Method and system for evaluating the corrosion state of an electrical contact switch contact

Abstract

The application discloses a kind of electric contact switch contact corrosion state evaluation method and system, belong to electric power equipment state monitoring and fault diagnosis technical field.The method includes: obtaining the multi-source feature data of electric contact switch, including opening and closing times, contact resistance, arc electrical characteristics and vibration signal;Tensor is constructed;The input tensor is input to the pre-trained deep learning model, the model uses dual-channel time series encoder to extract deep features, the first channel is based on Mamba to capture time domain degradation trend, and the second channel is based on Fourier neural operator to enhance frequency domain features;Determine the contact corrosion grade after fusing dual-channel features.The application realizes the complementary enhancement of time domain and frequency domain features through dual-channel architecture, improves the sensitivity and evaluation accuracy of early abnormal signs of contact corrosion, and is suitable for online monitoring and predictive maintenance of electric power switch equipment.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, and in particular to a method and system for assessing the corrosion status of electrical contact switch contacts. Background Technology

[0002] Electrical contact switches (including high-voltage switches, circuit breakers, contactors, etc.) are critical components in power systems that require frequent or periodic operation under load. Their contacts are subjected to the combined effects of arc erosion, mechanical wear, and chemical corrosion during opening and closing. Contact deterioration is one of the main causes of switchgear failures and even power system outages. Statistics show that switchgear failures account for a significant proportion of power grid accidents; therefore, timely and accurate assessment of contact corrosion status is crucial for ensuring the safe operation of the power grid. Currently, scholars and engineers both domestically and internationally have developed various contact condition assessment techniques, including dynamic contact resistance measurement, vibration acoustic analysis, drive motor current analysis, infrared thermal imaging, and dissolved gas analysis in oil. However, all of these methods have limitations to varying degrees: dynamic resistance measurement requires power outages and cannot achieve continuous online monitoring; vibration acoustic analysis is sensitive to sensor installation location and easily affected by environmental noise; drive motor current analysis lacks direct sensitivity to contact erosion; infrared thermal imaging is insensitive to instantaneous high temperatures and is greatly affected by environmental factors; and dissolved gas analysis in oil exhibits significant time lag. Existing single-feature methods cannot fully utilize the nonlinear coupling relationship between multi-source features, making it difficult to establish accurate contact condition assessment models. Furthermore, most methods fail to effectively capture early signs of contact corrosion. Therefore, there is an urgent need for a method that can integrate multi-source features and employ advanced machine learning algorithms to achieve intelligent assessment of contact corrosion status, overcoming the shortcomings of existing technologies. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for assessing the corrosion status of electrical contact switch contacts; On the one hand, a method for assessing the corrosion status of electrical contact switch contacts is provided, including: Acquire multi-source characteristic data of the electric contact switch, including opening and closing frequency characteristics, contact resistance characteristics, arc voltage and current signal characteristics, and vibration signal characteristics; An input tensor is constructed based on the multi-source feature data; the input tensor includes the number of opening and closing operations, average contact resistance, time-domain kurtosis index of vibration signal, frequency-domain centroid frequency of vibration signal, energy proportion of wavelet packets in each sub-band, and average moving arc power. The input tensor is fed into a pre-trained deep learning model, which uses a dual-channel temporal encoder to extract deep features. The first channel uses the Mamba state-space model to dynamically adjust the state update rate based on the current number of opening and closing operations, capturing the long-term degradation trend of contact corrosion. The second channel uses a Fourier neural operator to generate an attention mask using frequency domain feature anchors and perform frequency domain global convolution to enhance the frequency domain features of vibration impact and electric arc. The output characteristics of the first and second channels are combined, and the contact corrosion level of the electrical contact switch is determined based on the combined characteristics.

