Isolating switch mechanical fault diagnosis method, system, equipment and medium based on multi-source data fusion

Abstract

This invention discloses a method, system, equipment, and medium for mechanical fault diagnosis of disconnecting switches based on multi-source data fusion, belonging to the field of disconnecting switch diagnosis technology. The method involves collecting vibration and power signals and steady-state leakage current signals during disconnecting switch operation; extracting image and sideband features from the vibration through time-frequency transformation and mode decomposition; extracting statistical features from the power signal in time-series segments; and extracting the effective value and waveform distortion features of the leakage current. A multi-branch fusion network is constructed to output fault categories, which are then trained using a combined loss function for real-time diagnosis. This invention achieves deep feature adaptive fusion and high-precision, high-sensitivity fault diagnosis under the heterogeneous representation of multi-source signals from disconnecting switches through hierarchical feature extraction and lightweight multi-branch network adaptive fusion, combined with a combined constraint strategy.

Description

Technical Field

[0001] This invention relates to the field of disconnector diagnostic technology, specifically to a method, system, equipment, and medium for diagnosing mechanical faults in disconnectors based on multi-source data fusion. Background Technology

[0002] With the increasing demands for equipment reliability in power systems, disconnect switches, as critical switching equipment in high-voltage transmission and transformation systems, directly impact the safe operation of the power grid due to their mechanical condition. Typical mechanical faults of disconnect switches include operational jamming, incomplete opening and closing, and transmission mechanism failure. These faults often stem from contact wear, lubrication deterioration, loose connecting rods, or foreign object obstruction. Traditional diagnostic methods primarily rely on mechanical characteristic tests during periodic power outages for maintenance, measuring parameters such as opening and closing speed, stroke, and operating torque to determine equipment status. However, current methods are offline detection methods, resulting in poor timeliness and high costs. In recent years, online monitoring technology has been gradually introduced, and diagnostic methods based on single physical quantities have been proposed. Examples include analyzing meshing impact characteristics through vibration signals, identifying changes in drive load through motor current or power signals, and assessing insulation and contact status through grounding leakage current. These methods have achieved status perception to some extent under uninterrupted power conditions, but each only reflects equipment status from a single dimension, resulting in insufficient information utilization and hindering further improvements in diagnostic accuracy and reliability.

[0003] Existing diagnostic methods based on single signals have the following limitations: First, vibration signals are sensitive to transient impacts and friction events, but are easily affected by ambient noise, propagation path attenuation, and differences in sensor installation location. Moreover, vibration alone is insufficient to distinguish between jamming and abnormal positioning. Motor power signals can reflect the overall load changes of the transmission chain, but are insensitive to high-frequency local impacts, and the power waveform also exhibits certain random fluctuations under normal operating conditions. Ground leakage current signals can reflect the contact state and insulation level of the contacts under steady-state closing conditions, but cannot provide dynamic mechanical information during operation, and the effective value is greatly affected by system voltage fluctuations and weather conditions. Second, the feature extraction of the above methods heavily relies on manually designed time-domain or frequency-domain indicators, such as peak value, root mean square, and sideband energy, which have limited feature expression capabilities and are difficult to generalize to different types or voltage levels of disconnect switches. Third, existing solutions often analyze the three types of signals independently, failing to construct an effective multi-source information fusion framework. This prevents the full utilization of the complementary characteristics between the signals to achieve comprehensive judgment, resulting in a high false negative rate in the early stages of jamming or under minor fault conditions. Summary of the Invention

[0004] In view of the above-mentioned existing problems, the present invention provides a method, system, equipment and medium for diagnosing mechanical faults of disconnect switches based on multi-source data fusion, in order to solve the problems of insufficient information utilization caused by independent analysis of multi-source signals, insufficient generalization ability of manual feature extraction, and high false negative rates of early jamming and minor faults due to the lack of a multi-source information fusion framework in the prior art.

[0005] To address the aforementioned technical issues, a method for diagnosing mechanical faults in disconnect switches based on multi-source data fusion is proposed, including: Vibration and drive power signals during the operation of the disconnecting switch, as well as ground leakage current signals in steady state, are collected. The vibration signals undergo time-frequency transformation to generate a first image. Adaptive mode decomposition is then performed on the vibration signals to extract sideband feature vectors from each modal component. The drive power signal is segmented according to time sequence, and time-domain statistical feature vectors of each segment are extracted. Effective value features and waveform distortion features are obtained from the ground leakage current signal. A multi-branch diagnostic network is constructed, which fuses the outputs of each branch and outputs the corresponding fault state category. The multi-branch diagnostic network is trained using a combined loss function, and fault diagnosis is performed on the real-time acquired signals based on the trained multi-branch diagnostic network.

[0006] As a preferred embodiment of the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion described in this invention, the time-frequency transformation processing includes scaling and time-shifting the wavelet basis and convolving it with the vibration signal to obtain a time-frequency distribution matrix, and mapping the amplitude of the time-frequency distribution matrix to a first color image. The adaptive mode decomposition includes establishing a constraint problem based on the vibration signal with the objective of minimizing the sum of the estimated bandwidths of all modes, and solving the problem by the Lagrange multiplier method and alternating iterations to decompose the vibration signal into a preset number of modal components. The extraction of sideband feature vectors includes calculating the correlation spectrum for each modal component under different frequency combinations, and statistically analyzing the energy concentration and broadening of the correlation spectrum within a preset frequency range on both sides of the center frequency to form a sideband feature vector.

[0007] As a preferred embodiment of the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion described in this invention, the multi-branch diagnostic network includes a first branch, a second branch, and a third branch. The first branch includes integrating the first image and the sideband feature vector at the input layer along the channel dimension to form multi-channel input data; The multi-channel input data is feature abstracted by multiple convolutional processing modules arranged in sequence. Each convolutional processing module adopts the operation of decomposing the standard convolution into independent spatial filtering by channel and point-by-point linear combination across channels, and adds a direct connection between the input and output of the module. The feature map after convolution is compressed into a one-dimensional form by pooling, and the first feature vector is obtained by fully connected mapping.

[0008] As a preferred embodiment of the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion described in this invention, the second branch includes inputting a time-domain statistical feature vector into at least one fully connected layer and performing nonlinear mapping to obtain a second feature vector. The third branch includes inputting the effective value features and waveform distortion features into the fully connected layer, performing nonlinear mapping, and obtaining the third feature vector.

[0009] As a preferred embodiment of the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion described in this invention, the fault state category to which the output belongs includes concatenating the first feature vector, the second feature vector, and the third feature vector to obtain a fused feature vector. The fused feature vector is input into at least one fully connected layer, and the probability value corresponding to each preset fault state category is output through a normalized exponential mapping operation.

[0010] As a preferred embodiment of the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion described in this invention, the combined loss function includes a classification error term, an intra-class clustering penalty term, and an inter-class balance weight term. The classification error term is calculated based on the cross-entropy between the probability value and the true label. The intra-class clustering penalty term is obtained by calculating the mean squared Euclidean distance between the fused feature vector and the feature center of its class. The inter-class balance weight term is determined by the number of samples in each class. The weights of each class are multiplied by the classification error term of the corresponding class and summed to form the weighted classification error term. The combined loss function is the sum of the weighted classification error term and the intra-class clustering penalty term.

[0011] As a preferred embodiment of the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion described in this invention, the fault diagnosis of the real-time acquired signals includes, during the training process, updating the feature centers of each category according to the fused feature vectors of each batch of samples using an exponential moving average method. During the steady state of the disconnecting switch, the ground leakage current signal is continuously collected, and the effective value characteristics and waveform distortion characteristics are updated after each collection. When the effective value characteristic and waveform distortion characteristic both exceed the preset normal reference range in a preset number of consecutive measurements, an anomaly is determined and an early warning is issued.

