Non-contact low-frequency current reconstruction method and system fusing reference disturbance calibration and physical prior guidance neural network

Abstract

The invention provides a non-contact low-frequency current reconstruction method and system fusing reference disturbance calibration and a physical prior guidance neural network, and the method comprises the steps: collecting the structure parameters of a probe in actual production, and generating a probe response principal component feature vector of multi-ring probe simulation through a three-dimensional electromagnetic simulation method; collecting a preprocessed low-frequency disturbance current signal and a non-contact probe output voltage, and forming a standard sampling data set in combination with the probe response principal component feature vector; training a neural network model by using the standard sampling data set; and deploying the trained neural network model to an electromagnetic anti-interference injection test site of an actual electric vehicle module, reconstructing a time-varying current of an injection end target track in a non-contact manner, and generating a non-contact reconstruction current of the target track through derivative feedback correction. According to the invention, a set of systematic new scheme with engineering deployment capability is provided for the field of EMC testing of electric vehicles.

Description

Technical Field

[0001] This invention belongs to the field of deep learning, and in particular relates to a non-contact low-frequency current reconstruction method and system that integrates reference perturbation calibration and physical prior guided neural network. Background Technology

[0002] In electromagnetic compatibility (EMC) testing of modern electric vehicle electrical systems, immunity assessment is a core component in ensuring the stability and safety of critical control modules. Especially under simulated standard pulse interference such as ISO 7637, measuring the time-varying response of the current injection rail is crucial for determining the module's immunity and functional failure boundaries. However, such current injection signals are often located in the low-frequency range of 1Hz to 1MHz, making it difficult for conventional current probing schemes to accurately reconstruct them under non-contact conditions. Current mainstream methods mainly include deconvolution reconstruction based on a fixed frequency domain response function, or calculating the current waveform through low-frequency sampling and regression modeling. However, the former is limited by the measurement blind zone of the probe's frequency response in the low-frequency band, and the latter lacks characterization of the impact of structural differences on reconstruction performance, both failing to meet the measurement robustness and generalization capabilities required in complex injection scenarios. Furthermore, due to various non-ideal factors such as height drift, attitude deviation, and PCB trace differences in actual deployment, the electromagnetic response of non-contact probes exhibits significant fluctuations. Current modeling methods typically lack mechanisms to effectively encode response changes caused by variations in structural parameters, leading to a significant degradation in the performance of traditional models in real-world tests. More importantly, current methods generally rely on ideal signals and simulation environments, and cannot provide joint structure-signal datasets under real perturbation conditions. As a result, even if neural network methods achieve good fitting in the offline stage, they often fail in actual deployment and cannot be widely applied.

[0003] Therefore, there is an urgent need for a new path for low-frequency current reconstruction that combines real injection experiments, structural response modeling, and neural network inference capabilities to achieve accurate recovery of target rail disturbance current and evaluation of disturbance rejection performance under non-contact conditions. Summary of the Invention

[0004] The purpose of this invention is to propose a non-contact low-frequency current reconstruction method and system that integrates reference disturbance calibration and physical prior guided neural network. By introducing reference track experiments, structural principal component modeling and fusion neural network reconstruction path, it achieves high-precision reconstruction and evaluation feedback of low-frequency immunity injection current for the first time. This breaks through the current technical bottlenecks in frequency, accuracy and adaptability, and provides a new systematic solution with engineering deployment capabilities for the field of electric vehicle EMC testing.

[0005] To achieve the above objectives, a first aspect of the present invention provides a non-contact low-frequency current reconstruction method that integrates reference perturbation calibration and a physically prior guided neural network, comprising the following steps: Collect the probe structure parameters from actual production, and generate the principal component feature vector of the probe response for multi-ring probe simulation using a three-dimensional electromagnetic simulation method; The preprocessed low-frequency disturbance current signal and the output voltage of the non-contact probe are collected, and a standard sampling dataset is constructed by combining the principal component feature vector of the probe response. Train a neural network model using the aforementioned standard sampled dataset; The trained neural network model was deployed to the electromagnetic interference injection test site of the actual electric vehicle module. The time-varying current of the target track at the injection end was reconstructed in a non-contact manner, and the non-contact reconstructed current of the target track was generated by correcting the derivative feedback.

[0006] Furthermore, the method also includes: The response of the electric vehicle module to electromagnetic interference injection is evaluated based on the non-contact reconfiguration current to determine whether it meets the design immunity standard, and the evaluation results are fed back to the test system for parameter adjustment.

[0007] Furthermore, the probe structure parameters include the number of layers of the multi-ring inductive probe, the inner and outer radii of each ring, the line width, the insulation distance between lines, and the number and distribution of copper vias used to connect each layer.

[0008] Furthermore, the step of collecting actual production probe structure parameters and generating the principal component feature vector of the probe response for multi-ring probe simulation using a three-dimensional electromagnetic simulation method specifically includes: A 3D simulation model was established based on the probe structure parameters from actual production. Several working state variables are introduced into the 3D simulation model, including the vertical distance between the probe and the signal line under test, the installation tilt angle, and the change in PCB trace width. Multiple probe frequency response samples under different installation and routing conditions are generated through simulation, wherein each sample contains a complex frequency response vector. Acquired Each set of probe response samples is a complex frequency response vector with a dimension of 1024. PCA processing is performed on the complex frequency response vector to obtain the principal component feature vector of the probe response, with dimension [missing information]. .

