A time-frequency-space domain joint feature extraction method of friction and wear electrostatic monitoring signal

Abstract

The application relates to the field of mechanical equipment operation state monitoring, and discloses a time-frequency-space domain joint feature extraction method of a friction and wear electrostatic monitoring signal. The original electrostatic signal is collected through an array type electrostatic sensor, band pass filtering, variational mode decomposition and independent component analysis are used for denoising, and a pure time domain electrostatic monitoring signal is obtained; time domain, frequency domain, physical correlation and basic space domain feature parameters of the signal are extracted for splicing, and correlation analysis and a physical constraint rule base are combined for dimension reduction, so that a time-frequency-space domain joint feature library is constructed; the feature library and graph structure data generated based on three-dimensional coordinates of the sensor are input into a graph neural network calculation module, and a three-dimensional charge density distribution mapping of a monitoring area is output; finally, extreme value distribution area data in the mapping is extracted, so that spatial coordinate parameters and wear degree quantitative parameters of a wear part are obtained, and the application realizes accurate positioning and state evaluation of the wear part.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment operation status monitoring, and in particular to a time-frequency-spatial joint feature extraction method for friction and wear electrostatic monitoring signals. Background Technology

[0002] Friction and wear are common phenomena during the operation of mechanical equipment. Electrostatic monitoring technology can monitor the wear condition of equipment by detecting changes in charge at the friction interface caused by the collision of micro-protrusions and material peeling. However, existing electrostatic monitoring technologies for friction and wear still have the following shortcomings in practical applications: In complex industrial environments, the raw signals collected by electrostatic sensors are easily affected by external environmental interference. The signals are usually mixed with power frequency baseline drift and high-frequency electromagnetic interference. In addition, electrostatic signals of different frequency bands generated by friction interfaces are prone to aliasing. Conventional single filtering methods are difficult to effectively remove these mixed noises and separate the aliasing components, making it difficult to obtain high-quality pure electrostatic signals, which affects the reliability of the underlying data source and the monitoring accuracy.

[0003] To comprehensively reflect the wear state of friction pairs, existing technologies typically extract multiple feature parameters from the signal. Current processing methods often directly combine features from different dimensions such as the time domain, frequency domain, and spatial domain, resulting in an excessively high dimensionality of the feature matrix. Among these multidimensional features, some features have equivalent relationships or overlapping mappings in physical definition and representation. Most existing dimensionality reduction methods rely on purely data-driven mathematical algorithms and lack a discrimination mechanism based on physical constraint rules, which makes it impossible to effectively remove overlapping and redundant terms at the physical level, significantly increasing the complexity and computational burden of subsequent state calculation models.

[0004] Existing electrostatic monitoring systems primarily rely on signals collected by discretely distributed sensors to make macroscopic judgments or trend assessments of the overall health status of friction pairs. In data processing and feature extraction, current technologies only analyze data from the time and frequency domains, neglecting the spatial filtering effect of electrostatic sensing principles and lacking mechanisms for extracting features from the spatial domain. This prevents the full exploitation of multi-dimensional information from electrostatic monitoring. Furthermore, these systems often analyze signals from each channel as independent time series, failing to delve into and utilize the inherent spatial adjacency relationships of the sensors' three-dimensional locations. Due to the lack of mechanisms to map discrete node features to continuous three-dimensional space, existing technologies cannot effectively construct the spatial charge distribution pattern of the monitoring area. This makes it difficult to extract more feature parameters to reflect wear locations and evolution patterns, and hinders the intuitive and accurate location of specific wear points, thus limiting the spatial perception capabilities of electrostatic monitoring technology. Summary of the Invention

[0005] To overcome the above shortcomings, this invention provides a time-frequency-spatial joint feature extraction method for triboelectric electrostatic monitoring signals, aiming to improve the existing triboelectric electrostatic monitoring technology, which has problems such as difficulty in effectively separating aliasing noise, lack of physical constraint mechanism leading to high feature data redundancy, lack of in-depth mining of spatial feature dimensions, and failure to accurately locate actual wear parts by utilizing spatial adjacency relationships.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of friction and wear, comprising the following steps: S1: Acquire the original electrostatic signal of the friction interface collected by the array of electrostatic sensors arranged around the friction pair monitoring area, amplify and anti-aliasing filter the original electrostatic signal to obtain the analog electrostatic signal, and convert the analog electrostatic signal into a digital electrostatic signal with a timestamp. At the same time, acquire the mechanical rotation speed data and normal load data of the externally input friction pair, as well as the three-dimensional spatial position coordinate data of the array of electrostatic sensors. S2: Perform bandpass filtering and signal dealiasing and noise reduction processing on the digital electrostatic signal to separate the noise components in the digital electrostatic signal and obtain a multi-channel pure time-domain electrostatic monitoring signal; S3: Based on the pure time-domain electrostatic monitoring signal, the mechanical rotation frequency parameters obtained by converting the mechanical rotation speed data, and the three-dimensional spatial position coordinate data, extract the time-domain feature parameters, frequency-domain feature parameters, physical correlation feature parameters, and basic spatial feature parameters that characterize the friction state; S4: The time-domain feature parameters, the frequency-domain feature parameters, the physical association feature parameters, and the basic spatial-domain feature parameters are concatenated and combined, and the concatenated and combined data is subjected to dimensionality reduction processing to obtain the dimensionality-reduced time-frequency-spatial joint feature library. S5: Generate spatial adjacency edge data between nodes of a graph neural network based on the three-dimensional spatial location coordinate data, input the dimensionality-reduced time-frequency-spatial joint feature library into the graph neural network model for calculation, and obtain the three-dimensional charge density distribution mapping of the monitoring area; S6: Scan and analyze the three-dimensional charge density distribution map, extract the charge peak distribution region data inside the three-dimensional charge density distribution map, and obtain the spatial coordinate parameters and wear degree quantification parameters of the wear part of the friction pair based on the charge peak distribution region data.

[0007] As a further description of the above technical solution: The specific steps for performing bandpass filtering and signal dealiasing and denoising processing on the digital electrostatic signal to separate the noise components in the digital electrostatic signal and obtain a multi-channel clean time-domain electrostatic monitoring signal include: The digital electrostatic signal is input into a bandpass filter module. The bandpass filter module is pre-set with a passband frequency range that matches the inherent frequency range of the effective electrostatic signal of the friction pair wear. It filters out the power frequency baseline drift data and high-frequency electromagnetic interference data in the digital electrostatic signal to obtain a preliminary filtered signal. The variational mode decomposition algorithm is executed to decompose the preliminary filtered signal into data, calculate the center frequency parameters of the preliminary filtered signal in different frequency bands, and decompose the preliminary filtered signal into multiple intrinsic mode function components located at different center frequency positions. The multiple intrinsic mode function components are combined to construct a multidimensional observation signal matrix. The independent component analysis algorithm is called to perform dealiasing calculation on the multidimensional observation signal matrix to separate the independent principal components and the residual mixed noise components. Calculate the cross-correlation matrix between the independent principal components and the multiple intrinsic mode function components. Compare the calculated values ​​in the cross-correlation matrix with a preset cross-correlation threshold, wherein the cross-correlation threshold is set between 0.75 and 0.85. Components whose calculated values ​​are lower than the cross-correlation threshold are identified as residual mixed noise components and removed. Independent principal components whose calculated values ​​are greater than or equal to the cross-correlation threshold are retained. A linear superposition reconstruction operation is performed on the retained independent principal components to obtain the pure time-domain electrostatic monitoring signal of the multi-channel system.

