Electrolytic capacitor testing method and system

Abstract

The invention discloses an electrolytic capacitor testing method and system, and particularly relates to the technical field of electrolytic capacitor testing, and the method comprises the steps: obtaining a time-frequency domain coupling waveform of electrolytic capacitor testing, and constructing an anti-interference reference feature library covering a process fluctuation interval based on a correlation function of an electrolyte dynamic impedance spectrum and an ultrasonic speed; based on a physical equation of electrolyte evaporation and polar plate corrosion, a topological structure updating rule of a graph neural network and a physically constrained graph topological structure are dynamically generated; the method comprises the following steps: extracting defect sensitive frequency band characteristics of a cross-modal signal through a differentiable frequency band selection network, extracting multi-scale fusion characteristics from a time-frequency domain coupling waveform based on the defect sensitive frequency band characteristics, inputting the multi-scale fusion characteristics into an edge keeping regularization layer, and adaptively adjusting reconstruction parameters according to the confidence coefficient of a defect gradient field so as to obtain an edge keeping regularization layer; and a three-dimensional defect form meeting the continuity constraint is generated, so that the problem of low detection precision of the internal defect of the electrolytic capacitor is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of electrolytic capacitor testing, and more particularly, to a method and system for testing an electrolytic capacitor. Background Art

[0002] Electrolytic capacitors are widely used energy storage components in electronic devices. Internal defects (such as electrolyte leakage and plate corrosion) pose a direct threat to device reliability. Traditional inspection methods rely on destructive disassembly or sampling inspection (such as anatomical analysis and X-ray imaging), which suffer from low efficiency, insufficient coverage, and high costs. Existing nondestructive testing technologies (such as single-mode ultrasonic scanning or conductivity measurement) are limited by signal sensitivity and feature decoupling capabilities, making it difficult to accurately identify micron-level defects and unable to achieve three-dimensional defect localization, resulting in significant blind spots in quality control.

[0003] In existing technologies, nondestructive testing of internal defects in electrolytic capacitors faces the following core challenges:

[0004] The mapping relationship between indirect signals and internal defects is complex, and traditional single-modal detection methods have difficulty extracting high-discrimination features, resulting in a high misjudgment rate.

[0005] The lack of an effective fusion mechanism for multi-source heterogeneous data (such as acoustic and electromagnetic signals) restricts the accuracy of defect classification and location;

[0006] There is an urgent need for an efficient, high-precision and cost-effective non-destructive testing method to break through the bottleneck of hidden analysis of complex defects and optimize electrolytic capacitor testing. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art, the present invention provides an electrolytic capacitor testing method and system, which solves the problems raised in the above-mentioned background technology through synchronous acquisition of multi-modal physical signals, dynamic compensation and physical constraint modeling.

[0008] To achieve the object of the present invention, the present invention provides a method for testing an electrolytic capacitor, comprising the following steps:

[0009] Step 1: Synchronous acquisition of multimodal physical coupling signals: Establish a phase-locked trigger acquisition mechanism for ultrasonic and electromagnetic waves, and synchronously control the generation of time-frequency domain coupled waveforms of dual-modal signals through bus-level hardware. Based on the correlation function Φ(f) between the dynamic impedance spectrum of the electrolyte and the ultrasonic velocity, an anti-interference benchmark feature library covering the process fluctuation range is constructed, and dynamic compensation for the ultrasonic propagation velocity is performed based on the ratio of the physical properties of the current electrolyte and the benchmark electrolyte.

[0010] Step 2: Cross-modal modeling of physical constraints: Based on the physical equations of electrolyte evaporation and plate corrosion, the topology update rules of the graph neural network are dynamically generated, and the graph topology of the physical constraints is output. The defect-sensitive frequency band features of the cross-modal signal are extracted through a differentiable band selection network, and dynamic attention weights are generated. The graph topology of the physical constraints is used to define the spatial correlation of the defect diffusion path.

[0011] Step 3. Edge-adaptive defect morphology inversion: Extract multi-scale fusion features from the time-frequency domain coupled waveform based on the defect-sensitive frequency band characteristics, input the multi-scale fusion features into the edge-preserving regularization layer, and adaptively adjust the reconstruction parameters according to the defect gradient field confidence to generate a three-dimensional defect morphology that meets the continuity constraint; based on the difference in propagation characteristics between ultrasonic waves and electromagnetic waves, optimize the physical consistency of the inversion results; (verify the physical consistency of the optimized inversion results by judging the physical laws of the ultrasonic propagation time difference and the electromagnetic wave phase difference; three-dimensional defect morphology refers to the voxel grid of internal defects of the electrolytic capacitor generated by the inversion network, and each voxel value represents the probability of a defect existing at the corresponding position).

[0012] Preferably, by traversing different gradient electrolyte formulas, packaging stresses, and temperature combinations, benchmark feature vectors are collected. Each set of benchmark feature vectors includes the corrected ultrasonic velocity, the correlation function Φ(f) between the dynamic impedance spectrum of the electrolyte and the attenuation of the ultrasonic velocity, the frequency domain amplitude spectrum extracted by Fourier transform of the ultrasonic amplitude sequence, and the frequency domain gradient of the electromagnetic wave impedance spectrum. The benchmark feature vectors are summarized to obtain an anti-interference benchmark feature library. Based on the dynamic impedance spectrum Z electrolyte (f) and ultrasonic velocity v g (ω), ω represents the angular frequency, f represents the frequency, and the correlation function Φ(f) is calculated by the following formula, which provides a physical basis for cross-modal modeling:

[0013]

[0014] The influence of process fluctuation on ultrasonic velocity is eliminated by dynamic compensation of electrolyte viscosity and dielectric constant; the current electrolyte viscosity and dielectric constant are recorded as η c , ε c , the viscosity and dielectric constant of the reference electrolyte are respectively denoted as η b , ε b The dynamic compensation formula is used to automatically correct the influence of process fluctuations (such as electrolyte batch differences) on ultrasonic propagation velocity to ensure the consistency of detection conditions. The dynamic compensation formula for ultrasonic velocity is:

[0015]

[0016] v comp Indicates the corrected ultrasonic velocity.

