Method and system for compensating for zero drift of an eddy current sensor

Abstract

The application relates to the technical field of zero-point drift compensation, and discloses an eddy current sensor zero-point drift compensation method and system, which has the technical scheme as follows: coil resistance, probe wear and multi-channel magnetic field signals corresponding to an eddy current sensor are acquired, and the eddy current sensor is multiple; independent interference components and total interference components are obtained according to the coil resistance and the multi-channel magnetic field signals; a plurality of drift components are obtained according to the total interference components and multi-field coupling drift corresponding to a current period; and drift compensation parameters are obtained according to the drift components and a graph neural network model. The application converts the coupled interference signals into independent components through nonlinear adaptive independent component analysis, provides accurate basis for quantifying the zero-point drift caused by interference, and accurately captures time sequence mutations such as sudden temperature rise and interference amplitude mutation and linear drift characteristics such as temperature and sensor wear through a time sequence convolutional neural network model and a graph neural network model, so that the compensation precision is improved.

Description

Technical Field

[0001] This invention relates to the field of zero-point drift compensation technology, and more specifically to a method and system for zero-point drift compensation of an eddy current sensor. Background Technology

[0002] Eddy current sensors are widely used in the detection of physical quantities such as displacement, vibration, thickness, and gaps due to their advantages, including non-contact measurement, fast response speed, and strong anti-interference capability. For example, in shaft vibration monitoring of mechanical equipment, eddy current sensors can capture the radial displacement changes of the shaft in real time; in semiconductor wafer manufacturing, they are used to monitor the thickness uniformity of wafers; and in power equipment, they are used to detect the wear degree of high-voltage switch contacts.

[0003] However, in practical applications, the zero point of eddy current sensors can drift due to various factors. These factors include changes in ambient temperature altering the electromagnetic properties of the sensor itself and the measured object, interference from complex external electromagnetic environments (such as industrial interference and radio frequency interference), wear and aging of sensor components, and nonlinear coupling between multiple physical fields such as temperature, pressure, and electromagnetic interference. Zero-point drift severely affects the measurement accuracy of eddy current sensors, adversely impacting equipment status assessment and process control based on the measurement results, and even leading to serious consequences such as equipment failure and product quality defects. Currently, compensation technologies for zero-point drift in eddy current sensors lack sufficient ability to resolve nonlinear interference and are inaccurate in capturing temporal abrupt changes and linear drift, thus existing technologies have shortcomings. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method and system for zero-point drift compensation of eddy current sensors. By using nonlinear adaptive independent component analysis, coupled interference signals are converted into independent components, providing an accurate basis for zero-point drift quantification of interference causes. Furthermore, by using a temporal convolutional neural network model, the invention accurately captures temporal abrupt changes such as sudden temperature rises and abrupt changes in interference amplitude, as well as linear drift characteristics such as temperature and sensor wear, thereby improving compensation accuracy.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a method for zero-point drift compensation of an eddy current sensor, comprising:

[0007] Acquire the coil impedance and multi-channel magnetic field signal corresponding to the eddy current sensor, wherein there are multiple eddy current sensors;

[0008] The independent interference component and the total interference component are obtained based on the coil impedance and the multi-channel magnetic field signal.

[0009] Based on the total interference component and the multi-field coupling drift corresponding to the current period, multiple drift components are obtained;

[0010] Based on the drift components and the graph neural network model, drift compensation parameters are obtained.

[0011] As a further improvement of the present invention, the independent interference component and the total interference component are obtained based on the coil impedance and the multi-channel magnetic field signal, including:

[0012] The original data matrix is ​​obtained based on the multi-channel magnetic field signal;

[0013] Wavelet packet decomposition is performed on the original data matrix to obtain amplitude vectors corresponding to multiple sub-frequency bands;

[0014] Based on the coil impedance and the amplitude vector, a global feature vector is obtained;

[0015] The independent interference components and the total interference components are obtained based on the global feature vector.

[0016] As a further improvement of the present invention, the independent interference components and the total interference components are obtained based on the global feature vector, including:

[0017] The global feature vector is nonlinearly mapped using a kernel function to obtain high-dimensional mapped features;

[0018] Based on the high-dimensional mapping characteristics and short-time Fourier transform, the time-frequency coupling coefficient is obtained;

[0019] The independent interference component and the total interference component are obtained based on the time-frequency coupling coefficient.

[0020] As a further improvement of the present invention, the independent interference component and the total interference component are obtained based on the time-frequency coupling coefficient, including:

[0021] If the time-frequency coupling coefficient is less than a preset value, the independent interference component and the total interference component are obtained according to the separation matrix and the high-dimensional mapping feature, and the separation matrix is ​​obtained according to the initial separation matrix;

[0022] If the time-frequency coupling coefficient is greater than or equal to a preset value, a time-frequency adaptive separation basis is obtained based on the separation matrix;

[0023] Based on the time-frequency adaptive separation basis and the high-dimensional mapping features, the independent interference components and the total interference components are obtained.

[0024] As a further improvement of the present invention, the step of obtaining the initial separation matrix includes:

[0025] Acquire multiple independent sub-band interference signals;

[0026] Independent amplitude vectors and mixed interference signals are obtained based on the sub-band interference signals;

[0027] The initial separation matrix is ​​obtained based on the independent amplitude vector, the mixed interference signal, and the independent component analysis.

[0028] As a further improvement of the present invention, multiple drift components are obtained based on the total interference component and the multi-field coupling drift corresponding to the current period, including:

[0029] The coupling strength is obtained based on the multi-field coupling drift corresponding to the current cycle and the multi-field coupling drift corresponding to the adjacent previous cycle.

[0030] The size of the convolution kernel and the void ratio are obtained based on the coupling strength.

[0031] Based on the size and dilation rate of the convolutional kernel and the temporal convolutional neural network model, high-dimensional coupling features are obtained;

[0032] The multiple drift components are obtained based on the high-dimensional coupling characteristics and the total interference components.

[0033] As a further improvement of the present invention, high-dimensional coupling features are obtained based on the size and dilatation rate of the convolutional kernel and the temporal convolutional neural network model, including:

[0034] The offset is obtained based on the global feature vector and the size of the convolution kernel;

[0035] The high-dimensional coupling feature is obtained based on the offset, the hole rate, and the convolutional layers in the temporal convolutional neural network model.