[0008] As a further description of the above technical solution: The time-domain feature parameters include mean parameters, root mean square parameters, electrostatic activity level parameters, and event rate parameters. The specific steps for extracting the time-domain feature parameters characterizing the friction state include: The pure time-domain electrostatic monitoring signal is segmented and truncated according to the preset time-domain calculation time window parameters to obtain the time series data of the pure time-domain electrostatic monitoring signal; The time series data is integrated to obtain integral data. The integral data is divided with the time domain calculation time window parameter to obtain the mean parameter. The time series data is squared to obtain square data. The square data is integrated to obtain square integral data. The square integral data is divided with the time domain calculation time window parameter to obtain average square data. The average square data is then squared to obtain the root mean square parameter. The absolute value data of the time series data is extracted and integrated to obtain the absolute value integral data. The absolute value integral data is divided with the time domain calculation time window parameter to obtain the absolute value average data. The absolute value average data is then compared with the root mean square parameter to obtain the electrostatic activity level parameter. The system is configured with a preset trigger threshold parameter. The signal amplitude data in the time series data is compared with the preset trigger threshold parameter. The number of pulses in the time domain calculation time window parameter where the signal amplitude data is greater than the preset trigger threshold parameter is counted. The event rate parameter is obtained based on the result of the division operation between the number of pulses and the time domain calculation time window parameter.

[0009] As a further description of the above technical solution: The frequency domain feature parameters include center frequency feature parameters, root mean square frequency feature parameters, standard deviation frequency feature parameters, and energy proportion feature parameters of different frequency bands. The specific steps for extracting the frequency domain feature parameters and physical correlation feature parameters characterizing the friction state include: Calculate the power spectral density data of the spectrum signal, and perform an integral operation on the power spectral density data within the range of zero frequency to the preset highest analysis frequency parameter of the system to obtain the total energy data of the spectrum signal; Within the same frequency range, the first-order frequency integral moment data is obtained by performing frequency-weighted integration operation on the power spectral density data and its corresponding frequency axis data, and the first-order frequency integral moment data is divided by the total energy data to obtain the center frequency characteristic parameter. The second-order frequency integral moment data is obtained by integrating the power spectral density data and the frequency square data. The second-order frequency integral moment data is then divided by the total energy data to obtain the second-order average frequency data. The root mean square frequency characteristic parameter is obtained by taking the square root of the second-order average frequency data. The standard deviation frequency characteristic parameter is obtained by taking the square root of the difference between the square of the root mean square frequency characteristic parameter and the square of the center frequency characteristic parameter. The physical correlation characteristic parameter is obtained by dividing the center frequency characteristic parameter by the mechanical frequency parameter. The main frequency harmonic frequency boundary parameter is set according to the multiples of the mechanical frequency parameters. The main frequency harmonic frequency boundary parameter is set according to 3 to 5 times the mechanical frequency parameters. The spectrum signal is divided into low-frequency spectrum data between zero frequency and the main frequency harmonic frequency boundary parameter, mid-frequency spectrum data between the main frequency harmonic frequency boundary parameter and the preset system inherent frequency boundary parameter, and high-frequency spectrum data between the system inherent frequency boundary parameter and the system's highest analysis frequency parameter. The power spectral density data is integrated on the low-frequency spectrum data, the mid-frequency spectrum data, and the high-frequency spectrum data to obtain the corresponding sub-band energy data. The ratio of each sub-band energy data to the total energy data is calculated, and the energy proportion characteristic parameter of each corresponding frequency band is obtained in sequence.

[0010] As a further description of the above technical solution: The basic spatial domain feature parameters include the electrostatic field gradient entropy parameter and the spatial distribution asymmetry index parameter. The specific steps for extracting the basic spatial domain feature parameters include: Based on the three-dimensional spatial position coordinate data, extract the potential difference data corresponding to the adjacent sensors in the pure time-domain electrostatic monitoring signal, obtain the potential difference vector between the corresponding time-domain signals of adjacent sensors, and calculate the root mean square effective value data of the potential difference vector. The root mean square effective value data is normalized to obtain normalized gradient magnitude data. The electrostatic field gradient entropy parameter is obtained by performing a logarithmic summation mapping calculation operation on the normalized gradient magnitude data. The geometric center and dividing line of the array electrostatic sensor are determined based on the three-dimensional spatial coordinate data. The array electrostatic sensor is divided into a first array group and a second array group on opposite sides along the dividing line. The pure time-domain electrostatic monitoring signals corresponding to the first array group and the second array group are integrated to obtain the overall time-history energy data of each group. The difference between the overall time-history energy data of the first array group and the second array group is calculated, and the sum of the overall time-history energy data of the two groups is calculated. The ratio of the difference to the sum is calculated to obtain the spatial distribution asymmetry index parameter.

[0011] As a further description of the above technical solution: The specific steps for performing dimensionality reduction processing on the spliced ​​and combined data to obtain the dimensionality-reduced time-frequency-spatial domain joint feature library include: inputting the initial high-dimensional joint feature vector composed of the spliced ​​and combined data into the extreme value normalization algorithm unit, performing boundary extreme value mapping operation on each feature dimension data to map it to a preset normalization target value range, wherein the normalization target value range is set to [0, 1] or [-1, 1], and obtaining the normalized feature vector; The timestamp corresponding to the digital electrostatic signal is obtained, and the current cumulative running time of the friction pair is calculated by combining the initial start timestamp of the friction pair recorded by the system. The mechanical frequency parameters obtained by converting the mechanical speed data and the normal load data are combined, and the pre-configured time stage and wear state mapping table is queried in a multi-dimensional joint query to assign matching wear state label data to the normalized feature vector and obtain the feature vector with state label. The feature vectors with state labels corresponding to test samples from multiple time dimensions are accumulated to form a feature matrix. The Pearson correlation coefficient between any two feature dimension data in the feature matrix is ​​calculated. The Pearson correlation coefficient is compared with a preset correlation judgment threshold parameter. Two feature dimension data with a Pearson correlation coefficient greater than the correlation judgment threshold parameter are extracted to form a pair, and suspected redundant feature pair data are obtained. Using a pre-stored basic mathematical definition association table containing equivalent identifier data, a query and matching process is performed to query the equivalent identifier data corresponding to the suspected redundant feature pairs. The equivalent identifier data representing the first state with a physical equivalence relationship is used as the calculation of overlapping terms. One feature dimension data is deleted from the feature vector with state labels. The equivalent identifier data representing the second state with independent mappings is used as the independent physical mapping term and retained. The data after the retention operation is combined and matrixed to obtain the dimensionality-reduced time-frequency-spatial joint feature library.