[0017] Preferably, based on the anti-interference benchmark feature library, the physical laws of electrolyte evaporation and plate corrosion are embedded in the topological structure update and feature fusion process of the graph neural network to solve the misjudgment problem caused by traditional multimodal fusion relying on statistical correlation;

[0018] The operation process of the topology update rule of the graph neural network includes:

[0019] The electrolyte evaporation process is modeled as a concentration diffusion equation, and the plate corrosion process is modeled as an electrochemical reaction equation;

[0020] The physical equations of electrolyte evaporation and plate corrosion include concentration diffusion equation and electrochemical reaction equation;

[0021] The concentration diffusion equation is combined with the electrochemical reaction equation, discretized into the ratio of the concentration gradient difference between nodes to the compensation speed, and then multiplied by the correlation function to generate the edge weight calculation rule. The edge weight w ij It is obtained by multiplying the ratio of the absolute value of the electrolyte concentration gradient difference to the compensated ultrasonic group velocity by the correlation function Φ(f);

[0022] Adjust edge weights based on actual process parameters to ensure that the message transmission path of the graph neural network conforms to the physical laws of material corrosion and diffusion;

[0023] The differentiable frequency band selection network is implemented as follows:

[0024] Input signal: The ultrasonic time domain signal and the frequency domain characteristics of the electromagnetic wave complex impedance spectrum collected in step 1;

[0025] Extraction of defect-sensitive frequency bands: Calculate the product of the absolute value of the frequency domain gradient of the ultrasonic amplitude spectrum and the absolute value of the frequency domain gradient of the electromagnetic impedance spectrum to generate the energy distribution ΔE(f) of the defect-sensitive frequency band;

[0026] Dynamic attention weight generation: Normalizes the energy distribution of defect-sensitive frequency bands and generates a band attention mask through preset threshold screening to suppress noise interference in non-defect frequency bands;

[0027] The multi-scale feature fusion and node update method is as follows:

[0028] Input features: The baseline feature vector constructed in step 1 is used as the initial node feature of the graph neural network;

[0029] Feature transfer rule: The feature update of node i depends on the weighted sum of the features of neighboring node j, with the weight being the edge weight w ij The product of the frequency band value corresponding to the frequency band attention mask;

[0030] Output data: The updated node features are passed to step 3 for defect morphology inversion.

[0031] Preferably, the topology structure update rule includes an edge weight update formula, and the edge weight update formula is obtained as follows:

[0032] Nodes i and j are used to represent the indexes of the eigenvectors at different spatial locations inside the electrolytic capacitor; the edge weights are obtained using the following formula:

[0033]

[0034] Where, ΔC ij represents the electrolyte concentration gradient difference between nodes i and j;

[0035] Sum up the edge weights to get the edge weight matrix.

[0036] Preferably, the multi-scale fusion feature is obtained in the following manner:

[0037] Perform wavelet transform on the ultrasonic time domain signal to extract multi-resolution features (the macroscale is used to capture overall corrosion, and the microscale is used to locate micro defects), and perform short-time Fourier transform on the electromagnetic wave complex impedance spectrum to obtain the local frequency domain energy distribution;

[0038] Based on the frequency band attention mask, the ultrasonic wavelet coefficients and the electromagnetic wave frequency domain energy are weightedly fused to enhance the signal strength of the defect-sensitive frequency band;

[0039] Through the message passing mechanism of graph neural networks, the weighted cross-modal features are hierarchically aggregated at different scales (macro, meso, and micro), and combined with the topological constraints of the electrolyte diffusion path to generate three-dimensional space-frequency-time fusion features, providing highly discriminative input for defect inversion.

[0040] Preferably, the process of acquiring the three-dimensional defect morphology includes:

[0041] Input multi-scale fusion features, calculate the spatial gradient of the multi-scale fusion features to generate the defect gradient field of the defect boundary, and quantify the confidence of the defect boundary;

[0042] Based on the encoder-decoder architecture, the voxel grid corresponding to the 3D defect morphology is obtained. The encoder compresses the fused features into a low-dimensional latent vector through a 3D convolution layer to extract the macro-defect distribution. The decoder gradually upsamples through a 3D deconvolution layer and updates the voxel grid in combination with the topological structure of the physical constraint graph. The adjacency matrix of the physical constraint graph topology is used as the spatial connection template of the decoder to constrain the diffusion path of the defect morphology during the upsampling process to be consistent with the actual physical laws of electrolyte corrosion. The defect probabilities of adjacent spatial locations are forced to transition smoothly to avoid isolated noise points.

[0043] The ultrasonic propagation time difference and electromagnetic wave phase difference are calculated, and the three-dimensional defect morphology is corrected based on the calculation results. If the ultrasonic propagation time difference exceeds the threshold, it indicates that the material density in the defect area is reduced or there are voids, and the voxel value (defect probability) at the corresponding position is increased. If the electromagnetic wave phase difference is lower than the requirement, it indicates that the electrolyte conductivity is reduced or the plate is corroded, and the boundary confidence of the corrosion area is enhanced.

[0044] Preferably, the regularization coefficient is dynamically adjusted based on the defect gradient field confidence, the non-equilibrium entropy production rate, or a fusion of the two, and the regularization coefficient is used to control the smoothing strength.

[0045] Preferably, the regularization coefficient is dynamically adjusted based on the confidence of the defect gradient field:

[0046] For high confidence areas (defect gradient field confidence ≥ confidence threshold), reduce the regularization coefficient and prioritize retaining the sharp features of the defect boundary;

[0047] For low confidence areas (defect gradient field confidence < confidence threshold), the regularization coefficient is enhanced to suppress artifact interference caused by noise;

[0048] The confidence threshold is adaptively set according to the statistical distribution of the gradient field of historical defect samples; after each batch of inspection is completed, the threshold is dynamically updated according to the confidence distribution of the defect gradient field of the current sample.

[0049] Preferably, the regularization coefficient is dynamically adjusted based on the non-equilibrium entropy production rate,

[0050] According to the electrolyte temperature field T(x, y, z) and the dissipation function σ(x, y, z), the formula Calculation of the non-equilibrium entropy production rate in the defect region The non-equilibrium entropy production rate represents the amount of energy dissipated in the defect area per unit time;

[0051] The electrolyte temperature field is measured by an infrared thermal imager or temperature sensor. The dissipation function is used to characterize the energy dissipation density of the material corrosion process; Ω represents the spatial range of the defect area;

[0052] If the non-equilibrium entropy production rate threshold is exceeded Increase the regularization coefficient to enhance the smoothing strength and suppress noise interference in high dissipation areas;

[0053] If the non-equilibrium entropy production rate threshold is not exceeded, the regularization coefficient is reduced to preserve the defect details in the low dissipation region.