[0012] The beneficial effects of this preferred technical solution are as follows: starting with optimizing the acquisition of multi-source signals, through hierarchical feature extraction and lightweight multi-branch network adaptive fusion, combined with a combined constraint strategy, efficient and accurate diagnosis of mechanical faults in disconnect switches is achieved.

[0013] As a preferred embodiment of the mechanical fault diagnosis system for disconnecting switches based on multi-source data fusion described in this invention, it is characterized by including a multi-source feature extraction module, a diagnostic network construction module, a combined loss training module, and a real-time diagnosis and early warning module.

[0014] The multi-source feature extraction module is used to collect vibration, power and leakage current signals of the disconnecting switch, perform time-frequency transformation and adaptive mode decomposition on the vibration signal to obtain time-frequency image and sideband features, extract statistical features of the power signal in time sequence segments, and obtain effective value and distortion features from the leakage current to provide multi-source input for diagnosis.

[0015] The diagnostic network construction module is used to construct a multi-branch diagnostic network. The first branch fuses vibration images and sideband features, and extracts features through separable convolution and residual structure. The second and third branches process power statistics features and leakage current features respectively. The outputs of each branch are spliced ​​together, and the fault state probability is output through a fully connected layer and normalized exponential mapping.

[0016] The combined loss training module is used to design a combined loss function consisting of a classification error term, an intra-class clustering penalty term, and an inter-class balance weight term. The total loss is formed by weighted summation to guide network training, enabling the clustering of similar features and the separation of dissimilar features. The feature centers of each category are dynamically updated using exponential moving average.

[0017] The real-time diagnostic and early warning module is used to perform fault diagnosis on the real-time acquired signals using the trained network, output the state category, continuously acquire leakage current and update characteristics under steady state, and determine insulation abnormality or poor contact when the effective value or distortion characteristics exceed the normal reference range and issue an early warning.

[0018] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of a method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for diagnosing mechanical faults in disconnecting switches based on multi-source data fusion.

[0020] The beneficial effects of this invention are as follows: This invention, through multi-source signal acquisition and comparative verification of sensor installation locations, places the vibration sensor at the insulating basin metal flange where signal attenuation is minimized and the signal-to-noise ratio is highest, ensuring effective pickup of vibration and impact components from the source. Simultaneously, by synchronously acquiring vibration, power, and leakage current signals, it breaks down the information barriers between mechanical response, driving load, and static electrical state, laying a complete data foundation for multi-dimensional fusion diagnosis. By simultaneously applying continuous wavelet transform to generate time-frequency diagrams and variational mode decomposition to extract sideband features from the vibration signals, forming a dual-channel input of image and vector, the network can learn the spatiotemporal distribution of impact events from the time-frequency diagrams and capture weak modulation information from the sideband features. For the power signal, dynamic time warping and piecewise statistical feature extraction are combined to extract macroscopic waveform distortion... The system characterizes fault modes at two levels: load variation and local critical stage. It extracts effective values ​​and harmonics from leakage current, providing independent criteria for determining contact status and insulation level under static conditions, effectively preserving fault information at different levels of abstraction. In the multi-branch network structure, the combination of depthwise separable convolution and residual connections makes the model lightweight and suitable for edge deployment. The three branches process heterogeneous features separately before converging at the fusion layer, allowing the network to adaptively allocate weights based on the sensitivity of each signal to specific faults, eliminating the need for manually defined fusion rules. The combined loss function improves the misclassification cost of a few fault categories through class-balanced weights, while the central loss constrains the compactness of similar samples in the feature space, maintaining high overall diagnostic accuracy while retaining high recognition sensitivity for fault types with scarce samples. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0022] Figure 1 This is a flowchart illustrating the overall process of a mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion, provided in one embodiment of the present invention.

[0023] Figure 2 The swimlane diagram for the synchronous acquisition of multi-source signals in a method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion, provided in an embodiment of the present invention.

[0024] Figure 3 A schematic diagram of the feature matrix constructed by bispectral analysis in a method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion, provided in an embodiment of the present invention. Figure 4The flowchart shows a system scheme for a mechanical fault diagnosis system for disconnecting switches based on multi-source data fusion, provided as an embodiment of the present invention. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for diagnosing mechanical faults in disconnect switches based on multi-source data fusion is provided, comprising: S100: Collects vibration and drive power signals during the operation of the disconnecting switch, as well as ground leakage current signals when in steady state.

[0027] S200: Performs time-frequency transformation processing on the vibration signal to generate the first image, performs adaptive mode decomposition on the vibration signal, extracts sideband feature vectors from each mode component, segments the driving power signal according to time sequence, extracts time-domain statistical feature vectors of each segment, and obtains effective value features and waveform distortion features from the ground leakage current signal.

[0028] S300: Construct a multi-branch diagnostic network. The multi-branch diagnostic network merges the outputs of each branch and outputs the corresponding fault state category.

[0029] S400: The multi-branch diagnostic network is trained using a combined loss function, and fault diagnosis is performed on the real-time acquired signals based on the trained multi-branch diagnostic network.

[0030] It should be noted that this invention acquires vibration, power and leakage current information through synchronous acquisition of multi-source signals and optimized deployment of sensors. It constructs a hierarchical extraction system for vibration characterization with dual-channel time-frequency diagrams and sideband features, macroscopic and piecewise statistical features of power curves, and static leakage current features. Furthermore, it achieves a high-precision and high-sensitivity mechanical fault diagnosis scheme for disconnecting switches by adaptively fusing multi-source features through a lightweight multi-branch network and combining equilibrium-like and center loss constraints.

[0031] Example 2, refer to Figure 1 - Figure 3 This is a second embodiment of the present invention, which provides a method for diagnosing mechanical faults in disconnect switches based on multi-source data fusion, including: like Figure 2In step S100, the acquisition of vibration signals and drive power signals during the operation of the disconnecting switch, as well as the ground leakage current signal in steady state, includes steps S101 to S104: S101: Vibration signals are extremely sensitive to transient events such as mechanical collisions, friction, and jamming, and are the main basis for diagnosing contact engagement impacts and abnormal friction in transmission mechanisms.

[0032] In this embodiment, the vibration sensor specifically adopts an IEPE (piezoelectric integrated circuit) type accelerometer with a range of ±5g, a sensitivity of not less than 100mV / g, and a frequency response range of 0.5Hz~5kHz. The selection of the ±5g range is based on the fact that, according to preliminary experimental tests, the peak vibration acceleration at the insulating flange of a 126kV GIS disconnector switch during normal opening and closing operations typically does not exceed ±2g; under artificially simulated severe jamming fault conditions, the measured peak maximum impact acceleration is approximately ±4.2g. Therefore, the ±5g range can cover all normal and fault conditions, with a margin of approximately 20%.

[0033] The installation location of the sensor was determined through comparative experiments, which included selecting three candidate locations: the outer shell of the operating mechanism box, the top of the outer shell of the disconnect switch body, and the flange bolt of the stationary contact fixing insulation basin.

[0034] It should be noted that vibration signals were continuously collected at each location during 50 normal closing operations. The quantitative evaluation indicators included (1) signal-to-noise ratio, i.e., the ratio of the effective value of the signal to the effective value of the background noise; and (2) the attenuation rate of the first wave impact amplitude, with the operating mechanism as the benchmark. Experimental data showed that the average signal-to-noise ratio at the flange bolt of the insulating basin was 34.5dB, which was significantly higher than the 22.1dB at the operating mechanism and the 18.6dB at the outer shell. Moreover, the signal propagation path was the shortest at this location, and the attenuation rate of the first wave impact amplitude was only 12%, which was much lower than the 45% at the operating mechanism. Based on the above quantitative results, the long path and multi-interface reflection at the operating mechanism and the large energy dissipation caused by the thin-walled structure at the outer shell were excluded, and the insulating basin was selected as the sensor installation location.