[0036] As a further improvement of the present invention, drift compensation parameters are obtained based on the drift components and the graph neural network model, including:

[0037] Based on the total interference components and the graph neural network model, multi-field coupling drift is obtained;

[0038] The total drift is obtained based on the drift components and the multi-field coupled drift.

[0039] The drift compensation parameters are obtained based on the total drift.

[0040] As a further improvement of the present invention, the drift compensation parameters are obtained based on the total drift, including:

[0041] Based on the total drift and frequency adjustment coefficients, the basic compensation parameters and coupling compensation parameters are obtained;

[0042] The drift compensation parameters are obtained based on the basic compensation parameters and the coupling compensation parameters.

[0043] This invention provides a zero-point drift compensation system for an eddy current sensor, comprising:

[0044] The acquisition module is used to acquire the coil impedance and multi-channel magnetic field signals corresponding to the eddy current sensor, wherein there are multiple eddy current sensors;

[0045] A separation module is used to obtain independent interference components and total interference components based on the coil impedance and multi-channel magnetic field signals;

[0046] The drift module is used to obtain multiple drift components based on the total interference component and the multi-field coupling drift corresponding to the current period;

[0047] The compensation module is used to obtain drift compensation parameters based on the drift components and the graph neural network model.

[0048] This invention uses adaptive independent component analysis to accurately decompose nonlinear interference within multi-channel magnetic field signals, transforming coupled interference signals into independent components. This provides an accurate basis for quantifying zero-point drift to determine the causes of interference. By combining graph neural network models and temporal convolutional neural network models, it accurately captures temporal abrupt changes such as sudden temperature rises and abrupt changes in interference amplitude, as well as linear drift characteristics such as temperature and sensor wear. It also accurately quantifies the contribution of multi-field coupling to zero-point drift, obtaining drift compensation parameters and improving drift compensation accuracy. Compared to traditional off-calibration methods, this invention compensates for zero-point drift by adjusting parameters during sensor operation, enabling the sensor to output accurate results. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method steps of the present invention;

[0050] Figure 2 A schematic diagram illustrating the steps to obtain the independent interference components and the total interference components;

[0051] Figure 3 A schematic diagram illustrating the steps to obtain drift compensation parameters;

[0052] Figure 4 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0053] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof.

[0054] The term "and / or" in the following text is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0055] like Figure 1 As shown in the figure, this application provides a method for zero-point drift compensation of an eddy current sensor, including:

[0056] Acquire the coil impedance and multi-channel magnetic field signals corresponding to the eddy current sensor; there are multiple eddy current sensors.

[0057] The independent interference components and the total interference components are obtained based on the coil impedance and the multi-channel magnetic field signal.

[0058] Based on the total interference component and the multi-field coupling drift corresponding to the current period, multiple drift components are obtained;

[0059] The drift compensation parameters are obtained based on the drift components and the graph neural network model.

[0060] In this embodiment, the eddy current sensor is located on the device under test, and the method provided in this embodiment is executed multiple times during the detection process, that is, once every preset period. This embodiment does not limit the device under test and the period interval, which can be determined by those skilled in the art. For example, since instantaneous interferences such as equipment ignition interference change at the millisecond level, in order to capture such interferences and avoid compensation lag, the period interval is also in milliseconds, such as 10ms or 50ms. Near-field probes are deployed around the eddy current sensor to collect multi-channel magnetic field signals. The multi-channel magnetic field signals contain various ultra-wideband interferences in industrial scenarios, such as industrial electromagnetic interference, radio frequency interference, and equipment ignition interference. By separating the multi-channel magnetic field signals, independent interference components are obtained, which provides a basis for subsequent interference drift calculation and ultimately achieves zero-point drift compensation of all causes.

[0061] Specifically, the core sensing element of an eddy current sensor is the coil. The coil impedance fluctuates with industrial environmental temperature (e.g., equipment heating, ambient temperature changes). The eddy current sensor probe needs to be in close proximity to the device under test for extended periods, leading to mechanical wear over time. Electromagnetic interference, radio frequency interference, and equipment ignition interference in industrial settings are essentially alternating electromagnetic fields, which diffuse in the space where the eddy current sensor is located through near-field radiation. These interferences couple with the sensor's effective signal, causing irregular fluctuations in coil impedance and output signal. Therefore, temperature changes, mechanical wear, and interference in industrial settings all cause zero-point drift in eddy current sensors. Furthermore, because the various types of interference in industrial settings are distributed in multiple directions and frequency bands, single-channel acquisition cannot fully cover them. This embodiment acquires multi-channel magnetic field signals to provide full-band data for subsequent analysis.

[0062] This embodiment acquires the coil impedance and multi-channel magnetic field signals of the eddy current sensor, performs independent component analysis to obtain independent interference components and total interference components, and achieves accurate quantification of interference cause drift. At the same time, it combines a time-series convolutional neural network model to output multiple drift components such as temperature, interference, and wear. Then, it uses a graph neural network model to model the nonlinear coupling effect of multi-physics fields. Finally, the drift compensation parameters output can adjust the sensor hardware state in real time and improve the zero-point drift compensation accuracy.

[0063] Furthermore, this embodiment provides a step for obtaining independent interference components and total interference components based on coil impedance and multi-channel magnetic field signals, including:

[0064] The original data matrix is ​​obtained based on the multi-channel magnetic field signals;

[0065] Wavelet packet decomposition is performed on the original data matrix to obtain the amplitude vectors corresponding to multiple sub-frequency bands;

[0066] Based on the coil impedance and amplitude vector, the global characteristic vector is obtained;

[0067] The independent interference components and the total interference components are obtained from the global feature vector.