[0054] Preferably, the regularization coefficient is dynamically adjusted based on the integration of the defect gradient field confidence and the non-equilibrium entropy production rate, including:

[0055] Combining the signal defect gradient field confidence with the non-equilibrium entropy production rate to overcome the limitations of a single dimension;

[0056] The regularization coefficient is denoted as λ, λ0 represents the initial regularization coefficient, and the defect gradient field confidence is denoted as T z ,

[0057]

[0058] Among them, tanh(·) is used to compress the entropy production rate ratio to the (-1, 1) interval to avoid extreme values that cause the regularization coefficient to run out of control.

[0059] To achieve the object of the present invention, the present invention provides an electrolytic capacitor testing system, comprising:

[0060] The test signal synchronization acquisition module synchronously collects ultrasonic and electromagnetic wave signals through a phase-locked trigger acquisition mechanism, generates a time-frequency domain coupled waveform data stream, and transmits it to the dynamic compensation module;

[0061] The dynamic compensation module builds an anti-interference reference feature library covering the process fluctuation range. It dynamically compensates the ultrasonic propagation velocity based on the ratio of the current and reference electrolyte physical parameters, calculates the corrected ultrasonic velocity, and transmits it to the pattern generation engine module. The anti-interference reference feature library is then transmitted to the frequency selection network module.

[0062] The graph engine module dynamically generates topology update rules for the graph neural network based on the physical equations of electrolyte evaporation and plate corrosion, and outputs a physically constrained graph topology.

[0063] The frequency selection network module is used to calculate the cross-modal frequency domain gradient product to obtain the energy distribution of the defect-sensitive frequency band, generate the frequency band attention mask after normalization, and transmit it to the fusion engine module;

[0064] The fusion engine module extracts multi-scale fusion features from the time-frequency domain coupled waveform based on the defect-sensitive frequency band characteristics, inputs the multi-scale fusion features into the edge-preserving regularization layer, adaptively adjusts the reconstruction parameters according to the confidence level of the defect gradient field, generates a three-dimensional defect morphology that meets the continuity constraint, and transmits it to the inversion optimization module;

[0065] The inversion optimization module generates a voxel grid through an encoder-decoder architecture and applies an edge-preserving regularization layer. The counterfactual reasoning loss function is obtained by weighted summation of the ultrasonic propagation time difference and the electromagnetic wave phase difference, and is transmitted to the graph engine module to update the three-dimensional convolution kernel weights to achieve closed-loop feedback.

[0066] Technical effects and advantages of the present invention:

[0067] (1) The invention provides an electrolytic capacitor testing method, establishes a phase-locked trigger synchronous acquisition mechanism for ultrasonic and electromagnetic waves, constructs an anti-interference benchmark feature library based on the correlation function between the dynamic impedance spectrum of the electrolyte and the ultrasonic velocity, and corrects the influence of process fluctuations on the ultrasonic propagation velocity through a dynamic compensation formula to ensure the authenticity and consistency of the multimodal signal; based on the physical equations of electrolyte evaporation and plate corrosion, the material corrosion diffusion law is embedded in the edge weight calculation rules of the graph neural network, constraining the message transmission path to be consistent with the actual physical process, avoiding the misjudgment problem of traditional multimodal fusion relying on statistical correlation.

[0068] (2) The invention provides an electrolytic capacitor testing method, which generates a three-dimensional defect morphology that meets the continuity constraint based on multi-scale fusion features and edge-preserving regularization technology; extracts macro-micro fusion features by wavelet transform of ultrasonic time domain signals and short-time Fourier transform of electromagnetic wave frequency domain features, and enhances defect-sensitive frequency bands by combining frequency band attention mask; reconstructs a three-dimensional voxel grid using an encoder-decoder architecture to force a smooth transition of defect probabilities to avoid isolated noise points; introduces a physical consistency verification mechanism of ultrasonic propagation time difference and electromagnetic wave phase difference; and balances defect detail retention and noise suppression by dynamically adjusting the regularization coefficient, ultimately achieving high-fidelity inversion of internal defects of electrolytic capacitors to meet the accuracy and reliability requirements of industrial detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] Figure 1 This is a flow chart of the electrolytic capacitor testing method of the present invention.

[0070] Figure 2 This is a structural block diagram of the electrolytic capacitor testing system of the present invention. DETAILED DESCRIPTION

[0071] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0072] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0073] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0074] Technologies, methods, and equipment known to ordinary technicians in the relevant art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0075] The existing non-destructive detection technology for internal defects of electrolytic capacitors has the following problems:

[0076] 1. Single-modal detection methods have difficulty extracting highly discriminative features, resulting in a high misjudgment rate. This is because the mapping relationship between indirect signals and internal defects is complex, and single-modal detection methods cannot obtain sufficient information to accurately identify defects.

[0077] 2. The lack of an effective fusion mechanism for multi-source heterogeneous data (such as acoustic and electromagnetic signals) limits the accuracy of defect classification and location. This is due to the lack of a unified processing and analysis method for different types of signal data, which prevents the full utilization of the information contained in the various signals. Based on this, the present invention provides the technical solution of Example 1.

[0078] Example 1, see Figure 1 The present invention provides a flow chart of the electrolytic capacitor testing method. Figure 1 A method for testing an electrolytic capacitor shown includes the following steps:

[0079] Step 1: Synchronous acquisition of multimodal physical coupling signals: Establish a phase-locked trigger acquisition mechanism for ultrasonic and electromagnetic waves, and generate a time-frequency domain coupled waveform of the dual-modal signal through bus-level hardware synchronization control. Based on the correlation function Φ(f) between the dynamic impedance spectrum of the electrolyte and the ultrasonic velocity, an anti-interference benchmark feature library covering the process fluctuation range is constructed. Dynamic compensation of the ultrasonic propagation velocity is performed based on the ratio of the physical properties of the current electrolyte and the benchmark electrolyte.

[0080] Explanation: By building a baseline feature library that is resistant to process interference, signal changes caused by actual defects can be identified more accurately without being disturbed by normal fluctuations in the production process; it helps to reduce false alarm rates and improve detection accuracy and reliability; for example, in the production process of electrolytic capacitors, the composition and concentration of the electrolyte may have slight batch differences, which will affect the propagation characteristics of ultrasound inside the capacitor. By establishing a baseline feature library containing ultrasound propagation parameters under different electrolyte formulas and concentrations, the detection parameters can be dynamically adjusted according to the characteristics of the electrolyte of the current batch; assuming that the viscosity of a batch of electrolyte is slightly higher, resulting in a slower ultrasonic propagation speed, this change can be automatically compensated based on the data in the baseline feature library to ensure the consistency and accuracy of the detection results.