[0035] The sensor is fixed to the bolts of the metal flange of the insulating basin by magnetic mounting or thread. The sampling rate is set to 10kHz or higher. In this embodiment, 20kHz is used to meet the Nyquist sampling requirements for high-frequency vibration components, such as wideband impacts caused by jamming. Each opening or closing operation triggers the acquisition and records the vibration waveform from the start to the end of the operation, denoted as v(t).

[0036] S102: The operating mechanism of the disconnecting switch is driven by a three-phase AC motor. The motor power is proportional to the motor output torque, and the torque is ultimately converted into resistance to overcome the contact system through components such as transmission rods, gears and racks.

[0037] When a jamming fault occurs, the frictional resistance increases, the motor load increases, and the power rises; when a transmission decoupling occurs, such as when the interphase connecting rod falls off, the load decreases and the power drops; when the opening and closing of the circuit breaker is not in place, the stroke ends prematurely, and the time length of the power curve is shortened. Therefore, the motor power signal can indirectly reflect the mechanical state of the entire transmission chain and contact system.

[0038] Power acquisition uses Hall voltage sensors and Hall current sensors to measure the motor line voltage U(t) and line current I(t) respectively, and calculates the instantaneous power P(t) = U(t)·I(t). For a three-phase AC motor, the total three-phase power is used, that is, the phase voltage and phase current of phases A, B, and C are collected separately, the instantaneous power is calculated, and the total power is obtained by summing them. The calculation formula is expressed as: in, This represents the total instantaneous power of the three phases. This represents the instantaneous value of phase voltage A. This represents the instantaneous value of the phase current in phase A. This is the instantaneous value of phase B voltage. This represents the instantaneous value of the phase current in phase B. This represents the instantaneous value of the C-phase voltage. This represents the instantaneous value of the C-phase current. It is a time variable.

[0039] The three-phase voltage and current signals are synchronously acquired by Hall sensors and sampled by the data acquisition card under the same clock trigger to ensure phase synchronization. The sampling rate is set to 2kHz, and the power signal frequency components are relatively low-frequency.

[0040] It should be noted that the frequency components of the power signal are mainly concentrated in the fluctuation frequency and harmonics of the motor output torque, which are usually below 500Hz. In addition, the low-pass filtering effect of the mechanical transmission mechanism, i.e. gears and connecting rods, means that the 2kHz sampling rate fully satisfies the Nyquist sampling theorem and can recover the fundamental frequency and main harmonic components within 1kHz. Therefore, it is not necessary to sample the vibration signal at a frequency above 10kHz. The acquisition trigger is also controlled by the opening and closing command.

[0041] S103: When the disconnecting switch is in the closed state, the high-voltage conductor is connected through the disconnecting switch contacts. There is distributed capacitance between the three-phase conductors and the metal casing, generating capacitive leakage current flowing to the ground. When the disconnecting switch is in the open state and the grounding switch is closed, the conductor is directly grounded, and the leakage current path changes. When the disconnecting switch is open and the grounding switch is not closed, the conductor is suspended, and the leakage current is very small. In addition, if the contacts are not in good contact or the insulating tray is damp, the amplitude and phase of the leakage current will also be distorted.

[0042] Therefore, a high-precision open-type Hall current sensor with a range of 0~100mA and an accuracy of 0.1mA is installed at the casing grounding down conductor. It can continuously monitor the grounding leakage current when the disconnecting switch is not activated. For three-phase common-box GIS, the casing grounding down conductor is usually a total grounding busbar after the three phases are combined. In this embodiment, the casing grounding down conductor specifically refers to the current position of the total grounding busbar. The measured value is the vector composite value of the three-phase leakage current.

[0043] The effective values ​​collected during stable operation are used as static judgment characteristics. Leakage current is still significant in the open state: when the circuit breaker is open and the grounding switch is not closed, the moving contact and conductor are at a floating potential, and the leakage current is minimal; if the grounding switch is mistakenly closed or there are defects in the insulation support, the leakage current will increase abnormally. By comparing the current differences between the open and closed states, the status of the grounding switch and the insulation condition can be determined. Therefore, static monitoring covers both open and closed steady states.

[0044] Under static conditions (i.e., when the device is not in operation), leakage current is collected every minute for a total of 50 times, with a sampling rate of 1kHz and a duration of 0.2 seconds.

[0045] It should be noted that 0.2 seconds corresponds to 10 complete cycles of a 50Hz power frequency, which is sufficient to accurately calculate the effective value and the fundamental / third harmonic amplitude ratio, i.e., the Fourier transform frequency resolution reaches 5Hz. It can accurately separate the 50Hz fundamental wave from the 150Hz third harmonic, fully meeting the requirements for static feature extraction.

[0046] The use of sampling at 1-minute intervals instead of continuous monitoring is mainly due to the storage and computing resource limitations of embedded systems. Continuous monitoring would accumulate a massive amount of redundant data; a 1-minute interval is sufficient to capture slow-changing latent faults such as contact oxidation or insulation dampness, while significantly reducing system power consumption.

[0047] S104: Uses a multi-channel synchronous data acquisition card to simultaneously acquire vibration, power and leakage current signals, with timestamp alignment. Each time a circuit breaker is opened or closed, a sample is generated, containing three-channel time series data. At the same time, the operation type (open / close) and the actual fault label (confirmed through manual setting or on-site inspection) are recorded.

[0048] In step S200, the multi-scale time-frequency analysis and modal decomposition of the vibration signal includes steps S201 to S205: S201: The time-frequency transformation process includes scaling and time-shifting the wavelet basis and convolving it with the vibration signal to obtain the time-frequency distribution matrix, and mapping the amplitude of the time-frequency distribution matrix into a color first image.

[0049] Vibration signal v(t) is a non-stationary transient signal with its frequency components changing rapidly over time. Traditional short-time Fourier transform is limited by a fixed window length and cannot simultaneously obtain high time resolution and high frequency resolution. However, continuous wavelet transform, through variable scaling and translation factors, can achieve multi-scale analysis and is particularly suitable for extracting the impact component in vibration signals.

[0050] Specifically, the time-frequency transformation process employs continuous wavelet transform, with the Morlet mother wavelet as the wavelet basis. This wavelet has a sinusoidal shape with a Gaussian envelope, similar to the waveform of a mechanical impact response. The Morlet wavelet formula is expressed as: in, For Morlet mother wavelet function, is the base of the natural logarithm. For the Gaussian envelope term, For carrier terms.

[0051] The formula for continuous wavelet transform of the vibration signal v(t) is expressed as: in, These are continuous wavelet transform coefficients. The original vibration signal, The scaling factor controls the wavelet scaling, corresponding to frequency information; This is the translation factor, corresponding to time information; The complex conjugate of the mother wavelet is used; the time-frequency distribution matrix is ​​obtained by scaling and time-shifting the wavelet basis and convolving it with the vibration signal.

[0052] S202: Under the condition of a sampling rate of 20kHz, the frequency range corresponding to the scale range 1~64 is approximately f. c / (2×scale), that is, from 20kHz / 128≈156Hz to 20kHz / 2≈10kHz, mainly covering the effective frequency band of vibration signals from 0.5kHz to 5kHz.

[0053] It should be noted that the selection of 64 scales is the optimal value after comprehensively considering frequency resolution and computational efficiency. According to comparative experiments, the frequency resolution is insufficient at 32 scales, and some sideband features are aliased; at 128 scales, the image size is too large (128×128) and the computation time increases by about 3 times, while the diagnostic accuracy is improved by less than 1%. 64×64 pixels achieve the optimal balance between resolution and computational load.