[0012] As a further description of the above technical solution: The specific steps for generating spatial adjacency edge data between nodes of the graph neural network based on the three-dimensional spatial position coordinate data include: The three-dimensional spatial coordinate data is defined as node position data, and the dimensionality-reduced time-frequency-spatial joint feature library is assigned to the corresponding node position data to generate node feature matrix data. Calculate the Euclidean distance between different node position data, determine the spatial adjacency relationship between the node position data based on the node spatial Euclidean distance data, and generate the spatial adjacency edge data; the node position data and the spatial adjacency edge data together constitute graph structure data.

[0013] As a further description of the above technical solution: The graph neural network model is configured with graph convolutional layer computation units, spatial attention mechanism units, and three-dimensional mesh mapping units. The specific steps for inputting the time-frequency-spatial joint feature library into the graph neural network model for computation to obtain the three-dimensional charge density distribution mapping of the monitoring area include: The node feature matrix data and the graph structure data are input together into the graph convolutional layer computing unit. A nonlinear feature aggregation calculation operation is performed using an aggregation formula that includes an adjacency matrix with added self-connections, a degree matrix, and a learnable weight matrix to obtain the hidden layer state feature data. The hidden layer state feature data is input into the spatial attention mechanism unit. The hidden layer state feature data is mapped through a linear network layer to generate query vectors and key vectors. Dynamic attention weight data is calculated by adding a normalized exponential function to the inner product of the query vector and the key vector, and then weighted summation is performed to obtain aggregated spatial representation feature data. The three-dimensional mesh mapping unit is configured with continuous three-dimensional spatial mesh coordinate data. The aggregated spatial characterization feature data and the continuous three-dimensional spatial mesh coordinate data are input into the three-dimensional mesh mapping unit. The nonlinear mapping model parameters are called to calculate the nonlinear mapping relationship from the discrete aggregated spatial characterization feature data to the continuous three-dimensional spatial mesh coordinate data, thereby obtaining the three-dimensional charge density distribution mapping of the monitoring area.

[0014] As a further description of the above technical solution: The parameters of the nonlinear mapping model are obtained in advance through an offline training process, which specifically includes: The aggregated spatial characterization data of historical monitoring samples are used as model input to establish a three-dimensional geometric model of the friction pair, and the normal load and relative sliding velocity boundary conditions corresponding to the actual working conditions are applied. Finite element simulation analysis was performed using multiphysics coupling calculations of solid mechanics and electrostatics to obtain the charge density values ​​at the nodes of the friction interface. These charge density values ​​were then used as label data of the real three-dimensional charge density distribution to construct a training set. The network weights of the three-dimensional mesh mapping unit are continuously updated through the backpropagation algorithm until the loss function converges, thereby obtaining the pre-trained parameters of the nonlinear mapping model.

[0015] As a further description of the above technical solution: The specific steps for obtaining the spatial coordinate parameters and wear degree quantification parameters of the worn parts of the friction pair based on the charge peak distribution region data include: Extract the spatial coordinate parameters of the wear area corresponding to the charge peak distribution region data; It is equipped with a quantitative correlation table of charge peak and wear degree constructed by friction and wear failure simulation test. The quantitative correlation table records the mapping data of static charge peak density and wear depth under different normal loads. The wear degree quantification parameters of the friction pair are obtained by matching the charge density peak value of the charge peak distribution area data with the quantification correspondence table.

[0016] The present invention has the following beneficial effects: 1. The data processing hub of this invention filters out power frequency baseline drift data and high-frequency electromagnetic interference data from digital electrostatic signals through a bandpass filter module, outputting a preliminary filtered signal. Subsequently, a variational mode decomposition algorithm is executed to decompose the preliminary filtered signal, and multiple intrinsic mode function components are combined to construct a multidimensional observation signal matrix. An independent component analysis algorithm is called to perform dealiasing calculation on the matrix, separating independent principal components and residual mixed noise components. By calculating the cross-correlation matrix between the independent principal components and multiple intrinsic mode function components, components with calculated values ​​lower than the cross-correlation threshold are identified as residual mixed noise components and removed. This process can effectively filter out interference noise under complex operating conditions and separate aliased signals, obtaining high-quality multi-channel clean time-domain electrostatic monitoring signals, thereby improving the accuracy of monitoring results.

[0017] 2. The data processing hub of this invention acquires mean parameters, root mean square parameters, electrostatic activity level parameters, event rate parameters, center frequency characteristic parameters, root mean square frequency characteristic parameters, standard deviation frequency characteristic parameters, physical correlation characteristic parameters, energy proportion characteristic parameters, electrostatic field gradient entropy parameters, and spatial distribution asymmetry index parameters. It then performs vector concatenation on all acquired characteristic parameters to output an initial high-dimensional joint feature vector. Subsequently, it uses an internally configured physical constraint rule base for query matching, searching for equivalent label data corresponding to suspected redundant feature pairs. Equivalent label data in the first state is determined as overlapping terms, and one feature dimension is deleted from the feature vector. Equivalent label data in the second state is determined as independent physical mapping terms and retained. This design can fully explore the operating state of the friction interface from multiple dimensions, while relying on physical constraint rules to eliminate overlapping and redundant terms, effectively reducing the model complexity of subsequent calculations.

[0018] 3. The data processing hub of this invention is equipped with a graph neural network computing module, which defines three-dimensional spatial coordinate data as node position data. Then, the dimensionality-reduced time-frequency-spatial joint feature library is allocated to the corresponding node position data to generate node feature matrix data, and the spatial Euclidean distance data between different node position data is calculated. Based on this, the spatial adjacency relationship between node position data is determined, and spatial adjacency edge data between internal nodes is generated. Then, the nonlinear mapping model parameters are called through the three-dimensional mesh mapping unit to calculate the nonlinear mapping relationship from discrete aggregated spatial characterization feature data to continuous three-dimensional spatial mesh coordinate data, and output the three-dimensional charge density distribution mapping of the monitoring area. Finally, the three-dimensional charge density distribution mapping is scanned and analyzed to extract the extreme value distribution area data inside it, and then the spatial coordinate parameters of the wear part corresponding to the extreme value distribution area data are extracted. This processing mechanism based on spatial adjacency relationship establishes a mapping from discrete signal to continuous spatial distribution, which can intuitively and accurately locate the actual wear part. Attached Figure Description

[0019] Figure 1 This is a flowchart of the digital electrostatic signal denoising process within the data processing center of this invention; Figure 2 This is a flowchart of the multidimensional joint feature extraction and dimensionality reduction elimination process of the present invention; Figure 3 This is a diagram illustrating the framework for the computation of the graph neural network model and the reconstruction of the three-dimensional charge density distribution in this invention. Figure 4 This is a flowchart for the quantitative analysis of wear locations and degrees in this invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Reference Figure 1 - Figure 4 This invention provides a time-frequency-spatial joint feature extraction system for electrostatic monitoring signals of friction and wear, including an array-type electrostatic sensor, a signal conditioning module, a data acquisition module, a data processing center, and an external operating condition parameter receiving module.