[0009] Furthermore, the step of acquiring the preprocessed low-frequency disturbance current signal and the non-contact probe output voltage, and combining them with the probe response principal component feature vector to construct a standard sampling dataset specifically includes: An equivalent current channel is selected from the actual injection path of the target electric vehicle module to generate a low-frequency disturbance current signal; wherein, the low-frequency disturbance current signal is a standard double exponential decay current wave; The probe is fixed directly above the track, and its movement is controlled using a three-dimensional high-precision platform. The non-contact probe output voltage is recorded at a height of mm with a bandwidth of ≥1GHz and a sampling rate of ≥20GSa / s. A response offset term is introduced as a structural compensation for the reference rail to obtain the non-contact probe output voltage. The response offset term represents the derivative term of the current rising edge, which is used to reflect the asymmetric response distortion caused by rapid changes.

[0010] Furthermore, the neural network model adopts a multimodal fusion encoder-decoder system, consisting of three parts: a structural encoder, a response encoder, and a shared decoder; wherein, The steps of training the neural network model using the standard sampling dataset specifically include: The structure encoder inputs the principal component feature vector of the probe response into a three-layer feedforward network and outputs a structure embedding vector. ; The response encoder inputs the output voltage of the non-contact probe into a three-layer one-dimensional convolutional network, followed by a bidirectional GRU to extract temporal features, and outputs a structural embedding vector. ; The shared decoder embeds the structure embedding vector. Copy and embed vectors with structure By concatenating the vectors along the time dimension, a fused vector is obtained. Then, the data is input into a four-layer temporal convolutional network to decode and obtain the predicted current sequence.

[0011] Furthermore, the neural network model is also configured such that, before the shared decoder decodes, a gating factor is designed at each time step to control the degree to which structural information flows into the shared decoder for decoding; The training loss function of the neural network model includes a standard prediction error, a derivative consistency term, and a gating offset constraint. The standard prediction error is the deviation between the predicted current sequence and the low-frequency disturbance current signal. The derivative consistency term is the difference between the derivatives of the predicted current sequence and the output voltage of the contact probe, used to ensure that the trend of the predicted current change is consistent with the derivative of the voltage response. The gating offset constraint is the deviation between the gating factor and the expected average strength of the structure guidance.

[0012] Furthermore, the steps of deploying the trained neural network model to the electromagnetic interference injection test site of the actual electric vehicle module, reconstructing the time-varying current of the target track at the injection end in a non-contact manner, and correcting the generation of the non-contact reconstructed current of the target track through derivative feedback specifically include: The experimental equipment was deployed on the injection rail of the target electric vehicle module, and a multi-ring non-contact probe, consistent with the training phase of the neural network model, was arranged to extract its coupling response spectrum of 1Hz–1MHz, and the structural principal component vector was extracted by PCA processing. Acquire probe voltage response signals from the target orbit; The principal component vector of the structure and the probe voltage response signal are fed into the trained neural network model to perform current reconstruction inference, so as to obtain the predicted current of the target rail and the confidence score at each time point under the current working condition. Based on the probe voltage response signal collected from the target track, the predicted current of the target track under the previous working condition is corrected by derivative feedback to obtain the non-contact reconfiguration current of the target track.

[0013] Furthermore, the step of evaluating whether the electric vehicle module's response to electromagnetic interference injection meets the design immunity standard based on the non-contact reconfiguration current, and feeding back the evaluation results to the test system for parameter adjustment, specifically includes: The non-contact reconstructed current is segmented according to a preset interference time interval. Current intensity feature indexes are extracted for each segment, and weighted by the confidence level of the current segment to generate a confidence-weighted disturbance current intensity index for the current segment. The total deviation index is calculated based on the confidence-weighted disturbance current intensity index to determine whether a potential electromagnetic interference failure has occurred. If the total deviation index is greater than a preset threshold, an automatic marking signal is sent to the test system to prompt a retry or change of the injected waveform parameters. The total deviation index is calculated based on the set immunity expectation intensity reference and the confidence-weighted disturbance current intensity index.

[0014] A second aspect of the invention provides a non-contact low-frequency current reconstruction system that integrates reference perturbation calibration and a physically prior guided neural network, the system comprising: A non-contact low-frequency current reconstruction system integrating reference perturbation calibration and a physically prior-guided neural network, characterized in that the system comprises: The probe structure parameter acquisition unit is used to acquire the probe structure parameters in actual production and generate the principal component feature vector of the probe response for multi-ring probe simulation through a three-dimensional electromagnetic simulation method. The sample construction unit is used to acquire preprocessed low-frequency disturbance current signals and non-contact probe output voltages, and combine them with the probe response principal component feature vectors to form a standard sampling dataset; The model training unit is used to train a neural network model using the standard sampled dataset; The model deployment unit is used to deploy the trained neural network model to the electromagnetic interference injection test site of the actual electric vehicle module, reconstruct the time-varying current of the target track at the injection end in a non-contact manner, and correct the generation of the non-contact reconstructed current of the target track through derivative feedback. The system also includes: An evaluation unit is deployed to evaluate whether the electric vehicle module's response to electromagnetic interference injection meets the design immunity standard based on the non-contact reconfiguration current, and to feed back the evaluation results to the test system for parameter adjustment.

