism, Severity: 2 (Z=0.65)}.

[0134] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0135] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based mechanical failure prediction method for switchgear, characterized by: The method includes the following steps: S1. Collect vibration and current signals during the operation of the switching equipment, and remove environmental vibration interference and electromagnetic interference noise through filtering to obtain the signal to be analyzed containing fault characteristics; S2. By combining independent component analysis with causal decoupling logic for mechanical faults in switching equipment, the signal to be analyzed is decoupled. By setting causal constraints between fault features, the superimposed multi-fault mixed features are decomposed into independent feature components that correspond one-to-one with a single fault. S3. Perform layered identification actions sequentially. The first layer is to count the number of independent feature components and determine whether the current fault is a multi-fault coupling scenario. The second layer is to extract the key characterization parameters of each independent feature component and match them with a preset single fault feature library to determine the specific fault category corresponding to each independent component. S4. Call the pre-built causal relationship map of the mechanical system of the switching equipment, perform correlation verification on each fault category obtained in S3, verify the logical rationality of the fault evolution, eliminate false fault features caused by environmental interference, and finally output the multi-fault source localization result; the multi-fault source localization result includes each fault type, corresponding fault source component and fault severity ranking.

2. The method of claim 1, wherein the method is based on deep learning. S1 includes the following: Vibration sensors are deployed at key mechanical components of the switchgear, and current sensors are connected in series in the power supply circuit of the opening and closing coils. The sensors are connected to the data acquisition terminal through shielded cables. Sampling parameters are set according to the Nyquist sampling theorem and the frequency range of mechanical fault characteristics of the switchgear. The sampling duration covers the complete opening and closing cycle. Simultaneously, the sensor analog signals are collected and converted into digital signals for storage to obtain the original vibration signal sequence and the original current signal sequence. The min-max normalization algorithm is used to map the amplitudes of the two types of original signals to the [0,1] interval, eliminating the amplitude imbalance caused by the difference in sensor range, and obtaining normalized vibration signal and normalized current signal. Interference is removed by combining wavelet filtering and notch filtering, specifically as follows: After decomposing the normalized vibration signal using wavelets, thresholding and inverse transformation are performed to remove environmental vibration interference. Spectral analysis is then conducted on the processed vibration signal and normalized current signal to locate the dominant interference frequency. A notch filter is used to specifically remove electromagnetic interference noise, and the resulting signal is then integrated to obtain the signal to be analyzed.

3. The method of claim 1, wherein the method further comprises: S2 includes the following: The signal to be analyzed is reconstructed into a multi-channel signal matrix to meet the input requirements of the decoupling algorithm; Based on the mechanical structure principle of switching equipment, the causal relationships between fault characteristics are sorted out, and causal direction constraints and feature uniqueness constraints are set, which are transformed into quantitative causal constraint conditions; the quantitative causal constraint conditions include causal correlation index threshold constraints and feature similarity threshold constraints. An improved independent component analysis objective function incorporating the aforementioned causal constraints is constructed, and the separation matrix is ​​iteratively updated using the natural gradient descent method until the objective function converges. Substituting the signal to be analyzed into the optimal separation matrix yields the initial decoupled components. Invalid components are eliminated, and the causal constraint satisfaction is verified. Finally, independent feature components corresponding one-to-one with a single fault are obtained.

4. The method of claim 1, wherein the method further comprises: S3 includes the following: The number of independent feature components is counted. When the number of components is 1, it is determined to be a single fault scenario. When the number of components is not less than 2, it is determined to be a multi-fault coupled scenario. For each independent feature component, extract key representation parameters in multiple dimensions in the time domain, frequency domain, and time-frequency domain, and construct the feature vector corresponding to each component; A single fault feature library is preset, the fault discrimination of each parameter is calculated based on mutual information entropy and weights are assigned, and a weighted similarity algorithm is used to calculate the matching degree between the feature vector and the standard feature vector in the fault feature library; Based on a preset matching threshold, the specific fault category corresponding to each independent component is determined, and a set of fault categories without repetition is obtained by combining the fault scenarios.

5. The method of claim 1, wherein the method further comprises: S4 includes the following: A causal relationship graph of a pre-set switchgear mechanical system is provided. The graph includes fault type nodes, component nodes, and directed edges representing the causal evolution of faults. The directed edges are accompanied by causal association weights and evolution time lag attributes. The set of fault categories obtained in S3 is associated and mapped with the fault nodes in the causal relationship graph; The mapping fault categories are validated for correlation based on three dimensions: causal direction, correlation strength, and evolutionary time sequence, and a comprehensive correlation strength index is calculated. The comprehensive correlation strength index is calculated by combining the map correlation weights and the similarity of fault feature matching. Set a correlation strength threshold, remove false fault features with a comprehensive correlation strength lower than the threshold, and obtain a valid fault set that passes the verification. Construct a fault severity index that integrates feature deviation and associated impact range to quantitatively rank the effective fault set; The system integrates the types of valid faults, their corresponding source components, and severity ranking results to output multi-fault source location results.