[0081] Step 2: Cross-modal modeling of physical constraints: Based on the physical equations of electrolyte evaporation and plate corrosion, the topology update rules of the graph neural network are dynamically generated, and the graph topology of the physical constraints is output. The defect-sensitive frequency band features of the cross-modal signal are extracted through a differentiable band selection network, and dynamic attention weights are generated. The graph topology of the physical constraints is used to define the spatial correlation of the defect diffusion path.

[0082] The explanation shows that by automatically selecting defect-sensitive frequency band features, signal changes related to defects are captured, reducing the interference of irrelevant information; based on dynamic attention weights, the importance of each modal signal can be adaptively adjusted according to different defect types and detection conditions, thereby improving the flexibility and accuracy of detection; assuming that when detecting electrolyte leakage of electrolytic capacitors, it is found that ultrasonic signals are most sensitive to leakage in the 20-30kHz frequency band, while electromagnetic wave signals react most obviously in the 1-2MHz frequency band; the differentiable frequency band selection network will automatically highlight the characteristics of these frequency bands while suppressing noise in other frequency bands; in addition, if it is detected that electrolyte leakage mainly affects the ultrasonic signal, dynamically increasing the weight of the ultrasonic mode and reducing the weight of the electromagnetic wave mode can ensure that the system relies more on the most relevant and reliable information to make judgments.

[0083] Step 3: Edge-adaptive defect morphology inversion: Based on the defect-sensitive frequency band characteristics, multi-scale fusion features are extracted from the time-frequency domain coupled waveform, and the multi-scale fusion features are input into the edge-preserving regularization layer. The reconstruction parameters are adaptively adjusted according to the confidence level of the defect gradient field to generate a three-dimensional defect morphology that meets the continuity constraint; based on the difference in propagation characteristics between ultrasonic and electromagnetic waves, the physical consistency of the inversion results is optimized.

[0084] Explanation: By judging the physical laws of ultrasonic wave propagation time difference and electromagnetic wave phase difference, the physical consistency of the optimized inversion results is verified; the three-dimensional defect morphology refers to the voxel grid of internal defects of the electrolytic capacitor generated by the inversion network, and each voxel value represents the probability of a defect existing at the corresponding position.

[0085] Explanation: Based on the differences in the propagation characteristics of ultrasound and electromagnetic waves, optimizing the physical consistency of the inversion results means that by considering the differences in the propagation characteristics of ultrasound and electromagnetic waves within electrolytic capacitors, information from different modes can be better interpreted and integrated, reducing errors caused by a single mode, and improving the accuracy and reliability of the reconstruction results. This is especially advantageous when dealing with complex or multiple defects. For example, when detecting corrosion on the plates of electrolytic capacitors, ultrasound is better at detecting physical structural changes caused by corrosion, while electromagnetic waves are more sensitive to changes in electrical characteristics caused by corrosion. Suppose the ultrasound data shows a 5mmx5mm structural anomaly in a certain area, while the electromagnetic wave data indicates that the electrical anomaly in the area is 6mmx6mm. Taking these two types of information into consideration, it may be concluded that the core area of the corrosion is 5mmx5mm, but its affected area extends to 6mmx6mm. This fusion method takes into account both physical structural changes and electrical characteristic changes, resulting in a more comprehensive and accurate description of the defect.

[0086] Furthermore, in order to solve the impact of strong electromagnetic interference on electromagnetic wave signal acquisition in the production line environment, an adaptive electromagnetic shielding and signal enhancement method is proposed, which includes the following steps:

[0087] Dynamic electromagnetic shielding: An adjustable multi-layer electromagnetic shielding structure is designed around the electromagnetic wave sensor. The multi-layer electromagnetic shielding structure consists of multiple independently controlled metal mesh layers. The aperture and conductivity of each mesh layer are adjusted in real time using micro-electromechanical system technology;

[0088] Real-time monitoring of electromagnetic interference intensity and spectrum distribution in the environment.

[0089] Adaptive shielding control: Based on environmental interference monitoring results, a fuzzy control algorithm is used to adjust the parameters of the multi-layer shielding structure in real time. For example, when strong interference is detected in a certain frequency band, the mesh aperture and conductivity of the specific layer are adjusted accordingly to maximize the suppression of interference in that frequency band.

[0090] Signal enhancement processing: Adaptively filter and enhance the collected electromagnetic wave signals; use wavelet transform to decompose the signals into different frequency bands, and apply Wiener filters to the signals in each frequency band to suppress noise based on the environmental interference monitoring results;

[0091] Dynamic parameter optimization: Establish a feedback loop to continuously optimize shielding structure parameters and signal processing algorithm parameters based on signal quality indicators (such as signal-to-noise ratio), and use genetic algorithms to achieve dynamic optimization of multi-layer electromagnetic shielding structure parameters to adapt to different interference environments.

[0092] The explanation states that by combining dynamically adjusted physical shielding and intelligent signal processing, electromagnetic interference in complex production environments can be suppressed and the quality of electromagnetic wave signal collection can be significantly improved. In this way, the quality of electromagnetic wave signal collection can be guaranteed in a production line environment with strong electromagnetic interference, thereby improving the accuracy and reliability of electrolytic capacitor defect detection.

[0093] What needs to be further explained in the embodiment of the present invention is that by traversing different gradient electrolyte formulas, packaging stresses, and temperature combinations, benchmark feature vectors are collected. Each set of benchmark feature vectors includes the corrected ultrasonic velocity, the correlation function Φ(f) between the dynamic impedance spectrum of the electrolyte and the attenuation of the ultrasonic velocity, the frequency domain amplitude spectrum extracted by Fourier transform of the ultrasonic amplitude sequence, and the frequency domain gradient of the electromagnetic wave impedance spectrum. The benchmark feature vectors are summarized to obtain an anti-interference benchmark feature library; based on the dynamic impedance spectrum Z of the electrolyte electrolyte (f) and ultrasonic velocity v g (ω), ω represents the angular frequency, f represents the frequency, and the correlation function Φ(f) is calculated by the following formula, which provides a physical basis for cross-modal modeling:

[0094]

[0095] The influence of process fluctuation on ultrasonic velocity is eliminated by dynamic compensation of electrolyte viscosity and dielectric constant; the current electrolyte viscosity and dielectric constant are recorded as η c , ε c , the viscosity and dielectric constant of the reference electrolyte are respectively denoted as η b , ε b The dynamic compensation formula is used to automatically correct the influence of process fluctuations (such as electrolyte batch differences) on ultrasonic propagation velocity to ensure the consistency of detection conditions. The dynamic compensation formula for ultrasonic velocity is:

[0096]

[0097] v comp Indicates the corrected ultrasonic velocity.

