Method and system for predicting magnetic field annealing process abnormality based on multivariate time series analysis

Abstract

This invention discloses a method and system for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis, belonging to the field of magnetic field annealing data processing. The method performs magnetic field annealing domain jump analysis, including adaptive mode decomposition and magnetic field annealing wavelet transform, on a multivariate time series dataset. The analysis results determine whether to perform dual-channel data reconstruction. When dual-channel data reconstruction is performed, multidimensional dynamic behavior features of the annealing process are extracted based on the reconstruction results. When dual-channel data reconstruction is not performed, multidimensional dynamic behavior features of the annealing process are directly extracted based on the magnetic field annealing domain jump analysis results. After the extraction of multidimensional dynamic behavior features, magnetic field annealing anomaly prediction, including an early warning of magnetic field annealing anomalies, is performed based on the extracted results. This solves the technical problem of insufficient accuracy in predicting magnetic field annealing process anomalies.

Description

Technical Field

[0001] This invention relates to the field of magnetic field annealing data processing technology, and in particular to a method and system for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis. Background Technology

[0002] With the rapid iteration of fields such as semiconductors, spintronics, and new energy power electronics, the application scenarios of magnetic materials, such as amorphous or nanocrystalline soft magnetic materials and magnetic thin films, are constantly expanding. Magnetic field rapid annealing, as a core process for controlling the magnetic properties of magnetic materials, can suppress grain coarsening and reduce interfacial diffusion through rapid heating and cooling. At the same time, it can induce the directional alignment of magnetic domains with the help of an external magnetic field, thereby improving key indicators such as permeability and coercivity of the material. It has become an indispensable key step in the preparation of high-end magnetic materials. The existing equipment for magnetic field annealing mainly includes a vacuum chamber, a high-temperature heating module, a magnetic field generating module, and a cooling module. The vacuum chamber is used to provide a vacuum environment; the high-temperature heating module is responsible for performing heating and cooling heat treatment on the sample material; the magnetic field generating module applies a constant or alternating external magnetic field during the annealing process to directionally control the magnetic domain structure; and the cooling module ensures the temperature stability and operational safety of all components of the equipment through forced heat dissipation.

[0003] The existing anomaly prediction method for magnetic field annealing processes mainly follows this process: First, multivariate process data, such as temperature, magnetic field strength, and current, are collected during the annealing process using sensors such as temperature sensors, current transformers, and Hall sensors. Next, the multivariate process data is preprocessed using methods such as linear interpolation, the 3σ principle, and standardization to obtain preprocessed multivariate process data. Then, feature extraction algorithms such as principal component analysis and fast Fourier transform are used to extract features from the preprocessed multivariate process data, identifying multivariate process features that reflect abnormal process states, such as temperature deviation rate and magnetic field fluctuation rate. Finally, based on prediction models such as LSTM autoencoders and graph attention networks, abnormal states in the annealing process are identified and evaluated, ultimately generating early warning information, such as abnormal annealing temperature and abnormal magnetic field fluctuations.

[0004] There are already relatively mature detection and analysis methods for the quality inspection of magnetic materials during magnetic field annealing. For example, Chinese invention patent CN112016205B discloses a method for evaluating and analyzing the annealing quality of autotransformer cores, which includes: fitting the coupling relationship between the magnetic field strength and magnetic flux density inside the silicon steel sheet of the core with high precision, and combining it with the calculation model of the eddy current field inside the silicon steel sheet to calculate the eddy current loss value of the core; annealing the core and calculating the measured value of the eddy current loss of the core after annealing; and finally comparing the calculated value with the measured value to judge the annealing quality of the core.

[0005] Existing quality inspection methods for magnetic materials during magnetic field annealing primarily focus on offline detection and post-assessment of eddy current losses after annealing. These methods fail to provide real-time monitoring and early warning of anomalies within the material during annealing. Furthermore, in magnetic field annealing processes for high-end magnetic materials such as amorphous, nanocrystalline, and magnetic thin films, the cavities of traditional discrete magnetic field annealing equipment and testing devices are independent. After annealing, samples must be removed from the annealing cavity and exposed to the atmospheric environment before being transferred to the testing cavity for performance characterization. This environmental transfer process may introduce surface oxidation, interface contamination, and structural relaxation caused by interruptions in thermal history. Additionally, there is a lack of perception and data protection for the microscopic states such as magnetic domain movement and defect evolution during annealing. Conventional smoothing filters are prone to losing key defect characteristic signals, making it difficult to support high-precision prediction of process anomalies.

[0006] During magnetic field annealing, when the sample material contains defects such as lattice defects, impurities, grain boundaries, or internal stress, these defects pin and depin the magnetic domain walls, causing sudden jumps in the domain wall motion driven by the external magnetic field. This step-like abrupt change in the domain walls induces transient induced voltage pulses in sensors such as Hall sensors through electromagnetic induction, accompanied by instantaneous changes in local magnetization. At the sensing signal level, this manifests as discontinuous transient characteristic signals such as abrupt changes, spikes, and steps superimposed on the gradually changing trend of the macroscopic magnetic field. The amplitude, frequency, energy distribution, and waveform of these discontinuous transient characteristic signals carry crucial information such as the density of defects within the sample material, the distribution of pinning strength, and the degree of micro-stress concentration. However, existing technologies typically only utilize solid-state... Preprocessing of multivariate time-series data, such as local magnetic field strength fluctuation time series acquired by Hall sensors and excitation current pulsation time series acquired by high-frequency current probes, using smoothing filtering methods such as fixed-window mean filtering and moving average, may result in the loss of real defect feature signals reflecting sudden jumps, spikes, and steps in the sensor signals. This leads to distortion of multivariate time-series data and missing key defect information. When feature extraction is performed based on distorted multivariate time-series data with missing key defect information, the extracted multivariate time-series features are further distorted. When predicting anomalies in the magnetic field annealing process based on distorted multivariate time-series features, the model cannot identify early anomalies induced by defects, resulting in missed reports and false alarms, ultimately leading to low accuracy in predicting anomalies in the magnetic field annealing process. Summary of the Invention

[0007] To address the low accuracy of anomaly prediction in existing magnetic field annealing processes, this invention provides a method and system for anomaly prediction based on multivariate time-series analysis. The technical solution is as follows: On the one hand, a method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis is provided. This method includes: performing magnetic field annealing domain jump analysis on a multivariate time series set, including adaptive mode decomposition and magnetic field annealing wavelet transform; deciding whether to perform dual-channel data reconstruction based on the analysis results; the multivariate time series set includes local magnetic field strength fluctuation time series and excitation current pulsation time series; when performing dual-channel data reconstruction, after the dual-channel data reconstruction is completed, extracting macroscopic process feature vectors, microscopic defect dynamic feature vectors, and cross-domain coupling feature vectors based on the dual-channel data reconstruction results to extract multidimensional dynamic behavior features of the annealing process; when not performing dual-channel data reconstruction, directly extracting multidimensional dynamic behavior features of the annealing process based on the magnetic field annealing domain jump analysis results; after the extraction of multidimensional dynamic behavior features of the annealing process is completed, performing magnetic field annealing anomaly prediction including magnetic field annealing anomaly warning based on the extraction results.

[0008] On the other hand, a magnetic field annealing process anomaly prediction system based on multivariate time series analysis is provided, including: a magnetic field annealing domain jump monitoring module, an annealing process multidimensional feature extraction module, and a magnetic field annealing anomaly monitoring module. The magnetic field annealing domain jump monitoring module performs magnetic field annealing domain jump analysis on the multivariate time series set, including adaptive mode decomposition and magnetic field annealing wavelet transform. Based on the analysis results, it decides whether to perform dual-channel data reconstruction. The multivariate time series set includes local magnetic field strength fluctuation time series and excitation current pulsation time series. The annealing process multidimensional feature extraction module is used to determine whether to perform dual-channel data reconstruction when dual-channel data reconstruction is performed. During channel data reconstruction, after the dual-channel data reconstruction is completed, the annealing process multidimensional dynamic behavior feature extraction is performed based on the dual-channel data reconstruction results, extracting macroscopic process feature vectors, microscopic defect dynamic feature vectors, and cross-domain coupling feature vectors. When dual-channel data reconstruction is not performed, the annealing process multidimensional dynamic behavior feature extraction is performed directly based on the magnetic domain jump analysis results. The magnetic field annealing anomaly monitoring module is used to predict magnetic field annealing anomalies, including magnetic field annealing anomaly warnings, based on the results of the annealing process multidimensional dynamic behavior feature extraction after the extraction is completed.

[0009] The beneficial effects of the technical solutions provided by the embodiments of the present invention include at least the following: 1. The magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided by this invention performs magnetic field annealing domain jump analysis on a multivariate time series set, including adaptive mode decomposition and magnetic field annealing wavelet transform. Based on the analysis results, it decides whether to perform dual-channel data reconstruction. This helps to faithfully separate and independently characterize the microscopic defect feature signals induced by domain wall jumps, which are discarded as noise by traditional smoothing filtering methods. This reduces the mutual interference between macroscopic process trends and microscopic defect transients from the source, ensuring that the subsequent data foundation has complete physical information fidelity. When performing dual-channel data reconstruction, after the dual-channel data reconstruction is completed, multidimensional dynamic behavior feature extraction of the annealing process is performed based on the dual-channel data reconstruction results, extracting macroscopic process feature vectors, microscopic defect dynamic feature vectors, and cross-domain coupling feature vectors. When performing dual-channel data reconstruction, the multi-dimensional dynamic behavior features of the annealing process are directly extracted based on the magnetic domain jump analysis results of magnetic field annealing. This helps to encapsulate the complete state information of the annealing process from three dimensions: macroscopic process execution quality, microscopic defect dynamic evolution, and the cross-domain coupling relationship between the two when magnetic domain jump activity is active. After the multi-dimensional dynamic behavior feature extraction of the annealing process is completed, magnetic field annealing anomaly prediction, which includes magnetic field annealing anomaly warning, is performed based on the extraction results. This helps to shift the triggering time of the anomaly warning to the early stage when the defect activity coupling mode undergoes a critical change. This provides a sufficient time window for process intervention before actual damage is formed inside the sample material, thereby improving the accuracy of magnetic field annealing process anomaly prediction and effectively solving the problem of low accuracy of magnetic field annealing process anomaly prediction in the prior art.