6. A deep learning-based mechanical fault prediction system for switchgear, applied to the deep learning-based mechanical fault prediction method for switchgear according to any one of claims 1-5, characterized in that: The system includes: a signal acquisition and preprocessing module, a multi-fault decoupling module, a hierarchical fault identification module, and a causality verification and location output module; The signal acquisition and preprocessing module acquires vibration and current signals during the operation of the switching equipment, and removes environmental vibration interference and electromagnetic interference noise through filtering to obtain the signal to be analyzed containing fault characteristics. The multi-fault decoupling module uses a combination of independent component analysis and causal decoupling logic for mechanical faults in switching equipment to perform decoupling processing on the signal to be analyzed. By setting causal constraints between fault features, the superimposed multi-fault mixed features are decomposed into independent feature components that correspond one-to-one with a single fault. The hierarchical fault identification module performs hierarchical identification actions in sequence. First, it counts the number of independent feature components to determine whether the current fault is a multi-fault coupling scenario. Then, it extracts the key characterization parameters of each independent feature component and matches them with a preset single fault feature library to determine the specific fault category corresponding to each independent component. The causal verification and location output module calls the pre-built causal relationship map of the mechanical system of the switching equipment to perform correlation verification on each identified fault category to verify the logical rationality of the fault evolution, eliminate false fault features caused by environmental interference, and finally output multi-fault source location results including each fault type, corresponding fault source component, and fault severity ranking.

7. The deep learning-based mechanical fault prediction system for switching equipment according to claim 6, characterized in that: The signal acquisition and preprocessing module includes a signal acquisition unit and a signal preprocessing unit; The signal acquisition unit deploys vibration sensors at key mechanical components of the switchgear and connects current sensors in series in the power supply circuit of the opening and closing coils. It connects to the data acquisition terminal through shielded cables, sets sampling parameters based on the Nyquist sampling theorem and the frequency range of mechanical fault characteristics of the switchgear, synchronously acquires analog signals from the sensors and converts them into digital signals for storage, thereby obtaining the original vibration signal sequence and the original current signal sequence. The signal preprocessing unit uses the min-max normalization algorithm to uniformly map the amplitudes of the two types of original signals to the [0,1] interval to eliminate amplitude imbalance. Then, it uses a combination of wavelet filtering and notch filtering to remove interference and integrate the signals to be analyzed.

8. The deep learning-based mechanical fault prediction system for switching equipment according to claim 6, characterized in that: The multi-fault decoupling module includes a constraint construction and objective function optimization unit and a decoupling operation and component selection unit; The constraint construction and objective function optimization unit reconstructs the signal to be analyzed into a multi-channel signal matrix, sorts out the causal relationship between fault features based on the mechanical structure principle of switching equipment, sets causal direction constraints and feature uniqueness constraints and transforms them into quantitative constraints, constructs an improved independent component analysis objective function incorporating the causal constraints, and iteratively updates the separation matrix through the natural gradient descent method until the objective function converges. The decoupling operation and component screening unit substitutes the signal to be analyzed into the optimal separation matrix to obtain the initial decoupling components, eliminates invalid components and verifies the causal constraint satisfaction, and finally obtains the independent feature components that correspond one-to-one with a single fault.

9. A deep learning-based mechanical fault prediction system for switching equipment according to claim 6, characterized in that: The hierarchical fault identification module includes a fault scenario determination unit and a feature extraction and matching unit; The fault scenario determination unit counts the number of independent feature components. When the number of components is 1, it is determined to be a single fault scenario. When the number of components is not less than 2, it is determined to be a multi-fault coupled scenario. The feature extraction and matching unit extracts key characterization parameters in the time domain, frequency domain, and time-frequency domain for each independent feature component and constructs a feature vector. It calculates the fault discrimination degree of each parameter based on mutual information entropy and assigns weights. It uses a weighted similarity algorithm to calculate the matching degree between the feature vector and the standard feature vector in the preset single fault feature library. It combines the preset matching threshold to determine the fault category and integrates them to obtain a set of fault categories without repetition.

10. A deep learning-based mechanical fault prediction system for switching equipment according to claim 6, characterized in that: The causality verification and location output module includes a multi-dimensional causality verification unit and a fault location and sorting output unit; The multi-dimensional causal verification unit pre-sets a causal relationship graph of a mechanical system of switching equipment containing fault type nodes, component nodes, and directed edges with attributes. It associates and maps the set of fault categories with the fault nodes in the graph, performs correlation verification from three dimensions: causal direction, correlation strength, and evolution time sequence, calculates a comprehensive correlation strength index that combines the correlation weight of the graph with the similarity of fault feature matching, and sets a threshold to remove false fault features to obtain an effective fault set. The fault location and ranking output unit constructs a fault severity index that integrates feature deviation and associated influence range, quantifies and ranks the effective fault set, integrates the type of effective fault, corresponding source component and severity ranking results, and outputs multi-fault source location results.

Images