[0098] It is worth noting that in the embodiments of the present invention, the physical laws of electrolyte evaporation and plate corrosion are embedded in the topological structure update and feature fusion process of the graph neural network based on the anti-interference benchmark feature library, thus solving the misjudgment problem caused by the traditional multimodal fusion reliance on statistical correlation.

[0099] The operation process of the topology update rule of the graph neural network includes:

[0100] The electrolyte evaporation process is modeled as a concentration diffusion equation, and the plate corrosion process is modeled as an electrochemical reaction equation;

[0101] The physical equations of electrolyte evaporation and plate corrosion include concentration diffusion equation and electrochemical reaction equation;

[0102] The concentration diffusion equation is combined with the electrochemical reaction equation, discretized into the ratio of the concentration gradient difference between nodes to the compensation speed, and then multiplied by the correlation function to generate the edge weight calculation rule. The edge weight w ij It is obtained by multiplying the ratio of the absolute value of the electrolyte concentration gradient difference to the compensated ultrasonic group velocity by the correlation function Φ(f);

[0103] Adjust edge weights based on actual process parameters to ensure that the message transmission path of the graph neural network conforms to the physical laws of material corrosion and diffusion;

[0104] The differentiable frequency band selection network is implemented as follows:

[0105] Input signal: The ultrasonic time domain signal and the frequency domain characteristics of the electromagnetic wave complex impedance spectrum collected in step 1;

[0106] Extraction of defect-sensitive frequency bands: Calculate the product of the absolute value of the frequency domain gradient of the ultrasonic amplitude spectrum and the absolute value of the frequency domain gradient of the electromagnetic impedance spectrum to generate the energy distribution ΔE(f) of the defect-sensitive frequency band;

[0107] Dynamic attention weight generation: Normalizes the energy distribution of defect-sensitive frequency bands and generates a band attention mask through preset threshold screening to suppress noise interference in non-defect frequency bands;

[0108] The multi-scale feature fusion and node update method is as follows:

[0109] Input features: The baseline feature vector constructed in step 1 is used as the initial node feature of the graph neural network;

[0110] Feature transfer rule: The feature update of node i depends on the weighted sum of the features of neighboring node j, with the weight being the edge weight w ij The product of the frequency band value corresponding to the frequency band attention mask;

[0111] Output data: The updated node features are passed to step 3 for defect morphology inversion.

[0112] In a possible embodiment, the topology structure update rule includes an edge weight update formula, and the edge weight update formula is obtained as follows:

[0113] Nodes i and j are used to represent the indexes of the eigenvectors at different spatial locations inside the electrolytic capacitor; the edge weights are obtained using the following formula:

[0114]

[0115] Where, ΔC ij represents the electrolyte concentration gradient difference between nodes i and j;

[0116] Sum up the edge weights to get the edge weight matrix.

[0117] It needs to be further explained in the embodiment of the present invention that the multi-scale fusion feature is obtained in the following manner:

[0118] Perform wavelet transform on the ultrasonic time domain signal to extract multi-resolution features (the macroscale is used to capture overall corrosion, and the microscale is used to locate micro defects), and perform short-time Fourier transform on the electromagnetic wave complex impedance spectrum to obtain the local frequency domain energy distribution;

[0119] Based on the frequency band attention mask, the ultrasonic wavelet coefficients and the electromagnetic wave frequency domain energy are weightedly fused to enhance the signal strength of the defect-sensitive frequency band;

[0120] Through the message passing mechanism of graph neural networks, the weighted cross-modal features are hierarchically aggregated at different scales (macro, meso, and micro), and combined with the topological constraints of the electrolyte diffusion path to generate three-dimensional space-frequency-time fusion features, providing highly discriminative input for defect inversion.

[0121] It needs to be further explained in the embodiment of the present invention that the process of obtaining the three-dimensional defect morphology includes:

[0122] Input multi-scale fusion features, calculate the spatial gradient of the multi-scale fusion features to generate the defect gradient field of the defect boundary, and quantify the confidence of the defect boundary;

[0123] Based on the encoder-decoder architecture, the voxel grid corresponding to the 3D defect morphology is obtained. The encoder compresses the fused features into a low-dimensional latent vector through a 3D convolution layer to extract the macro-defect distribution. The decoder gradually upsamples through a 3D deconvolution layer and updates the voxel grid in combination with the topological structure of the physical constraint graph. The adjacency matrix of the physical constraint graph topology is used as the spatial connection template of the decoder to constrain the diffusion path of the defect morphology during the upsampling process to be consistent with the actual physical laws of electrolyte corrosion. The defect probabilities of adjacent spatial locations are forced to transition smoothly to avoid isolated noise points.

[0124] Calculate the ultrasonic propagation time difference and electromagnetic wave phase difference, and drive the correction of three-dimensional defect morphology based on the calculation results; if the ultrasonic propagation time difference exceeds the threshold, it indicates that the material density in the defect area is reduced or there is a void, and the voxel value (defect probability) at the corresponding position is increased. If the electromagnetic wave phase difference is lower than the requirement, it indicates that the electrolyte conductivity is reduced or the plate is corroded, and the boundary confidence of the corroded area is enhanced;

[0125] In one possible embodiment, the weighted error between the ultrasonic wave propagation time difference and the electromagnetic wave phase difference is recorded as the counterfactual reasoning loss function, which is transmitted to the encoder through the deconvolution layer of the decoder and fed back to the inversion network to update the three-dimensional convolution kernel weights. In one possible embodiment, the weights of the convolution kernels of the first three layers of the encoder are fixed, and only the parameters of the last two layers and the decoder are fine-tuned to prevent overfitting; the decoder parameters are adjusted so that the generated three-dimensional defect morphology gradually approaches the actual physical laws.