[0010] 2. This invention, through adaptive mode decomposition and magnetic field annealing wavelet transform, obtains intrinsic mode function components. This helps to completely separate the domain jumping transient signal superimposed on the gradually changing trend of the macroscopic magnetic field from the low-frequency process trend signal, overcoming the defect of existing smoothing filtering methods that irreversibly filter out high-frequency transient signals as noise. It ensures the complete and faithful preservation of the physical information of internal defect activity in the sample material from the signal source. Identifying low-frequency trend modes and high-frequency transient modes based on intrinsic mode function components helps reduce the mixing of information from two different physical sources in a single data channel, which is problematic in existing technologies. To address the issue of ambiguous physical meaning of subsequent analysis objects, this method uses low-frequency trend mode sets and high-frequency transient mode sets to determine whether dual-channel data reconstruction of multivariate time series sets is necessary. This helps to make adaptive decisions based on the actual activity level of magnetic domain jumping during annealing, completely separating macroscopic process trends and microscopic defect transients into two independent data streams. This overcomes the shortcomings of existing technologies that filter out high-frequency signals indiscriminately or use fixed processing strategies under all circumstances, resulting in insufficient adaptability and low reliability of anomaly detection. As a result, it achieves high fidelity and high discriminative power in predicting annealing process anomalies from the data layer.

[0011] 3. When the non-stationarity of the local magnetic field strength fluctuation time series and the excitation current pulsation time series in the multivariate time series set exceeds the preset stationarity index, the spectral segmentation boundary on which the empirical wavelet transform relies shifts with the sliding of the analysis time window under strong non-stationary signal conditions. If a fixed spectral segmentation boundary based on global fast Fourier transform is still used for uniform mode decomposition throughout the entire time period, it may lead to the correct separation of low-frequency trends and high-frequency transients in one time period resulting in frequency band aliasing in another time period, thus causing subsequent mode classification errors and misjudgments of jump events. Therefore, an alternative scheme of adaptive mode decomposition needs to be implemented. By obtaining the recursion rate index sequence, it is helpful to quantitatively evaluate the non-stationarity of the sensing signal throughout the magnetic field annealing process, thereby automatically identifying the time periods in which the signal statistical characteristics change, and for the signal segment within each boundary segmentation time period, independently... A Fast Fourier Transform (FFT) is performed to obtain the local spectral distribution of the signal segment. Based on each local spectral distribution, the corresponding local spectral segmentation boundary is obtained. This helps to dynamically adjust the spectral segmentation boundary according to the time-varying statistical characteristics of the signal itself, overcoming the defect of frequency band mismatch under strong non-stationary conditions caused by global fixed spectral segmentation. This ensures that the frequency band segmentation in each time period always matches the actual spectral structure of the signal in that time period. Within each boundary segmentation time period, a magnetic field annealing wavelet transform is performed based on the local spectral segmentation boundary to obtain the set of local intrinsic mode function components. This helps to ensure that the low-frequency trend components and high-frequency transient components extracted in each time period are correctly separated in the frequency domain, reducing cross-time period frequency band aliasing and mode missegmentation caused by global segmentation boundary offset. Thus, intrinsic mode function components with consistent frequency band segmentation and clear physical meaning are obtained throughout the entire time period. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart outlining the general overview of the magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in the embodiments of this application; Figure 2 A logic diagram of magnetic domain jump analysis for a magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in this application embodiment; Figure 3 A monitoring curve of surface temperature fluctuation of a coplanar waveguide for a magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in this application embodiment; Figure 4A schematic diagram of the sample delivery process of the low-temperature testing equipment on which the magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in the embodiments of this application is based; Among them, 1. Annealing area; 2. Detection area; 3. Sample; Figure 5 A schematic diagram of a desktop cryogenic testing device for a method of predicting anomalies in magnetic field annealing process based on multivariate time series analysis provided in this application embodiment; Among them, 4. Vacuum system; 5. Magnetic field annealing system; 6. In-situ testing system; 7. Telescopic arm; 8. Movable magnet; 9. Heating device; 10. Temperature sensor; 11. Quartz tube; 12. Thermal insulation material; 13. Test probe. Detailed Implementation

[0014] The following provides explanations for some of the terms used in this application. It should be noted that these explanations are for the convenience of those skilled in the art and do not constitute a limitation on the scope of protection claimed in this application.

[0015] The magnetic field annealing process anomaly prediction method and system based on multivariate time series analysis provided in this application relies on a pre-constructed magnetic field annealing process anomaly prediction knowledge base. This knowledge base includes core judgment parameters verified by process experts, such as preset peak quantity, preset noise floor level, preset cutoff frequency, and preset energy threshold. It also includes standardized pre-processed historical operating data of the magnetic field annealing process, covering local magnetic field strength fluctuation time series data and excitation current pulsation time series data under different magnetic material types, different annealing process curves, and different anomaly modes, along with corresponding anomaly prediction results. Furthermore, it includes model iteration optimization data, such as dual-channel data reconstruction effect evaluation and magnetic domain dynamic behavior feature space. The knowledge base for predicting anomalies in magnetic field annealing processes employs a domain-specific heterogeneous storage architecture. It utilizes a relational database to store structured index data such as the time series of local magnetic field strength fluctuations, the time series of excitation current pulsations, and the intrinsic mode function component parameters after modal decomposition. It also utilizes a non-relational database to store unstructured data such as waveform fragments and sequences of magnetic domain jumping events. Process engineers can periodically verify and dynamically iterate and adjust various preset parameters stored in the knowledge base based on real-time collected multivariate monitoring data of the magnetic field annealing process and feedback on anomaly prediction performance. This ensures that the knowledge base can continuously adapt to the complex and varied magnetic field annealing operating conditions under different magnetic material systems and different annealing process windows, such as the crystallization annealing window period of amorphous alloys.

[0016] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0017] In Example 1, further, during the magnetic field annealing process, a magnetic domain jump analysis, including adaptive mode decomposition and magnetic field annealing wavelet transform, is performed on the multivariate time series set. Adaptive mode decomposition is used to adaptively separate low-frequency and high-frequency components of the multivariate time series signal. Magnetic field annealing wavelet transform is used to achieve joint characterization of the multivariate time series signal related to the magnetic field annealing process in the frequency and time domains. Based on the analysis results, a decision is made on whether to perform dual-channel data reconstruction to completely separate the sources of the multivariate time series signal and form two independent, physically complete parallel data streams. The multivariate time series set includes the local magnetic field strength fluctuation time series and the excitation current pulsation time series. The multivariate time series signal represents the combined result of the local magnetic field strength fluctuation time series and the excitation current pulsation time series, which is used to completely describe the magnetization state and excitation response dynamics of the sample material during the magnetic field annealing process. Multivariate time-series observation data with correlation characteristics; when performing dual-channel data reconstruction, after the dual-channel data reconstruction is completed, multi-dimensional dynamic behavior feature extraction of annealing process is performed based on the dual-channel data reconstruction results to extract macroscopic process feature vectors, microscopic defect dynamic feature vectors, and cross-domain coupling feature vectors. When dual-channel data reconstruction is not performed, multi-dimensional dynamic behavior feature extraction of annealing process is directly performed based on the magnetic domain jump analysis results of magnetic field annealing. Multi-dimensional dynamic behavior feature extraction of annealing process is used to extract dynamic features of annealing process from multiple dimensions simultaneously and quantify the cross-domain coupling relationship between the two. After the multi-dimensional dynamic behavior feature extraction of annealing process is completed, magnetic field annealing anomaly prediction including magnetic field annealing anomaly warning is performed based on the multi-dimensional dynamic behavior feature extraction results. Magnetic field annealing anomaly prediction is used to detect risks such as grain anomaly risk, annealing brittle critical transition, and interface diffusion failure in advance during magnetic field annealing process.

[0018] In this embodiment, through magnetic domain jump analysis of magnetic annealing, extraction of multi-dimensional dynamic behavior features of the annealing process, and prediction of magnetic annealing anomalies, it is helpful to improve the entire chain of anomaly perception capabilities, from signal fidelity separation to defect feature reconstruction and cross-modal fusion prediction, and to advance the early warning time window for magnetic annealing anomalies. Specifically, magnetic domain jump analysis of magnetic annealing solves the problem of loss of key defect information caused by smoothing filtering from the data source, and provides a complete and undamaged microscopic transient data source for extraction of multi-dimensional dynamic behavior features of the annealing process. The extraction of multi-dimensional dynamic behavior features of the annealing process directly reflects the magnetic domain dynamic behavior feature space of defect density, pinning strength and micro-stress concentration. The prediction of magnetic annealing anomalies shifts the anomaly criterion from the exceeding of macroscopic parameters to the moment when the coupling mode of the two changes critically, thereby achieving high-precision early warning of magnetic annealing process anomalies.

[0019] like Figure 1 The diagram shown is a flowchart outlining the general overview of the magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in this application embodiment. The method includes: performing magnetic domain jump analysis during magnetic field annealing and obtaining the total number of magnetic domain jump events; determining whether the total number of magnetic domain jump events exceeds a preset event count threshold; if so, performing dual-channel data reconstruction; otherwise, extracting multidimensional dynamic behavior features of the annealing process and obtaining a multidimensional dynamic behavior feature vector; predicting magnetic field annealing anomalies based on the multidimensional dynamic behavior feature vector and obtaining the annealing process anomaly prediction result; issuing a magnetic field annealing anomaly warning based on the annealing process anomaly prediction result; determining whether the comprehensive score value of the annealing anomaly exceeds a preset warning threshold; if so, triggering an anomaly warning signal; otherwise, continuing to collect multivariate time series data at the next sampling time and repeatedly executing subsequent analysis steps.