[0004] On the other hand, a corrosion status assessment system for electrical contact switch contacts is provided, comprising: The feature extraction unit acquires multi-source feature data of the electric contact switch, including opening and closing frequency features, contact resistance features, arc voltage and current signal features, and vibration signal features. Tensor construction unit constructs input tensors based on the multi-source feature data; the input tensors include the number of opening and closing, average contact resistance, time-domain kurtosis index of vibration signal, frequency-domain centroid frequency of vibration signal, energy proportion of wavelet packets in each sub-band, and average moving arc power. A dual-channel encoding unit is used to input the input tensor into a pre-trained deep learning model. The deep learning model uses a dual-channel temporal encoder to extract deep features. The first channel uses the Mamba state-space model to dynamically adjust the state update rate according to the current number of opening and closing operations, capturing the long-term degradation trend of contact corrosion. The second channel uses a Fourier neural operator to generate an attention mask using frequency domain feature anchors and perform frequency domain global convolution to enhance the frequency domain features of vibration impact and electric arc. The level determination unit integrates the output characteristics of the first channel and the second channel, and determines the contact corrosion level of the electrical contact switch based on the integrated characteristics.

[0005] Furthermore, an electronic device is also provided, including: Memory, used for non-transitory storage of computer-readable instructions; and Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0006] In another aspect, a computer-readable storage medium is also provided for non-transitory storage of computer-readable instructions, wherein when the non-transitory computer-readable instructions are executed by a computer, the method described in the first aspect is performed.

[0007] In another aspect, a computer program product is also provided, including a computer program that, when run on one or more processors, is used to implement the method described in the first aspect above.

[0008] The above technical solution has the following advantages or beneficial effects: This invention is the first to deeply integrate the number of opening and closing cycles, contact resistance, arc electrical characteristics, and vibration signals to construct a multi-source feature set that comprehensively reflects the mechanical wear, electrical wear, and corrosion state of the contacts, overcoming the shortcomings of existing single-feature methods that do not fully utilize information. Simultaneously, this invention constructs a dual-channel time-series encoder architecture. The first channel is based on the Mamba state-space model, enabling the state update rate to adaptively change with the number of opening and closing cycles, thus tracking the long-term evolution of corrosion. The second channel is based on Fourier neural operators, performing global convolution in the time-frequency dimension and introducing frequency domain centroids and sub-band energy as physical anchors to generate attention masks, enhancing the frequency domain characteristics of vibration impact and arc. The two channels extract complementary features from the time and frequency domains respectively, overcoming the shortcomings of single models in coupling multiple types of signals and improving the sensitivity and assessment accuracy of early abnormal signs of contact corrosion. Attached Figure Description

[0009] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0010] Figure 1 This is a flowchart of the method in Example 1. Detailed Implementation

[0011] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0012] Example 1 This embodiment provides a method for assessing the corrosion status of electrical contact switch contacts, including: S1. Obtain multi-source characteristic data of the electric contact switch.

[0013] Specifically, the multi-source feature data includes opening and closing frequency features, contact resistance features, arc voltage and current signal features, and vibration signal features.

[0014] The number of opening and closing features includes the cumulative number of opening and closing operations. n And the number of new opening and closing times Δ in each detection cycle n The reading is obtained through the mechanical action counter built into the circuit breaker.

[0015] The contact resistance characteristic is the average contact resistance of the contact resistance time series Rc(t) under the steady state of switch closure, which can reflect the corrosion, oxidation and wear state of the contact surface; Preferably, the contact resistance characteristic adopts the dynamic contact resistance measurement principle. In this embodiment, the measurement is started 20ms after the switch is closed and stabilized. The voltage drop across the contact is measured by the wire constant current source method. The injected constant current is 50 A, and 100 contact resistance data points are continuously collected at a sampling rate of 1kHz to form a contact resistance time series. For the collected contact resistance time series, outliers exceeding 3 standard deviations are first removed, and then the average contact resistance is calculated using the following formula:

[0016] In the formula, T To measure the window time, t The initial time after the closure stabilizes; the calculated... R c This serves as the fundamental characteristic of the operation, and its relationship with the initial value is also recorded. R c0 The ratio of .