[0022] An array of electrostatic sensors is arranged on the outer structure of the friction pair monitoring area. Preferably, the array of electrostatic sensors is distributed in a ring at equal intervals along the outer periphery of the friction pair monitoring area. The sensing end face of the array of electrostatic sensors faces the friction interface and is used to collect the original electrostatic signal generated by the friction interface. It should be noted that although this embodiment uses a 4-channel sensor array as an example to illustrate the spatial feature extraction logic, those skilled in the art will understand that the number of array of electrostatic sensors can be systematically expanded according to the size of the monitoring area (such as expanding to 6 channels, 8 channels, etc.), without departing from the protection scope of this invention.

[0023] The signal conditioning module is connected to the array-type electrostatic sensor to receive the raw electrostatic signal. The signal conditioning module amplifies and performs anti-aliasing filtering on the raw electrostatic signal, converts the sensed electrostatic charge / current signal into an analog signal in voltage form, and outputs the analog electrostatic signal.

[0024] The data acquisition module is connected to the signal conditioning module and is used to receive the analog electrostatic signal output by the signal conditioning module and convert the analog electrostatic signal into a digital electrostatic signal with a timestamp. Preferably, the data acquisition module has a multi-channel synchronous sampling function to ensure that the output digital electrostatic signals of each channel have strict time consistency. The digital electrostatic signal is then transmitted to the data processing center.

[0025] The external operating condition parameter receiving module is connected to the data processing center to receive externally input friction pair mechanical speed and normal load data, and transmit them to the equipment operating parameter storage unit configured inside the data processing center.

[0026] The data processing center receives digital electrostatic signals. In some specific embodiments, the hardware carrier of the data processing center may include computing devices such as industrial control computers, microprocessors (MCUs), digital signal processors (DSPs), or workstations equipped with graphics processing units (GPUs). The data processing center processes the digital electrostatic signals through bandpass filtering, variational mode decomposition, and independent component analysis to separate noise components in the digital electrostatic signals and finally outputs multi-channel clean time-domain electrostatic monitoring signals.

[0027] Subsequently, the data processing center calculates time-domain characteristic parameters based on the pure time-domain electrostatic monitoring signal. These time-domain characteristic parameters include the mean parameter, root mean square parameter, electrostatic activity level parameter, and event rate parameter of the pure time-domain electrostatic monitoring signal.

[0028] The data processing center also performs a fast Fourier transform on the pure time-domain electrostatic monitoring signal, outputting a spectral signal. The data processing center extracts the center frequency characteristic parameter, root mean square frequency characteristic parameter, and standard deviation frequency characteristic parameter of this spectral signal. It also extracts the physical correlation characteristic parameter by combining the mechanical frequency parameters of the friction pair equipment obtained from the equipment operation parameter storage unit configured inside the data processing center. In addition, the data processing center divides the spectral signal into low-frequency band spectral data, mid-frequency band spectral data, and high-frequency band spectral data, and extracts the energy proportion characteristic parameter of each spectral data. The aforementioned center frequency characteristic parameter, root mean square frequency characteristic parameter, standard deviation frequency characteristic parameter, and energy proportion characteristic parameter of each spectral data together constitute the frequency domain characteristic parameters.

[0029] The data processing center calculates basic spatial characteristic parameters based on pure time-domain electrostatic monitoring signals. These basic spatial characteristic parameters include electrostatic field gradient entropy parameters and spatial distribution asymmetry index parameters.

[0030] Next, the data processing center concatenates and combines the aforementioned time-domain feature parameters, frequency-domain feature parameters, physical correlation feature parameters, and basic spatial-domain feature parameters. Then, it performs dimensionality reduction on the concatenated feature data using correlation algorithms and physical constraint rules, outputting a dimensionality-reduced time-frequency-spatial joint feature library.

[0031] The data processing center pre-stores the three-dimensional spatial position coordinate data of the array electrostatic sensor. The data processing center is equipped with a graph neural network computing module. The data processing center generates spatial adjacency edge data between nodes inside the graph neural network computing module based on the three-dimensional spatial position coordinate data, and inputs the time-frequency-spatial joint feature library into the graph neural network computing module. The graph neural network computing module outputs the three-dimensional charge density distribution mapping of the monitoring area.

[0032] Finally, the data processing center scans and analyzes the three-dimensional charge density distribution map, extracts the extreme value distribution area data within it, and extracts the spatial coordinate parameters and wear degree quantification parameters of the wear area based on the extreme value distribution area data.

[0033] Furthermore, the specific steps of the aforementioned data processing center in processing digital electrostatic signals through bandpass filtering, variational mode decomposition, and independent component analysis include: the data processing center acquires the digital electrostatic signal and inputs it into an internally configured bandpass filtering module. The bandpass filtering module is pre-set with a passband frequency range that matches the inherent frequency range of the effective electrostatic signal of the friction pair wear. This bandpass filtering module filters out power frequency baseline drift data and high-frequency electromagnetic interference data from the digital electrostatic signal, outputting a preliminary filtered signal. Next, the data processing center executes a variational mode decomposition algorithm to decompose the preliminary filtered signal. This algorithm is pre-set with parameters for the number of decomposition modes and a penalty factor, and calculates the center frequency parameters of the preliminary filtered signal in different frequency bands. The preliminary filtered signal is decomposed into multiple intrinsic mode function components located at different center frequency positions. Subsequently, the data processing center combines these multiple intrinsic mode function components to construct... The system generates a multidimensional observation signal matrix and uses an independent component analysis (ICA) algorithm to perform dealiasing calculations on the matrix, separating independent principal components (IPCs) from residual mixed noise components. For precise removal, a cross-correlation coefficient threshold is preset in the data processing center. Preferably, the cross-correlation coefficient threshold is set between 0.75 and 0.85. The data processing center calculates the cross-correlation coefficient matrix between the IPCs and multiple intrinsic mode function (IMF) components. The data processing center compares the calculated values ​​in the cross-correlation coefficient matrix with the cross-correlation coefficient threshold. Components with calculated values ​​lower than the cross-correlation coefficient threshold are identified as residual mixed noise components and removed. Meanwhile, IPCs with calculated values ​​greater than or equal to the cross-correlation coefficient threshold are retained. The data processing center performs a linear superposition reconstruction operation on the retained IPCs, outputting a multi-channel clean time-domain electrostatic monitoring signal. This clean time-domain electrostatic monitoring signal serves as the basic data source for subsequent extraction of various feature parameters.