[0020] like Figure 2 The diagram shown is a logic diagram of magnetic domain jump analysis in the magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in this application embodiment. Figure 2 It can be seen that: magnetic domain jumping analysis is performed during magnetic field annealing, adaptive mode decomposition is executed, and continuous sub-frequency bands are obtained. Based on the continuous sub-frequency bands, magnetic field annealing wavelet transform is performed to obtain intrinsic mode function components. Based on the intrinsic mode function components, low-frequency trend modes and high-frequency transient modes are identified, and peak-to-average power is obtained. It is determined whether the peak-to-average power ratio is greater than a preset ratio threshold. If it is, the corresponding transient energy pulse event is identified as a magnetic domain jumping event. Otherwise, the corresponding transient energy pulse event is identified as a false transient component caused by random electrical noise and is removed. At the same time, the total number of magnetic domain jumping events is obtained, and it is determined whether the total number of magnetic domain jumping events is greater than a preset event number threshold. If it is, dual-channel data reconstruction is performed. Otherwise, multi-dimensional dynamic behavior features of the annealing process are extracted.

[0021] Furthermore, the specific process of magnetic domain jump analysis during magnetic field annealing is as follows: A multivariate time series set of the sample material is collected during the magnetic field annealing process; the multivariate time series set includes the local magnetic field strength fluctuation time series and the excitation current pulsation time series; the local magnetic field strength fluctuation time series represents the induced voltage sequence signal collected by the Hall sensor throughout the entire magnetic field annealing process, reflecting the dynamic evolution of the local magnetization state of the sample material. For example, the local magnetic field strength fluctuation time series is represented as {h(t1), h(t2), ..., h(tN)}, where h(tk) represents the induced voltage value at the k-th sampling time, k = 1, 2, ..., N, k represent the sampling times of the induced voltage value, and N represents the total number of sampling points. This voltage value is proportional to the component of the local magnetic induction intensity of the sample material in the direction of the sensor's sensitive axis. The excitation current pulsation sequence represents the current signal sequence collected by the high-frequency current probe during the entire magnetic field annealing process, reflecting the dynamic changes of the excitation current. For example, the excitation current pulsation sequence is represented as {i(t0), i(t1), ..., i(tN)}, where i(tk) represents the current value at the kth sampling time, which reflects the actual excitation current intensity passing through the magnetic field generating coil.

[0022] Adaptive mode decomposition is performed on a multivariable time series set, specifically as follows: A Fast Fourier Transform (FFT) is applied to the multivariable time series set to obtain the global spectral distribution. This global spectral distribution includes the magnetic field fluctuation spectral distribution obtained after applying the FFT to the local magnetic field strength fluctuation time series within the multivariable time series, and the current pulsation spectral distribution obtained after applying the FFT to the excitation current pulsation time series within the multivariable time series. The global spectral distribution describes the frequency components of the local magnetic field strength fluctuation time series or the excitation current pulsation time series from zero frequency to the Nyquist frequency range. Amplitude distribution; obtaining spectral segmentation boundaries based on global spectral distribution; it should be noted that the process of obtaining the spectral segmentation boundaries for magnetic field fluctuation spectral distribution and current pulsation spectral distribution is the same and carried out independently and in parallel. The following description uses one of the spectral distributions as an example; the specific process of obtaining the spectral segmentation boundaries is as follows: sort the amplitudes corresponding to each frequency point in one of the global spectral distributions in descending order, obtain the first preset number of peak values, and take them as spectral peaks. For example, if the preset number of peak values ​​is M, then the set of obtained spectral peaks is represented as {f1, f2, ..., f...} m}, where f m This represents the frequency value corresponding to the m-th spectral peak, where m = 1, 2, ..., M, and m represents the index of the spectral peak, taking values ​​from 1 to M in descending order of amplitude. f1 is the frequency of the highest spectral peak. mLet f1 be the frequency of the Mth highest spectral peak, where the preset number of peaks is pre-set by a pre-defined team. The median frequency between two adjacent spectral peaks is used as the spectral boundary between those peaks. Within the low-frequency band before the first peak, starting from the frequency position of the first peak, a search is performed progressively along the frequency axis towards zero frequency. The starting boundary frequency is determined when the spectral amplitude gradually decreases with decreasing frequency and first drops to the level of the preset noise floor. For example, if the first peak is f1, the search proceeds from f1 towards zero frequency within the frequency range below f1, marking the frequency point where the spectral amplitude first decreases to the preset noise floor level as the starting boundary frequency. ,in, This indicates the starting boundary frequency where signal energy in the low-frequency band can be ignored. The preset noise floor level is used to distinguish between effective signal components and background noise components in the spectrum distribution, and is set in advance by the preset personnel.

[0023] The low-frequency band represents the frequency range from zero frequency to the frequency corresponding to the first spectral peak in the global spectrum; the global spectrum represents the frequency axis and the amplitude corresponding to each frequency in the global spectrum distribution; the frequency axis represents the continuous frequency coordinate axis from zero frequency to the Nyquist frequency in the global spectrum distribution obtained after the fast Fourier transform of the sensing signal, and each coordinate point on the frequency axis corresponds to a specific frequency value; in the high-frequency band after the last spectral peak, starting from the frequency position of the last spectral peak, searching along the frequency axis towards the Nyquist frequency, the position where the spectral amplitude gradually decreases with increasing frequency and first drops to the level equal to the preset noise floor level is determined as the corresponding termination boundary frequency point. For example, the last spectral peak is f m At frequencies higher than f m Within the interval f m Searching in the direction of the Nyquist frequency, the frequency point where the spectral amplitude first decays to the preset noise floor level is marked as the termination boundary frequency point. ,Should The terminator represents the termination boundary frequency where the signal energy in the high-frequency band can be ignored; the high-frequency band represents the frequency range in the global spectrum from the last spectral peak to the Nyquist frequency; based on the spectrum segmentation boundary, the starting boundary frequency, and the ending boundary frequency, the global spectrum is adaptively segmented into continuous sub-bands. The specific segmentation process is as follows: starting from the starting boundary frequency and ending at the ending boundary frequency, the frequency range defined by two adjacent spectrum segmentation boundaries is selected as a continuous sub-band, resulting in multiple continuous sub-bands with unequal bandwidths, which do not overlap and together completely cover the global spectrum.

[0024] The specific process of magnetic field annealing wavelet transform is as follows: In each continuous sub-frequency band, based on the wavelet transform method, empirical wavelet functions and empirical scaling functions are obtained; the empirical scaling function is then multiplied by the sensing signal currently undergoing adaptive mode decomposition in the multivariate time series set, i.e., either the local magnetic field strength fluctuation time series or the excitation current pulsation time series, to output the wavelet approximation coefficients corresponding to the lowest frequency band of that sensing signal; the lowest frequency band represents the continuous sub-frequency band from zero frequency to the first spectral division boundary in the global spectrum; the wavelet approximation coefficients are used to carry the sensing signal within the lowest frequency band. Low-frequency, slowly varying trend information provides low-frequency basis components for subsequent reconstruction of macroscopic process trend data sequences. Each empirical wavelet function is inner-producted with one of the corresponding multivariate time series signals—either the local magnetic field strength fluctuation time series or the excitation current pulsation time series—to output wavelet detail coefficients for each continuous sub-band, thus obtaining a set of intrinsic mode function components. These wavelet detail coefficients carry high-frequency details and transient change information of the corresponding sensor signal within each continuous sub-band, providing a multi-band high-frequency component source for subsequent identification and separation of magnetic domain jump events. The product operation represents the integration of the integrand function over the continuous time domain along the time axis. Since the actual signal being processed is a discrete sampled signal, this inner product operation is specifically transformed into the summation of the product of the corresponding discrete values ​​at the sampling time points. For example, the sequence after discretization of the empirical scaling function is multiplied point by point with the discrete signal sequence and then summed. The integrand function represents the product of the conjugate of the empirical scaling function or empirical wavelet function and the sensing signal. The intrinsic mode function components are obtained by convolving the wavelet approximation coefficients with the empirical scaling function and convolving the wavelet detail coefficients with the corresponding empirical wavelet function. The result of the convolution operation, followed by summation, represents the physical meaning of the complete time-domain signal components carried within a specific continuous sub-band. Its expression is an amplitude time-series sequence with the same sampling rate and time axis length as the input local magnetic field strength fluctuation time-series or excitation current pulsation time-series. The convolution operation represents a mathematical operation that uses wavelet approximation coefficients or wavelet detail coefficients as weighting coefficients to perform a weighted shift and superposition of the empirical wavelet function or empirical scaling function on the time axis. For example, for a wavelet detail coefficient sequence {d1, d2, ..., d...} corresponding to a certain continuous sub-band... p The result of the convolution operation is expressed as follows:} and the empirical wavelet function ψ(t) on this sub-band, where p represents the total length of the wavelet detail coefficient sequence corresponding to the continuous sub-band. , where d k Let ψ(t-τ) represent the wavelet detail coefficient weights corresponding to the k-th sampling time. k The expression represents shifting the empirical wavelet function along the time axis to the delay position τ corresponding to the k-th sampling time. k The resulting shift basis functions, k=1,2,…,L, where L represents the total number of wavelet detail coefficients, τk This represents the delay time corresponding to the k-th sampling moment. Its value is equal to the absolute time coordinate of the k-th sampling moment on the time axis. The weighted shift superposition process is essentially to regard each wavelet detail coefficient as the basis function amplitude scaling factor of its corresponding moment, and to linearly superimpose all the scaled shift basis functions on the time axis to reconstruct the complete time domain waveform within the continuous sub-band.