[0017] The arc voltage and current signal is characterized by the average moving arc power during the opening and closing process; The specific steps for obtaining the average moving arc power are as follows: Transient voltage and current waveforms are acquired by three-phase voltage transformers and current transformers installed on the input and output sides of the transformer, with a sampling rate of not less than 50kHz; the waveforms can cover the entire arcing process. Instantaneous arc power is calculated using the instantaneous differential power method.

[0018] In the formula, These are the instantaneous voltage and current values ​​on the input side. These are the instantaneous voltage and current values ​​on the output side. The start and end times of arcing are determined by detecting sudden power fluctuations. t 0 and t 1. Calculate the cumulative arc energy:

[0019] Average arc power:

[0020] In the formula, T arc = t 1-t0, T arc This refers to the duration of the arc. To smooth out single random fluctuations, the average moving arc power of the most recent M switching events (M is usually set to 100) is taken:

[0021] The vibration signal characteristics are the time-domain, frequency-domain, and time-frequency-domain characteristics of the vibration acceleration signal during the switching operation process; The acceleration signal is acquired by an acceleration sensor installed on the switch body or operating mechanism; The time-domain features are used to extract the peak value, root mean square value, kurtosis index, and waveform factor of the vibration waveform; The formula for calculating the kurtosis index K is as follows:

[0022] in m a The mean, s a The standard deviation is kurtosis, which is sensitive to impact pulses and can effectively characterize contact collision anomalies. The frequency domain characteristics are obtained by performing a fast Fourier transform on the vibration signal to obtain the power spectral density PSD(f), and then extracting the centroid frequency. f c Frequency standard deviation f std And the energy ratio of low, medium and high frequency bands; The formula for centroid frequency is:

[0023] In the time-frequency domain, the vibration signal was decomposed into eight sub-bands using the db4 wavelet with a three-level wavelet packet decomposition. The energy of each sub-band was calculated and normalized to form an eight-dimensional energy spectrum feature vector. The specific energy calculation formula is as follows:

[0024] In the formula, c i,k These are the wavelet packet coefficients.

[0025] S2. Construct an input tensor based on the multi-source feature data.

[0026] Step S2 is used to convert the multi-source heterogeneous feature vectors extracted in each detection into a standardized temporal input tensor acceptable to the deep learning model, specifically including the following steps: S201. Assemble the multi-source physical features extracted from each detection into an original feature vector in a fixed order; The features extracted in each detection include: number of opening and closing operations. Average contact resistance vibration signal time-domain kurtosis index Vibration signal frequency domain centroid frequency Energy proportion of wavelet packets in each sub-band ( , (Number of sub-bands), average moving arc power The above features are assembled into an original feature vector in a fixed order. .

[0027] S202. Use a sliding window to organize the feature vectors of multiple consecutive historical detections into a temporal tensor; Set window length to , the current moment and before The feature vectors from each detection are stacked sequentially to construct a three-dimensional input tensor. ( (as a feature dimension); this operation enables subsequent models to perceive the dynamic evolution trend of the corrosion state over time.

[0028] S203. Group the features in the tensor according to their physical meaning and perform independent standardization; Due to the significant differences in the dimensions of the features, directly inputting them into the model would lead to training difficulties. The features are divided into three groups: the degenerate group... Accumulative normalization is used to preserve its monotonically increasing trend; vibration group Standard BatchNorm is used; energy group It uses the standard BatchNorm. Each group is equipped with independent learnable scaling and offset parameters.

[0029] S204. Introduce non-uniform timestamp position encoding; Since the time interval between two inspections in industrial settings is not fixed, simply encoding by index position will introduce timing deviations. Therefore, it is necessary to calculate the actual time interval between the two inspections. The input tensor is mapped to a time interval embedding vector using a multilayer perceptron. This vector is then added element-wise to a standard sinusoidal position code to form the final position code, which is superimposed on the input tensor. This allows the model to perceive the density of the actual sampling. The final input tensor is then output. .

[0030] S3. Determine the contact corrosion level of the electrical contact switch based on the input tensor and the pre-trained deep learning model.

