Non-contact current sensing non-linear distortion correction method, system, device and medium

Abstract

The invention discloses a non-contact current sensing non-linear distortion correction method, system and device and a medium, and the method comprises the steps: placing an input probe at a current input end of a power transmission line for collection to obtain ideal reference data, and placing an output probe at a current output end of the power transmission line for collection to obtain distortion data; preprocessing the ideal reference data and the distortion data to obtain a training data set, training a signal restoration module through the training data set, and deploying the trained signal restoration module in a production environment for application; performing current detection on the to-be-detected electric wire through an output probe to obtain a distortion signal, and preprocessing the distortion signal; and inputting the pre-processed distorted signal to a signal restoration module to obtain corrected current signal data. The method solves the problems that an existing correction method is difficult to process complex nonlinear mapping, neglects signal time sequence dependency and cannot effectively separate linear and nonlinear components, so that correction precision is insufficient and adaptability is poor.

Description

Technical Field

[0001] This invention relates to the field of current detection technology, and in particular to a non-contact current sensing nonlinear distortion correction method, system, device and medium. Background Technology

[0002] With the continuous improvement of the intelligence level of power systems, non-contact current sensing technology has been widely used in power monitoring, fault diagnosis, and energy management due to its advantages such as no need for circuit breaking, convenient installation, and low maintenance costs. However, during non-contact current measurement, the acquired current signal often suffers from significant nonlinear distortion due to factors such as magnetic core saturation, temperature drift, and external electromagnetic interference. These distortions seriously affect the measurement accuracy and system reliability, becoming a key bottleneck restricting the further promotion and application of non-contact current sensing technology.

[0003] Existing distortion correction techniques primarily rely on physical model-based compensation algorithms or traditional signal processing methods. While these methods can mitigate the impact of linear errors to some extent, they often perform poorly when faced with complex and variable nonlinear distortions. Due to the coupling of multiple distortion factors, such as amplitude nonlinear compression introduced by magnetic core saturation, gain fluctuations caused by temperature drift, and high-frequency noise superimposed by external interference, traditional methods struggle to accurately establish a complete nonlinear mapping relationship, thus limiting correction effectiveness. Furthermore, these methods lack in-depth analysis of the signal's temporal dependencies when processing distorted signals, failing to fully utilize historical information to assist in current correction decisions, thereby affecting the ability to identify nonlinear distortion patterns with long-term correlations. In addition, existing technologies often employ a uniform processing strategy to correct the entire signal, failing to specifically separate and process linear background components and nonlinear distortion components. This coarse-grained processing not only increases computational complexity but also limits the accuracy and fidelity of the final reconstructed signal, making it difficult to meet the practical needs of high-precision current measurement scenarios. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a non-contact current sensing nonlinear distortion correction method, system, device, and medium to solve the problems of existing correction methods being unable to handle complex nonlinear mappings, ignoring signal timing dependencies, and failing to effectively separate linear and nonlinear components, resulting in insufficient correction accuracy and poor adaptability.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a non-contact current sensing nonlinear distortion correction method, comprising: placing an input probe at the current input end of a power transmission line to acquire ideal reference data, and simultaneously placing an output probe at the current output end of the power transmission line to acquire distortion data; preprocessing the ideal reference data and distortion data to obtain a training dataset, training a signal repair module using the training dataset, and deploying the trained signal repair module in a production environment for application; using the output probe to detect the current of the power line under test to obtain a distortion signal, and preprocessing the distortion signal; inputting the preprocessed distortion signal to the signal repair module to obtain corrected current signal data.

[0007] As a preferred embodiment of the nonlinear distortion correction method for non-contact current sensing described in this invention, the steps of training a signal restoration module using the training dataset and deploying the trained signal restoration module in a production environment include: inputting the distortion data in the training dataset into an encoding network, and performing layer-by-layer transformation and feature extraction on the distortion data through the encoding network to obtain a feature vector; inputting the feature vector into a decoding network, and performing layer-by-layer inverse transformation and signal reconstruction on the feature vector through the decoding network to obtain a time-domain waveform sequence; calculating the mean square error between the time-domain waveform sequence and the corresponding ideal reference data in the training dataset; optimizing the encoding network and decoding network based on the mean square error and an error backpropagation algorithm and iterating until the mean square error converges to a preset threshold range, thus completing the training of the signal restoration module; and deploying the trained signal restoration module to an actual measurement device for distortion correction.

[0008] The beneficial effects of this preferred technical solution are as follows: By performing layer-by-layer transformation and feature extraction on distorted data through an encoding network, key features in the distorted data can be obtained to form a feature vector. The decoding network then performs layer-by-layer inverse transformation and signal reconstruction based on this feature vector to obtain a time-domain waveform sequence. The mean square error of the obtained time-domain waveform sequence and the ideal reference data is calculated, and the encoding and decoding networks are optimized and iterated using an error backpropagation algorithm until the mean square error converges to a preset threshold range, thus completing the signal repair module training. After the trained module is deployed to the actual measurement device, it can directly process the distorted data generated during non-contact current sensing. Through the synergistic effect of feature extraction by the encoding network and signal reconstruction by the decoding network, the deviation between the output time-domain waveform sequence and the ideal reference data is kept controllable, achieving targeted correction of nonlinear distortion in non-contact current sensing, ensuring the consistency between the output signal of the actual measurement device and the ideal reference data, effectively repairing the distorted data, and meeting the signal correction requirements in non-contact current sensing scenarios.