[0025] Low-frequency trend modes and high-frequency transient modes are identified based on intrinsic mode function components. The specific process for identifying low-frequency trend modes and high-frequency transient modes is as follows: Power spectral density distribution curves of each intrinsic mode function component are obtained based on Welch's method for power spectral density estimation; on the power spectral density distribution curve, the frequency coordinate corresponding to the maximum power density is marked as the center frequency of the intrinsic mode function component; the result of the ratio calculation between the power density integral value corresponding to the center frequency and the total power density integral value is used as the energy ratio value. The power density integral value is represented by the result of definite integration of the power spectral density distribution curve of the intrinsic mode function component within the frequency interval defined by the center frequency and the adjacent spectral division boundaries on the left and right; the total power density integral value is represented by the intrinsic mode function components... The power spectral density curve is represented by the area integral of the power spectral density curve across the entire frequency band. Based on the energy proportion, the intrinsic mode function components are divided into a low-frequency trend mode set for characterizing the gradual change of macroscopic magnetic field and temperature process, and a high-frequency transient mode set for carrying transient signals induced by magnetic domain jumps. The low-frequency trend mode set includes intrinsic mode function components with a center frequency less than a preset cutoff frequency and an energy proportion greater than a preset energy threshold. The preset cutoff frequency is represented by the average center frequency over a historical period, and the preset energy threshold is represented by the average energy proportion over a historical period. The high-frequency transient mode set includes intrinsic mode function components with a center frequency not less than the preset cutoff frequency and an energy proportion exhibiting short-term burst characteristics. Intrinsic mode function components other than those in the low-frequency trend mode set and the high-frequency transient mode set are marked as residual components and discarded.

[0026] The short-term burst characteristic is measured as follows: the signal energy amplitude of the intrinsic mode function component is obtained within a sliding time window; the signal energy amplitude is represented by the sum of the squares of the amplitudes of the intrinsic mode function component at each sampling time within the sliding time window; if the duration for which the signal energy amplitude is greater than a preset burst energy threshold is less than a preset duration threshold, and the burst energy ratio is greater than a preset burst ratio threshold within that continuous duration, then the intrinsic mode function component is determined to have short-term burst characteristics. The preset burst energy threshold is represented by the average value of the signal energy amplitude over a historical time period, the preset duration threshold is represented by the average duration for which the signal energy amplitude is greater than the preset burst energy threshold over a historical time period, and the preset burst ratio threshold is represented by the average value of the burst energy ratio over a historical time period; the burst energy ratio is represented by the ratio of the peak signal energy within the continuous duration to the average signal energy of adjacent time periods outside the coverage of the sliding time window; based on the low-frequency trend mode set and the high-frequency transient mode set, it is determined whether dual-channel data reconstruction of the multivariate time series set is required.

[0027] The specific process for determining whether dual-channel data reconstruction is needed for a multivariate time series set is as follows: First, for each intrinsic mode function component in the high-frequency transient mode set, the instantaneous energy curve of each intrinsic mode function component on the time axis of the magnetic domain jump analysis period is obtained based on the Hilbert transform method; the first derivative operation is performed on the instantaneous energy curve to obtain the first derivative sequence, and the zero-crossing sampling points in the first derivative sequence where the sign changes from positive to negative are marked as local maxima, and the time corresponding to the local maxima is marked as the occurrence time of the transient energy pulse event. The peak time and peak energy of each transient energy pulse event are recorded; the peak time is the time coordinate corresponding to the local maxima; the peak energy is the amplitude value of the instantaneous energy curve at the peak time; then, within a preset time window centered on each peak time, the peak-to-average power ratio of the transient energy pulse event is obtained, where the preset time window is set in advance by preset personnel; the peak-to-average power ratio is obtained by measuring the peak signal value within the time window. The result of the ratio calculation between the power and the average power within the window is represented; when the peak-to-average power ratio is greater than the preset ratio threshold, the transient energy pulse event is determined to be a domain jump event induced by a sudden jump of the domain wall; otherwise, it is determined to be a false transient component caused by random electrical noise and is removed. The preset ratio threshold is represented by the average value of the peak-to-average power ratio over a historical time period. Then, the total number of domain jump events within the domain jump analysis time period is counted. When the total number is greater than the preset event number threshold, dual-channel data reconstruction is performed; otherwise, it is determined that the domain jump activity in the current magnetic field annealing process is insufficient to constitute an independent signal source that can effectively characterize the dynamics of material defects. Only all intrinsic mode function components contained in the low-frequency trend mode set are summed and reconstructed into a macroscopic process trend data sequence. Based on this macroscopic process trend data sequence, multidimensional dynamic behavior features of the annealing process are extracted. The preset event number threshold is represented by the average value of the total number of domain jump events over a historical time period.The summation and reconstruction of all intrinsic mode function components contained in the low-frequency trend mode set is represented on the time axis of the magnetic domain jump analysis period. For all intrinsic mode function components contained in the low-frequency trend mode set, the algebraic summation operation of the amplitude is performed point by point according to the discrete sampling time. The summation result of each sampling time is used as the amplitude value of the reconstructed macroscopic process trend data sequence at that time. For example, the local magnetic field strength fluctuation time series or excitation current pulsation time series in the multivariate time series set is decomposed into two low-frequency trend modes IMF1={a1, b1, c1, d1} and IMF2={a2, b2, c2, d2}. The macroscopic process trend data sequence generated after summation and reconstruction is {(a1+a2), (b1+b2), (c1+c2), (d2)}. 1+d2), where a1 represents the signal amplitude of the intrinsic mode function component IMF1 at the first sampling time, b1 represents the signal amplitude of the intrinsic mode function component IMF1 at the second sampling time, c1 represents the signal amplitude of the intrinsic mode function component IMF1 at the third sampling time, d1 represents the signal amplitude of the intrinsic mode function component IMF1 at the fourth sampling time, a2 represents the signal amplitude of the intrinsic mode function component IMF2 at the first sampling time, b2 represents the signal amplitude of the intrinsic mode function component IMF2 at the second sampling time, c2 represents the signal amplitude of the intrinsic mode function component IMF2 at the third sampling time, and d2 represents the signal amplitude of the intrinsic mode function component IMF2 at the fourth sampling time.

[0028] In this embodiment, magnetic domain jump analysis during magnetic field annealing helps to completely preserve the physical fingerprint information of material defects carried by domain wall jumps from the data source, reduce the loss rate of key defect information caused by smoothing filtering preprocessing, reduce the false alarms and misjudgments caused by data distortion, improve the sensing sensitivity of transient induced voltage pulse signals induced by sudden domain wall jumps, and realize the faithful preservation of physical information of material defect activity at the sensing signal level during magnetic field annealing. This provides an independent microscopic transient data source with clear physical meaning and unaffected by macroscopic process trends for subsequent multidimensional dynamic behavior feature extraction and cross-modal fusion anomaly prediction of annealing process.

[0029] Furthermore, the specific process of dual-channel data reconstruction is as follows: First data channel and second data channel are constructed; the first data channel represents a one-dimensional time-series data channel obtained by summing the amplitudes of all intrinsic mode function components in the low-frequency trend mode set at each sampling time along the time axis of the magnetic domain jump analysis period. Its essence is a linear synthesis of each low-frequency mode component in the time domain. The synthesized signal only contains the slowly varying trend components related to the applied magnetic field and thermal process, and no longer contains any high-frequency transient fluctuations caused by magnetic domain jumps; the amplitudes of all intrinsic mode function components contained in the low-frequency trend mode set are summed at each sampling time along the time axis of the magnetic domain jump analysis period, and the summation result is used as the amplitude value at each sampling time in the first data channel; in the next magnetic domain jump analysis period, the data sequence output by the first data channel is automatically captured as a macroscopic process trend data sequence, retaining only the slowly varying trends related to the applied magnetic field and thermal process. For example, the macroscopic process trend data sequence is represented as {m(t1), m(t2), ..., m(tN)}, where m(t1), m(t2), ..., m(tN)}. k The summation of the macroscopic trend values ​​at the k-th sampling time indicates that the summation only reflects the slowly changing trend components determined by the applied magnetic field and the thermal process. k = 1, 2, ..., N, where N is the total number of sampling points. The transient signal segments of all magnetic domain jump events in the high-frequency transient mode are arranged in chronological order of their peak times on the time axis to obtain the second data channel. The essence of the second data channel is to integrate the isolated magnetic domain jump events scattered in the multi-band high-frequency modes into a time-ordered event stream. Each record in this event stream corresponds to an independent magnetic domain wall detachment jump event, and the occurrence time, energy intensity, and waveform morphology of each jump event are fully preserved.

[0030] During the next domain hopping analysis period, the data sequence automatically captured from the second data channel is the domain hopping data sequence. This sequence is represented as a hopping event sequence composed of several real domain hopping event records arranged chronologically. Each hopping event record contains at least the peak time, peak energy, peak-to-average power ratio, and a hopping waveform segment. For example, for the q-th real domain hopping event, its domain hopping event record is represented as {tq, Aq, Rq, Wq}, where tq represents the peak time of the domain hopping event, Aq represents the peak energy of the domain hopping event, Rq represents the peak energy of the domain hopping event, and Rq represents the peak energy of the domain hopping event. The peak-to-average power ratio of the domain jump event is shown, and Wq represents the jump waveform segment of the domain jump event. The jump waveform segment is the original amplitude time sequence segment obtained by taking a preset duration forward and backward from the peak time of the jump event. This segment completely preserves the abrupt waveform shape of the domain wall depin jump at the sensing signal level. The preset duration is set in advance by preset personnel. The macroscopic process trend data sequence and the domain jump data sequence are output as dual-channel parallel data. Based on the output macroscopic process trend data sequence and the domain jump data sequence, the multi-dimensional dynamic behavior features of the annealing process are extracted.