[0034] The specific steps of the data processing center in calculating time-domain characteristic parameters based on the pure time-domain electrostatic monitoring signal include: The data processing center presets a time-domain calculation window parameter, and segments the pure time-domain electrostatic monitoring signal according to this time-domain calculation window parameter to obtain time-series data of the pure time-domain electrostatic monitoring signal; the data processing center performs integration on the time-series data to obtain integral data; it then divides the integral data with the time-domain calculation window parameter to output the mean parameter; simultaneously, it performs squaring on the time-series data to obtain squared data, integrates the squared data to obtain squared integral data, divides the squared integral data with the time-domain calculation window parameter to obtain the mean square data, and performs square root calculation on the mean square data to output the root mean square parameter of the time-series data. In addition, the data processing center extracts the absolute value data of the time series data and performs integration to obtain the absolute value integral data. It then divides this absolute value integral data with the time-domain calculation time window parameter to obtain the absolute value average data. The absolute value average data is then compared with the root mean square parameter to calculate the electrostatic activity level parameter. The data processing center is also configured with a preset trigger threshold parameter. The signal amplitude data in the time series data is compared with the preset trigger threshold parameter, and the number of pulses with signal amplitude data greater than the preset trigger threshold parameter within the time-domain calculation time window parameter is counted. Based on the result of the division operation between the pulse count data and the time-domain calculation time window parameter, the event rate parameter is output. Finally, the data processing center combines the mean parameter, root mean square parameter, electrostatic activity level parameter, and event rate parameter to construct the time-domain feature parameters.

[0035] The specific steps of the data processing center performing a Fast Fourier Transform (FFT) operation on the clean time-domain electrostatic monitoring signal to obtain relevant characteristic parameters include: the data processing center performs a FFT operation on the clean time-domain electrostatic monitoring signal to obtain the spectral signal, and calculates the power spectral density data of the spectral signal. The data processing center has a preset system maximum analysis frequency parameter, and performs an integral operation on the power spectral density data within the range from zero frequency to the system maximum analysis frequency parameter to obtain the total energy data of the spectral signal; simultaneously, within the same frequency range, it performs a frequency-weighted integral operation based on the power spectral density data and its corresponding frequency axis data. The first-order frequency integral moment data is obtained and divided by the total energy data to obtain the center frequency characteristic parameter of the spectral signal. Further, within the same frequency range, the data processing center performs an integral operation on the power spectral density data and the frequency square data to obtain the second-order frequency integral moment data. This second-order average frequency data is obtained by dividing the power spectral density data and the frequency square data, and the root mean square frequency characteristic parameter is obtained by taking the square root of the second-order average frequency data. Subsequently, the standard deviation frequency characteristic parameter is obtained by taking the square root of the difference between the square of the root mean square frequency characteristic parameter and the square of the center frequency characteristic parameter.

[0036] The data processing center is equipped with a device operating parameter storage unit, and obtains the mechanical frequency parameters of the friction pair device from it. The data processing center performs a division operation between the center frequency characteristic parameter and the mechanical frequency parameter, and outputs the physical correlation characteristic parameter. This physical correlation characteristic parameter is used to characterize the frequency deviation state of the center frequency characteristic parameter relative to the mechanical frequency parameter.

[0037] The data processing center is pre-set with main frequency harmonic frequency boundary parameters and system inherent frequency boundary parameters. Preferably, the main frequency harmonic frequency boundary parameters are set according to the 3rd to 5th harmonics of the mechanical frequency parameters. Based on this, the data processing center divides the spectrum signal into low-frequency band spectrum data, mid-frequency band spectrum data, and high-frequency band spectrum data. The frequency range of the low-frequency band spectrum data is between zero frequency and the main frequency harmonic frequency boundary parameters; the frequency range of the mid-frequency band spectrum data is between the main frequency harmonic frequency boundary parameters and the system inherent frequency boundary parameters; and the frequency range of the high-frequency band spectrum data is between the system inherent frequency boundary parameters and the system's highest analysis frequency parameter. The data processing center performs power spectral density data integration on the above three spectrum data segments to obtain the corresponding sub-band energy data, calculates the ratio of each sub-band energy data to the total energy data, and then sequentially outputs the energy proportion characteristic parameters of the low-frequency band spectrum data, the mid-frequency band spectrum data, and the high-frequency band spectrum data.

[0038] The specific steps of the aforementioned data processing center in calculating the basic spatial characteristic parameters based on the pure time-domain electrostatic monitoring signal include: the data processing center calls the three-dimensional spatial position coordinate data of the array electrostatic sensor, extracts the potential difference data corresponding to the time-domain electrostatic monitoring signal between adjacent sensors based on the coordinate data, obtains the potential difference vector between the corresponding time-domain signals of adjacent sensors, and calculates the root mean square effective value data of the potential difference vector. The data processing center normalizes the effective value data to obtain normalized gradient amplitude data, and performs a logarithmic summation mapping calculation operation on the normalized gradient amplitude data to output the electrostatic field gradient entropy parameter. Furthermore, the data processing center... Based on the three-dimensional spatial coordinate data, the geometric center and axis of symmetry of the sensor array are determined (for example, when the sensors are distributed in a ring, the axis of symmetry is the axis passing through the center of the ring distribution). The array-type electrostatic sensor is divided into a first array group and a second array group on opposite sides along the axis of symmetry. The pure time-domain electrostatic monitoring signals corresponding to the first array group and the second array group are integrated to obtain the overall time-history energy data of each group. Then, the difference between the overall time-history energy data of the first array group and the second array group is calculated, and the sum of the overall time-history energy data of the two groups is calculated. The ratio of the difference to the sum is calculated to output the spatial distribution asymmetry index parameter.