[0054] For each sample, the magnitude of the continuous wavelet transform coefficient matrix is ​​normalized and mapped to an RGB pseudo-color image, i.e., the first image, with a uniform size of 64×64 pixels. The one-dimensional vibration signal is converted into a two-dimensional time-frequency map, which serves as the input to the first branch of the subsequent deep network. The current time-frequency map intuitively shows the distribution of vibration energy on the time-frequency plane. Normal meshing impact is manifested as a bright spot concentrated in the low-frequency band with a narrow time width. However, jamming faults will lead to the enhancement of high-frequency broadband energy, which is manifested as continuous diffusion in the high-frequency region on the time-frequency map.

[0055] S203: Adaptive mode decomposition includes establishing a constraint problem based on the vibration signal with the objective of minimizing the sum of the estimated bandwidths of all modes, and solving the problem by the Lagrange multiplier method and alternating iterations to decompose the vibration signal into a preset number of modal components.

[0056] To further quantify the energy distribution and sideband characteristics of each frequency component in the vibration signal, this embodiment introduces variational mode decomposition (VMD). Unlike empirical mode decomposition, VMD adaptively decomposes the signal into a specified number K band-limited eigenmode functions by solving variational problems. It has a good mathematical theoretical basis and avoids mode aliasing.

[0057] It should be noted that adaptive mode decomposition specifically employs variational mode decomposition, assuming that the original vibration signal v(t) is decomposed into K modal components o. k (t), where the center frequency of each mode is ω. k The VMD algorithm iteratively updates each mode and its center frequency to minimize the sum of the estimated bandwidths of all modes. The constrained variational problem constructed is expressed as: in, To and Find the minimum value. It is a set of modal components. The set of center frequencies The number of modal decompositions. Let be the partial derivative with respect to time t. For the Dirac function, The imaginary unit, For Hilbert transform kernel, For convolution operations, For frequency shift term, Let be the center frequency of the k-th mode. The square of the L2 norm, For modal indexing.

[0058] The physical meaning of the objective function is for each mode o k(t) After obtaining the analytic signal through Hilbert transform, it is modulated to the baseband. The sum of the squares of the L2 norm of the gradient is minimized, that is, each mode is as compact as possible around its center frequency in the frequency domain. By introducing the quadratic penalty factor and Lagrange multipliers, the constrained problem is transformed into an unconstrained problem, and the alternating direction multiplier method is used to solve it in the frequency domain. By solving the problem through the Lagrange multiplier method and alternating iteration, the vibration signal is decomposed into a preset number of modal components.

[0059] It should be noted that in this embodiment, K is set to 16, which means that it is decomposed into 16 band-limited intrinsic mode function components, covering the frequency band from low frequency (<100Hz) to high frequency (>5kHz). After decomposition, each component represents the vibration mode of a specific frequency band.

[0060] It should be further explained that K=16 was determined through the center frequency observation method. Initially, K=10~20 was set, and the K value was gradually increased while observing the center frequency distribution of each mode. When K<14, some modes aliased in the frequency domain, and the center frequencies were too close. When K=16, the center frequencies of each mode were approximately uniformly distributed in the range of 0Hz~10kHz without overlap. When K was further increased to K=18 and above, excessive decomposition occurred, producing false low-frequency components. Sensitivity analysis showed that the model diagnosis accuracy fluctuated by <0.5% when K was in the range of 15~17. In this embodiment, the median value of 16 was taken. The optimal K value varies slightly under different fault types. For jammed faults, which are biased towards high frequencies, K needs to be slightly larger, while for normal faults, which are biased towards low frequencies, K needs to be slightly smaller. However, K=16 is taken as the globally optimal compromise value, taking into account all operating conditions. The second penalty factor is taken as the default value of 2000, and the Lagrange multiplier is taken as 0, that is, the noise tolerance is 0.

[0061] S204: Extracting sideband feature vectors includes calculating the correlation spectrum for each modal component under different frequency combinations, and statistically analyzing the energy concentration and broadening of the correlation spectrum within a preset frequency range on both sides of the center frequency to form a sideband feature vector.

[0062] like Figure 3 For each band-limited intrinsic mode function component, feature extraction is required to reflect the modulation characteristics. Mechanical jamming faults often cause the high-frequency carrier in the vibration signal to be modulated by low-frequency impulses, generating sideband frequencies. Bispectrum is the Fourier transform of triple integral, which can detect secondary phase coupling and nonlinear modulation, and is a powerful tool for analyzing mechanical modulation faults.

[0063] In this embodiment, the correlation spectrum under different frequency combinations specifically adopts a bispectral design, which is defined as follows: in, For signal o k (t) at frequency Fourier transform at the point, It is a bispectral function. As the first frequency variable, As the second frequency variable, For signal o k (t) at frequency Fourier transform at the point, For the signal at frequency The complex conjugate of the Fourier transform at that point, As the mathematical expectation operator, the bispectrum is a two-dimensional function that reflects frequency. and The degree of coupling between them.

[0064] Furthermore, to simplify the feature dimension, a sideband feature extraction method is proposed. For each modality, the center frequency is ω. k Calculate the sideband regions on both sides of the center frequency ([ω k -Δf, ω k The sum of the bispectral energy within the range of +Δf], and the preset frequency range is specifically taken as Δf=200Hz.

[0065] It should be noted that the basis for Δf=200Hz is that the motor speed of the disconnector operating mechanism is usually between 1450 and 3000 rpm, corresponding to a rotational frequency of about 24 to 50 Hz. The fault modulation sideband interval is generally 1 to 4 times the rotational frequency, that is, not exceeding 200 Hz. According to the statistics of the bispectral energy distribution of different modes, the sideband energy is mainly concentrated in the range of ±200 Hz of the center frequency. The sideband energy outside the current range accounts for less than 5% of the total sideband energy and can be ignored. Different modes can uniformly use Δf=200Hz, because the modulation frequency depends on the motor rotational frequency rather than the carrier frequency.

[0066] S205: The energy concentration of the statistical correlation spectrum is specifically expressed by the sideband energy ratio, which is given by the following formula: in, Let be the sideband energy ratio of the k-th mode, which is dimensionless. The center frequency of the k-th mode, in Hz. The width of the sideband. This is the lower limit frequency of the sideband region. This represents the upper limit frequency of the sideband region.

[0067] The broadening of the statistical correlation spectrum is specifically measured by the sideband width, defined as the width from the center frequency ω. k The single-side bandwidth decreases to 1 / e of the peak value in the bispectral amplitude; when there is a jamming fault, the sideband energy of the high-frequency mode will increase significantly and the sideband width will be lengthened. In this embodiment, the sideband energy ratio and sideband width of each mode are extracted to form a vibration feature vector, specifically a 16×2=32-dimensional feature vector.

[0068] The extracted sideband feature vectors, i.e. vibration feature vectors, are combined with the original time-frequency image to form the multi-channel input of the first branch, i.e., image + feature vector, which enhances the network's expressive power. Specifically, in the subsequent input layer of the first branch, the time-frequency image is used as the image channel input, and the sideband feature vectors are used as auxiliary feature channels for integration in the channel dimension.

[0069] Furthermore, in step S200, obtaining the effective value features and waveform distortion features includes steps S211~S213: S211: The motor power signal q(t) is relatively smooth compared to the vibration signal, but the fault information it contains is mainly reflected in the waveform shape and statistical characteristics. This embodiment uses two methods to process the power signal: one is dynamic time warping for similarity matching; the other is time-domain statistical feature extraction for network input.