[0031] In this embodiment, dual-channel data reconstruction helps reduce the mutual interference caused by the superposition of macroscopic process trend information and microscopic defect transient information in a single data channel, reduces the degree of mixing of macroscopic gradual components and microscopic transient components in the subsequent feature extraction stage, and realizes the data flow paradigm upgrade from single-channel mixed monitoring to dual-channel parallel sensing.

[0032] Furthermore, the specific process for extracting the multidimensional dynamic behavior features of the annealing process is as follows: Macroscopic process features are extracted from the macroscopic process trend data sequence based on a piecewise linear fitting algorithm. These macroscopic process features include the average heating rate during the heating phase, the temperature variation coefficient during the holding phase, and the average cooling rate during the cooling phase. First-order linear fitting is performed on the macroscopic process trend data sequence within three preset data intervals: the heating phase, the holding phase, and the cooling phase. The fitting slope value for each interval is obtained. The preset heating phase, holding phase, and cooling phase are pre-set by the pre-set personnel. The average heating rate during the heating phase is obtained by taking the absolute value of the fitting slope value for the preset heating phase data interval, used to quantify the speed at which the external heating field heats the sample material during the heating phase of the annealing process. The temperature variation coefficient during the holding phase is represented by the ratio of the standard deviation of the temperature sequence to the mean temperature within the preset holding phase data interval, used to quantify the fluctuation amplitude and stability of the temperature field near the preset holding point during the holding phase of the annealing process. The average cooling rate during the cooling phase is obtained by taking the absolute value of the fitting slope value for the cooling phase data interval. This is used to quantify the control precision of the heat dissipation and cooling rate of sample materials during the cooling stage of the annealing process. Macroscopic process features are arranged according to a preset process sequence to form a one-dimensional vector, resulting in a macroscopic process feature vector. For example, if the average heating rate in the heating stage is vup, the temperature variation coefficient in the holding stage is csoak, and the average cooling rate in the cooling stage is vdown, then the macroscopic process feature vector obtained after arranging the heating, holding, and cooling stages according to the preset process sequence is represented as [vup, csoak, vdown]. The first dimension of the vector is the average heating rate in the heating stage, the second dimension is the temperature variation coefficient in the holding stage, and the third dimension is the average cooling rate in the cooling stage. The preset process sequence is set in advance by the personnel. The macroscopic process feature vector characterizes the overall execution quality of the external magnetic field and thermal process during the annealing process. It also acquires the dynamic characteristics of microscopic defects in the magnetic domain jump data sequence. These dynamic characteristics include the mean sequence of magnetic domain jump event frequencies, the statistical distribution skewness and kurtosis of the peak energy of magnetic domain jumps, and the mean normalized cross-correlation coefficient between adjacent jump waveform segments.

[0033] The mean sequence of magnetic domain jump events, used to quantify the activity level of defect activity and its evolution trend over time, is obtained as follows: The magnetic domain jump data sequence is divided into several consecutive time windows according to a preset sliding window length. The total number of magnetic domain jump event records contained in each time window is counted to obtain the jump event frequency sequence. Then, a sliding mean filter is applied to the jump event frequency sequence based on a sliding filter algorithm. The filtered result is used as the mean sequence of magnetic domain jump event frequencies. The preset sliding window is set in advance by a preset team. The statistical distribution skewness and kurtosis of the peak energy of magnetic domain jumps are used to quantify the distribution characteristics of the energy released by a single magnetic domain jump event, reflecting the uniformity of the pinning intensity distribution and the existence of jump events with extreme intensity. The specific acquisition process is as follows: Extracting magnetic domains. The peak energy values ​​of all jump events recorded in the jump data sequence constitute a peak energy sample set. Skewness values ​​are calculated using the third-order standard moment algorithm, and kurtosis values ​​are calculated using the fourth-order standard moment algorithm. The mean normalized cross-correlation coefficient between adjacent jump waveform segments is used to quantify the similarity of different jump events in waveform morphology, reflecting the consistency of domain wall detachment jump behavior patterns. It is obtained as follows: the mean normalized cross-correlation coefficient is calculated for each temporally adjacent jump waveform segment in the domain jump data sequence using a zero-mean normalized cross-correlation algorithm; the arithmetic mean of the cross-correlation coefficients calculated for all adjacent jump waveform segments is then used as the mean normalized cross-correlation coefficient.

[0034] The dynamic features of micro-defects are concatenated end-to-end according to a preset dimension order to obtain a dynamic feature vector of micro-defects. The first-to-last concatenation represents the sequential connection of the mean sequence of magnetic domain jump event frequencies, the statistical distribution skewness and kurtosis of magnetic domain jump peak energy, and the mean normalized cross-correlation coefficients between adjacent jump waveform segments, into a one-dimensional vector. The dimension of this one-dimensional vector is equal to the sum of the dimensions of the aforementioned dynamic features of micro-defects, where the preset dimensions are pre-defined by the personnel. The dynamic feature vector of micro-defects characterizes the dynamic evolution of the internal defect density, pinning strength distribution, and micro-stress concentration of the material during the annealing process. The Pearson correlation coefficient between the macro-process feature sequence and the dynamic feature sequence of micro-defects is obtained based on the Pearson correlation algorithm. The maximum information coefficient algorithm is used to obtain the macro-process feature sequence... The maximum information coefficient between the macroscopic process feature sequence and the microscopic defect dynamic feature sequence is obtained; the dynamic time warping distance between the macroscopic process feature sequence and the microscopic defect dynamic feature sequence is obtained based on the dynamic time warping algorithm; the obtained Pearson correlation coefficient, maximum information coefficient and dynamic time warping distance are arranged in window time order to obtain the cross-domain coupling feature vector; the cross-domain coupling feature vector represents the dynamic coupling strength and time delay relationship between the macroscopic external physical field and the microscopic defect activity inside the sample material; the macroscopic process feature vector, the microscopic defect dynamic feature vector and the cross-domain coupling feature vector are serially spliced ​​to obtain the annealing process multidimensional dynamic behavior feature vector; the serial splicing means that the macroscopic process feature vector, the microscopic defect dynamic feature vector and the cross-domain coupling feature vector are connected in sequence to form a single high-dimensional feature vector with a dimension equal to the sum of the dimensions of the three.

[0035] In this embodiment, the extraction of multidimensional dynamic behavior features of the annealing process helps to capture the magnetic domain dynamic behavior features of the internal defect density, pinning strength distribution and micro-stress concentration of the material, reduce the dimensional redundancy and physical semantic ambiguity of the multidimensional dynamic behavior feature vector of the annealing process, and improve the sensitivity and discrimination ability of feature expression for early abnormal precursors induced by defects.

[0036] Furthermore, the specific process for predicting magnetic field annealing anomalies is as follows: The multi-dimensional dynamic behavior feature vector of the annealing process is input into a preset anomaly prediction model, outputting the annealing process anomaly prediction result at the current sampling time, including the comprehensive annealing anomaly score and the probability distribution of annealing anomaly types. The comprehensive annealing anomaly score reflects the degree to which the current annealing process state deviates from the normal process state; the closer the comprehensive annealing anomaly score is to 1, the higher the probability of an annealing process anomaly occurring. The probability distribution of annealing anomaly types represents a probability vector composed of the probability values ​​corresponding to each known anomaly category. The sum of all probability values ​​in this probability vector is 1, and each probability value corresponds to the confidence level of an occurrence of a known anomaly category at the current sampling time. Based on the annealing process anomaly prediction results, The magnetic field annealing anomaly early warning process is as follows: When the comprehensive score of annealing anomaly exceeds the preset early warning threshold, it is determined that there is an abnormal risk in the current annealing process, triggering an anomaly early warning signal. Conversely, if the score is below the preset threshold, the current annealing process is determined to be normal, and multivariate time-series data at the next sampling time is collected, and subsequent analysis steps are executed iteratively. The preset early warning threshold is represented by the average value of the comprehensive score of annealing anomaly over a historical time period. When an anomaly early warning signal is triggered, the anomaly category corresponding to the maximum probability value in the probability distribution of annealing anomaly types is determined as the anomaly type for the current warning, and the current annealing process anomaly prediction result is sent to the magnetic field annealing data monitoring center. Anomaly categories include grain anomaly risk, annealing brittle critical transition, interface diffusion failure, etc.

[0037] It should be noted that the preset anomaly prediction model used in this embodiment is specifically a temporal classification and regression model based on a cascaded long short-term memory network and a multi-head self-attention mechanism. Alternatively, a hybrid model based on a temporal convolutional network and a graph attention network, a sequence labeling model based on a Transformer encoder and a conditional random field, or a fusion model based on a deep residual shrinkage network and a gated recurrent unit can also be used. The specific training process of the preset anomaly prediction model is as follows: First, multiple sets of historical annealing process operation data pre-stored in the magnetic field annealing process anomaly prediction knowledge base are acquired. Each set of historical annealing process operation data includes the complete time series and complete event stream of the first data channel, the complete time series and complete event stream of the second data channel, and anomaly label data obtained by process experts who have labeled the anomaly type and scored the anomaly severity of the historical annealing process operation. A sliding window is performed on the first data channel, and the process trend feature vector within each time window is extracted. Simultaneously, the second data channel is divided according to phase... Simultaneously, the dynamic behavior feature vector of the magnetic domain is extracted within the time window. The process trend feature vector corresponding to each time window is concatenated with the dynamic behavior feature vector of the magnetic domain to form a training sample with the time window as the basic unit. Each training sample carries the anomaly label data of its corresponding time window as a supervision signal. Then, the training samples are divided into training set and validation set according to the preset partitioning rules. The preset anomaly prediction model is trained by multiple tasks using the training set. The preset partitioning rules are set in advance by preset personnel. During the training process, the model output of each batch of training samples is calculated by forward propagation. The overall loss function value is calculated by combining the anomaly label data. Then, the weight parameters of each layer of the model are updated by the backpropagation algorithm. The validation set is used to periodically evaluate the anomaly scoring error and anomaly type classification accuracy of the model on the validation samples during the training process. When the overall loss function value on the validation set does not decrease during the training period, the early stop mechanism is triggered and the optimal model parameters are saved, resulting in the pre-trained preset anomaly prediction model.