[0015] The beneficial technical effects of the present invention are at least as follows: This invention proposes a non-contact time-varying current reconstruction system for low-frequency immunity testing of electric vehicles, establishing a complete closed-loop process from probe physical modeling to on-site injection experiments, and then to a structure-aware reconstruction network and immunity assessment feedback. First, by performing full-parameter three-dimensional simulation of a multi-ring non-contact probe, the frequency domain response characteristics are constructed, and the structural response vector is obtained by combining principal component compression, thus solving the problem of difficulty in measuring the frequency domain response in the low-frequency band.

[0016] Subsequently, a reference perturbation track system was designed. Based on a standard pulse source, a known perturbation current was injected into the equivalent path of the structure. The non-contact voltage response was collected through a high-bandwidth sampling system to form a structure-voltage-current triplet training sample, thus building a modeling foundation that combines data-driven and physical constraints in a real environment.

[0017] In the neural network model, a dual-branch architecture integrating structure encoding and response timing is proposed, and a dynamic structure influence adjustment method based on a gating mechanism is introduced. By learning the contribution of structure to the response at each time step, the prediction accuracy under complex perturbation conditions is effectively improved. Simultaneously, a consistency regularization term for voltage and current derivatives is added to the loss function to constrain the physical dynamic relationship, guiding the model to learn the differential coupling characteristics between structure and response, and enhancing low-frequency reconstruction capabilities. In the experimental stage, the trained model is used to perform non-contact current measurement of the injection response into the target track, and a derivative offset feedback mechanism is used to post-correct the reconstructed current, further improving the stability of on-site reconstruction.

[0018] Finally, by combining prediction confidence and immunity reference indicators, a performance deviation index is constructed that can be used to automatically evaluate the immunity margin of modules. This supports the test system in issuing automatic feedback signals for abnormal states, forming a complete closed-loop measurement mechanism for deployment environments. While maintaining non-contact measurement, this solution, through the introduction of reference track experiments, structural principal component modeling, and a fusion neural network reconstruction path, achieves for the first time high-precision reconstruction and evaluation feedback of low-frequency immunity injection current. This breakthrough overcomes the current technological bottlenecks in frequency, accuracy, and adaptability, providing a new, systematic solution with engineering-deployable capabilities for electric vehicle EMC testing. Attached Figure Description

[0019] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the steps of the non-contact low-frequency current reconstruction method that integrates reference perturbation calibration and physical prior guided neural network of the present invention. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0022] In one or more embodiments, such as Figure 1 As shown, a non-contact low-frequency current reconstruction method integrating reference perturbation calibration and a physical prior-guided neural network is disclosed. The method includes the following steps: S1. Collect the probe structure parameters from actual production, and generate the principal component feature vector of the probe response for multi-ring probe simulation using a three-dimensional electromagnetic simulation method.

[0023] Specifically, this step is used to construct a probe physical response model for low-frequency non-contact current measurement. The core objective is to obtain high-precision response features in the low-frequency range (1Hz to 1MHz) as the physical prior input for the subsequent neural network reconstruction model. To achieve this goal, we employ a three-dimensional electromagnetic simulation method combined with principal component analysis (PCA) to reduce the dimensionality of the response data, thereby extracting low-dimensional physical feature vectors representing the probe's electromagnetic behavior.

[0024] First, an accurate 3D simulation model is established based on the actual probe structure parameters from production. These structural parameters include: the number of layers in the multi-ring inductor probe (e.g., a three-layer coaxial coupling loop), the inner and outer radii of each ring (e.g., 1.0mm / 1.5mm), the line width (0.2mm), the inter-line insulation distance (0.3mm), and the number and distribution of copper vias connecting the layers. This data comes from the probe PCB's Gerber file and process specifications and can be directly imported into CAD tools (such as SolidWorks) for 3D geometric modeling. Next, the geometric model is imported into electromagnetic simulation software (such as ANSYS HFSS or CST Studio Suite), and material parameters are set (the dielectric constant of the FR4 substrate is approximately 4.7, and the loss tangent is approximately 0.02; the conductivity of the copper conductor is...). The simulation environment includes S / m (S / m), open boundary conditions, grounding layer, port impedance matching structure, etc.