[0039] The specific steps of the aforementioned data processing center in constructing a dimensionality-reduced time-frequency-spatial joint feature library based on basic spatial feature parameters and other feature parameters include: the data processing center acquires the mean parameter, root mean square parameter, electrostatic activity level parameter, event rate parameter, center frequency feature parameter, root mean square frequency feature parameter, standard deviation frequency feature parameter, physical correlation feature parameter, energy proportion feature parameter, electrostatic field gradient entropy parameter, and spatial distribution asymmetry index parameter; and performs vector concatenation operation on all the acquired feature parameters to output an initial high-dimensional joint feature vector. The data processing center is internally configured with an extreme value normalization algorithm unit and a preset normalization target value range, preferably set to [0, 1] or [-1, ... [1] The initial high-dimensional joint feature vector is input into the algorithm unit to perform boundary extreme value mapping operation on the data of each feature dimension, thereby mapping to the normalized target value range and outputting a normalized feature vector. In addition, the data processing center has preset wear state label data, including running-in period label, stable wear period label and severe wear period label, as well as a mapping table of time stage and wear state obtained through historical calibration experiments. The historical calibration experiment refers to: performing a full life cycle wear test on the same type of friction pair under standard working conditions, and recording the actual wear amount under different cumulative running time using a surface profilometer or weighing method. Based on this, the running-in period, stable wear period and severe wear period are divided, thereby constructing a mapping table of time stage and wear state. By obtaining the timestamp corresponding to the digital electrostatic signal and combining it with the initial start timestamp of the friction pair recorded by the system, the current cumulative running time of the friction pair is calculated. Preferably, the data processing center also combines the currently obtained mechanical frequency parameters and normal load data, and through multi-dimensional joint query of the mapping table, assigns matching wear state label data to the normalized feature vector and outputs a feature vector with state label.

[0040] The data processing center is also equipped with a correlation analysis module. The center accumulates feature vectors with state labels corresponding to test samples across multiple time dimensions to form a feature matrix. The correlation analysis module calculates the Pearson correlation coefficient between any two feature dimensions in the feature matrix and outputs the feature correlation matrix data. The center has a preset correlation judgment threshold parameter. By comparing the Pearson correlation coefficient with this threshold parameter, it extracts two feature dimensions whose Pearson correlation coefficient is greater than the threshold parameter, forming a pair and outputting suspected redundant feature pairs. Subsequently, the center uses an internally configured physical constraint rule base for query matching. This rule base stores a basic mathematical definition association table containing equivalent identifier data. Specifically, when the first feature is defined as the signal mean and the second feature is defined as the zero-frequency component, they are physically equivalent, and this equivalent identifier data is labeled as the first state (i.e., representing the existence of a physically equivalent relationship). For example, when the first feature is defined as the signal root mean parameter and the second feature is defined as... When dealing with total energy data, the two are strongly equivalent in characterizing signal energy intensity and are labeled as the first state. Similarly, if the time-domain event rate parameter and the high-frequency energy proportion feature parameter are determined to be strongly correlated, they map the same micro-convex body collision physical process and are also labeled as the first state. Conversely, if there is no physical equivalence, they are labeled as the second state (i.e., representing mutually independent mappings). The data processing center queries the equivalent label data corresponding to suspected redundant feature pairs. The equivalent label data in the first state is determined to be an overlapping term, and one feature dimension data is deleted from the feature vector. The equivalent label data in the second state is determined to be an independent physical mapping term and is retained. After completing the feature deletion and retention operations, the dimensionality-reduced feature vector data is output. That is, after dimensionality reduction and redundancy removal, the 13-dimensional joint feature vector composed of 13 key dimensions such as the mean and root mean square is finally retained. Subsequently, the data processing center combines the dimensionality-reduced feature vector data corresponding to different samples into a matrix and outputs the dimensionality-reduced time-frequency-spatial joint feature library.

[0041] The specific steps of the data processing center in constructing a graph neural network model and performing a three-dimensional charge density distribution reconstruction calculation based on the dimensionality-reduced time-frequency-spatial joint feature library include: the data processing center is equipped with a graph neural network calculation module, and the called three-dimensional spatial coordinate data is defined as node position data. Then, the dimensionality-reduced time-frequency-spatial joint feature library is allocated to the corresponding node position data to generate node feature matrix data. The data processing center calculates the node spatial Euclidean distance data between different node position data, thereby determining the spatial adjacency relationship between node position data and generating spatial adjacency edge data between internal nodes. The node position data and the spatial adjacency edge data together constitute the graph structure data of the graph neural network calculation module.

[0042] The data processing hub, within the graph neural network computation module, is equipped with a graph convolutional layer computation unit and a spatial attention mechanism unit. Node feature matrix data and graph structure data are input together into the graph convolutional layer computation unit to perform a nonlinear feature aggregation computation operation. This operation utilizes an aggregation formula for calculation.

[0043] In the formula, This represents the layer index of a graph neural network. For the first The node feature matrix of the layer; A self-connected adjacency matrix (where A is the original adjacency matrix and I is the identity matrix) is added to preserve the node's own characteristics during aggregation; for The corresponding degree matrix; The weight matrix is ​​a learnable weight matrix; The function is a nonlinear activation function, outputting hidden layer state feature data. This data is then input into a spatial attention mechanism unit, where the hidden layer state feature data is mapped through linear network layers to generate query vectors and key vectors. Dynamic attention weight data is calculated by adding a normalized exponential function to the inner product of the query vectors and key vectors, followed by weighted summation to output aggregated spatial representation feature data. The graph neural network computation module also includes a 3D mesh mapping unit with preset continuous 3D spatial mesh coordinate data. The aggregated spatial representation feature data and the continuous 3D spatial mesh coordinate data are input into this unit. This 3D mesh mapping unit has preset nonlinear mapping model parameters trained and converged using a backpropagation algorithm. The preferred offline training process for these nonlinear mapping model parameters involves obtaining the aggregated space of historical monitoring samples. Characteristic data is used as model input, and the corresponding real three-dimensional charge density distribution is obtained by finite element simulation analysis as label data to construct a training set. The finite element simulation analysis specifically includes: establishing a three-dimensional geometric model of the friction pair, applying normal load and relative sliding velocity boundary conditions corresponding to the actual working conditions, using multi-physics coupling calculation of solid mechanics and electrostatic field to obtain the charge density values ​​on the friction interface nodes as label data of the real three-dimensional charge density distribution, continuously updating the network weight of the three-dimensional mesh mapping unit through the backpropagation algorithm until the loss function converges to obtain the parameter, and the three-dimensional mesh mapping unit calls the parameter to calculate the nonlinear mapping relationship from discrete aggregated spatial characteristic data to continuous three-dimensional spatial mesh coordinate data, and outputting the three-dimensional charge density distribution mapping of the monitoring area.

[0044] Finally, the data processing center scans and analyzes the three-dimensional charge density distribution map, extracts the extreme value distribution region data within it, and then extracts the spatial coordinate parameters of the wear part corresponding to the extreme value distribution region data. The data processing center has a preset quantitative correlation table between charge peak and wear degree constructed through friction and wear failure simulation test. This correlation table records the positive correlation mapping curve or discrete data pairs between static charge peak density and wear depth under different normal loads. The data processing center performs quantitative calculation matching based on the charge density peak of the extreme value distribution region data and the quantitative correlation table, and outputs the quantitative parameters of the wear degree of the friction pair.