[0031] Step S3 receives the input tensor constructed in step S2. The erosion level and confidence level at the current moment are determined through forward inference using a pre-trained deep learning model. Specifically, the steps include: S301 uses a dual-channel time encoder to extract depth features.

[0032] The input tensor is fed in parallel into two encoding branches to extract complementary high-level semantic features: Branch A employs the Mamba state-space model, whose recursive parameters can dynamically change with input features. Specifically, this model uses the current number of opening and closing cycles, *n*, as part of the input features. Through neural network layers, it generates dynamic parameters in real time, including the time step parameter Δ, the input control matrix B, and the output control matrix C. The magnitude of Δ directly determines the speed of state updates. When *n* is large, the model generates a larger Δ to accelerate state updates and focus on the latest input, achieving rapid forgetting. When *n* is small, it generates a smaller Δ to smooth updates and retain historical memory, thus achieving on-demand adaptive adjustment of the forgetting rate. This mechanism makes the Mamba model particularly suitable for capturing the long-term degradation characteristics of contact resistance accumulated with the number of opening and closing cycles, while also taking into account short-term abrupt changes.

[0033] Preferably, based on the electrical life range of a typical point contact switch, the initial monitoring range of n can be set to 0 to 5000 times.

[0034] Branch B employs the Fourier Neural Operator (FNO) to perform a global convolution operation on the sequence in the frequency domain. It utilizes the existing frequency domain centroid in the input. and the energy of each sub-band As "frequency domain anchors," the model is guided to adaptively focus on key frequency bands. Specifically, the input time-domain signal is first transformed to the time-frequency domain using a short-time Fourier transform or wavelet packet transform, resulting in a time-frequency spectrum containing two-dimensional information of time and frequency. Subsequently, the frequency domain centroid (i.e., the frequency location where signal energy is concentrated) and the energy distribution of each sub-band are automatically extracted from this time-frequency spectrum, serving as anchors describing the current signal's frequency domain characteristics. Next, these two anchor information are input into a small neural network, which adaptively outputs a set of weight values ​​equal to the number of frequency points, forming a frequency domain attention mask—frequency points close to the frequency domain centroid and located in high-energy sub-bands receive high weights close to 1 (indicating "focused attention"), while frequency points far from the centroid receive low weights close to 0 (indicating "ignored"). This mask is multiplied point-by-point with the original time-frequency spectrum, i.e., suppressing non-target frequency bands and preserving key frequency bands, equivalent to performing a dynamic bandpass filter in the frequency domain, thereby effectively suppressing irrelevant noise and enhancing the frequency domain characteristics of the target signal. Subsequently, a global convolution is performed on the weighted time-domain spectrum along the frequency dimension. Since each point in the frequency domain contains information from the entire time domain, a single frequency-domain convolution operation can capture the global dependency between any two time points, maximizing the receptive field, and its computational efficiency is significantly better than point-by-point time-domain convolution. Finally, the time-domain signal is restored from the masked weighted and frequency-domain convolutional processed signal to the time-domain signal through an inverse transform. The output signal is a clean signal with selectively enhanced frequency-domain features and effectively suppressed noise, which can be directly used for subsequent fault diagnosis or degradation trend analysis. This mechanism uses a frequency-domain anchor point forced model to apply only to... The bias is updated with high weights in frequency bands within a certain range, and irrelevant frequency band noise is suppressed, thereby enhancing the frequency domain characteristics of vibration shock and electric arc, and significantly improving the model's sensitivity to early abnormal signs of contact corrosion.

[0035] Based on the current kurtosis Generating soft-gating coefficients and dynamically fusing dual-branch features. When When the value is large, the model focuses on the FNO frequency domain branch to capture impulse characteristics; when When the size is small, the model focuses on the Mamba branch to track the steady degradation trend.

[0036] S302, Global Pooling and Softmax Probability Output.