[0025] Furthermore, to make the simulation results more closely resemble real-world applications, we introduced several operating state variables for modeling: including the vertical distance between the probe and the signal line under test (typically set to 0.1 mm), the mounting tilt angle (±5°), and the variation in PCB trace width (0.8–1.5 mm). Using the parameter scanning function, the simulation generated multiple probe frequency response sample data under different mounting and routing conditions. Each sample contains a complex frequency response vector. Its definition is: ; in, Indicates the probe at frequency Complex response coefficients under these conditions; This represents the complex voltage spectrum at the probe output port at this frequency. This is the complex current spectrum of the input microstrip signal line. These data points cover the entire frequency band from 1 Hz to 1 GHz to ensure that the low-frequency characteristics are fully captured. For each operating condition sample, we use equal frequency points to uniformly sample its response, often using a logarithmic uniform sampling method, for example, selecting 1024 frequency points in the range of 1 Hz to 1 GHz.

[0026] After the simulation was completed, we obtained The sample data consists of groups of probe response samples, each group being a complex frequency response vector of dimension 1024. Since subsequent AI models are not well-suited for directly processing high-dimensional complex vectors, and the probe responses to different frequencies often exhibit highly redundant and coordinated variations, PCA is used to reduce the dimensionality of the response data. The PCA processing flow is as follows: First, for all... The samples are normalized and centered, then the covariance matrix is ​​calculated, and eigenvalue decomposition is performed. The top eigenvalues ​​are selected. Principal component directions (usually) ), forming the transformation matrix Finally, each response vector is projected onto a low-dimensional principal component space to obtain the principal component feature vectors. The calculation formula is as follows: ; in, It is the final output feature vector. The conjugate transpose of the principal component direction matrix. This is the current probe response sample. It is the mean of the responses of all samples.

[0027] For a practical example, simulations of a set of actual probes at different positions and angles generated 40 sets of response data with a sampling frequency of 1024. After PCA processing, the first 6 principal components were extracted, which cumulatively explained more than 96% of the variance. These principal component vectors not only captured the response differences of the probes in the low-frequency range, but also provided physical constraints for the subsequent neural network model, effectively improving the reconstruction accuracy and generalization ability.

[0028] The final output is the probe response principal component feature vector. Its dimensions are These can be directly used as important physical prior features for perturbation trajectory test modeling and neural network training input in subsequent steps.

[0029] The innovation of this step lies primarily in its adoption of a fusion strategy combining full-parameter simulation, multi-condition response modeling, and PCA feature compression. This approach is the first to perform physically interpretable modeling of the response characteristics of non-contact probes in the low-frequency band, without relying on low-frequency response functions that are difficult to obtain through actual measurements. This ensures the integrity of the frequency domain characteristics and optimizes the input quality of subsequent algorithm models. This method significantly improves the accuracy and stability of reconstructed current in low-frequency immunity testing, and is particularly suitable for the 1Hz–1MHz frequency range, which is difficult to measure directly.

[0030] S2. Acquire the pre-processed low-frequency disturbance current signal and the output voltage of the non-contact probe, and combine them with the principal component feature vector of the probe response to form a standard sampling dataset.

[0031] Specifically, in completing the multi-ring probe simulation in the first step, the probe response principal component eigenvector ( Following this, this step focuses on how to construct high-fidelity, physically interpretable training samples, with the core being the collection of a set of physically constrained data pairs. And ensure that they are consistent with the principal component feature vectors of the probe response. This consistency ensures a reproducible and generalizable training foundation for building a neural network model that integrates physical structure information. Unlike traditional methods that rely solely on measured or simulated results, we propose an integrated acquisition method combining a reference perturbation track, waveform injection, and structural response. We also introduce an original equivalent response offset term for the reference track using an enhanced calibration mechanism to correct response deviations caused by minor assembly errors or environmental factors, thereby improving sample consistency.

[0032] First, referencing the disturbance track design phase, an equivalent current path is selected from the actual injection path of the target electric vehicle PCB module. In a laboratory environment, the disturbance path is replicated with the same electrical length, material, and wiring method. The injection interface (SMA), ground plane, transmission line, and terminating load are pre-configured to provide injection control capabilities. We use ISO7637 standard waveform sources (such as Pulse1 and Pulse3A) to generate representative low-frequency disturbance current signals. Its form is a standard double-exponential decaying current wave: ; in, It is the duration The unit window function within, , , , These represent the pulse amplitude and time decay parameters, respectively, both of which can be adjusted via the injector to simulate the characteristics of different interference sources. This waveform is input to the reference disturbance track via a coaxial connector and grounded by a matched terminal to ensure closed-loop current conduction.

[0033] Next, the probe is fixed directly above the track, and its movement is controlled using a three-dimensional high-precision platform. The voltage remains stable at a height of mm. The acquisition device (such as a Keysight DSOS054A oscilloscope) records the non-contact probe output voltage at a bandwidth of ≥1 GHz and a sampling rate of ≥20 GSa / s. Simultaneously record the current source output at the injection port. (This can be obtained through bypass current measurement or conversion from standard resistor voltage). To address waveform shifts in measured data caused by noise and contact instability, we introduce a response offset term into the dataset. As structural compensation for the reference rail: ; in, This is an adjustable hyperparameter used to compensate for the phase shift caused by the nonlinear delay in probe flux capture. The derivative term representing the rising edge of the current reflects the asymmetric response distortion caused by rapid changes. This term cannot be directly corrected by filtering at the hardware level. Therefore, we perform numerical estimation and inverse compensation during data preprocessing, which significantly improves the timing consistency and amplitude relative error stability of the dataset.