[0068] Specifically, firstly, an original data matrix is ​​obtained based on the multi-channel magnetic field signals collected in the current cycle. The number of rows in the original data matrix corresponds to the number of channels, and the number of columns corresponds to the number of sampling points. Then, wavelet packet decomposition is performed on each row of the original data matrix. The number of wavelet packet decomposition layers is determined based on the frequency band coverage of the multi-channel magnetic field signals and the main frequency band width of different types of interference. This embodiment does not impose limitations on this. For example, for different interference types such as equipment ignition interference, signals can be collected in advance through simulation experiments and spectrum analysis can be performed to obtain the main frequency band width corresponding to different interference types. Then, a reference width is determined based on multiple main frequency band widths, such as the mean or mode of multiple main frequency band widths. After obtaining the multi-channel magnetic field signals, the specific number of decomposition layers is determined based on the coverage of the multi-channel magnetic field signals and the reference width. For example, if the multi-channel magnetic field signals cover 0.1MHz-1GHz, and the reference width is 0.1GHz, then through 8 layers of wavelet packet decomposition, the entire frequency band can be divided into 8 sub-bands, so that the bandwidth of each sub-band matches the main frequency band range of different interference types, ensuring accurate capture of the characteristics of each type of interference. Taking 8-layer wavelet packet decomposition as an example, for each eddy current sensor, after acquiring multi-channel magnetic field signals, wavelet packet decomposition is first performed on the multi-channel magnetic field signals. At this time, the signal of each channel is split into 8 sub-bands. For each channel, the amplitude corresponding to each sub-band is concatenated to obtain the amplitude vector corresponding to each channel. Then, the eddy current sensor performs comprehensive processing on the amplitude vector corresponding to each channel within the sensor, such as merging according to the weight of the interference direction or taking the average of the amplitudes of multiple channels, integrating the independent signals of multiple channels into the amplitude vector corresponding to the sensor. Through wavelet packet decomposition and processing within the eddy current sensor, the chaotic multi-channel original signal is transformed into accurate and usable features, avoiding the problems of interference feature aliasing and insufficient analysis accuracy.

[0069] Then, based on the deviation of the coil impedance of multiple eddy current sensors, normal sensors are selected. Taking three sensors A, B, and C as an example, the difference in coil impedance in each pairwise combination can be calculated. If each difference is less than or equal to a preset difference, then all three sensors are considered normal sensors. If the difference in coil impedance between sensors A and B, and the difference in coil impedance between sensors B and C are both greater than the preset difference, but the difference in coil impedance between sensors A and C is less than the preset difference, then since sensor B appears twice, sensors A and C are considered normal sensors. If only one set of sensors has a coil impedance difference greater than the preset difference, it indicates that there may be a temporary abnormal fluctuation. It is determined that there is no clearly abnormal sensor, and all sensors are considered normal sensors. If the difference in coil impedance between all three sets of sensors is greater than the preset difference, it indicates that there is no consistency in the data between the three sensors. This may be caused by a sudden change in the environment or a sudden failure of multiple sensors. In this case, the error will be large when performing subsequent analysis based on the data collected by these sensors, and the detection needs to be stopped in time for repair.

[0070] Then, the amplitude vectors corresponding to the normal sensors are obtained and weighted and fused to obtain the global feature vector. The weights can be determined based on the standard deviation of the coil impedance of each normal sensor within a preset number of cycles. The smaller the standard deviation, the greater the weight. Then, independent component analysis is performed based on the global feature vector to generate independent interference components and total interference components.

[0071] This embodiment constructs an original data matrix from multi-channel magnetic field signals, then obtains multi-band amplitude vectors through wavelet packet decomposition, and combines this with coil impedance to obtain a global feature vector. Finally, independent component analysis is used to extract independent interference components and total interference components, achieving precise decomposition of nonlinear interference. This provides accurate independent and total interference amplitude data for zero-point drift quantification of interference causes, and simultaneously provides multi-band interference feature inputs for subsequent graph neural network models, effectively improving the calculation accuracy of interference-type drift components. This is a key prerequisite for achieving high-precision compensation for all causes of zero-point drift in eddy current sensors, laying a solid foundation of interference features for the accuracy of the entire compensation scheme.

[0072] Furthermore, this embodiment provides a step for obtaining independent interference components and total interference components based on global feature vectors, including:

[0073] High-dimensional mapped features are obtained by performing a nonlinear mapping on the global feature vector using a kernel function;

[0074] The time-frequency coupling coefficient is obtained based on the high-dimensional mapping characteristics and short-time Fourier transform.

[0075] The independent interference components and the total interference components are obtained based on the time-frequency coupling coefficient.

[0076] Specifically, firstly, a nonlinear mapping is performed on the global feature vector based on the kernel function to obtain high-dimensional mapped features. for:

[0077]

[0078] in, For global feature vectors, For An exponential function with base 0. For the closest The core center, express and The Euclidean distance and kernel center can be obtained through training. Specifically, multiple labeled data are first acquired as a training set. Each labeled data includes the global feature vector corresponding to the standard signal and its multiple independent interference components decomposed from it. The labeled data can be divided into two categories: steady-state samples and abrupt samples. Steady-state samples include two types: no interference and steady-state interference. Abrupt samples include one type: abrupt interference. Steady-state interference is regular and predictable, while abrupt interference is irregular. For example, electromagnetic interference generated by a continuously running motor is steady-state interference, while a temporarily started radio frequency transmitter is abrupt interference. Furthermore, since industrial interference has different frequency bands and different amplitudes, and a kernel function can only match one fixed interference mode, multiple kernel functions need to be set to cover all possible interference modes. Each kernel function is different only in its kernel center. This embodiment does not limit the specific number. Those skilled in the art can determine it based on the specific interference type. For example, since the entire frequency band is divided into 8 sub-bands by 8-layer wavelet packet decomposition, so that the bandwidth of each sub-band matches the main frequency band range of different interference types, 8 kernel functions are set here.

[0079] Next, an initial separation matrix is ​​obtained, and an initial kernel width parameter is randomly generated. Then, for each labeled data's corresponding global feature vector, its high-dimensional mapping feature with respect to each kernel function is calculated. The high-dimensional mapping features of each kernel function are then concatenated to obtain a high-dimensional mapping feature vector. The initial separation matrix is ​​then multiplied by the high-dimensional mapping feature vector to obtain the predicted interference component. The squared error between the predicted interference component and the corresponding real multiple independent interference components is calculated as the loss function. Then, each kernel center and the initial separation matrix are updated using gradient descent. The principle of gradient descent is a technique understood by those skilled in the art, and will not be elaborated upon in this embodiment. The above steps are then repeated based on the updated kernel centers and the initial separation matrix. After the number of repetitions reaches a preset number of cycles (e.g., 10 times), training is paused, and the testing phase begins.

[0080] During the testing phase, multiple labeled data sets need to be acquired as test sets. These labeled data sets are of the same type as the labeled data sets mentioned above, including steady-state samples and mutation samples. Then, for the two sets of samples, multiple kernel width parameters are tested respectively. The loss function corresponding to each kernel width parameter is calculated using each kernel center updated after the 10th iteration and the initial separation matrix. Finally, a kernel width parameter corresponding to the minimum loss function for steady-state samples and a kernel width parameter corresponding to the minimum loss function for mutation samples are obtained.