[0126] It needs to be further explained in the embodiments of the present invention that the regularization coefficient is dynamically adjusted based on the defect gradient field confidence, the non-equilibrium entropy production rate, or a fusion of the two, and the regularization coefficient is used to control the smoothing strength.

[0127] It is necessary to further explain in the embodiment of the present invention that the regularization coefficient is dynamically adjusted based on the confidence level of the defect gradient field:

[0128] For high confidence areas (defect gradient field confidence ≥ confidence threshold), reduce the regularization coefficient and prioritize retaining the sharp features of the defect boundary;

[0129] For low confidence areas (defect gradient field confidence < confidence threshold), the regularization coefficient is enhanced to suppress artifact interference caused by noise;

[0130] The confidence threshold is adaptively set according to the statistical distribution of the gradient field of historical defect samples; after each batch of inspection is completed, the threshold is dynamically updated according to the confidence distribution of the defect gradient field of the current sample;

[0131] In a possible embodiment, the mean μ and standard deviation σ of the defect gradient field confidence are calculated based on a historical defect sample library, and the initial threshold is set to μ-2σ; the defect gradient field confidence maps the spatial gradient of the fusion feature to a probability value through a Sigmoid function, directly reflecting the reliability of the defect boundary.

[0132] It is necessary to further explain in the embodiment of the present invention that the regularization coefficient is dynamically adjusted based on the non-equilibrium entropy production rate.

[0133] According to the electrolyte temperature field T(x, y, z) and the dissipation function σ(x, y, z), the formula Calculation of the non-equilibrium entropy production rate in the defect region The non-equilibrium entropy production rate represents the amount of energy dissipated in the defect area per unit time;

[0134] The electrolyte temperature field is measured by an infrared thermal imager or temperature sensor. The dissipation function is used to characterize the energy dissipation density of the material corrosion process; Ω represents the spatial range of the defect area;

[0135] If the non-equilibrium entropy production rate threshold is exceeded Increase the regularization coefficient to enhance the smoothing strength and suppress noise interference in high dissipation areas;

[0136] If the non-equilibrium entropy production rate threshold is not exceeded, the regularization coefficient is reduced to preserve the defect details in the low dissipation region.

[0137] It is necessary to further explain in the embodiment of the present invention that the regularization coefficient is dynamically adjusted based on the integration of the defect gradient field confidence and the non-equilibrium entropy production rate, including:

[0138] Combining the signal defect gradient field confidence with the non-equilibrium entropy production rate to overcome the limitations of a single dimension;

[0139] The regularization coefficient is denoted as λ, λ0 represents the initial regularization coefficient, and the defect gradient field confidence is denoted as T z ,

[0140]

[0141] Among them, tanh(·) is used to compress the entropy production rate ratio to the (-1, 1) interval to avoid extreme values that cause the regularization coefficient to run out of control.

[0142] Summary: Example 1 achieves high-precision electrolytic capacitor defect detection and process optimization through the combination of multimodal signal acquisition and deep learning technology. The specific working principle is as follows:

[0143] The multimodal physical coupling signal synchronous acquisition mechanism ensures the time and frequency synchronization of ultrasonic and electromagnetic wave signals, providing rich defect feature information, which solves the problem of insufficient feature extraction of single-modal detection methods; physically constrained cross-modal modeling utilizes the physical properties of the electrolyte and the plate, dynamically updates the graph neural network structure, and extracts defect-sensitive features through a differentiable frequency band selection network, realizing the effective fusion of multi-source heterogeneous data and improving the accuracy of defect classification and positioning; the edge-adaptive defect morphology inversion technology achieves high-precision three-dimensional defect morphology reconstruction through multi-scale feature fusion and edge-preserving regularization, further improving the accuracy of defect detection.

[0144] Background: The separation of the inspection process from the production process makes it impossible to dynamically optimize manufacturing parameters through real-time data feedback, resulting in a difficulty in systematically reducing the defect rate. This is because the existing technology cannot effectively feed back inspection results into the production process, making it impossible to achieve closed-loop optimization of inspection and production. Based on this, the present invention provides the solution of Example 2.

[0145] Example 2: The difference between this embodiment of the present invention and Example 1 is that it further includes a production process optimization step, including:

[0146] Step 4: Causal-driven process closed-loop optimization: Construct a multi-objective decision-making model with defect confidence and material property drift as observation states, and generate a Pareto optimal process solution set by modeling the causal relationship between process parameters and defect formation.

[0147] Furthermore, an adversarial training strategy is used to jointly optimize the contents of Example 1 and Example 2. Through adversarial training, Example 1 is forced to learn more robust and generalized features, while Example 2 is trained to cope with more diverse production situations, which can help the system better handle unseen defect types or production conditions, thereby improving the effectiveness and reliability of the entire quality control and process optimization process. In practical applications, an adversarial training framework is set up, in which Example 1 (understood as a detection model) acts as a generator for adversarial training, attempting to generate defect patterns that are difficult to be corrected by the process controller, while Example 2 (process controller) acts as a discriminator for adversarial training, matching the best process parameter adjustment scheme for each detected defect. For example, Example 1 identifies a complex electrolyte uneven distribution pattern that is difficult to correct by simple process parameter adjustments; Example 2 learns to solve the existing electrolyte uneven distribution pattern by combining adjustments to multiple parameters (such as temperature, pressure, electrolyte injection rate, etc.); through multiple rounds of training, the adversarial training framework can remain efficient and stable in the face of various complex production situations.

[0148] Explanation: Through adversarial training, Example 1 is forced to learn more robust and generalized features, and the process controller is trained to cope with more diverse production situations. This method can help the system better handle unseen defect types or production conditions, thereby improving the effectiveness and reliability of the entire quality control and process optimization process. In practical applications, by designing an adversarial training framework, attempts are made to generate defect patterns that are difficult to correct by the process controller, and the process controller, as a "discriminator", attempts to find the best process parameter adjustment solution for each detected defect. For example, Example 1 may identify a complex pattern of uneven electrolyte distribution that is difficult to correct by simple process parameter adjustments. The process controller needs to learn how to cope with this situation by combining fine adjustments of multiple parameters (such as temperature, pressure, electrolyte injection rate, etc.). Through multiple rounds of training, Example 1 becomes more sensitive to potential problem patterns, and the process controller is able to develop more refined and effective adjustment strategies, so that the entire system can remain efficient and stable in the face of various complex production situations.