[0034] Ultimately, the samples form the following set of triples: The probe response principal component feature vector obtained in the previous step (e.g., obtained through PCA decomposition) Principal components, , which represents the main response structure of the probe, is obtained by principal component extraction after modeling the S-parameters of the probe in different frequency bands in HFSS or CST simulation software; : The measured non-contact response signal after offset term correction; : The known current waveform generated by the standard signal source in the reference disturbance track.

[0035] Among them, the entire triplet It provides high-quality, physically consistent samples corresponding to multiple structures, multiple responses, and multiple waveforms for subsequent neural network model training, and is the core support for realizing joint structure-signal modeling.

[0036] The output of this step is the collection of the structured datasets mentioned above. ,in The number of experimental combinations covers different injection waveforms, probe structure parameters, and attitude combinations.

[0037] S3. Train the neural network model using the standard sampling dataset.

[0038] Specifically, this step aims to build upon the standard sampling dataset constructed in step two. A neural network model that integrates the principal components of the probe structure and the voltage response of the non-contact probe is trained to achieve the control of perturbation injection current. The model must not only be able to learn complex nonlinear mapping relationships, but also maintain physical consistency and have broad adaptability to different structure probes and injection waveforms.

[0039] Input data includes: principal component vectors of probe structure response The main structural information reflecting the probe's magnetic coupling capability at different frequencies is obtained from the first step of simulation dimensionality reduction; typically, the dimension is 6 to 10. The probe output voltage after offset correction... Obtained in the second step, this reflects the probe response sequence after standard perturbation injection, with a uniform length of [length missing]. (For example, 2000 points); Furthermore, the output target is the reference injection current. ,and Time alignment, also for length .

[0040] The neural network model adopts a multimodal fusion encoder-decoder architecture, consisting of three parts: a structural encoder, a response encoder, and a shared decoder. Structural encoder: Input a three-layer feedforward network (each layer has 64, 64, and 128 nodes, with ReLU activation function), output a structure embedding vector. .

[0041] Response encoder: The input is a three-layer one-dimensional convolutional network (Conv1D, kernel width 5, number of channels 32→64→128, stride 2), followed by a bidirectional GRU to extract temporal features, and the output is... .

[0042] Decoder: converts the structure vector Copy and By concatenating the vectors along the time dimension, a fused vector is obtained. Then, the data is input into a four-layer temporal convolutional network (TCN) to decode and obtain the predicted current sequence. .

[0043] Furthermore, one of the model's innovations lies in its design of a structure-response cross-gating mechanism, which dynamically adjusts the influence weights of the structure embedded in time series modeling. That is, at each time step... Introducing gating factors The extent to which structural information flows into the decoder is defined as follows: ; ; in, This represents the Sigmoid function. and For trainable parameters, This is the temporal convolution decoding function. This indicates the strength of the influence of the principal structural components on the decoder at that moment. This gating mechanism can automatically adjust the proportion of structure sensing introduced based on the dynamic characteristics of the probe response at each moment, avoiding static overfitting of structural information and improving the ability to capture local disturbance events.

[0044] Furthermore, regarding the training objective function, a combined loss function with a physical consistency regularization term is adopted: ; Among them: the first item The standard prediction error; the second term The derivative consistency term introduced in the second step ensures that the trend of the predicted current change is consistent with the derivative of the voltage response, effectively capturing nonlinear abrupt changes such as the rising and falling edges of the current; the third term is the gated offset constraint. , The expected average strength (e.g., 0.5) is guided by the structure to prevent the network from completely ignoring or over-relying on structural features under certain perturbations, thus maintaining generalization balance.

[0045] The canonical design reflects two special characteristics of the patent scenario: First, the signal is very weak in the 1Hz–1MHz range, and the structural coupling characteristics are more important than in the high-frequency range, so it is necessary to dynamically enhance the structural information; Second, there are amplitude drift and nonlinear phase mismatch in actual measurements, and derivative consistency canonicalization can effectively suppress the propagation of such errors.

[0046] Finally, this step outputs the trained structure-aware neural network model. It can accept the principal components of the probe structure. Corresponding non-contact response The combined input outputs the corresponding current reconstructed waveform. It has high accuracy and physical consistency, and can be directly deployed in subsequent real-world scenarios for electromagnetic interference response diagnosis of unknown modules.

[0047] S4. Deploy the trained neural network model to the electromagnetic interference injection test site of the actual electric vehicle module, reconstruct the time-varying current of the target track at the injection end in a non-contact manner, and generate the non-contact reconstructed current of the target track through derivative feedback correction.

[0048] The purpose of this step is to complete the structure-aware neural network model. After training, it was deployed to the electromagnetic interference immunity injection test site of an actual electric vehicle module, and the time-varying current of the target track at the injection end was reconstructed in a non-contact manner. This achieves the ultimate goal of this patent—to accurately estimate the anti-interference injection current curve without the need for a contact current measurement device.