[0081] Then, using the kernel width parameters corresponding to the steady-state and mutation samples obtained in this testing phase, the randomly generated initial kernel width parameters are updated. Each kernel center is trained again, and the kernel width parameters are updated again through the training phase until the preset number of iterations is reached. The updating of the kernel width parameters, each kernel center, and the initial separation matrix is ​​then terminated, and the final updated initial separation matrix is ​​recorded as the separation matrix. In this embodiment, there are no restrictions on the preset number of iterations, the preset number of cycles, and the kernel width parameters used in the testing phase. However, the setting of the number of iterations and the preset number of cycles needs to consider parameter convergence and computational power. Optionally, the preset number of iterations is 50, and the same kernel width parameters are used in multiple testing phases to ensure the stability of the testing rules.

[0082] Furthermore, after inputting the global feature vector, the kernel center with the smallest Euclidean distance to the global feature vector is selected from each kernel center obtained during training. The kernel center with the smallest Euclidean distance is most similar to the interference pattern of the global eigenvector. Then, a short-time Fourier transform is performed on the magnetic field signal corresponding to each channel of the global eigenvector to obtain the time-frequency domain energy distribution. Next, the frequency is divided according to the number of sub-bands to obtain the original energy amplitude of each sub-band, and these are summed to obtain the original total interference amplitude corresponding to the current cycle. Then, the original total interference amplitude corresponding to the adjacent previous cycle is obtained, and the rate of change between the two original total interference amplitudes is calculated. If the current cycle is the first cycle, there is no original total interference amplitude corresponding to the adjacent previous cycle. The reference interference amplitude obtained by calibrating the eddy current sensor in a noise-free laboratory environment can be used as the original total interference amplitude corresponding to the adjacent previous cycle to calculate the rate of change. The rate of change is then used to determine... For example, when the rate of change is greater than a preset rate of change, it indicates a significant abrupt change in the signal amplitude, classifying it as a strong dynamic interference scenario. This indicates that the global feature vector corresponds to the abrupt interference type. In this case, the kernel width parameter corresponding to the abrupt sample is selected from the kernel width parameters obtained during training. To preserve mutation details in high-dimensional mapping, when the rate of change is less than or equal to a preset rate of change, it indicates that the amplitude change is gradual, which is judged as a weak dynamic interference scenario, indicating that the global feature vector corresponds to a steady-state interference type. At this time, the kernel width parameter corresponding to the steady-state sample is selected from the kernel width parameters obtained during training. This method can smooth noise and integrate continuous interference features, avoiding over-focusing on local areas and losing the overall coupling pattern. Since the interference features of steady-state samples are stable over a long period with minimal changes, while those of mutation samples are sudden and drastic, the kernel width parameter corresponding to steady-state samples is usually larger than that corresponding to mutation samples to avoid misjudgment. The preset change rate can be obtained from the test set. This embodiment does not limit this. For example, after the last update of the kernel width parameter, the test set used in the current iteration is obtained, and the mean change rates of the steady-state samples and mutation samples included in it are calculated respectively. Finally, the median of the two means is used as the preset change rate to distinguish the sample types.

[0083] Then, a short-time Fourier transform is performed on the high-dimensional mapping features to obtain a time-frequency matrix. The number of rows in the time-frequency matrix equals the number of frequency points, and the number of columns equals the number of time frames. Specifically, the frequency resolution of the short-time Fourier transform is determined by the window length and the sampling rate of the multi-channel magnetic field signal. Based on the frequency resolution and frequency band coverage, the number of frequency points can be calculated. Based on the frequency resolution, window length, and overlap rate, the time interval between adjacent windows can be calculated. Then, based on the duration and time interval of the multi-channel magnetic field signal, the number of time frames can be calculated. The window length and overlap rate are determined according to the type of window function, such as the Hanning window, which has a window length of 256 points and an overlap rate of 50%.

[0084] The time-frequency coupling coefficient is then calculated based on the time-frequency matrix as follows:

[0085]

[0086] in, The first element in the time-frequency matrix Each frequency point, For the first The average amplitude at each frequency point for and covariance, This represents the variance corresponding to the frequency point.

[0087] Before obtaining the independent interference components and the total interference components, it is first necessary to obtain the initial separation matrix. This embodiment provides a step for obtaining the initial separation matrix, including:

[0088] Acquire multiple independent sub-band interference signals;

[0089] Independent amplitude vectors and mixed interference signals are obtained from sub-band interference signals;

[0090] The initial separation matrix is ​​obtained based on the independent amplitude vector, the mixed interference signal, and the independent component analysis.

[0091] The step of obtaining the initial separation matrix should be performed before the kernel width parameter, each kernel center and the separation matrix are obtained in the above training. Only then can the initial separation matrix be applied to the above training process. The above training process should be performed before the eddy current sensor is turned on.

[0092] Specifically, the first step is to obtain multiple independent sub-band interference signals. Continuing with the previous example, if eight sub-bands are needed, eight uncoupled sub-band interference signals can be artificially generated, and independent amplitude vectors can be formed by combining the amplitudes of the interference signals. Each individual interference signal is then superimposed to obtain a mixed interference signal. The initial separation matrix was then obtained through independent component analysis. , making .

[0093] Furthermore, such as Figure 2 As shown, this embodiment provides a step for obtaining independent interference components and total interference components based on the time-frequency coupling coefficient, including:

[0094] If the time-frequency coupling coefficient is less than the preset value, the independent interference components and the total interference components are obtained based on the separation matrix and high-dimensional mapping characteristics.

[0095] If the time-frequency coupling coefficient is greater than or equal to the preset value, the time-frequency adaptive separation basis is obtained based on the separation matrix;

[0096] Based on the time-frequency adaptive separation basis and high-dimensional mapping characteristics, the independent interference components and the total interference components are obtained.

[0097] Based on the above analysis, the separation matrix is ​​obtained by training and updating the initial separation matrix.