[0037] The fused temporal features are then globally averaged and pooled along the time dimension to compress them into a fixed-length feature vector. This vector is then fed into a single fully connected classification layer, outputting four levels of raw Logits, which are then normalized to a probability distribution using the Softmax function. P=[P I , P II , P III , P IV ]

[0038] in For the first The Logits class. Each element corresponds to one of the four corrosion status levels: "Normal (I)", "Caution (II)", "Danger (III)" and "Critical (IV)".

[0039] S303, Calculation of overall confidence level.

[0040] The final confidence level is calculated by integrating three indicators: C conf = p max × r density × l mono The first item is the maximum probability in Softmax. ; Second item r density The third term is the K-nearest neighbor density ratio in the embedding space, calculated as the ratio of the average distance between the nearest neighbors of the same class and the average distance between the nearest neighbors of different classes in the embedding space. l mono As a temporal monotonicity penalty factor, it utilizes the physical irreversibility of the erosion state progressing unidirectionally from I→II→III→IV with increasing opening and closing times. It checks whether the historical inference level within the window satisfies monotonically non-decreasing; if level regression occurs, a penalty is applied. The final confidence level is the product of the above three factors.

[0041] S304. Determine the level based on the final grading criteria and execute the three-branch decision output.

[0042] First, the final classification standard for contact corrosion adopted in this invention is as follows:

[0043] This table defines the complete physical characteristic range corresponding to each corrosion level, and is the basis for the final determination of the corrosion level in this invention. The level output by the model must fall within the corresponding range defined in this table to be valid.

[0044] Based on overall confidence level Perform hierarchical output: when When the value exceeds the high threshold (preferably 0.85), the level corresponding to the highest probability in the model's Softmax output is taken as the final corrosion level. At this point, it is necessary to verify whether this level falls within the feature interval of the corresponding level in the table above: if all measured values ​​of each feature fall within the level interval, then the level and confidence level are directly output; if some features fall within adjacent level intervals, but the highest probability level is still the mainstream judgment, then the level is output with the note "Some features have level overlap, attention is recommended"; if the highest probability level is significantly inconsistent with the interval in the table above (e.g., the model judges it as level I but...), then... If the measured value exceeds the initial value by 200%, then a delayed decision will be forced.

[0045] when When the value falls between the low threshold (preferably 0.60) and the high threshold, a delayed decision is triggered, indicating "boundary state, manual review recommended". The current measured values ​​of each feature are also provided for comparison with the intervals in the table above, for experts to make a comprehensive judgment.

[0046] when When the data falls below a low threshold, mandatory intervention is triggered, and the data is marked as "abnormal data / sensor failure, requiring manual investigation," without outputting a grade prediction. For samples with delayed decisions and mandatory interventions, the system automatically records them and incorporates them into the incremental learning dataset for subsequent model updates and iterations.

[0047] S305. The pre-training method of the deep learning model is as follows: Model training employs an end-to-end supervised learning approach. The labels on the training dataset are annotated by experts according to the final grading criteria in steps S3-4. The loss function is standard cross-entropy loss plus an L2 regularization term. The optimizer used is AdamW, with an initial learning rate of... The learning rate scheduling adopts a cosine annealing heating and restart strategy.

[0048] S306, the recommended value ranges for each hyperparameter in the method are as follows: Sliding window length The recommended value range is 8~20, with 12 being preferred; number of wavelet packet frequency bands Recommended value range is 4~16, with 8 being the preferred value; Hidden dimension The recommended value range is 128~512, preferably 256; the recommended value range for the high confidence threshold is 0.80~0.90, preferably 0.85; the recommended value range for the low confidence threshold is 0.50~0.70, preferably 0.60; the recommended value range for the L2 regularization coefficient is 1×10. -5 ~1×10 -3 1×10 is preferred -4 The parameters listed above can be adjusted adaptively according to the specific application scenario and the characteristics of the collected data, and are not limited to the above numerical ranges.