[0149] Working Principle: Example 2 establishes a correlation between defect detection results and production process parameters based on causal-driven closed-loop process optimization. By generating optimal process parameters through a multi-objective decision-making model, this method achieves closed-loop optimization of the detection and production processes, contributing to a systematic reduction in defect rates. This synergistic effect with Example 1 effectively addresses existing technical challenges, enabling efficient, high-precision, and cost-effective electrolytic capacitor defect detection and process optimization.

[0150] Example 3, see Figure 2 The invention provides a structural block diagram of an electrolytic capacitor testing system, comprising:

[0151] The test signal synchronization acquisition module synchronously collects ultrasonic and electromagnetic wave signals through a phase-locked trigger acquisition mechanism, generates a time-frequency domain coupled waveform data stream, and transmits it to the dynamic compensation module;

[0152] The dynamic compensation module builds an anti-interference reference feature library covering the process fluctuation range. It dynamically compensates the ultrasonic propagation velocity based on the ratio of the current and reference electrolyte physical parameters, calculates the corrected ultrasonic velocity, and transmits it to the pattern generation engine module. The anti-interference reference feature library is then transmitted to the frequency selection network module.

[0153] The graph engine module dynamically generates topology update rules for the graph neural network based on the physical equations of electrolyte evaporation and plate corrosion, and outputs a physically constrained graph topology.

[0154] The frequency selection network module is used to calculate the cross-modal frequency domain gradient product to obtain the energy distribution of the defect-sensitive frequency band, generate the frequency band attention mask after normalization, and transmit it to the fusion engine module;

[0155] The fusion engine module extracts multi-scale fusion features from the time-frequency domain coupled waveform based on the defect-sensitive frequency band characteristics, inputs the multi-scale fusion features into the edge-preserving regularization layer, adaptively adjusts the reconstruction parameters according to the confidence level of the defect gradient field, generates a three-dimensional defect morphology that meets the continuity constraint, and transmits it to the inversion optimization module;

[0156] The inversion optimization module generates a voxel grid through an encoder-decoder architecture and applies an edge-preserving regularization layer. The counterfactual reasoning loss function is obtained by weighted summation of the ultrasonic propagation time difference and the electromagnetic wave phase difference, and is transmitted to the graph engine module to update the three-dimensional convolution kernel weights to achieve closed-loop feedback.

[0157] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for testing an electrolytic capacitor, characterized in that: The following steps are involved: Step 1: Synchronous acquisition of multi-modal physical coupling signals: Establish a phase-locked trigger acquisition mechanism for ultrasonic and electromagnetic waves to generate a time-frequency domain coupled waveform of the dual-modal signal; Based on the correlation function Φ(f) between the dynamic impedance spectrum of the electrolyte and the ultrasonic velocity, construct an anti-interference benchmark feature library covering the process fluctuation range, and dynamically compensate for the ultrasonic propagation velocity based on the ratio of the physical properties of the current electrolyte and the benchmark electrolyte; Step 2: Cross-modal modeling of physical constraints: Based on the physical equations of electrolyte evaporation and plate corrosion, the topology update rules of the graph neural network are dynamically generated, and the graph topology of the physical constraints is output. The defect-sensitive frequency band features of the cross-modal signal are extracted through a differentiable frequency band selection network, and dynamic attention weights are generated. Step 3. Edge-adaptive defect morphology inversion: Based on the defect-sensitive frequency band characteristics, multi-scale fusion features are extracted from the time-frequency domain coupled waveform, and the multi-scale fusion features are input into the edge-preserving regularization layer. The reconstruction parameters are adaptively adjusted according to the confidence level of the defect gradient field to generate a three-dimensional defect morphology; based on the difference in propagation characteristics between ultrasonic waves and electromagnetic waves, the physical consistency of the inversion results is optimized.

2. The electrolytic capacitor testing method according to claim 1, wherein: By traversing different gradient electrolyte formulas, packaging stress and temperature combinations, the benchmark feature vectors are collected. Each set of benchmark feature vectors includes the corrected ultrasonic velocity, the correlation function Φ(f) between the dynamic impedance spectrum of the electrolyte and the ultrasonic velocity attenuation, the frequency domain amplitude spectrum extracted by Fourier transform of the ultrasonic amplitude sequence, and the frequency domain gradient of the electromagnetic wave impedance spectrum. The benchmark feature vectors are summarized to obtain the anti-interference benchmark feature library; based on the dynamic impedance spectrum Z electrolyte (f) and ultrasonic velocity v g (ω), ω represents the angular frequency, f represents the frequency, and the correlation function Φ(f) is calculated by the following formula: Get the current electrolyte viscosity and dielectric constant, denoted as η c , ε c , the viscosity and dielectric constant of the reference electrolyte are respectively denoted as η b , ε b The influence of process fluctuations on ultrasonic propagation velocity is automatically corrected through the dynamic compensation formula to ensure the consistency of detection conditions. The dynamic compensation formula for ultrasonic velocity is: Among them, v comp Indicates the corrected ultrasonic velocity.

3. The electrolytic capacitor testing method according to claim 1, wherein: Based on the anti-interference benchmark feature library, the physical laws of electrolyte evaporation and plate corrosion are embedded in the topological structure update and feature fusion process of the graph neural network. The operation process of the topological structure update rule of the graph neural network includes: The electrolyte evaporation process is modeled as a concentration diffusion equation, and the plate corrosion process is modeled as an electrochemical reaction equation; The physical equations of electrolyte evaporation and plate corrosion include concentration diffusion equation and electrochemical reaction equation; The concentration diffusion equation is combined with the electrochemical reaction equation, discretized into the ratio of the concentration gradient difference between nodes to the compensation speed, and then multiplied by the correlation function to generate the edge weight calculation rule. The edge weight w ij The absolute value of the electrolyte concentration gradient difference and the ratio of the compensated ultrasonic group velocity are multiplied by the correlation function Φ(f; Adjust edge weights based on actual process parameters to ensure that the message transmission path of the graph neural network conforms to the physical laws of material corrosion and diffusion; The differentiable frequency band selection network is implemented as follows: Input signal: ultrasonic time domain signal and frequency domain characteristics of electromagnetic wave complex impedance spectrum; Extraction of defect-sensitive frequency bands: Calculate the product of the absolute value of the frequency domain gradient of the ultrasonic amplitude spectrum and the absolute value of the frequency domain gradient of the electromagnetic impedance spectrum to generate the energy distribution ΔE(f) of the defect-sensitive frequency band; Dynamic attention weight generation: Normalizes the energy distribution of defect-sensitive frequency bands and generates a band attention mask through preset threshold screening to suppress noise interference in non-defect frequency bands; The multi-scale feature fusion and node update method is as follows: Input features: The constructed baseline feature vector is used as the initial node feature of the graph neural network; Feature transfer rule: The feature update of node i depends on the weighted sum of the features of neighboring node j, with the weight being the edge weight w ij The product of the frequency band value corresponding to the frequency band attention mask; Output data: The updated node features are passed to step 3 for defect morphology inversion.