[0098] Specifically, if the time-frequency coupling coefficient is less than a preset value, weak coupling processing is performed. The high-dimensional mapping features are decomposed using a separation matrix to obtain the interference component matrix:

[0099]

[0100] Each row in the interference component matrix represents an independent interference component. If the time-frequency coupling coefficient is greater than or equal to a preset value, strong coupling processing is performed. First, the time-frequency adaptive separation basis is obtained through the separation matrix:

[0101]

[0102] in, The average amplitude of each column in the time-frequency matrix. The matrix is ​​a diagonal matrix. Then, based on the time-frequency adaptive separation basis, the interference component matrix is ​​obtained as follows:

[0103]

[0104] After obtaining each independent interference component from the interference component matrix, each independent interference component is weighted to obtain the total interference component. In this embodiment, the specific value of the weight is not limited. For example, the weight of each independent interference component can be set to be the same to ensure that the contribution of each interference component is counted equally.

[0105] Among these, strong coupling and weak coupling, along with steady state and abrupt change mentioned earlier, are two independent separation dimensions. Steady state and abrupt change are distinguished based on the rate of change of amplitude, while strong / weak coupling and steady state / interference type are classified based on the degree of overlap of different interference frequency bands. For both steady state and abrupt change types, there are corresponding preset values. The following describes the process of determining the preset values ​​using the steady state type as an example. First, multiple steady state samples are obtained, and their corresponding time-frequency matrices and time-frequency coupling coefficients are obtained based on the above steps. Then, for each steady state sample, the predicted interference component is obtained based on the above separation matrix, and the squared error between the predicted interference component and the corresponding multiple real independent interference components is calculated. Therefore, the corresponding squared error can be obtained for each time-frequency coupling coefficient. Next, each steady state sample is sorted in ascending order according to its time-frequency coupling coefficient, and all samples are divided into multiple intervals according to the time-frequency coupling coefficient. For example, for every 0.2 increase in the time-frequency coupling coefficient, a new interval is defined (this embodiment does not impose restrictions on this). Then, the mean squared error of all steady-state samples within each interval is calculated. Since the overlap between different interference frequency bands decreases with increasing time-frequency coupling coefficient, the predicted interference components are more accurate, and the squared error is smaller. Then, based on the time-frequency coupling coefficient, the mean squared error of each interval is iterated from smallest to largest to determine the first mean squared error greater than a preset error. The sample interval corresponding to this mean squared error is then determined, and the maximum time-frequency coupling coefficient included in the sample interval is used as the preset value corresponding to the steady-state type. Similarly, the preset value corresponding to the mutation type can be obtained. Therefore, after obtaining the time-frequency coupling coefficient, an appropriate preset value needs to be selected for coupling processing according to the specific type.

[0106] The preset error is the acceptable accuracy of the separation matrix in separating interference components. When the time-frequency coupling coefficient is less than the preset value, a high accuracy can be achieved by using only the separation matrix. When the time-frequency coupling coefficient is greater than or equal to the preset value, the error is large when using only the separation matrix. Therefore, this embodiment further sets a time-frequency adaptive separation basis. The essence of the time-frequency adaptive separation basis is to add the weight of time-frequency features to the initial separation matrix, so as to more accurately distinguish different interference components in the overlapping frequency band.

[0107] This embodiment performs both strong and weak coupling processing to adapt to the signal coupling characteristics under different interference scenarios, achieving accurate interference decomposition. In weak coupling scenarios, a separation matrix is ​​used to efficiently decompose interference with low coupling, while in strong coupling scenarios, a time-frequency adaptive separation basis is used to specifically enhance the separation capability for highly coupled interference. This embodiment obtains independent interference components and total interference components by dynamically switching the separation strategy, providing an accurate basis for subsequent interference drift calculations. Simultaneously, it provides detailed interference features for multi-field coupling modeling of graph neural networks, improving the quantization accuracy of interference-type drift.

[0108] Furthermore, this embodiment provides a step for obtaining multiple drift components based on the total interference component and the multi-field coupling drift corresponding to the current period, including:

[0109] The coupling strength is obtained based on the multi-field coupling drift corresponding to the current cycle and the multi-field coupling drift corresponding to the adjacent previous cycle.

[0110] The size of the convolution kernel and the hole ratio are obtained based on the coupling strength;

[0111] Based on the size and dilatation rate of the convolutional kernel and the temporal convolutional neural network model, high-dimensional coupling features are obtained;

[0112] Based on the high-dimensional coupling characteristics and the total interference components, multiple drift components are obtained.

[0113] For each cycle, the multi-field coupling drift corresponding to the current cycle and the multi-field coupling drift corresponding to the adjacent previous cycle are first obtained. The rate of change of the multi-field coupling drift corresponding to the current cycle relative to the multi-field coupling drift corresponding to the adjacent previous cycle is recorded as the coupling strength. Since the multi-field coupling drift corresponding to the current cycle has not yet been calculated, the multi-field coupling drift corresponding to the two adjacent previous cycles can be obtained, and the rate of change between these two multi-field coupling drifts is calculated. Then, based on the multi-field coupling drift corresponding to the adjacent previous cycle and the rate of change, an approximate multi-field coupling drift corresponding to the current cycle is obtained and used to calculate the coupling strength. If the current cycle is the first cycle, the coupling strength can be initialized to 0% for subsequent calculations. The reason for choosing multi-field coupling drift in this embodiment is that multi-field coupling drift is an additional zero-point drift generated by the mutual coupling of multiple physical quantities such as temperature, interference, and probe wear. It directly reflects the strength and degree of change of the interaction between physical quantities. A single drift component only reflects the independent effect of a single physical quantity and cannot characterize the dynamic change of the coupling effect. Therefore, based on multi-field coupling drift, the degree of change of the coupling timing can be accurately quantified, providing a basis for subsequent dynamic adjustment.