[0049] Example 2 This embodiment provides a system for assessing the corrosion status of electrical contact switch contacts, including: The feature extraction unit acquires multi-source feature data of the electric contact switch, including opening and closing frequency features, contact resistance features, arc voltage and current signal features, and vibration signal features. Tensor construction unit constructs input tensors based on the multi-source feature data; the input tensors include the number of opening and closing, average contact resistance, time-domain kurtosis index of vibration signal, frequency-domain centroid frequency of vibration signal, energy proportion of wavelet packets in each sub-band, and average moving arc power. A dual-channel encoding unit is used to input the input tensor into a pre-trained deep learning model. The deep learning model uses a dual-channel temporal encoder to extract deep features. The first channel uses the Mamba state-space model to dynamically adjust the state update rate according to the current number of opening and closing operations, capturing the long-term degradation trend of contact corrosion. The second channel uses a Fourier neural operator to generate an attention mask using frequency domain feature anchors and perform frequency domain global convolution to enhance the frequency domain features of vibration impact and electric arc. The level determination unit integrates the output characteristics of the first channel and the second channel, and determines the contact corrosion level of the electrical contact switch based on the integrated characteristics.

[0050] Example 3 This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0051] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0052] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0053] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0054] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0055] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0056] Example 4 This embodiment also provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the method described in Embodiment 1.

[0057] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0058] A computer program product includes software code, wherein the program in the software code performs the steps described in Embodiment 1 of the present invention.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for assessing the corrosion status of electrical contact switch contacts, characterized in that, include: Acquire multi-source characteristic data of the electric contact switch, including opening and closing frequency characteristics, contact resistance characteristics, arc voltage and current signal characteristics, and vibration signal characteristics; An input tensor is constructed based on the multi-source feature data; the input tensor includes the number of opening and closing operations, average contact resistance, time-domain kurtosis index of vibration signal, frequency-domain centroid frequency of vibration signal, energy proportion of wavelet packets in each sub-band, and average moving arc power. The input tensor is fed into a pre-trained deep learning model, which uses a dual-channel temporal encoder to extract deep features. The first channel uses the Mamba state-space model to dynamically adjust the state update rate based on the current number of opening and closing operations, capturing the long-term degradation trend of contact corrosion. The second channel uses a Fourier neural operator to generate an attention mask using frequency domain feature anchors and perform frequency domain global convolution to enhance the frequency domain features of vibration impact and electric arc. The output characteristics of the first and second channels are combined, and the contact corrosion level of the electrical contact switch is determined based on the combined characteristics.

2. The method for assessing the corrosion status of electrical contact switch contacts according to claim 1, characterized in that, The opening and closing frequency feature includes the cumulative opening and closing frequency and the newly added opening and closing frequency in each detection cycle; the contact resistance feature is the average contact resistance of the contact resistance time series under steady-state switch closure, reflecting the corrosion, oxidation and wear state of the contact surface.

3. The method for assessing the corrosion status of electrical contact switch contacts according to claim 1, characterized in that, The arc voltage and current signal characteristics are the average moving arc power during the opening and closing process; calculated based on transient voltage and current waveforms; the vibration signal characteristics are the time domain, frequency domain, and time-frequency domain characteristics of the vibration acceleration signal during the switching action. The time-domain feature is to extract the kurtosis index of the vibration waveform; the frequency-domain feature is to perform a fast Fourier transform on the vibration signal to obtain the power spectral density and extract the centroid frequency; the time-frequency domain feature is to use the db4 wavelet to perform a 3-level wavelet packet decomposition on the vibration signal to obtain 8 sub-frequency bands, calculate the energy of each sub-frequency band and normalize it to form an 8-dimensional energy spectrum feature vector.

4. The method for assessing the corrosion status of electrical contact switch contacts according to claim 1, characterized in that, The method of dynamically adjusting the state update rate based on the current number of opening and closing operations includes: The current number of opening and closing operations, n, is used as part of the input features and fed together with the current input state into the learnable neural network layer of the Mamba state space model. The time step parameter Δ, input control matrix B, and output control matrix C are generated in real time through the neural network layer. The size of the time step parameter Δ determines the speed of state updates: when the current number of opening and closing operations n is greater than a preset threshold, the neural network layer generates a larger Δ to accelerate state updates, focus on the latest input, and achieve rapid forgetting; when the current number of opening and closing operations n is less than or equal to the preset threshold, the neural network layer generates a smaller Δ to smoothly update the state and retain historical memory.