[0049] The input for this step is the trained neural network model, which is the output of the third step. The model accepts two input variables: the probe structure principal component. probe voltage response signal acquired from the target orbit The output is the corresponding current prediction sequence. Therefore, to ensure the continuity of the process, this step requires first obtaining two types of data that match the model input.

[0050] First, a multi-ring non-contact probe, identical to that used in the training phase, is deployed on the injection rail of the target electric vehicle module using experimental equipment. The probe structure (number of coil turns, arrangement, size) must be consistent with the data from previous simulations and training, or a new structure can be modeled in the frequency domain using simulation tools (such as HFSS) to extract its coupling response spectrum from 1Hz to 1MHz. The structural principal component vectors are extracted using the same principal component analysis method (such as PCA) as in the training phase. For example, if 10 principal components are selected, the final input vector will be... .

[0051] Secondly, voltage response Acquired by a high-speed oscilloscope (bandwidth ≥ 1 GHz, sampling rate ≥ 20 MSa / s), the time window should cover the entire immunity injection pulse process (e.g., the typical ISO7637 Pulse1 length is 5 ms). The signal needs to be preprocessed by a 1 Hz–100 MHz bandpass filter to suppress DC and high-frequency interference, and the amplitude should be normalized and the length unified (e.g., unified to 2048 points) to ensure consistency with the model input format.

[0052] Then, the two input variables are fed into the trained neural network model. The current reconstruction inference is performed, and the expression is as follows: ; in, The fusion structure-response neural network trained in step three consists of a structure encoder, a temporal encoder, and a temporal decoder, and its output is... This refers to the predicted current of the target rail under the current operating conditions. (Model) Simultaneously output the confidence score at each time point. For example, by adding a Sigmoid activation function to the last layer and outputting a confidence map, the relative reliability of the current prediction value at each point can be reflected.

[0053] In real-world applications, there may be slight offsets or rotational errors in the position of the probe and the track, leading to... Since the response distribution deviates from that in the training set, we introduce a practical calibration enhancement mechanism during the inference phase: dynamic correction based on derivative consistency offset, to improve the matching degree of the reconstruction result to the actual perturbation response. ; in, The proportionality coefficient is selected based on experience (e.g., between 0.05 and 0.2). This item actually serves as a post-feedback mechanism. When the predicted current trend deviates too much from the voltage response derivative, it automatically guides the model output to return to the input dynamic characteristics, thus playing a role in physical correction.

[0054] Furthermore, the method also includes: S5. Evaluate whether the response of the electric vehicle module to electromagnetic interference injection meets the design immunity standard based on the non-contact reconfiguration current, and feed back the evaluation results to the test system for parameter adjustment.

[0055] This step is optional, and the goal is to complete the non-contact reconfiguration of the target rail current in the fourth step. After non-contact reconstruction, the results are used to evaluate whether the electric vehicle module's response to electromagnetic interference injection meets the design immunity standard, and the evaluation results are fed back to the test system for parameter adjustment.

[0056] The input for this step comes from the two output variables of step four: the non-contact reconfiguration current. The non-contact reconfiguration current is corrected by derivative feedback; confidence curve Reflecting the non-contact reconfiguration current Prediction reliability at each point in time.

[0057] Furthermore, to achieve immunity assessment, we first tested the non-contact reconfiguration current. Interference time intervals as defined by injection specifications (such as ISO 7637-2) The pulse is segmented, with typical segments including the pulse leading edge (0–200µs), the steady-state pulse (200–1000µs), and the pulse decay pulse (>1000µs). Current intensity characteristics are extracted from each segment and weighted according to confidence levels, calculated as follows: ; In this formula, Indicates the first The confidence-weighted disturbance current intensity index of the section. and All results come directly from the output of the previous step, and no modification to the model structure is required.

[0058] To determine whether a potential electromagnetic interference failure has occurred, we introduce an anomaly assessment formula for relative deviation, defined as follows: ; in: This represents the reference for the expected immunity strength set during the design phase of the module (derived from historical sample statistics or simulation modeling). Set the segment weight (e.g., set the leading segment to 0.5); To prevent the stable terms from being divided by zero (such as...) ); For the total deviation index, when A value >1.0 indicates a risk of disturbance rejection failure.

[0059] To improve the engineering reproducibility of the solution, we provide a sample process for a typical test application: Apply a typical interference waveform (such as ISO7637Pulse1) to the CAN communication port of the target module controller. Acquire the response of the injected orbital probe Combining structural principal components Input neural network get ; Derivative feedback correction is obtained ; use Calculated using the method described above and Thus, the disturbance rejection deviation index is obtained; like (If the threshold is set to 1.0), an automatic marking signal will be sent to the test system to prompt a retry or change the injected waveform parameters.