[0114] Next, the preset interval to which the coupling strength belongs is determined, thereby obtaining the corresponding case type. Case types include normal case, mild mutation case, and severe mutation case. This embodiment does not limit the specific coupling strength values ​​for dividing the preset intervals, where the coupling strength is lowest in the normal case and highest in the severe mutation case. For example, multiple coupling drifts over the past 1-3 months can be extracted and the rate of change between adjacent periods can be calculated. A histogram of the rate of change distribution can be plotted for analysis. The 95th percentile is used as the upper limit for the normal case (e.g., if 95% of the rate of change is ≤6%, the preset interval for the normal case can be set to 0-6%). The 99th percentile is used as the boundary between mild and severe mutation cases (e.g., if 99% of the rate of change is ≤18%, the preset interval for mild mutation cases can be set to 6%-18%, and the preset interval for severe mutation cases can be set to >18%). This makes the coupling strength interval more closely match the actual drift characteristics of a specific scenario. Then, based on the coupling strength, a pre-calibrated parameter table is called to obtain the convolution kernel size. and void ratio kernel size and void ratio Together, these factors determine the receptive field of the convolution to match the feature characteristics of different situations. For example, in normal cases, since the drift changes are gradual, it is only necessary to focus on the gradual changes in adjacent cycles and avoid expanding the receptive field to introduce noise. Therefore, a smaller convolution kernel and dilation rate (e.g., K=3, d=1) are set to accurately capture slow drifts while ensuring computational efficiency. In cases of mild mutations, since the drift has small jumps and is accompanied by a gradual trend, it is necessary to cover a longer temporal context to fully track the changes before and after the jump. Therefore, a medium convolution kernel and dilation rate (e.g., K=5, d=2) are set to achieve complete temporal feature extraction of small abnormal drifts and avoid feature loss. In cases of severe mutations, since the drift is violent, long-lasting, and complex, it is necessary to cover a wider temporal range to capture the complete mutation process and subsequent development. Therefore, a larger convolution kernel and dilation rate (e.g., K=7, d=3) are set to achieve comprehensive perception of violent mutation drifts and avoid missing key faults.

[0115] The parameter table can be obtained based on sample data, which includes steady-state samples, interference samples, and samples belonging to strong and weak coupling. For example, for each sample, the sample corresponding to each case type is obtained according to the corresponding coupling strength. Then, for each case type, the size and dilatation rate of multiple sets of convolutional kernels are set. Then, the difference between the sum of the drift components of the final output of the model and the total drift amount actually measured is calculated under different convolutional kernel sizes and dilatation rates. The total drift amount actually measured is the difference between the real-time measurement value and the steady-state reference value. Finally, the size and dilatation rate of the convolutional kernel with the smallest difference are taken as the size and dilatation rate of the convolutional kernel corresponding to the case type and put into the parameter table. In this embodiment, there is no limitation on the size and dilatation rate of each set of convolutional kernels or the number of sets. For example, the number of sets is set to 3 to cover the three case types: normal case, mild mutation case, and severe mutation case.

[0116] Furthermore, this embodiment provides a step for obtaining high-dimensional coupled features based on the size and dilatation rate of the convolutional kernel and a temporal convolutional neural network model, including:

[0117] The offset is obtained based on the global feature vector and the size of the convolution kernel;

[0118] High-dimensional coupling features are obtained based on the offset, the hole rate, and the convolutional layers in the temporal convolutional neural network model.

[0119] Specifically, after obtaining the global feature vector corresponding to the current cycle, the temperature, pressure, and probe wear corresponding to the current moment are obtained. Probe wear can be represented by the difference between the current thickness of the sensor probe and the initial thickness. The temperature, pressure, and probe wear are then dimensionally processed and concatenated with the global feature vector to obtain the fused feature vector corresponding to the current cycle. Next, the fused feature vector corresponding to each cycle within a preset time period needs to be obtained. In this embodiment, the length of the preset time period is not limited. The fused feature vectors corresponding to each cycle are concatenated to obtain the fused temporal feature, which is a matrix. The fused temporal feature is then input into the offset prediction branch of the temporal convolutional neural network model to obtain the offset. The offset prediction branch is a convolutional layer, and the offset represents the positional offset required for each convolutional kernel sampling point, thereby focusing the sampling points on the region with the greatest change in coupling features (such as moments of sudden temperature rise or sudden changes in interference amplitude). After obtaining the offset, the fixed sampling positions of conventional convolution are used... and offset The offset position is obtained as .

[0120] Next, based on the offset, dilation rate, and convolutional layers in the temporal convolutional neural network model, deformable convolution is performed to obtain high-dimensional coupled features. The specific formula for deformable convolution is:

[0121]

[0122] in, The size of the convolution kernel. This refers to the index of the sampling points within the convolution kernel. Indicates the first convolution kernel Sampling points at each location, For convolution kernel weights, To integrate temporal features, To round down, The first in the high-dimensional coupling feature The values ​​at each position are then input into the output layer. The output layer consists of multiple independent branches, ultimately outputting three physically meaningful drift components. Each component corresponds to a drift cause (temperature, interference, wear). Through the branching structure, the output layer ensures that each branch learns only one mapping relationship between the cause and the high-dimensional coupling feature, thus separating out multiple drift components.

[0123] The temporal convolutional neural network model needs to be trained. For example, time-series data covering multiple operating conditions can be collected, including global feature vectors under different temperature changes, interference, and probe wear scenarios, as well as corresponding temperature, pressure, and probe wear data. The actual drift components under the corresponding operating conditions are obtained through experimental calibration. Data preprocessing is then performed to obtain fused feature vectors and fused temporal feature matrices, which are then divided into training and validation sets, such as in an 8:2 ratio. The model includes an input layer, an offset prediction branch, deformable convolutional layers, and an output layer. The input layer receives the fused temporal feature matrix. The offset prediction branch consists of one or more convolutional layers that output offsets based on the fused temporal features. The deformable convolutional layers perform deformable sampling of the fused temporal features based on the offsets, outputting high-dimensional coupled features. The output layer includes three independent branches. During training, the fused temporal feature matrix from the training set is first input, and the output results of the three branches are obtained. The average mean squared error of the three branches is calculated based on the output results and the true drift components as the loss function. The parameters in the model are updated through the loss function. The parameters in the model include the bias terms and weights in the convolutional layers and the kernel weights of the deformable convolutional layers. When updating the parameters, the Adam optimizer can be selected, and the learning rate is set to 0.001. The training termination condition is reaching the preset number of iterations, such as 500 times.

[0124] This embodiment achieves adaptive capture of temporal abrupt changes and linear drift features by dynamically adjusting the convolution kernel size and deformable convolution based on offset prediction. It can accurately extract linear permeability drift caused by slow temperature changes and mechanical drift caused by sensor wear accumulation, and focus on the nonlinear features of interference amplitude abrupt changes and multi-physics coupling, outputting independent interference drift. This provides a foundation for subsequent multi-field coupling modeling of graph neural network models and ensures compensation accuracy.

[0125] Furthermore, this embodiment provides a step for obtaining drift compensation parameters based on drift components and a graph neural network model, including:

[0126] Based on the total interference components and the graph neural network model, the multi-field coupling drift is obtained;

[0127] The total drift is obtained based on the drift components and multi-field coupled drift.