[0038] In this embodiment, the prediction of anomalies caused by magnetic field annealing helps to reduce the missed detection of defect-induced anomalies caused by monitoring only a single macroscopic variable, reduce the frequency of false alarms caused by data distortion due to smoothing filtering, improve the ability to perceive and classify material microstructure degradation events in advance, and realize the upgrade of anomaly prediction from post-event threshold alarm to pre-event coupled anomaly early warning.

[0039] Preferably, the magnetic field annealing process anomaly prediction system based on multivariate time series analysis includes: a magnetic field annealing domain jump monitoring module, an annealing process multidimensional feature extraction module, and a magnetic field annealing anomaly monitoring module. The magnetic field annealing domain jump monitoring module performs magnetic field annealing domain jump analysis on the multivariate time series set, including adaptive mode decomposition and magnetic field annealing wavelet transform, and decides whether to perform dual-channel data reconstruction based on the analysis results. The magnetic field annealing domain jump monitoring module helps to physically separate macroscopic process trend signals from microscopic defect transient signals from the data source, reducing the loss rate of key defect information caused by traditional smoothing filtering preprocessing.

[0040] The annealing process multidimensional feature extraction module is used to extract macroscopic process feature vectors, microscopic defect dynamic feature vectors, and cross-domain coupling feature vectors based on the dual-channel data reconstruction results after the dual-channel data reconstruction is performed. When dual-channel data reconstruction is not performed, the multidimensional dynamic behavior feature extraction of the annealing process is directly based on the magnetic domain jump analysis results of magnetic field annealing. The annealing process multidimensional feature extraction module helps to transform the discrete magnetic domain jump event flow into a defect dynamic feature space that can quantify defect density, pinning strength, and stress concentration, thereby improving the sensitivity of feature expression for early abnormal precursors induced by defects.

[0041] The magnetic field annealing anomaly monitoring module is used to predict magnetic field annealing anomalies, including early warnings, based on the results of the multi-dimensional dynamic behavior feature extraction of the annealing process. This module helps reduce the frequency of false alarms caused by data distortion and improves the ability to proactively detect and classify material microstructure degradation events such as abnormal grain growth risk, annealing brittle critical transition, and interface diffusion failure.

[0042] In this embodiment, the magnetic domain jump monitoring module, the multi-dimensional feature extraction module for the annealing process, and the magnetic annealing anomaly monitoring module help to fundamentally reverse the passive situation of information breakage caused by smoothing filtering in traditional methods. Specifically, the magnetic domain jump monitoring module provides an independent microscopic transient data source for the multi-dimensional feature extraction module for the annealing process, the multi-dimensional feature extraction module for the annealing process provides high-discrimination input features for the magnetic annealing anomaly monitoring module, and the magnetic annealing anomaly monitoring module feeds back the anomaly prediction results to the knowledge base for continuous optimization of the parameters of the first two modules, thereby forming a closed-loop iterative anomaly prediction system that combines data-driven and physical cognition.

[0043] Example 2, as an alternative to the adaptive mode decomposition in Example 1, addresses the situation where the non-stationarity of the local magnetic field strength fluctuation time series and the excitation current pulsation time series in the multivariate time series set exceeds a preset stationarity index. Since the spectral segmentation boundary relied upon by the empirical wavelet transform shifts under strong non-stationary signal conditions as the analysis time window slides, if a fixed spectral segmentation boundary based on the global fast Fourier transform is still used for unified mode decomposition across the entire time period, it may lead to frequency band aliasing between the correctly separated low-frequency trend and high-frequency transient in one time period, resulting in subsequent mode classification errors and misjudgments of jump events. Therefore, adaptive mode decomposition is required. The specific process is as follows: Obtain the local magnetic field strength fluctuation time series and the excitation current pulsation time series from the multivariate time series set along the time axis... The recursion rate index sequence is dynamic; the recursion rate index sequence includes the magnetic field fluctuation recursion rate index sequence and the excitation current recursion rate index sequence; the recursion rate index is obtained in the following way: within a preset length window sliding along the time axis, the phase space calculation of the local magnetic field strength fluctuation time series or the excitation current pulsation time series is performed based on the delayed coordinate embedding algorithm to obtain the phase space trajectory matrix. For example, for a local magnetic field strength fluctuation time series segment {h(t1), h(t2), h(t3), h(t4), h(t5)} within a preset length window, if the embedding dimension m=3 and the delay time τ=1, then the phase space trajectory matrix X obtained by delayed coordinate embedding within the preset length window is represented as [v1, v2, v3], where v1=[h(t1), h(t2), h(t3)] T Let v2 be the three-dimensional delay vector corresponding to the first sampling time, where v2 = [h(t2), h(t3), h(t4)]. T The three-dimensional delay vector corresponding to the second sampling time is v3=[h(t3), h(t4), h(t5)]. T h(t) is the three-dimensional delay vector corresponding to the third sampling time. k) represents the induced voltage value at the k-th sampling time, T represents the transpose of the vector, and each row of matrix X corresponds to a state point at a time in the phase space, reflecting the joint evolution trajectory of the signal amplitude in the three-dimensional phase space within the preset length window. The preset length window is set in advance by preset personnel. The Euclidean distance between any two phase points in the phase space trajectory matrix is ​​obtained based on the Euclidean distance algorithm. The ratio of the number of phase point pairs whose Euclidean distance is less than the preset neighborhood radius to the total number of phase point pairs is counted, and this ratio is used as the recursion rate index value at the center time of the preset length window. The preset neighborhood radius is represented by the average Euclidean distance between any two phase points in the phase space trajectory matrix of the historical time period. When the cumulative proportion of recursive sampling points is greater than the preset proportion threshold, the sensing signal of the corresponding time series in the multivariate time series set is determined to be a strongly non-stationary signal, and adaptive spectrum segmentation boundary is performed. Otherwise, conventional adaptive mode decomposition is performed. The preset stationary range includes both endpoints and is set in advance by preset personnel. The preset proportion threshold is represented by the average of the cumulative proportion of recursive sampling points in the historical time period. The specific process of obtaining the cumulative proportion of recursive sampling points is as follows: count the number of recursion rate index values ​​in the recursion rate index sequence that exceed the preset stationary range, and use the result of the proportion calculation between this number and the total number of sampling points in the recursion rate index sequence as the cumulative proportion of recursive sampling points.

[0044] The specific process of adaptive spectrum segmentation boundaries is as follows: The magnetic domain hopping analysis time period is divided into multiple boundary segmentation time periods according to a preset window length and a preset overlap rate. For example, if the total duration of the magnetic domain hopping analysis time period is S, the preset window length is W, and the preset overlap rate is α, then the first boundary segmentation time period is [0, W], the second boundary segmentation time period is [...]. , The third boundary segmentation time period is [ , ], and so on, the j-th boundary segmentation time period is represented as [ , ], where j = 1, 2, ..., J, and J is the total number of boundary segmentation time periods, whose values ​​are equal to satisfying The smallest integer, the overlapping region formed between adjacent boundary segmentation time periods due to the existence of a preset overlap rate α, the length of the overlapping region is equal to In this process, the preset overlap rate is pre-set by a preset team. Based on the Fast Fourier Transform (FFT) algorithm, each signal segment within a boundary segmentation time period (i.e., the sensor signal segment between the start and end times of that boundary segmentation time period) is independently subjected to FFT to obtain its local spectral distribution. The corresponding local spectral segmentation boundary is then obtained based on each local spectral distribution. It should be noted that the method for obtaining the local spectral segmentation boundary is the same as that in Example 1, except that the input for each operation is a local spectral distribution rather than a global spectral distribution. It should also be noted that multiple boundary segmentation time periods are obtained within the magnetic domain jump analysis time period according to the preset window length and preset overlap rate. Adjacent boundary segmentation time periods form an overlapping region due to the preset overlap rate, and the length of the overlapping region is the product of the preset window length and the preset overlap rate.

[0045] Within each boundary segmentation time period, a magnetic field annealing wavelet transform is performed based on the local spectral segmentation boundary to obtain a set of local intrinsic mode function (EMF) components. In the overlapping region of adjacent boundary segmentation time periods, the local EMF components of the overlapping region between the previous and subsequent boundary segmentation time periods are fused using a weighted gradient splicing method to obtain continuous EMF components whose frequency band division adaptively adjusts according to the non-stationary characteristics of the signal. For any two temporally adjacent boundary segmentation time periods, the end of the previous boundary segmentation time period and the beginning of the subsequent boundary segmentation time period overlap on the time axis; this overlapping portion is the overlapping region. A preset window length and a preset overlap rate are then used to... The value obtained from the product operation is used as the overlap length. The specific process of local component fusion is as follows: within the overlapping region, the local intrinsic mode function (IMF) components in the previous boundary segmentation time period are assigned component weighting coefficients that decrease linearly from front to back, while the local IMF components in the next boundary segmentation time period are assigned component weighting coefficients that increase linearly from front to back. The weighted local IMF components are then summed point by point at the corresponding sampling points. The summation result is the amplitude value of the fused local IMF component at that sampling point. The component weighting coefficients are used to quantify the contribution ratio of the local IMF components generated in the previous or next boundary segmentation time period to the final local component fusion result.