[0070] A1: The fingerprint database selection criteria must meet three conditions simultaneously: the equipment has been confirmed to be in good mechanical condition after power outage maintenance, the power curve is consistent with the shape of the historical standard curve (i.e., correlation coefficient > 0.95), and there are no abnormal impacts in the vibration signal; the number of normal curves n in the fingerprint database is 30 to 50, and in this embodiment, 40 are specifically selected to cover normal fluctuations under different seasons and different operating voltages; the fingerprint database adopts a sliding window update strategy: for every 10 newly added power curves diagnosed as "normal", the 10 oldest curves in the database are removed to keep the database capacity constant, so that the fingerprint database can adapt to the normal drift caused by long-term wear and tear of the equipment.

[0071] The power curves of different opening and closing operations may have slight scaling on the time axis. Euclidean distance cannot accurately measure their similarity. DTW calculates the shortest distance between two sequences through nonlinear regularization, which can tolerate time axis offset.

[0072] It should be noted that the power curve fingerprint library Q under normal conditions was constructed. normal ={q1(t), q2(t), ..., q n For the power signal q to be tested, (t)} test Calculate the DTW distance with each curve in the fingerprint database, and take the average distance as the anomaly index. The DTW distance formula is as follows: in, Let be the dynamic time-warped distance between sequences q and h. The power signal sequence to be tested. This is a normal power curve from the fingerprint database. To align the path, and For sequence index, For paired points, Let be the magnitude of the d-th point in sequence q. Let g be the amplitude of the g-th point in sequence h. When the DTW distance exceeds the threshold, which is three times the standard deviation of the normal mean, the power curve is judged to be abnormal. However, DTW can only give whether it is abnormal or not, and cannot refine the fault type. Further feature extraction is needed for deep learning.

[0073] A2: In order to capture local details in the power curve, this embodiment divides the power signal into 4 segments at equal time intervals, corresponding to the start-up stage, before contact engagement, engagement stage, and end-of-stroke buffer stage, respectively. The division into 4 segments is based on the typical action sequence of the disconnector switch operating mechanism: the start-up stage is from motor start-up to the start of movement of the moving contact, the no-load stroke stage is before contact engagement, the contact engagement stage is the interaction between the moving and stationary contacts, and the end-of-stroke buffer stage is the buffer action at the end of the stroke.

[0074] Equal time interval segmentation is the preferred simplification scheme, that is, taking the total action time T, the four segment boundaries are 0~T / 4, T / 4~T / 2, T / 2~3T / 4, and 3T / 4~T. More precise differentiation can be achieved by combining stroke sensors to determine the boundaries of each stage. However, in order to reduce the number of sensors, this embodiment adopts equal interval simplification. The segmentation schemes for opening and closing operations are symmetrical, both divided into the above four segments, only the action directions are opposite.

[0075] For the entire curve and each segment, the following statistical features are extracted to form a time-domain statistical feature vector: (1) Peak power This reflects the maximum load and increases significantly when there is jamming.

[0076] in, Peak power, This is the maximum value operator. This is an instantaneous power signal.

[0077] (2) Average value This reflects average energy consumption, among which, This is the average power value. This represents the total motion time.

[0078] (3) Waveform factor , It is used to detect changes in waveform flatness, where, For waveform factor, This is the effective value of power. It is the square of the instantaneous power.

[0079] (4) Sample entropy SampEn(m, r) measures the sequence complexity. When stuck, the sample entropy increases due to the increase of frictional noise. In this embodiment, the embedding dimension m is 2 and the similarity tolerance r is 0.2 times the signal standard deviation.

[0080] The value is the standard recommended parameter for the sample entropy algorithm, which is widely used in the complexity analysis of mechanical vibration signals and can effectively distinguish between regular impact signals (normal) and random friction noise signals (jamming fault).

[0081] (5) Rise time t rise The time from 10% peak to 90% peak reflects the startup response.

[0082] (6) Amplitude P of meshing wave peak meshing The height of the characteristic wave peak generated when the contacts just make contact.

[0083] By concatenating all features, a power feature vector is obtained, which has a dimension of 4×5+2=22. Specifically, the peak power, average value, waveform factor, and sample entropy of each of the four segments total 4×4=16 dimensions. Adding the rise time and meshing peak amplitude of the entire curve (2 dimensions), and the peak power, average value, waveform factor, and sample entropy of the entire curve (4 dimensions), the total is 22 dimensions.

[0084] S212: For static operating conditions, the ground leakage current waveform is sampled every 1 minute at a sampling rate of 1kHz for 0.2 seconds. The effective value, i.e., the effective value characteristic, is calculated using the following formula: in, This is the effective value of the grounding leakage current. This is the instantaneous value of the ground leakage current. It is the square of the instantaneous value of the leakage current.

[0085] Simultaneously calculate the fundamental frequency. With the amplitude of the third harmonic The ratio is the waveform distortion characteristic. ;when Exceeding the normal range (normal value is 5-10mA, exceeding by 20%) and When the value increases, it indicates that there may be an increase in contact resistance or moisture in the insulation. This feature is used as the input scalar of the third branch.

[0086] It should be noted that in this embodiment, the effective value and harmonic ratio are used as the core input features of the third branch, but not the only features; the network input is only 2-dimensional because the information has been highly purified in the preprocessing stage: the effective value reflects the energy loss caused by the change in contact resistance, and the harmonic ratio reflects the degree of nonlinear distortion. The two complement each other and can characterize most static insulation / contact faults.

[0087] To improve sensitivity, phase angle change is introduced into the warning logic. If the acquisition card supports it, it can be extracted as an auxiliary criterion without changing the network structure. The network can still classify effectively under 2D input and avoid overfitting caused by high-dimensional sparse features.

[0088] S213: After the above steps, each operation sample obtains the following preprocessing results: the time-frequency diagram (i.e., the first image) obtained from the vibration signal through continuous wavelet transform, 64×64×3, and the 32-dimensional sideband feature vector obtained through variational mode decomposition and bispectral analysis; the 22-dimensional time-domain statistical feature vector obtained from the power signal through segmented statistics; and the effective value and harmonic ratio calculated from the leakage current signal, totaling 2 dimensions.

[0089] All preprocessing results will be input into the corresponding branches of the multi-branch diagnostic network in step S300.

[0090] In step S300, the fault state category to which the output belongs includes, in order to achieve high-precision fault classification while avoiding the degradation problem and overfitting of deep networks, a multi-branch network structure is designed to process different types of extracted features separately and fuse them in the later stage, specifically including steps S301~S305: S301: The total number of parameters in the multi-branch network is approximately 0.82M, of which the first branch is approximately 0.75M, the second branch is approximately 0.05M, and the third branch is approximately 0.02M. Under the same input size, the number of parameters in the 7-module residual network constructed using standard 3×3 convolution is approximately 7.2M. In comparison, the number of parameters in this embodiment is reduced to approximately 1 / 9. On the ARM Cortex-A53 edge processor with a main frequency of 1.2GHz, the measured single inference time is approximately 35ms, and the memory usage is approximately 45MB, which meets the requirements of real-time diagnostics.

[0091] S302: The first branch vibration time-frequency image processing branch includes receiving a 64×64×3 color time-frequency image and a 32-dimensional vibration feature vector as input; in the input layer, the time-frequency image, i.e. the first image, and the sideband feature vector are integrated in terms of channel dimension to form multi-channel input data.

[0092] Specifically, the time-frequency graph serves as the main channel (64×64×3), and the sideband feature vectors are copied and expanded to serve as auxiliary feature channels. These are then spliced ​​together along the channel dimension to form an enhanced input.