[0128] The drift compensation parameters are obtained based on the total drift.

[0129] The graph neural network model simulates the coupling relationship between multiple physics fields through nodes and edges. Each node corresponds to a physical quantity, including independent interference components, total interference components, temperature, pressure, and probe wear. Edge weights are used to measure the coupling relationship between different physical quantities. The larger the weight, the more significant the mutual influence between the two physical quantities. The specific value of the edge weight can be optimized through experimental simulation or through historical data samples. This embodiment does not impose any restrictions on this, and not every two nodes have an edge. Those skilled in the art can determine the value based on the actual situation.

[0130] After inputting the physical quantity corresponding to the current cycle into the graph neural network model, the graph neural network model obtains the multi-field coupling drift through message passing and feature aggregation. Specifically, each node combines its own physical quantity with the edge weights. The physical quantity corresponding to each node needs to be converted into a dimensionless value of the same dimension. Then, it passes the coupling information to each neighboring node. The neighboring node is the node connected to the node through an edge. Each node may have multiple neighboring nodes. The coupling information passed to a neighboring node is the product of its own value and the edge weight between the node and the neighboring node. After receiving the coupling information from the neighboring node, each node averages and aggregates it with its own value and then passes it again. After multiple rounds of message passing and aggregation, the output layer maps the values ​​corresponding to all nodes to the contribution of multi-field coupling to the zero-point drift, i.e., multi-field coupling drift. Then, the three components of temperature drift, disturbance drift, and wear drift are added to the multi-field coupling drift to obtain the total drift.

[0131] This embodiment sets up nodes for each core factor (physical quantity) causing zero-point drift, and there are real industrial coupling relationships between these physical quantities. Multiple rounds of information transmission allow each node to gradually integrate the coupling information of all related physical quantities, thereby accurately restoring the coupling process of nested multi-physical fields. Furthermore, since the total drift in an industrial scenario is the sum of the independent contribution of a single physical quantity and the additional contribution of mutual influence between physical quantities, this embodiment obtains the total drift by adding the three components of temperature drift, disturbance drift, and wear drift with the multi-field coupling drift. This can cover all drift sources and reduce drift compensation errors under complex working conditions.

[0132] Furthermore, such as Figure 3 As shown, this embodiment provides a step for obtaining drift compensation parameters based on total drift, including:

[0133] Based on the total drift and frequency adjustment coefficients, the basic compensation parameters and coupling compensation parameters are obtained;

[0134] The drift compensation parameters are obtained based on the basic compensation parameters and the coupling compensation parameters.

[0135] After obtaining the total drift, it is necessary to obtain the total drift corresponding to the adjacent previous cycle. Then, compare the absolute value of the current total drift with the absolute threshold. If it is greater, proceed to the step of obtaining drift compensation parameters. If it is less than or equal to the absolute threshold, calculate the absolute value of the difference between the current total drift and the total drift corresponding to the adjacent previous cycle. If the absolute value is greater than the relative threshold and the current total drift is more significant than the previous cycle, proceed to the step of obtaining drift compensation parameters. Otherwise, do not proceed to the step of obtaining drift compensation parameters. The absolute threshold refers to the maximum drift allowed by the sensor measurement accuracy, and the relative threshold is a warning value for a sudden change in drift, indicating that the total drift value is changing rapidly and the parameters need to be adjusted to prevent further deterioration.

[0136] Specifically, in the total drift, the sum of temperature drift, interference drift, and wear drift constitutes the base drift. The base drift is the overall drift caused by all factors. The base compensation parameters include the excitation frequency adjustment and the amplification gain adjustment. Multiplying the base drift by the frequency adjustment coefficient yields the excitation frequency adjustment. The frequency adjustment coefficient can be calibrated experimentally. For example, if the experiment shows that 0.0004MHz adjustment is required for every 0.001Ω drift, then the frequency adjustment coefficient is 0.4. Multiplying the base drift by the gain adjustment coefficient yields the amplification gain adjustment. The gain adjustment coefficient can also be calibrated experimentally.

[0137] In the total drift, multi-field coupling drift is an additional drift caused by the interaction of multiple physical quantities. It cannot be covered by basic compensation. When the multi-field coupling drift is greater than the trigger threshold, it indicates that the multi-field coupling drift is strong, and the excitation frequency needs to be adjusted again. At this time, the multi-field coupling drift needs to be multiplied by the coupling frequency adjustment coefficient to obtain a second excitation frequency adjustment. The coupling frequency adjustment coefficient can also be calibrated experimentally and is denoted as the coupling compensation parameter. The basic compensation parameter and the coupling compensation parameter are collectively referred to as the drift compensation parameter. Finally, the excitation frequency and amplification gain are adjusted according to the drift compensation parameter to compensate for the zero-point offset. If the multi-field coupling drift is less than or equal to the trigger threshold, there is no need to calculate the coupling compensation parameter. In this case, the basic compensation parameter is the drift compensation parameter. In this embodiment, the adjustment direction of the drift compensation parameter needs to be determined according to the change direction of the current total drift value relative to the total drift corresponding to the adjacent previous cycle.

[0138] In this embodiment, the trigger threshold is not limited. Those skilled in the art can determine it according to the accuracy requirements. For example, multi-field coupling drift data from the past 1-3 months can be extracted, and its distribution histogram can be plotted to obtain the 95th percentile (i.e., 95% of the coupling drift data are below this value). The selection of this quantile must simultaneously satisfy the requirements of covering the normal coupling fluctuation range in the field and not exceeding the upper limit of the allowable error of the eddy current sensor measurement accuracy. This quantile is used as the new trigger threshold (for example, if 95% of the coupling drift in a certain workshop is ≤0.0006Ω, and this value does not exceed the allowable error range of the sensor measurement accuracy, then the trigger threshold can be updated to 0.0006Ω). This makes the threshold more suitable for the actual drift characteristics of the specific field, while ensuring that the measurement results always meet the industrial accuracy requirements. Furthermore, since only the excitation frequency can match the changes in electromagnetic characteristics caused by coupling, and the amplification gain cannot solve the physical errors such as the distortion of electromagnetic interaction caused by coupling, this embodiment only considers the excitation frequency when calculating the coupling compensation parameter.