[0060] In one or more embodiments, another embodiment of the present invention provides a non-contact low-frequency current reconstruction system that integrates reference perturbation calibration and a physically prior guided neural network, the system comprising: A non-contact low-frequency current reconstruction system integrating reference perturbation calibration and a physically prior-guided neural network, characterized in that the system comprises: The probe structure parameter acquisition unit is used to acquire the probe structure parameters in actual production and generate the principal component feature vector of the probe response for multi-ring probe simulation through a three-dimensional electromagnetic simulation method. The sample construction unit is used to acquire preprocessed low-frequency disturbance current signals and non-contact probe output voltages, and combine them with the probe response principal component feature vectors to form a standard sampling dataset; The model training unit is used to train a neural network model using the standard sampled dataset; The model deployment unit is used to deploy the trained neural network model to the electromagnetic interference injection test site of the actual electric vehicle module, reconstruct the time-varying current of the target track at the injection end in a non-contact manner, and correct the generation of the non-contact reconstructed current of the target track through derivative feedback. The system also includes: An evaluation unit is deployed to evaluate whether the electric vehicle module's response to electromagnetic interference injection meets the design immunity standard based on the non-contact reconfiguration current, and to feed back the evaluation results to the test system for parameter adjustment.

[0061] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0062] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0063] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0064] The various illustrative logic blocks, modules, and circuits described in conjunction with the embodiments disclosed herein can be implemented or performed using a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternatives, it may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more microprocessors cooperating with a DSP core, or any other such configuration.

[0065] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of both. The software module may reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor such that the processor can read and write information to / from the storage medium. In an alternative, the storage medium may be integrated into the processor. The processor and storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and storage medium may reside as discrete components in the user terminal.

[0066] In one or more exemplary embodiments, the described functionality may be implemented in hardware, software, firmware, or any combination thereof. If implemented in software as a computer program product, the functionality may be stored or transmitted as one or more instructions or code on or through a computer-readable medium. A computer-readable medium includes both computer storage media and communication media, encompassing any medium that facilitates the transfer of a computer program from one location to another. A storage medium may be any available medium accessible to a computer. By way of example and not limitation, such a computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disc storage, disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and is accessible to a computer. Any connection is also legitimately referred to as a computer-readable medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio, and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave are included in the definition of a medium. As used in this article, disk and disc include compact discs (CDs), laser discs, optical discs, digital multi-purpose discs (DVDs), floppy disks, and Blu-ray discs. Disks typically reproduce data magnetically, while discs reproduce data optically using lasers. Combinations of these should also be included within the scope of computer-readable media.

[0067] The prior description of this disclosure is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein may be applied to other variations without departing from the spirit or scope of this disclosure. Therefore, this disclosure is not intended to be limited to the examples and designs described herein, but should be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A non-contact low-frequency current reconstruction method integrating reference perturbation calibration and physically prior-guided neural networks, characterized in that, Includes the following steps: Collect the probe structure parameters from actual production, and generate the principal component feature vector of the probe response for multi-ring probe simulation using a three-dimensional electromagnetic simulation method; The preprocessed low-frequency disturbance current signal and the output voltage of the non-contact probe are collected, and a standard sampling dataset is constructed by combining the principal component feature vector of the probe response. Train a neural network model using the aforementioned standard sampled dataset; The trained neural network model was deployed to the electromagnetic interference injection test site of the actual electric vehicle module. The time-varying current of the target track at the injection end was reconstructed in a non-contact manner, and the non-contact reconstructed current of the target track was generated by correcting the derivative feedback.

2. The non-contact low-frequency current reconstruction method according to claim 1, which integrates reference perturbation calibration and physical prior-guided neural network, is characterized in that... The method further includes: The response of the electric vehicle module to electromagnetic interference injection is evaluated based on the non-contact reconfiguration current to determine whether it meets the design immunity standard, and the evaluation results are fed back to the test system for parameter adjustment.

3. The non-contact low-frequency current reconstruction method according to claim 1, which integrates reference perturbation calibration and physical prior-guided neural network, is characterized in that... The probe structure parameters include the number of layers of the multi-ring inductor probe, the inner and outer radii of each ring, the line width, the insulation distance between lines, and the number and distribution of copper vias used to connect each layer.

4. The non-contact low-frequency current reconstruction method according to claim 1, which integrates reference perturbation calibration and physical prior-guided neural network, is characterized in that... The steps of collecting actual production probe structure parameters and generating the principal component feature vector of the probe response for multi-ring probe simulation using a three-dimensional electromagnetic simulation method specifically include: A 3D simulation model was established based on the probe structure parameters from actual production. Several working state variables are introduced into the 3D simulation model, including the vertical distance between the probe and the signal line under test, the installation tilt angle, and the change in PCB trace width. Multiple probe frequency response samples under different installation and routing conditions are generated through simulation, wherein each sample contains a complex frequency response vector. Acquired Each set of probe response samples is a complex frequency response vector with a dimension of 1024. PCA processing is performed on the complex frequency response vector to obtain the principal component feature vector of the probe response, with dimension [missing information]. .