[0093] The network structure draws inspiration from residual networks and depthwise separable convolutions, containing seven convolutional processing modules. In each module, standard convolutions are replaced with depthwise separable convolutions to significantly reduce the number of parameters. The standard convolution is decomposed into independent spatial filtering by channel and pointwise linear combination across channels, specifically depthwise separable convolutions, which consist of two steps: (1) Channel-by-channel convolution: Each input channel uses an independent two-dimensional convolution kernel, and the number of output channels is equal to the number of input channels. The computational cost is... ,in, Where M is the kernel size and M is the number of input channels. The size of the feature map.

[0094] (2) Pointwise convolution: A 1×1 convolution is used to map the number of channels from M to N, where N is the number of output channels. This is a pointwise linear combination across channels, and the computational cost is... .

[0095] Furthermore, the total computational cost of depthwise separable convolution is: The computational cost of standard convolution is: The ratio of the two is: when When N is large, such as N=128, the ratio is about 1 / 9, that is, the computational cost is reduced to 1 / 9 of that of standard convolution. In this embodiment, all convolutional layers use 3×3 depth separable convolution.

[0096] To prevent gradient vanishing and network degradation, direct connections are added between the input and output of the modules, and residual connections are added to each convolutional module. The basic form of the residual block is as follows: in, Let be the input feature map of the l-th convolutional module. This is the output feature map of the l-th convolutional module. For the residual function, This is the set of learnable weight parameters in the l-th module.

[0097] It should be noted that residual connections enable the network to learn identity mappings, allowing the construction of deeper networks without performance saturation and degradation. Specifically, residual connections involve adding the input of each module to the module output element-wise through skip connections and feeding it into the next module. Modules 1-4 are responsible for downsampling and channel expansion, while modules 5-7 further abstract the deep feature space.

[0098] The Batch Normalization (BN) layer is located after the convolution operation and before the activation function, i.e., input → convolution → batch normalization → ReLU → output. During the training phase, the BN layer uses the mean and variance of the current batch, and during the inference phase, it uses the mean and variance of the global moving average recorded during training. All three branches use the ReLU activation function formula F(x) = max(0, x), which is simple to calculate and can effectively alleviate gradient vanishing. Different branches do not distinguish activation functions to ensure training consistency.

[0099] The feature map after convolution is compressed into a one-dimensional (256-dimensional) form by global average pooling, and the first feature vector is obtained through fully connected mapping. 128 dimensions.

[0100] It should be noted that when the sideband feature vector is input into the multi-branch diagnostic network, it is not simply concatenated with the first image. Instead, it is first converted into a feature map with the exact same spatial dimensions as the first image through spatial dimension replication and expansion, and then integrated in the channel dimension. Specifically, the sideband feature vector contains 32 dimensions, which are composed of two features: the sideband energy ratio and the sideband width extracted from each of the 16 modal components.

[0101] During the transformation, each scalar value in the 32-dimensional vector is independently copied and expanded into a 64×64 two-dimensional matrix. This means each scalar value appears repeatedly in all 4096 spatial locations of the matrix, generating a complete auxiliary feature map for each scalar value. The first image itself is a 64×64 three-channel color time-frequency image. Each of these 32 auxiliary feature maps is treated as an independent auxiliary channel and concatenated with the original three channels in the channel dimension, ultimately forming a 64×64×35 dimensional multi-channel input data. The first three channels are the RGB channels of the original time-frequency image, and the latter 32 channels are the auxiliary feature channels formed by copying and expanding the elements of the sideband features.

[0102] In this structure, since all spatial locations of each auxiliary feature map have the same value, each element of the sideband feature vector has a one-to-one correspondence with each spatial location of the current auxiliary feature map. That is, each feature value provides the same feature response at all pixels. This design enables subsequent convolutional layers to obtain both the local texture information of the time-frequency map corresponding to that location and the global sideband modulation information when processing each spatial location, thereby achieving effective parallel input and fusion of local spatial features and global spectral features.

[0103] S303: The second branch, the power feature processing branch, includes a 22-dimensional power feature vector as input and a temporal statistical feature vector extracted in step S200. Due to the low feature dimension, convolution is not used; instead, two fully connected layers are used for non-linear mapping: the first layer maps 22 dimensions to 64 dimensions, and the second layer maps 64 dimensions to 32 dimensions. Each layer is followed by ReLU activation and Dropout (0.3), ultimately outputting a 32-dimensional second feature vector. .

[0104] S304: The third branch, the static leakage current processing branch, includes inputs of scalar (RMS) features and harmonic (ratio waveform distortion) features, totaling 2 dimensions. These are mapped to 8 dimensions through a fully connected layer, outputting the third feature vector. .

[0105] S305: Concatenate the output feature vectors of the three branches, that is, concatenate the first feature vector, the second feature vector, and the third feature vector to obtain the fused feature vector: The dimensions are 128+32+8=168.

[0106] A fully connected layer maps the data to 7 dimensions, corresponding to 7 states. Softmax normalization is then used to normalize the exponential mapping, outputting the probability values ​​for each preset fault state category. The formula is as follows: in, Let w be the predicted probability of the w-th type of fault state. For category index number, The output of the w-th neuron in the fully connected layer. To Take the natural index, The index number is used for summation.

[0107] It should be noted that the seven diagnostic states include: Normal closing: The opening and closing command is closed, the actual contacts are reliably closed, the power curve waveform is normal, and the vibration time-frequency diagram shows a clear single impact.

[0108] Normal tripping: The tripping command is executed in place, the power curve is normal, and the vibration signal shows obvious impact during the tripping process.

[0109] Closing jamming: During the closing process, the frictional resistance increases abnormally, the power peak rises by more than 20% above the normal value, the high-frequency component of the vibration signal (>3kHz) increases in energy, and the VMD sideband widens.

[0110] Opening jamming: Similar to closing jamming, but occurs in the opening direction.

[0111] Incomplete closing: Due to premature action of the limit switch or mechanical jamming, the moving contact does not reach the rated position, the power curve time is shortened, and the time interval from the initial impact to the end of the movement in the vibration signal is less than 80% of the normal value.

[0112] Incomplete tripping: similar to the previous situation, the contacts did not completely separate after tripping.

[0113] Transmission decoupling: The interphase connecting rod falls off or the gear strips, causing one or two phases to stop operating. The overall amplitude of the power curve drops by more than 30%, the operating time is shortened, and the vibration signal is weak and there is no typical impact.

[0114] It should be further noted that the power peak increase exceeding 20% ​​is based on statistical analysis of power peak fluctuations under 50 sets of normal operation. The 3σ upper limit is 115% of the normal average, rounded to 20% as a safety margin to distinguish anomalies. In the increase of high-frequency components, i.e., energy >3kHz, 3kHz is the inherent upper limit of normal meshing impact. The measured main frequency of normal meshing impact is concentrated in 0.5~2.8kHz. The jamming fault will excite energy in the 3~6kHz frequency band. The judgment threshold is that the current frequency band energy exceeds 3 times the normal average standard deviation. In the case of the power curve time shortening less than 80% of the normal value, 80% corresponds to the lower limit of the rated stroke setting of the micro switch. That is, when the stroke is less than 80% of the rated stroke, the mechanical characteristics have obviously deviated from the normal. In the case of the amplitude decrease exceeding 30%, the transmission decoupling causes the load torque to decrease to 60%~70% of the normal value. After multiple simulation tests, the median value of 70%, i.e., a decrease of 30%, is taken as the criterion.

[0115] In step S400, training the multi-branch diagnostic network using the combined loss function includes steps S401 to S406: S401: In fault diagnosis tasks, the commonly used cross-entropy loss function, when used alone, can only minimize the distribution difference between the prediction and the label, and cannot directly control the intra-class distance and inter-class distance. Moreover, when the number of samples of each class in the training set is unbalanced, for example, there are 150 normal samples and only 50 faulty samples, the model will be biased towards predicting the majority class, resulting in a low fault detection rate. A combined loss function is proposed, which includes three terms: classification error term, i.e., cross-entropy loss; intra-class clustering penalty term, i.e., center loss; and inter-class balance weight term, i.e., class balance loss.