[0139] In this embodiment, when the multi-field coupling drift exceeds the trigger threshold, an additional frequency adjustment is added to improve the compensation accuracy. Through the synergy of basic compensation and coupling compensation, high-precision real-time compensation under all causes and all scenarios is achieved, effectively solving the shortcomings of existing technologies in multi-physics field coupling drift compensation.

[0140] Furthermore, such as Figure 4 As shown, this application provides an eddy current sensor zero-point drift compensation system, comprising:

[0141] The acquisition module is used to acquire the coil impedance and multi-channel magnetic field signals corresponding to the eddy current sensor; there are multiple eddy current sensors.

[0142] The separation module is used to obtain the independent interference component and the total interference component based on the coil impedance and multi-channel magnetic field signal;

[0143] The drift module is used to obtain multiple drift components based on the total interference component and the multi-field coupling drift corresponding to the current period;

[0144] The compensation module is used to obtain drift compensation parameters based on the drift components and the graph neural network model.

[0145] The acquisition module, separation module, drift module, and compensation module are all located in the server.

[0146] This application provides a zero-point drift compensation method and system for an eddy current sensor. Through adaptive independent component analysis, it accurately decomposes nonlinear interference within multi-channel magnetic field signals, converting coupled interference signals into independent components. This provides an accurate basis for quantifying the zero-point drift caused by interference. Furthermore, by combining graph neural network and temporal convolutional neural network models, it accurately captures temporal abrupt changes such as sudden temperature rises and abrupt changes in interference amplitude, as well as linear drift characteristics such as temperature and sensor wear. It also accurately quantifies the contribution of multi-field coupling to zero-point drift, obtaining drift compensation parameters and improving drift compensation accuracy.

[0147] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for zero-point drift compensation of an eddy current sensor, characterized in that, include: Acquire the coil impedance and multi-channel magnetic field signal corresponding to the eddy current sensor, wherein there are multiple eddy current sensors; The independent interference component and the total interference component are obtained based on the coil impedance and the multi-channel magnetic field signal. Based on the total interference component and the multi-field coupling drift corresponding to the current period, multiple drift components are obtained; Based on the drift components and the graph neural network model, drift compensation parameters are obtained.

2. The zero-point drift compensation method for an eddy current sensor according to claim 1, characterized in that, Based on the coil impedance and multi-channel magnetic field signal, the independent interference components and the total interference components are obtained, including: The original data matrix is ​​obtained based on the multi-channel magnetic field signal; Wavelet packet decomposition is performed on the original data matrix to obtain amplitude vectors corresponding to multiple sub-frequency bands; Based on the coil impedance and the amplitude vector, a global feature vector is obtained; The independent interference components and the total interference components are obtained based on the global feature vector.

3. The zero-point drift compensation method for an eddy current sensor according to claim 2, characterized in that, The independent interference components and the total interference components are obtained based on the global feature vector, including: The global feature vector is nonlinearly mapped using a kernel function to obtain high-dimensional mapped features; Based on the high-dimensional mapping characteristics and short-time Fourier transform, the time-frequency coupling coefficient is obtained; The independent interference component and the total interference component are obtained based on the time-frequency coupling coefficient.

4. The zero-point drift compensation method for an eddy current sensor according to claim 3, characterized in that, The independent interference components and the total interference components are obtained based on the time-frequency coupling coefficient, including: If the time-frequency coupling coefficient is less than a preset value, the independent interference component and the total interference component are obtained according to the separation matrix and the high-dimensional mapping feature, and the separation matrix is ​​obtained according to the initial separation matrix; If the time-frequency coupling coefficient is greater than or equal to a preset value, a time-frequency adaptive separation basis is obtained based on the separation matrix; Based on the time-frequency adaptive separation basis and the high-dimensional mapping features, the independent interference components and the total interference components are obtained.

5. The zero-point drift compensation method for an eddy current sensor according to claim 4, characterized in that, The steps to obtain the initial separation matrix include: Acquire multiple independent sub-band interference signals; Independent amplitude vectors and mixed interference signals are obtained based on the sub-band interference signals; The initial separation matrix is ​​obtained based on the independent amplitude vector, the mixed interference signal, and the independent component analysis.

6. The zero-point drift compensation method for an eddy current sensor according to claim 2, characterized in that, Based on the total interference component and the multi-field coupling drift corresponding to the current period, multiple drift components are obtained, including: The coupling strength is obtained based on the multi-field coupling drift corresponding to the current cycle and the multi-field coupling drift corresponding to the adjacent previous cycle. The size of the convolution kernel and the void ratio are obtained based on the coupling strength. Based on the size and dilation rate of the convolutional kernel and the temporal convolutional neural network model, high-dimensional coupling features are obtained; The multiple drift components are obtained based on the high-dimensional coupling characteristics and the total interference components.

7. The zero-point drift compensation method for an eddy current sensor according to claim 6, characterized in that, Based on the size and dilatation rate of the convolutional kernel and the temporal convolutional neural network model, high-dimensional coupling features are obtained, including: The offset is obtained based on the global feature vector and the size of the convolution kernel; The high-dimensional coupling feature is obtained based on the offset, the hole rate, and the convolutional layers in the temporal convolutional neural network model.

8. The zero-point drift compensation method for an eddy current sensor according to claim 1, characterized in that, Based on the drift components and the graph neural network model, drift compensation parameters are obtained, including: Based on the total interference components and the graph neural network model, multi-field coupling drift is obtained; The total drift is obtained based on the drift components and the multi-field coupled drift. The drift compensation parameters are obtained based on the total drift.

9. A zero-point drift compensation method for an eddy current sensor according to claim 8, characterized in that, The drift compensation parameters are obtained based on the total drift, including: Based on the total drift and frequency adjustment coefficients, the basic compensation parameters and coupling compensation parameters are obtained; The drift compensation parameters are obtained based on the basic compensation parameters and the coupling compensation parameters.

10. A zero-point drift compensation system for an eddy current sensor, characterized in that, include: The acquisition module is used to acquire the coil impedance and multi-channel magnetic field signals corresponding to the eddy current sensor, wherein there are multiple eddy current sensors; A separation module is used to obtain independent interference components and total interference components based on the coil impedance and multi-channel magnetic field signals; The drift module is used to obtain multiple drift components based on the total interference component and the multi-field coupling drift corresponding to the current period; The compensation module is used to obtain drift compensation parameters based on the drift components and the graph neural network model.

Images