[0046] It is important to note that the component weighting coefficients involved in this embodiment were pre-constructed by process experts based on the modal decomposition and reconstruction effects of a large amount of historical magnetic field annealing data, and stored in the magnetic field annealing process anomaly prediction knowledge base. This provides a core basis for the smooth connection of local components in overlapping intervals during adaptive modal decomposition. Specifically, the construction process of these component weighting coefficients is as follows: First, local magnetic field strength fluctuation time series and excitation current pulsation time series samples under different annealing process curves, different magnetic material types, and different annealing durations are collected. For each sample, a sliding window is divided to generate adjacent boundary segmentation time periods with overlapping regions. The local intrinsic mode function components generated by the previous boundary segmentation time period and the next boundary segmentation time period within each group of adjacent boundary segmentation time periods are recorded. Subsequently, for each group of adjacent local components... Amplitude deviation analysis is performed point-by-point within the overlapping region. Different linear weighting combinations are statistically analyzed, i.e., local components in the previous boundary segmentation time period are assigned weighting coefficients that decrease linearly from 1 to 0, and local components in the subsequent boundary segmentation time period are assigned weighting coefficients that increase linearly from 0 to 1. The reconstruction error between the fusion result and the original unsegmented continuous signal is analyzed. The optimal weighting coefficient change rate is selected with the goal of minimizing the reconstruction error, and the characteristic parameters of the annealing process scene corresponding to the optimal weighting coefficient combination are recorded simultaneously. Correlation analysis, such as Spearman rank correlation coefficient, is used to eliminate unstable weighting coefficient values ​​caused by instantaneous anomalies in the sensor signal, and the correspondence between the component weighting coefficients with statistical consistency and the characteristic parameters of the process scene is retained, thus forming a complete component weighting coefficient parameter system.

[0047] In this embodiment, adaptive mode decomposition helps to reduce the risk of frequency band aliasing between low-frequency trend modes and high-frequency transient modes at different times when the local magnetic field strength fluctuation time sequence and the excitation current pulsation time sequence exhibit strong non-stationary characteristics. This reduces the misjudgment and missed detection of transient jump events caused by mode aliasing, and improves the ability of adaptive mode decomposition to track and adapt to the time-varying signal spectrum caused by the dynamic evolution of physical parameters such as material permeability and resistivity during the annealing process.

[0048] like Figure 3The figure shows the monitoring curve of coplanar waveguide surface temperature fluctuation in the magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in this application embodiment. As can be seen from the figure, the horizontal axis represents time in minutes, and the vertical axis represents temperature in Kelvin. During the entire time series monitoring process from 0 to 6 minutes, the original time-domain fluctuation amplitude of the coplanar waveguide surface temperature is effectively constrained within an envelope of ±0.005K. This envelope corresponds to the narrow channel pattern in the figure with a constant central baseline as the axis of symmetry, and flat and basically equal width at the upper and lower boundaries. This indicates that the local thermal environment maintained by the low-temperature testing equipment when the high-frequency waveguide is in close contact with the sample has excellent short-term stability and low noise characteristics. Thus, before performing magnetic field annealing domain jump analysis, the temperature sensing noise and the transient signal induced by the real magnetic domain jump can be effectively distinguished in the frequency domain from the physical signal source. This provides a clean temperature reference baseline for high-fidelity acquisition of the global spectrum distribution in adaptive mode decomposition.

[0049] like Figure 4 The figure shows a schematic diagram of the sample delivery process of the low-temperature testing equipment on which the magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in this application embodiment relies. In the figure, the horizontal x-axis represents the spatial position change of sample 3 between annealing area 1 and detection area 2, and the vertical y-axis represents the operating space of telescopic arm 7 in the vertical direction. As shown in the figure, during the annealing process, the sample is initially placed in annealing area 1. The heating device is quickly moved to the vicinity of the sample by the telescopic arm to achieve rapid annealing. During this period, the movable magnet slides to the annealing center to provide a magnetic field environment for the annealing process. After annealing, through the coordinated operation of multiple telescopic arms, the sample is transferred non-destructively from annealing area 1 to detection area 2 through the intermediate sample transfer path. Then, the telescopic arm, together with test options such as coplanar waveguide, optical probe or mechanical probe, performs in-situ testing of multiple physical properties of the sample. The sample transfer process can achieve non-destructive transfer of the sample between the annealing area and the detection area through the cooperation of multiple telescopic arms, reducing the risks of oxidation, contamination and structural relaxation introduced by the transfer of the sample between different cavities.

[0050] like Figure 5The diagram shows a desktop cryogenic testing device structure for a magnetic field annealing process anomaly prediction method based on multivariate time series analysis provided in this application embodiment. The implementation of this method relies on a desktop cryogenic testing device. The main structure of this device includes a vacuum system 4, a magnetic field annealing system 5, an in-situ testing system 6, a telescopic arm 7, a movable magnet 8, a heating device 9, and a temperature sensor 10. A quartz tube 11 connects to the vacuum system to form an internal vacuum environment. The quartz tube connects the annealing area and the detection area, providing internal connection for the telescopic arm. Optional internal connecting pipes can be made of other high-temperature resistant materials, including but not limited to quartz, alumina, and sapphire, to extend the operating space. The non-destructive transfer of samples between the annealing and testing areas is achieved through the telescopic movement of the telescopic arm. The movement of the telescopic arm can be selected in manual or automatic modes, and the extension can be driven by a lead screw. The power transmission can be manual or driven by a servo motor, stepper motor, or piezoelectric motor. The quartz tube is wrapped with thermal insulation material 12 to reduce heat loss during annealing. The heating device and temperature sensor are integrated and installed at the end of the telescopic arm, and the temperature is measured by the displacement of the telescopic arm. The current heating device allows for rapid approach and separation from the sample, enabling rapid temperature control of the sample area. The fastest temperature change rate can reach 100K / s. Heating devices include, but are not limited to, Joule heating, laser heating, and infrared heating, with a maximum heating temperature of 1700K. Temperature sensors include, but are not limited to, thermistors, thermocouples, and infrared thermometers, covering a test temperature range of 0.3K-1700K. A retractable arm carries heat and cold sources to provide different test temperatures. Heat sources include, but are not limited to, heating rods, heating plates, and heating wires, while cold sources include, but are not limited to, Peltier sensors and liquid nitrogen spray guns. The range of K varies from 0.3K to 600K. A movable magnet is placed outside the quartz tube and can slide freely between the annealing area and the detection area, providing a magnetic field environment for the annealing process and the in-situ testing process, respectively. The movable magnet includes, but is not limited to, permanent magnets, electromagnets, superconducting magnets, vector magnets, etc. The in-situ testing system is set in the detection area and uses a telescopic arm with different test probes to achieve the testing of various physical properties. The test probe 13 includes, but is not limited to, coplanar waveguides, optical probes, mechanical probes, etc. Among them, the coplanar waveguides used for high-frequency testing have an insulating coating on their surface, and the waveguides and samples can achieve gapless contact.

[0051] This application provides a desktop low-temperature testing device that integrates a magnetic field annealing system and an in-situ testing system into the same vacuum chamber. A robotic arm enables non-destructive transfer of samples between the annealing and testing areas, avoiding the risks of oxidation, contamination, and structural relaxation introduced by traditional discrete devices due to sample transfer between different chambers. This improves experimental efficiency and the authenticity of test data, and enhances the temperature change rate and process flexibility.

Claims

1. A method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis, characterized in that, Includes the following steps: A magnetic field annealing domain jump analysis, including adaptive mode decomposition and magnetic field annealing wavelet transform, is performed on a multivariate time series set. Based on the analysis results, a decision is made on whether to perform dual-channel data reconstruction. The multivariate time series set includes local magnetic field strength fluctuation time series and excitation current pulsation time series. When performing dual-channel data reconstruction, after the dual-channel data reconstruction is completed, the annealing process multidimensional dynamic behavior feature extraction is performed based on the dual-channel data reconstruction results, extracting macroscopic process feature vectors, microscopic defect dynamic feature vectors, and cross-domain coupling feature vectors. When dual-channel data reconstruction is not performed, the annealing process multidimensional dynamic behavior feature extraction is performed directly based on the magnetic domain jump analysis results of magnetic field annealing. After the multidimensional dynamic behavior feature extraction of the annealing process is completed, magnetic field annealing anomaly prediction, including magnetic field annealing anomaly early warning, is performed based on the multidimensional dynamic behavior feature extraction results of the annealing process.

2. The method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis as described in claim 1, characterized in that, The specific process of the magnetic field annealing domain jump analysis is as follows: Collect a multivariable time series of sample materials during magnetic field annealing; Adaptive mode decomposition is performed on the multivariate time series set, and the specific process is as follows: The multivariate time series set is subjected to Fast Fourier Transform (FFT) algorithm to obtain the global spectral distribution. The global spectrum distribution includes the magnetic field fluctuation spectrum distribution obtained by performing a fast Fourier transform on the local magnetic field strength fluctuation time series in the multivariable time series set and the current pulsation spectrum distribution obtained by performing a fast Fourier transform on the excitation current pulsation time series in the multivariable time series set. Spectrum segmentation boundaries are obtained based on global spectrum distribution; In the low-frequency range before the first spectral peak, starting from the frequency position of the first spectral peak, the search is carried out step by step along the frequency axis towards zero frequency. The position where the spectral amplitude gradually decreases as the frequency decreases and first drops to the position equal to the preset noise floor level is determined as the corresponding starting boundary frequency point. The low-frequency band refers to the frequency range in the global spectrum from zero frequency to the frequency corresponding to the first spectral peak. In the high-frequency band after the last spectral peak, starting from the frequency position of the last spectral peak, search along the frequency axis in the direction of the Nyquist frequency, and determine the corresponding termination boundary frequency point when the spectral amplitude gradually decreases with the frequency and first drops to the level of the preset noise floor. Based on the aforementioned spectrum segmentation boundaries, starting boundary frequency points, and ending boundary frequency points, the global spectrum is adaptively segmented into continuous sub-bands. The specific segmentation process is as follows: Starting from zero frequency, the frequency intervals defined by two adjacent spectrum division boundaries are selected as a continuous sub-band. Magnetic field annealing wavelet transform based on continuous sub-bands.