[0045] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for joint time-frequency and spatial feature extraction of electrostatic monitoring signals for triboelectric wear, characterized in that, Includes the following steps: S1: Acquire the original electrostatic signal of the friction interface collected by the array of electrostatic sensors arranged around the friction pair monitoring area, amplify and anti-aliasing filter the original electrostatic signal to obtain the analog electrostatic signal, and convert the analog electrostatic signal into a digital electrostatic signal with a timestamp. At the same time, acquire the mechanical rotation speed data and normal load data of the externally input friction pair, as well as the three-dimensional spatial position coordinate data of the array of electrostatic sensors. S2: Perform bandpass filtering and signal dealiasing and noise reduction processing on the digital electrostatic signal to separate the noise components in the digital electrostatic signal and obtain a multi-channel pure time-domain electrostatic monitoring signal; S3: Based on the pure time-domain electrostatic monitoring signal, the mechanical rotation frequency parameters obtained by converting the mechanical rotation speed data, and the three-dimensional spatial position coordinate data, extract the time-domain feature parameters, frequency-domain feature parameters, physical correlation feature parameters, and basic spatial feature parameters that characterize the friction state; S4: The time-domain feature parameters, the frequency-domain feature parameters, the physical association feature parameters, and the basic spatial-domain feature parameters are concatenated and combined, and the concatenated and combined data is subjected to dimensionality reduction processing to obtain the dimensionality-reduced time-frequency-spatial joint feature library. S5: Generate spatial adjacency edge data between nodes of a graph neural network based on the three-dimensional spatial location coordinate data, input the dimensionality-reduced time-frequency-spatial joint feature library into the graph neural network model for calculation, and obtain the three-dimensional charge density distribution mapping of the monitoring area; S6: Scan and analyze the three-dimensional charge density distribution map, extract the charge peak distribution region data inside the three-dimensional charge density distribution map, and obtain the spatial coordinate parameters and wear degree quantification parameters of the wear part of the friction pair based on the charge peak distribution region data.

2. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of friction and wear according to claim 1, characterized in that, The specific steps for performing bandpass filtering and signal dealiasing and denoising processing on the digital electrostatic signal to separate the noise components in the digital electrostatic signal and obtain a multi-channel clean time-domain electrostatic monitoring signal include: The digital electrostatic signal is input into a bandpass filter module. The bandpass filter module is pre-set with a passband frequency range that matches the inherent frequency range of the effective electrostatic signal of the friction pair wear. It filters out the power frequency baseline drift data and high-frequency electromagnetic interference data in the digital electrostatic signal to obtain a preliminary filtered signal. The variational mode decomposition algorithm is executed to decompose the preliminary filtered signal into data, calculate the center frequency parameters of the preliminary filtered signal in different frequency bands, and decompose the preliminary filtered signal into multiple eigenmode function components located at different center frequency positions. The multiple intrinsic mode function components are combined to construct a multidimensional observation signal matrix. The independent component analysis algorithm is called to perform dealiasing calculation on the multidimensional observation signal matrix to separate the independent principal components and the residual mixed noise components. Calculate the cross-correlation matrix between the independent principal components and the multiple intrinsic mode function components. Compare the calculated values ​​in the cross-correlation matrix with a preset cross-correlation threshold, wherein the cross-correlation threshold is set between 0.75 and 0.

85. Components whose calculated values ​​are lower than the cross-correlation threshold are identified as residual mixed noise components and removed. Independent principal components whose calculated values ​​are greater than or equal to the cross-correlation threshold are retained. A linear superposition reconstruction operation is performed on the retained independent principal components to obtain the pure time-domain electrostatic monitoring signal of the multi-channel system.

3. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of friction and wear according to claim 1, characterized in that, The time-domain feature parameters include mean parameters, root mean square parameters, electrostatic activity level parameters, and event rate parameters. The specific steps for extracting the time-domain feature parameters characterizing the friction state include: The pure time-domain electrostatic monitoring signal is segmented and truncated according to the preset time-domain calculation time window parameters to obtain the time series data of the pure time-domain electrostatic monitoring signal; The time series data is integrated to obtain integral data. The integral data is divided with the time domain calculation time window parameter to obtain the mean parameter. The time series data is squared to obtain square data. The square data is integrated to obtain square integral data. The square integral data is divided with the time domain calculation time window parameter to obtain average square data. The average square data is then squared to obtain the root mean square parameter. The absolute value data of the time series data is extracted and integrated to obtain the absolute value integral data. The absolute value integral data is divided with the time domain calculation time window parameter to obtain the absolute value average data. The absolute value average data is then compared with the root mean square parameter to obtain the electrostatic activity level parameter. The system is configured with a preset trigger threshold parameter. The signal amplitude data in the time series data is compared with the preset trigger threshold parameter. The number of pulses in the time domain calculation time window parameter where the signal amplitude data is greater than the preset trigger threshold parameter is counted. The event rate parameter is obtained based on the result of the division operation between the number of pulses and the time domain calculation time window parameter.

4. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of friction and wear according to claim 1, characterized in that, The frequency domain feature parameters include center frequency feature parameters, root mean square frequency feature parameters, standard deviation frequency feature parameters, and energy proportion feature parameters of different frequency bands. The specific steps for extracting the frequency domain feature parameters and physical correlation feature parameters characterizing the friction state include: Calculate the power spectral density data of the spectrum signal, and perform an integral operation on the power spectral density data within the range of zero frequency to the preset highest analysis frequency parameter of the system to obtain the total energy data of the spectrum signal; Within the same frequency range, the first-order frequency integral moment data is obtained by performing frequency-weighted integration operation on the power spectral density data and its corresponding frequency axis data, and the first-order frequency integral moment data is divided by the total energy data to obtain the center frequency characteristic parameter. The second-order frequency integral moment data is obtained by integrating the power spectral density data and the frequency square data. The second-order frequency integral moment data is then divided by the total energy data to obtain the second-order average frequency data. The root mean square frequency characteristic parameter is obtained by taking the square root of the second-order average frequency data. The standard deviation frequency characteristic parameter is obtained by taking the square root of the difference between the square of the root mean square frequency characteristic parameter and the square of the center frequency characteristic parameter. The physical correlation characteristic parameter is obtained by dividing the center frequency characteristic parameter by the mechanical frequency parameter. The main frequency harmonic frequency boundary parameter is set according to the multiples of the mechanical frequency parameters. The main frequency harmonic frequency boundary parameter is set according to 3 to 5 times the mechanical frequency parameters. The spectrum signal is divided into low-frequency spectrum data between zero frequency and the main frequency harmonic frequency boundary parameter, mid-frequency spectrum data between the main frequency harmonic frequency boundary parameter and the preset system inherent frequency boundary parameter, and high-frequency spectrum data between the system inherent frequency boundary parameter and the system's highest analysis frequency parameter. The power spectral density data is integrated on the low-frequency spectrum data, the mid-frequency spectrum data, and the high-frequency spectrum data to obtain the corresponding sub-band energy data. The ratio of each sub-band energy data to the total energy data is calculated, and the energy proportion characteristic parameter of each corresponding frequency band is obtained in sequence.

5. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of friction and wear according to claim 1, characterized in that, The basic spatial domain feature parameters include the electrostatic field gradient entropy parameter and the spatial distribution asymmetry index parameter. The specific steps for extracting the basic spatial domain feature parameters include: Based on the three-dimensional spatial position coordinate data, extract the potential difference data corresponding to the adjacent sensors in the pure time-domain electrostatic monitoring signal, obtain the potential difference vector between the corresponding time-domain signals of adjacent sensors, and calculate the root mean square effective value data of the potential difference vector. The root mean square effective value data is normalized to obtain normalized gradient magnitude data. The electrostatic field gradient entropy parameter is obtained by performing a logarithmic summation mapping calculation operation on the normalized gradient magnitude data. The geometric center and dividing line of the array electrostatic sensor are determined based on the three-dimensional spatial coordinate data. The array electrostatic sensor is divided into a first array group and a second array group on opposite sides along the dividing line. The pure time-domain electrostatic monitoring signals corresponding to the first array group and the second array group are integrated to obtain the overall time-history energy data of each group. The difference between the overall time-history energy data of the first array group and the second array group is calculated, and the sum of the overall time-history energy data of the two groups is calculated. The ratio of the difference to the sum is calculated to obtain the spatial distribution asymmetry index parameter.

6. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of friction and wear according to claim 1, characterized in that, The specific steps for performing dimensionality reduction processing on the spliced ​​and combined data to obtain the dimensionality-reduced time-frequency-spatial domain joint feature library include: inputting the initial high-dimensional joint feature vector composed of the spliced ​​and combined data into the extreme value normalization algorithm unit, performing boundary extreme value mapping operation on each feature dimension data to map it to a preset normalization target value range, wherein the normalization target value range is set to [0, 1] or [-1, 1], and obtaining the normalized feature vector; The timestamp corresponding to the digital electrostatic signal is obtained, and the current cumulative running time of the friction pair is calculated by combining the initial start timestamp of the friction pair recorded by the system. The mechanical frequency parameters obtained by converting the mechanical speed data and the normal load data are combined, and the pre-configured time stage and wear state mapping table is queried in a multi-dimensional joint query to assign matching wear state label data to the normalized feature vector and obtain the feature vector with state label. The feature vectors with state labels corresponding to test samples from multiple time dimensions are accumulated to form a feature matrix. The Pearson correlation coefficient between any two feature dimension data in the feature matrix is ​​calculated. The Pearson correlation coefficient is compared with a preset correlation judgment threshold parameter. Two feature dimension data with a Pearson correlation coefficient greater than the correlation judgment threshold parameter are extracted to form a pair, and suspected redundant feature pair data are obtained. Using a pre-stored basic mathematical definition association table containing equivalent identifier data, a query and matching process is performed to query the equivalent identifier data corresponding to the suspected redundant feature pairs. The equivalent identifier data representing the first state with a physical equivalence relationship is used as the calculation of overlapping terms. One feature dimension data is deleted from the feature vector with state labels. The equivalent identifier data representing the second state with independent mappings is used as the independent physical mapping term and retained. The data after the retention operation is combined and matrixed to obtain the dimensionality-reduced time-frequency-spatial joint feature library.

7. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of triboelectric wear according to claim 1, characterized in that, The specific steps for generating spatial adjacency edge data between nodes of the graph neural network based on the three-dimensional spatial position coordinate data include: The three-dimensional spatial coordinate data is defined as node position data, and the dimensionality-reduced time-frequency-spatial joint feature library is assigned to the corresponding node position data to generate node feature matrix data. Calculate the Euclidean distance between different node position data, determine the spatial adjacency relationship between the node position data based on the node spatial Euclidean distance data, and generate the spatial adjacency edge data; the node position data and the spatial adjacency edge data together constitute graph structure data.

8. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of triboelectric wear according to claim 7, characterized in that, The graph neural network model is configured with graph convolutional layer computation units, spatial attention mechanism units, and three-dimensional mesh mapping units. The specific steps for inputting the time-frequency-spatial joint feature library into the graph neural network model for computation to obtain the three-dimensional charge density distribution mapping of the monitoring area include: The node feature matrix data and the graph structure data are input together into the graph convolutional layer computing unit. A nonlinear feature aggregation calculation operation is performed using an aggregation formula that includes an adjacency matrix with added self-connections, a degree matrix, and a learnable weight matrix to obtain the hidden layer state feature data. The hidden layer state feature data is input into the spatial attention mechanism unit. The hidden layer state feature data is mapped through a linear network layer to generate query vectors and key vectors. Dynamic attention weight data is calculated by adding a normalized exponential function to the inner product of the query vector and the key vector, and then weighted summation is performed to obtain aggregated spatial representation feature data. The three-dimensional mesh mapping unit is configured with continuous three-dimensional spatial mesh coordinate data. The aggregated spatial characterization feature data and the continuous three-dimensional spatial mesh coordinate data are input into the three-dimensional mesh mapping unit. The nonlinear mapping model parameters are called to calculate the nonlinear mapping relationship from the discrete aggregated spatial characterization feature data to the continuous three-dimensional spatial mesh coordinate data, thereby obtaining the three-dimensional charge density distribution mapping of the monitoring area.

9. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of tribology and wear according to claim 8, characterized in that, The parameters of the nonlinear mapping model are obtained in advance through an offline training process, which specifically includes: The aggregated spatial characterization data of historical monitoring samples are used as model input to establish a three-dimensional geometric model of the friction pair, and the normal load and relative sliding velocity boundary conditions corresponding to the actual working conditions are applied. Finite element simulation analysis was performed using multiphysics coupling calculations of solid mechanics and electrostatics to obtain the charge density values ​​at the nodes of the friction interface. These charge density values ​​were then used as label data of the real three-dimensional charge density distribution to construct a training set. The network weights of the three-dimensional mesh mapping unit are continuously updated through the backpropagation algorithm until the loss function converges, thereby obtaining the pre-trained parameters of the nonlinear mapping model.

10. The time-frequency-spatial joint feature extraction method for electrostatic monitoring signals of friction and wear according to claim 1, characterized in that, The specific steps for obtaining the spatial coordinate parameters and wear degree quantification parameters of the worn parts of the friction pair based on the charge peak distribution region data include: Extract the spatial coordinate parameters of the wear area corresponding to the charge peak distribution region data; It is equipped with a quantitative correlation table of charge peak and wear degree constructed by friction and wear failure simulation test. The quantitative correlation table records the mapping data of static charge peak density and wear depth under different normal loads. The wear degree quantification parameters of the friction pair are obtained by matching the charge density peak value of the charge peak distribution area data with the quantification correspondence table.

Images