5. The non-contact low-frequency current reconstruction method according to claim 1, which integrates reference perturbation calibration and physical prior-guided neural network, is characterized in that... The steps of acquiring the preprocessed low-frequency disturbance current signal and the non-contact probe output voltage, and combining them with the probe response principal component feature vector to construct a standard sampling dataset specifically include: An equivalent current channel is selected from the actual injection path of the target electric vehicle module to generate a low-frequency disturbance current signal; wherein, the low-frequency disturbance current signal is a standard double exponential decay current wave; The probe is fixed directly above the track, and its movement is controlled using a three-dimensional high-precision platform. The non-contact probe output voltage is recorded at a height of mm with a bandwidth of ≥1GHz and a sampling rate of ≥20GSa / s. A response offset term is introduced as a structural compensation for the reference rail to obtain the non-contact probe output voltage. The response offset term represents the derivative term of the current rising edge, which is used to reflect the asymmetric response distortion caused by rapid changes.

6. The non-contact low-frequency current reconstruction method according to claim 1, which integrates reference perturbation calibration and physical prior-guided neural network, is characterized in that... The neural network model employs a multimodal fusion encoder-decoder architecture, consisting of three parts: a structural encoder, a response encoder, and a shared decoder. The steps of training the neural network model using the standard sampling dataset specifically include: The structure encoder inputs the principal component feature vector of the probe response into a three-layer feedforward network and outputs a structure embedding vector. ; The response encoder inputs the output voltage of the non-contact probe into a three-layer one-dimensional convolutional network, followed by a bidirectional GRU to extract temporal features, and outputs a structural embedding vector. ; The shared decoder embeds the structure embedding vector. Copy and embed vectors with structure By concatenating the vectors along the time dimension, a fused vector is obtained. Then, the data is input into a four-layer temporal convolutional network to decode and obtain the predicted current sequence.

7. The non-contact low-frequency current reconstruction method according to claim 6, which integrates reference perturbation calibration and physical prior-guided neural network, is characterized in that... The neural network model is further configured such that, before the shared decoder decodes, a gating factor is designed at each time step to control the degree to which structural information flows into the shared decoder for decoding; The training loss function of the neural network model includes a standard prediction error, a derivative consistency term, and a gating offset constraint. The standard prediction error is the deviation between the predicted current sequence and the low-frequency disturbance current signal. The derivative consistency term is the difference between the derivatives of the predicted current sequence and the output voltage of the contact probe, used to ensure that the trend of the predicted current change is consistent with the derivative of the voltage response. The gating offset constraint is the deviation between the gating factor and the expected average strength of the structure guidance.

8. The non-contact low-frequency current reconstruction method according to claim 1, which integrates reference perturbation calibration and physical prior-guided neural network, is characterized in that... The steps of deploying the trained neural network model to the electromagnetic interference injection test site of the actual electric vehicle module, reconstructing the time-varying current of the target track at the injection end in a non-contact manner, and correcting the generated non-contact reconstructed current of the target track through derivative feedback specifically include: The experimental equipment was deployed on the injection rail of the target electric vehicle module, and a multi-ring non-contact probe, consistent with the training phase of the neural network model, was arranged to extract its coupling response spectrum of 1Hz–1MHz, and the structural principal component vector was extracted by PCA processing. Acquire probe voltage response signals from the target orbit; The principal component vector of the structure and the probe voltage response signal are fed into the trained neural network model to perform current reconstruction inference, so as to obtain the predicted current of the target rail and the confidence score at each time point under the current working condition. Based on the probe voltage response signal collected from the target track, the predicted current of the target track under the previous working condition is corrected by derivative feedback to obtain the non-contact reconfiguration current of the target track.

9. The non-contact low-frequency current reconstruction method according to claim 1, which integrates reference perturbation calibration and physical prior-guided neural network, is characterized in that... The steps of evaluating whether the electric vehicle module's response to electromagnetic interference injection meets the design immunity standard based on the non-contact reconfiguration current, and feeding back the evaluation results to the test system for parameter adjustment, specifically include: The non-contact reconstructed current is segmented according to a preset interference time interval. Current intensity feature indexes are extracted for each segment, and weighted by the confidence level of the current segment to generate a confidence-weighted disturbance current intensity index for the current segment. The total deviation index is calculated based on the confidence-weighted disturbance current intensity index to determine whether a potential electromagnetic interference failure has occurred. If the total deviation index is greater than a preset threshold, an automatic marking signal is sent to the test system to prompt a retry or change of the injected waveform parameters. The total deviation index is calculated based on the set immunity expectation intensity reference and the confidence-weighted disturbance current intensity index.

10. A non-contact low-frequency current reconstruction system integrating reference perturbation calibration and a physical prior-guided neural network, characterized in that, The system includes: The probe structure parameter acquisition unit is used to acquire the probe structure parameters in actual production and generate the principal component feature vector of the probe response for multi-ring probe simulation through a three-dimensional electromagnetic simulation method. The sample construction unit is used to acquire preprocessed low-frequency disturbance current signals and non-contact probe output voltages, and combine them with the probe response principal component feature vectors to form a standard sampling dataset; The model training unit is used to train a neural network model using the standard sampled dataset; The model deployment unit is used to deploy the trained neural network model to the electromagnetic interference injection test site of the actual electric vehicle module, reconstruct the time-varying current of the target track at the injection end in a non-contact manner, and generate the non-contact reconstructed current of the target track through derivative feedback correction.

Images