3. The method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis as described in claim 2, characterized in that, The specific process of the magnetic field annealing wavelet transform is as follows: In each of the continuous sub-frequency bands, empirical wavelet functions and empirical scaling functions are obtained based on the wavelet transform method; The empirical scaling function is combined with the sensor signal currently undergoing adaptive mode decomposition in the multivariate time series set to perform an inner product operation, and the wavelet approximation coefficients corresponding to the lowest frequency band of the sensor signal are output. Each of the empirical wavelet functions is combined with one of the sensing signals in the corresponding multivariable time series set, namely the local magnetic field strength fluctuation time series or the excitation current pulsation time series, and the wavelet detail coefficients corresponding to each continuous sub-frequency band are output to obtain a set of intrinsic mode function components. Low-frequency trend modes and high-frequency transient modes are identified based on intrinsic mode function components.

4. The method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis as described in claim 3, characterized in that, The specific process for identifying the low-frequency trend mode and the high-frequency transient mode is as follows: The power spectral density distribution curves of each intrinsic mode function component are obtained by power spectral density estimation based on the Welch method. On the power spectral density distribution curve, the frequency coordinate corresponding to the maximum power density is marked as the center frequency of the intrinsic mode function component; The result of the ratio calculation between the power density integral value corresponding to the center frequency and the total power density integral value is used as the energy ratio value. Determine whether dual-channel data reconstruction is needed for the multivariate time series based on the low-frequency trend mode set and the high-frequency transient mode set; The low-frequency trend mode set includes intrinsic mode function components with a center frequency less than a preset cutoff frequency and an energy percentage greater than a preset energy threshold. The high-frequency transient mode set includes intrinsic mode function components with a center frequency not less than a preset cutoff frequency and an energy proportion exhibiting short-term burst characteristics.

5. The method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis as described in claim 4, characterized in that, The specific process for determining whether dual-channel data reconstruction of the multivariate time series set is needed is as follows: For each intrinsic mode function component in the high-frequency transient mode set, the instantaneous energy curve of each intrinsic mode function component on the time axis during the magnetic domain jump analysis period is obtained based on the Hilbert transform method; Within a preset time window centered on each peak moment, the peak-to-average power ratio of transient energy pulse events is obtained; The peak-to-average power ratio is expressed as the result of calculating the ratio of the peak power of the signal within the time window to the average power within the window. When the peak-to-average power ratio is greater than a preset ratio threshold, the transient energy pulse event is determined to be a domain jump event induced by a sudden jump of the domain wall; otherwise, it is determined to be a false transient component caused by random electrical noise and is eliminated. The total number of domain jump events within the domain jump analysis period is counted. When the total number is greater than a preset event number threshold, dual-channel data reconstruction is performed. Otherwise, all intrinsic mode function components contained in the low-frequency trend mode set are summed and reconstructed into a macroscopic process trend data sequence. Based on this macroscopic process trend data sequence, multidimensional dynamic behavior features of the annealing process are extracted.

6. The method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis as described in claim 5, characterized in that, The specific process of dual-channel data reconstruction is as follows: Construct the first and second data channels; The amplitudes of all intrinsic mode function components contained in the low-frequency trend mode set are summed point by point on the time axis of the magnetic domain jump analysis period according to the sampling time, and the summation result is used as the amplitude value of each sampling time in the first data channel. During the next magnetic domain jump analysis period, the data sequence output by the first data channel is automatically captured as a macroscopic process trend data sequence. The transient signal segments of all magnetic domain jump events in the high-frequency transient mode set are arranged in chronological order of their peak times on the time axis to obtain the second data channel; In the next magnetic domain jump analysis period, the data sequence output by the second data channel is automatically captured as the magnetic domain jump data sequence; The macroscopic process trend data sequence and the magnetic domain jump data sequence are output as dual-channel parallel data, and multi-dimensional dynamic behavior features of the annealing process are extracted based on the output macroscopic process trend data sequence and magnetic domain jump data sequence.

7. The method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis as described in claim 6, characterized in that, The specific process for extracting the multidimensional dynamic behavior features of the annealing process is as follows: Macro-process features are extracted from macro-process trend data sequences based on a piecewise linear fitting algorithm. The macroscopic process features are arranged in a preset process order to obtain the macroscopic process feature vector; To obtain the dynamic characteristics of micro-defects in magnetic domain jump data sequences; The dynamic features of micro-defects are concatenated end to end according to a preset dimensional order to obtain the dynamic feature vector of micro-defects. Obtain the Pearson correlation coefficient, maximum information coefficient, and dynamic time warp distance between the macroscopic process feature sequence and the microscopic defect dynamic feature sequence; The obtained Pearson correlation coefficient, maximum information coefficient, and dynamic time warp distance are arranged in order of window time to obtain the cross-domain coupling feature vector; The macroscopic process feature vector, the microscopic defect dynamic feature vector, and the cross-domain coupling feature vector are sequentially spliced ​​together to obtain the multidimensional dynamic behavior feature vector of the annealing process. Predicting magnetic field annealing anomalies based on multidimensional dynamic behavior feature vectors of annealing process.

8. The method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis as described in claim 7, characterized in that, The specific process for predicting magnetic field annealing anomalies is as follows: Input the multidimensional dynamic behavior feature vector of the annealing process into the preset anomaly prediction model, and output the annealing process anomaly prediction result at the current sampling time, including the comprehensive score value of annealing anomaly and the probability distribution of annealing anomaly type; Based on the prediction results of annealing process anomalies, an early warning of magnetic field annealing anomalies is performed. The specific process is as follows: When the comprehensive score of the annealing anomaly is greater than the preset warning threshold, an anomaly warning signal is triggered; otherwise, multivariate time series data at the next sampling time are collected and subsequent analysis steps are executed in a loop. When an abnormal warning signal is triggered, the abnormal category corresponding to the maximum probability value in the probability distribution of the annealing abnormality type is determined as the current abnormality type, and the current annealing process abnormality prediction result is sent to the magnetic field annealing data monitoring center.

9. The method for predicting anomalies in magnetic field annealing processes based on multivariate time series analysis as described in claim 2, characterized in that, The adaptive mode decomposition further includes: Obtain the recursion rate index sequence of the local magnetic field strength fluctuation time series and the excitation current pulsation time series sliding along the time axis in a multivariate time series set; The recursion rate index sequence includes the magnetic field fluctuation recursion rate index sequence and the excitation current recursion rate index sequence. The recursion rate metric is obtained through the following methods: Within a preset length window that slides along the time axis, phase space calculations are performed on the local magnetic field strength fluctuation time series or excitation current pulsation time series based on the delayed coordinate embedding algorithm to obtain the phase space trajectory matrix; The Euclidean distance between any two phase points in the phase space trajectory matrix is ​​obtained based on the Euclidean distance algorithm. The ratio of the number of phase point pairs whose Euclidean distance is less than the preset neighborhood radius to the total number of phase point pairs is calculated, and this ratio is used as the recursion rate index value at the center time of the preset length window. When the cumulative proportion of recursive sampling points is greater than the preset proportion threshold, adaptive spectrum segmentation boundary is performed; otherwise, conventional adaptive mode decomposition is performed. The specific process of adaptive spectrum segmentation boundary is as follows: The magnetic domain jump analysis time period is divided into multiple boundary segmentation time periods according to the preset window length and preset overlap rate; For each boundary segment within a time period, the signal segment is independently subjected to Fast Fourier Transform (FFT) to obtain the local spectral distribution of the signal segment. Based on each local spectral distribution, the corresponding local spectral segmentation boundary is obtained. Within each boundary segmentation time period, magnetic field annealing wavelet transform is performed based on the local spectrum segmentation boundary to obtain the set of local intrinsic mode function components. In the overlapping region within the adjacent boundary segmentation time period, the local intrinsic mode function components of the overlapping region within the previous boundary segmentation time period and the next boundary segmentation time period are fused locally based on the weighted gradual splicing method to obtain the intrinsic mode function components. The specific process of local component fusion is as follows: Within the overlapping region, the local intrinsic mode function (IMF) components within the previous boundary segmentation time period are assigned component weighting coefficients that decrease linearly from front to back, while the local IMF components within the subsequent boundary segmentation time period are assigned component weighting coefficients that increase linearly from front to back. The weighted local IMF components are then summed point by point at the corresponding sampling points, and the summation result is the amplitude value of the fused local IMF component at that sampling point.

10. A magnetic field annealing process anomaly prediction system based on multivariate time series analysis, characterized in that, include: Magnetic domain jump monitoring module for magnetic annealing, multi-dimensional feature extraction module for annealing process, and magnetic annealing anomaly monitoring module: The magnetic field annealing domain jump monitoring module is used to perform magnetic field annealing domain jump analysis on a multivariate time series set, including adaptive mode decomposition and magnetic field annealing wavelet transform. Based on the analysis results, it decides whether to perform dual-channel data reconstruction. The multivariate time series set includes local magnetic field strength fluctuation time series and excitation current pulsation time series. The annealing process multidimensional feature extraction module is used to extract the macroscopic process feature vector, microscopic defect dynamic feature vector, and cross-domain coupling feature vector of the annealing process based on the dual-channel data reconstruction results after the dual-channel data reconstruction is performed. When dual-channel data reconstruction is not performed, the multidimensional dynamic behavior feature extraction of the annealing process is directly based on the magnetic domain jump analysis results of magnetic field annealing. The magnetic field annealing anomaly monitoring module is used to predict magnetic field annealing anomalies, including early warnings, based on the results of the multi-dimensional dynamic behavior feature extraction of the annealing process after the multi-dimensional dynamic behavior feature extraction is completed.

Images