[0116] S402: The classification error term, also known as the cross-entropy loss, is calculated based on the cross-entropy between the probability value and the true label, and is defined as follows: in, This represents the cross-entropy loss value. For batch size, This refers to the sample index number in the batch. The total number of fault state categories, J=7. For category index number, One-hot encoding of the real label. To predict probabilities for the model, The loss is the natural logarithm of the predicted probability, which encourages the model to classify correctly.

[0117] S403: The intra-class clustering penalty term, also known as the center loss, is obtained by calculating the mean squared Euclidean distance between the fused feature vector and the feature centers of its respective class. The center loss is defined as follows: in, The center loss value, For the first The fused feature vector of each sample (features before Softmax). For the first The feature centers of each sample's category; during training, the feature centers of each category are updated using an exponential moving average method based on the fused feature vectors of each batch of samples, specifically as momentum updates. The momentum update formula is: in, For the first The first round of updates Class feature center, For the first The first round of updates Class feature center, For the first Update gradient of class center, momentum coefficient Take 0.95, For iteration round index, For the first The true class label of each sample For the category number that is being updated, Dirac indicator function; feature center The initial value is set to a 0 vector, and the update frequency is once per batch, so that the center converges quickly, rather than being updated every epoch.

[0118] The center loss forces features of similar samples to cluster around the center, thereby reducing intra-class distance and indirectly increasing inter-class distance. Experiments show that after adding the center loss, the confusion between fault categories is significantly reduced.

[0119] S404: The inter-class balance weight term, also known as the class balance loss, is determined by the number of samples in each class. To address imbalanced samples, class weights are introduced based on the cross-entropy. These weights are inversely proportional to the square root of the number of samples in each class. in, For the first The number of samples of a class in the training set. For the first The category weights for each type of fault state are calculated by multiplying the weights of each category by the corresponding classification error term and summing the results to form a weighted classification error term.

[0120] Weight Before training begins, the class distribution based on the entire training set is pre-calculated to maintain the stability of the training process. If the number of samples of a certain class... , weight Set it to 0, and invalidate the current term in the loss calculation of the current batch, i.e., do not participate in gradient backpropagation. At the same time, in practical applications, avoid training the model in the extreme case where it does not contain any samples of a certain type.

[0121] The class-balanced cross-entropy loss, i.e., the weighted classification error term, is: in, To balance the cross-entropy loss value, the loss increases the cost of misclassification for rare fault categories, forcing the model to focus more on fault samples. S405: The combined loss function is the sum of the weighted classification error term and the intra-class clustering penalty term. The total loss is obtained by weighted summation of these three terms. in, This represents the total combined loss. and To balance the hyperparameters, this embodiment selects... , The optimal performance is achieved when λ=0; when λ=0, the center loss has no effect and the intra-class distance is large; when λ is too large, such as 0.01, the model focuses too much on compressing the intra-class distance and ignores the inter-class separation, resulting in a decrease in accuracy.

[0122] It should be noted that the reason λ takes a minimum value of 0.0005 is due to the central loss. The magnitude is usually in the range of 10 0 ~10 1 Between, and cross-entropy loss The order of magnitude is approximately 10 -1 ~10 0If λ is too large, the center loss dominates the training, forcing the model to over-compress intra-class distances and sacrifice classification boundaries. Through parameter scanning experiments, when λ=0.0005, the model's overall accuracy on the validation set reached 97.2%; when λ=0 (no center loss), the accuracy was 92.6%; when λ=0.01, the accuracy dropped to 89.3% due to insufficient inter-class separation. Similarly, the selection of [a specific factor] balances the contributions of the three losses.

[0123] The optimizer uses Adam with an initial learning rate of 1×10⁻⁶. -3 The learning rate is adjusted using a stepped decay strategy, with a value of 0.5 every 30 epochs. The batch size is 32, and the number of training epochs is 100. An early stopping strategy is adopted, and training is stopped when the validation set loss does not decrease for 10 consecutive epochs. The weights are initialized using He initialization, which is suitable for the ReLU activation function. Data augmentation is performed by randomly shifting the time-frequency graph horizontally by ±5 pixels and randomly rotating it by ±5° to enhance the model's generalization ability.

[0124] S406: Divide the dataset into a ratio of training set:validation set:test set = 7:1.5:1.5, that is, the training set accounts for 70%, the validation set accounts for 15%, and the test set accounts for 15%. The validation set is used for hyperparameter tuning and early stopping judgment, and the test set is only used for final performance evaluation to ensure the objectivity of the results. The distribution of each category in each set is maintained in a proportion consistent with the original dataset by stratified sampling.

[0125] Overall diagnostic accuracy: 97.2%.

[0126] Recall rates for each category: normal closing 98.1%, normal opening 97.6%, closing jamming 92.3%, opening jamming 91.8%, incomplete closing 96.2%, incomplete opening 95.7%, and transmission decoupling 97.9%.

[0127] The F1 scores for each category are 0.98, 0.97, 0.93, 0.92, 0.96, 0.96, and 0.97, respectively.

[0128] Confusion matrix: The main misjudgments occurred between "closing jam" and "opening jam", with a confusion rate of about 4.2%, and between "closing incomplete" and "opening incomplete", with a confusion rate of about 2.8%. The confusion rate among the remaining categories was <1%.

[0129] Compared with existing methods: On the same dataset, the overall accuracy of traditional support vector machine method is 78.3%, and that of standard convolutional neural network is 86.7%. The accuracy of the method of this invention is improved by 18.9% and 10.5% respectively, and the recall rate is improved by an average of about 25% for fault classes.

[0130] Furthermore, in step S400, the fault diagnosis of the real-time acquired signal includes steps S411~S413: S411: Based on the trained multi-branch diagnostic network, fault diagnosis is performed on the real-time acquired signals. The test samples are converted into network input according to the above preprocessing process. The probability distribution of 7 categories is obtained through forward propagation, and the category with the highest probability is taken as the diagnosis result. At the same time, the intermediate features of each branch are output for interpretability analysis, that is, the feature distribution is visualized by t-SNE dimensionality reduction.

[0131] S412: During the steady state of the disconnecting switch, the ground leakage current signal is continuously collected, and the effective value characteristics and waveform distortion characteristics are updated after each collection. When the effective value characteristics and waveform distortion characteristics exceed the preset normal reference range in a preset number of consecutive measurements, an abnormality is determined and an early warning is issued.

[0132] Specifically, when the effective value of static leakage current deviates from the normal range for 10 consecutive measurements, such as >12mA or <3mA and the equipment does not have an operation command, an "insulation abnormality or poor contact" warning is issued; ±20% is based on the statistical distribution of the effective value of leakage current under normal conditions, with a mean of ±2.5 times the standard deviation, corresponding to an approximately 99% confidence interval; 0.15 is based on the maximum observed value of harmonic ratio under normal conditions, normally ≤0.1, and is set after considering margin; the normal leakage current reference value is different for different voltage levels and environments, but the relative fluctuation ratio of ±20% and the absolute value threshold of harmonic ratio of 0.15 are universal.

[0133] It should be noted that the detailed rules for 10 consecutive measurements include: if at least 8 of the most recent 10 measurements exceed the warning threshold, an alarm is triggered. The time window is 10 sampling points counted from the first abnormality, with each sampling point spaced 1 minute apart, for a total window of approximately 10 minutes. If the abnormality is not reached 8 times and the system returns to normal, the counter is reset and the count restarts. The current rules effectively prevent false alarms caused by external interference in a single measurement.