5. The method for assessing the corrosion status of electrical contact switch contacts according to claim 1, characterized in that, The second channel uses a Fourier neural operator to generate an attention mask using frequency domain feature anchors and performs a frequency domain global convolution, including: The input time-domain signal is converted to the time-frequency domain through short-time Fourier transform or wavelet packet transform to obtain the time spectrum containing two-dimensional information of time and frequency. Extract the frequency domain centroid and energy distribution of each sub-band at each moment from the time spectrum, and use them as frequency domain feature anchor points; The frequency domain centroid and the energy distribution of each sub-band are input into a small neural network, and the small neural network adaptively outputs a set of weight values ​​equal to the number of frequency points to form a frequency domain attention mask. In the frequency domain attention mask, frequency points that are close to the frequency domain centroid and located in high-energy sub-bands are given a high weight close to 1, and frequency points that are far from the frequency domain centroid are given a low weight close to 0. The frequency domain attention mask is multiplied point by point with the time spectrum to suppress non-target frequency bands and retain key frequency bands, thereby achieving dynamic bandpass filtering. Perform a global convolution on the weighted time spectrum in the frequency dimension to capture the global dependency between any two time points; The time-domain signal is restored by inverse transformation after masking weighting and frequency domain convolution, and the output is a clean signal with selectively enhanced frequency domain features.

6. The method for assessing the corrosion status of electrical contact switch contacts according to claim 1, characterized in that, The determination of the contact corrosion level of the electrical contact switch based on the fused characteristics includes, The probability distribution of each corrosion level is obtained by global pooling and normalization using the Softmax function; Calculate the overall confidence score, which is the product of the Softmax maximum probability, the K-nearest neighbor density ratio in the embedding space, and the time-series monotonicity penalty factor. When the overall confidence level is higher than the preset high threshold, the level corresponding to the highest probability is taken as the final corrosion level, and it is verified whether the level falls into the feature range of the corresponding level in the preset grading standard. After the verification is passed, the level and confidence level are output. When the overall confidence level is between a preset low threshold and a high threshold, a delayed decision is triggered, prompting manual review. When the overall confidence level is lower than a preset low threshold, forced intervention is triggered, and no level prediction is output.

7. A system for assessing the corrosion status of electrical contact switch contacts, characterized in that, include: The feature extraction unit acquires multi-source feature data of the electric contact switch, including opening and closing frequency features, contact resistance features, arc voltage and current signal features, and vibration signal features. Tensor construction unit constructs input tensors based on the multi-source feature data; the input tensors include the number of opening and closing, average contact resistance, time-domain kurtosis index of vibration signal, frequency-domain centroid frequency of vibration signal, energy proportion of wavelet packets in each sub-band, and average moving arc power. A dual-channel encoding unit is used to input the input tensor into a pre-trained deep learning model. The deep learning model uses a dual-channel temporal encoder to extract deep features. The first channel uses the Mamba state-space model to dynamically adjust the state update rate according to the current number of opening and closing operations, capturing the long-term degradation trend of contact corrosion. The second channel uses a Fourier neural operator to generate an attention mask using frequency domain feature anchors and perform frequency domain global convolution to enhance the frequency domain features of vibration impact and electric arc. The level determination unit integrates the output characteristics of the first channel and the second channel, and determines the contact corrosion level of the electrical contact switch based on the integrated characteristics.

8. An electronic device, comprising: A memory and a processor, wherein the memory is used to store a computer program; and the processor is used to execute, when the computer program is invoked, a method for assessing the corrosion status of an electrical contact switch contact as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for assessing the corrosion status of an electrical contact switch contact as described in any one of claims 1 to 6.

10. A computer program product, wherein the computer program product comprises a computer program that, when executed by a processor, implements a method for assessing the corrosion status of an electrical contact switch contact as described in any one of claims 1 to 6.

Images