4. The electrolytic capacitor testing method according to claim 1, wherein: The multi-scale fusion feature is obtained as follows: Perform wavelet transform on ultrasonic time domain signals to extract multi-resolution features, and perform short-time Fourier transform on electromagnetic wave complex impedance spectrum to obtain local frequency domain energy distribution; Based on the frequency band attention mask, the ultrasonic wavelet coefficients and the electromagnetic wave frequency domain energy are weightedly fused to enhance the signal strength of the defect-sensitive frequency band; Through the message passing mechanism of graph neural networks, the weighted cross-modal features are hierarchically aggregated at different scales, and combined with the topological constraints of the electrolyte diffusion path to generate three-dimensional space-frequency-time fusion features, providing highly discriminative input for defect inversion.

5. The electrolytic capacitor testing method according to claim 1, wherein: The process of obtaining the three-dimensional defect morphology includes: Input multi-scale fusion features, calculate the spatial gradient of the multi-scale fusion features to generate the defect gradient field of the defect boundary, and quantify the confidence of the defect boundary; Based on the encoder-decoder architecture, the voxel grid corresponding to the 3D defect morphology is obtained. The encoder compresses the fused features into a low-dimensional latent vector through a 3D convolution layer to extract the macro-defect distribution. The decoder gradually upsamples through a 3D deconvolution layer and updates the voxel grid in combination with the topological structure of the physical constraint graph. The adjacency matrix of the physical constraint graph topology is used as the spatial connection template of the decoder to constrain the diffusion path of the defect morphology during the upsampling process to be consistent with the actual physical laws of electrolyte corrosion. The defect probabilities of adjacent spatial locations are forced to transition smoothly to avoid isolated noise points. The ultrasonic wave propagation time difference and electromagnetic wave phase difference are calculated, and the three-dimensional defect morphology is corrected based on the calculation results.

6. The electrolytic capacitor testing method according to claim 5, characterized in that: The regularization coefficient is dynamically adjusted based on the defect gradient field confidence, the non-equilibrium entropy production rate, or a fusion of the two, and the regularization coefficient is used to control the smoothing strength.

7. The electrolytic capacitor testing method according to claim 6, characterized in that: Dynamically adjust the regularization coefficient based on the defect gradient field confidence: For high confidence areas, reduce the regularization coefficient and prioritize retaining the sharp features of the defect boundary; For low confidence areas, the regularization coefficient is enhanced to suppress artifact interference caused by noise; The confidence threshold is adaptively set according to the statistical distribution of the gradient field of historical defect samples; after each batch of inspection is completed, the threshold is dynamically updated according to the confidence distribution of the defect gradient field of the current sample.

8. The electrolytic capacitor testing method according to claim 7, characterized in that: Dynamically adjust the regularization coefficient based on the non-equilibrium entropy production rate: According to the electrolyte temperature field T(x, y, z) and the dissipation function σ(x, y, z), the formula Calculation of the non-equilibrium entropy production rate in the defect region The non-equilibrium entropy production rate represents the amount of energy dissipated in the defect area per unit time; The electrolyte temperature field is measured by an infrared thermal imager or temperature sensor. The dissipation function is used to characterize the energy dissipation density of the material corrosion process; Ω represents the spatial range of the defect area; If the non-equilibrium entropy production rate threshold is exceeded Increase the regularization coefficient to enhance the smoothing strength and suppress noise interference in high dissipation areas; If the non-equilibrium entropy production rate threshold is not exceeded, the regularization coefficient is reduced to preserve the defect details in the low dissipation region.

9. The electrolytic capacitor testing method according to claim 8, characterized in that: The regularization coefficient is dynamically adjusted based on the integration of defect gradient field confidence and non-equilibrium entropy production rate, including: The regularization coefficient is denoted as λ, λ0 represents the initial regularization coefficient, and the defect gradient field confidence is denoted as T z , Among them, tanh(·) is used to compress the entropy production rate ratio to the (-1, 1) interval.

10. An electrolytic capacitor testing system, used to implement the electrolytic capacitor testing method according to claim 1, characterized in that: include: The test signal synchronization acquisition module synchronously collects ultrasonic and electromagnetic wave signals through a phase-locked trigger acquisition mechanism, generates a time-frequency domain coupled waveform data stream, and transmits it to the dynamic compensation module; The dynamic compensation module builds an anti-interference reference feature library covering the process fluctuation range, dynamically compensates for the ultrasonic propagation velocity based on the ratio of the current and reference electrolyte physical parameters, calculates the corrected ultrasonic velocity, and transmits it to the pattern generation engine module; Transmitting the anti-interference reference feature library to the frequency selection network module; The graph engine module dynamically generates topology update rules for the graph neural network based on the physical equations of electrolyte evaporation and plate corrosion, and outputs a physically constrained graph topology. The frequency selection network module is used to calculate the cross-modal frequency domain gradient product to obtain the energy distribution of the defect-sensitive frequency band, generate the frequency band attention mask after normalization, and transmit it to the fusion engine module; The fusion engine module extracts multi-scale fusion features from the time-frequency domain coupled waveform based on the defect-sensitive frequency band characteristics, inputs the multi-scale fusion features into the edge-preserving regularization layer, adaptively adjusts the reconstruction parameters according to the confidence level of the defect gradient field, generates a three-dimensional defect morphology that meets the continuity constraint, and transmits it to the inversion optimization module; The inversion optimization module generates a voxel grid through an encoder-decoder architecture and applies an edge-preserving regularization layer. The counterfactual reasoning loss function is obtained by weighted summation of the ultrasonic propagation time difference and the electromagnetic wave phase difference, and is transmitted to the graph engine module to update the three-dimensional convolution kernel weights to achieve closed-loop feedback.

Images