[0134] S413: In this embodiment, under static conditions, when the disconnecting switch is closed and making good contact, the effective value of the ground leakage current is measured to be 6.2mA and the harmonic ratio is 0.08. After artificially reducing the contact pressure of the A-phase contact by 30%, the leakage current drops to 4.1mA and the harmonic ratio rises to 0.25. The system successfully issues an "A-phase poor contact warning", which is consistent with the actual situation.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0136] Example 3, referring to Figure 4 This is the third embodiment of the present invention. This embodiment provides a mechanical fault diagnosis system for disconnecting switches based on multi-source data fusion, including a multi-source feature extraction module, a diagnostic network construction module, a combined loss training module, and a real-time diagnosis and early warning module.

[0137] The multi-source feature extraction module is used to collect vibration, power and leakage current signals of the disconnecting switch, perform time-frequency transformation and adaptive mode decomposition on the vibration signal to obtain time-frequency image and sideband features, extract statistical features of the power signal in time sequence segments, and obtain effective value and distortion features from the leakage current to provide multi-source input for diagnosis.

[0138] The diagnostic network construction module is used to construct a multi-branch diagnostic network. The first branch fuses vibration images and sideband features, and extracts features through separable convolution and residual structure. The second and third branches process power statistics features and leakage current features respectively. The outputs of each branch are spliced ​​together, and the fault state probability is output through a fully connected layer and normalized exponential mapping.

[0139] The combined loss training module is used to design a combined loss function consisting of a classification error term, an intra-class clustering penalty term, and an inter-class balance weight term. The total loss is formed by weighted summation to guide network training, enabling the clustering of similar features and the separation of dissimilar features. The feature centers of each category are dynamically updated using exponential moving average.

[0140] The real-time diagnostic and early warning module is used to perform fault diagnosis on the real-time acquired signals using the trained network, output the state category, continuously acquire leakage current and update characteristics under steady state, and determine insulation abnormality or poor contact when the effective value or distortion characteristics exceed the normal reference range and issue an early warning.

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0142] Example 4, the fourth embodiment of the present invention, differs from the previous three embodiments in that: If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0144] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0145] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

Claims

1. A method for diagnosing mechanical faults in disconnect switches based on multi-source data fusion, characterized in that: include, The vibration signal and drive power signal during the operation of the disconnecting switch, as well as the ground leakage current signal when in steady state, are collected. The vibration signal is processed by time-frequency transformation to generate the first image. Adaptive mode decomposition is performed on the vibration signal to extract the sideband feature vectors from each mode component. The driving power signal is segmented according to time sequence, and the time domain statistical feature vectors of each segment are extracted. The effective value features and waveform distortion features are obtained from the ground leakage current signal. A multi-branch diagnostic network is constructed, which merges the outputs of each branch and outputs the corresponding fault state category. A multi-branch diagnostic network is trained using a combined loss function, and fault diagnosis is performed on the real-time acquired signals based on the trained multi-branch diagnostic network.

2. The method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion as described in claim 1, characterized in that: The time-frequency transformation process includes scaling and time-shifting the wavelet basis and convolving it with the vibration signal to obtain the time-frequency distribution matrix, and mapping the amplitude of the time-frequency distribution matrix to a color first image. The adaptive mode decomposition includes establishing a constraint problem based on the vibration signal with the objective of minimizing the sum of the estimated bandwidths of all modes, and solving the problem by the Lagrange multiplier method and alternating iterations to decompose the vibration signal into a preset number of modal components. The extraction of sideband feature vectors includes calculating the correlation spectrum for each modal component under different frequency combinations, and statistically analyzing the energy concentration and broadening of the correlation spectrum within a preset frequency range on both sides of the center frequency to form a sideband feature vector.

3. The method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion as described in claim 2, characterized in that: The multi-branch diagnostic network includes a first branch, a second branch, and a third branch; The first branch includes integrating the first image and the sideband feature vector at the input layer along the channel dimension to form multi-channel input data; The multi-channel input data is feature abstracted by multiple convolutional processing modules arranged in sequence. Each convolutional processing module adopts the operation of decomposing the standard convolution into independent spatial filtering by channel and point-by-point linear combination across channels, and adds a direct connection between the input and output of the module. The feature map after convolution is compressed into a one-dimensional form by pooling, and the first feature vector is obtained by fully connected mapping.

4. The method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion as described in claim 3, characterized in that: The second branch includes inputting the time-domain statistical feature vector into at least one fully connected layer and performing nonlinear mapping to obtain a second feature vector; The third branch includes inputting the effective value features and waveform distortion features into the fully connected layer, performing nonlinear mapping, and obtaining the third feature vector.

5. The method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion as described in claim 4, characterized in that: The fault state category to which the output belongs includes concatenating the first feature vector, the second feature vector, and the third feature vector to obtain a fused feature vector; The fused feature vector is input into at least one fully connected layer, and the probability value corresponding to each preset fault state category is output through a normalized exponential mapping operation.

6. The method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion as described in claim 5, characterized in that: The combined loss function includes a classification error term, an intra-class clustering penalty term, and an inter-class balance weight term; The classification error term is calculated based on the cross-entropy between the probability value and the true label. The intra-class clustering penalty term is obtained by calculating the mean squared Euclidean distance between the fused feature vector and the feature center of its class. The inter-class balance weight term is determined by the number of samples in each class. The weights of each class are multiplied by the classification error term of the corresponding class and summed to form the weighted classification error term. The combined loss function is the sum of the weighted classification error term and the intra-class clustering penalty term.

7. The method for diagnosing mechanical faults of disconnecting switches based on multi-source data fusion as described in claim 6, characterized in that: The training of the multi-branch diagnostic network includes updating the feature centers of each category during the training process by fusing feature vectors based on each batch of samples using an exponential moving average method. During the steady state of the disconnecting switch, the ground leakage current signal is continuously collected, and the effective value characteristics and waveform distortion characteristics are updated after each collection. When the effective value characteristic and waveform distortion characteristic both exceed the preset normal reference range in a preset number of consecutive measurements, an anomaly is determined and an early warning is issued.

8. A mechanical fault diagnosis system for disconnecting switches based on multi-source data fusion, employing the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion as described in any one of claims 1 to 7, characterized in that, It includes a multi-source feature extraction module, a diagnostic network construction module, a combined loss training module, and a real-time diagnostic and early warning module; The multi-source feature extraction module is used to collect vibration, power and leakage current signals of the disconnecting switch, perform time-frequency transformation and adaptive mode decomposition on the vibration signal to obtain time-frequency image and sideband features, extract statistical features of the power signal in time sequence segments, and obtain effective value and distortion features from the leakage current to provide multi-source input for diagnosis. The diagnostic network construction module is used to construct a multi-branch diagnostic network. The first branch fuses vibration images and sideband features, and extracts features through separable convolution and residual structure. The second and third branches process power statistics features and leakage current features respectively. The outputs of each branch are spliced ​​together, and the probability of the fault state is output through a fully connected layer and normalized exponential mapping. The combined loss training module is used to design a combined loss function consisting of a classification error term, an intra-class clustering penalty term, and an inter-class balance weight term. The total loss is formed by weighted summation to guide network training, enabling the clustering of similar features and the separation of dissimilar features. The exponential moving average is used to dynamically update the feature centers of each category. The real-time diagnostic and early warning module is used to perform fault diagnosis on the real-time acquired signals using the trained network, output the state category, continuously acquire leakage current and update characteristics under steady state, and determine insulation abnormality or poor contact when the effective value or distortion characteristics exceed the normal reference range and issue an early warning.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the mechanical fault diagnosis method for disconnecting switches based on multi-source data fusion as described in any one of claims 1 to 7.

Images