[0009] As a preferred embodiment of the nonlinear distortion correction method for non-contact current sensing described in this invention, the step of performing layer-by-layer transformation and feature extraction on the distorted data through the coding network includes: decomposing the distorted data into linear components and nonlinear residual components, extracting local features of different granularities from the residual components through multi-scale convolution and fusing them to obtain fused features; and performing temporal memory integration on the fused features to output a feature vector.

[0010] The beneficial effects of this preferred technical solution are as follows: decomposing distorted data into linear components and nonlinear residual components enables targeted processing of distorted components with different properties. Multi-scale convolution applied to the nonlinear residual components can extract local features of different granularities, comprehensively capturing the distortion-related information hidden in the residual components. Then, feature fusion integrates local features of different granularities into a unified fused feature, avoiding the limitations of single-granularity features. Subsequent temporal memory integration processes the fused features, preserving the temporal correlation between local features. This ensures that the output feature vector contains multi-dimensional detailed information of the nonlinear residual components while also taking into account the temporal characteristics of the data, accurately corresponding to the original signal features of the distorted data. This provides precise feature support for the layer-by-layer inverse transformation and signal reconstruction of the decoding network, directly assisting in the effective correction of nonlinear distortion in non-contact current sensing, ensuring that the deviation between the time-domain waveform sequence and the ideal reference data is controlled within a reasonable range, and achieving accurate repair of distorted data.

[0011] As a preferred embodiment of the nonlinear distortion correction method for non-contact current sensing described in this invention, the step of decomposing the distortion data into linear components and nonlinear residual components includes: constructing a linear dynamic system composed of K parallel first-order state units, inputting the distortion data into the linear dynamic system, using the calculated linear response as the linear component; and subtracting the linear component from the distortion data to obtain the nonlinear residual component.

[0012] The beneficial effects of this preferred technical solution are as follows: It can accurately capture the linear dynamic characteristics in distorted data through the synergistic effect of multiple first-order state units. The calculated linear response can truly reflect the linear components in the data, avoiding mutual interference between linear and nonlinear components. By directly subtracting the linear component from the distorted data, the nonlinear residual component can be clearly separated, allowing the two different types of components to be presented independently. This ensures the integrity of the linear component while highlighting the unique characteristics of the nonlinear residual component. Subsequent multi-scale convolutional feature extraction of the residual component can directly focus on nonlinear distortion-related information without needing to filter target information from the mixed components. This achieves direct localization and separation of the root cause of nonlinear distortion, making the processing of different distortion components more targeted throughout the distortion correction process. It ensures that feature extraction and signal reconstruction can revolve around the true distortion components, achieving accurate correction of nonlinear distortion in non-contact current sensing.

[0013] As a preferred embodiment of the nonlinear distortion correction method for non-contact current sensing described in this invention, the step of performing temporal memory integration on the fused features and outputting a feature vector includes: inputting the fused features into a recurrent neural network in ascending time order for sequential processing, and passing the output result of the current time to the next time as an auxiliary input, thereby accumulating and generating a feature sequence containing historical information; performing global importance weighting and summation on the feature sequence to obtain a feature vector characterizing the characteristics of the entire time period.

[0014] The beneficial effects of this preferred technical solution are as follows: The fused features are input into the recurrent neural network in ascending chronological order for sequential processing. The output of the current time is passed to the next time as an auxiliary input, ensuring that the feature processing at each time point incorporates historical information. This accumulates and generates a feature sequence containing historical information, avoiding the one-sidedness of features at a single time point. Global importance weighting and summing of the feature sequence integrates the correlation value of features at different time points, converging the fused features scattered across various time dimensions into a feature vector representing the characteristics of the entire time period. This feature vector encompasses both the details of local fused features at each time point and fully reflects the feature correlations across time dimensions, accurately matching the temporal characteristics and global features of distorted data. The feature vector output by the encoding network comprehensively reflects the essence of the data, achieving the integration of fused features. This directly corresponds to the temporal changes of nonlinear residual components and the overall correlation of multi-granularity local features, promoting the correction of nonlinear distortion in non-contact current sensing. This ensures that subsequent signal reconstruction can be based on complete temporal and global features, effectively controlling the deviation between the time-domain waveform sequence and the ideal reference data.

[0015] As a preferred embodiment of the nonlinear distortion correction method for non-contact current sensing described in this invention, the step of performing layer-by-layer inverse transformation and signal reconstruction on the feature vector through the decoding network includes: expanding the feature vector into an initial time-domain feature sequence through linear mapping, wherein the calculation formula for the initial time-domain feature sequence is expressed as: ; In the formula, This is the initial time-domain feature sequence. t For time indexing, z For feature vectors, It is a linear mapping matrix. The bias vector is used to perform multi-layer convolution processing on the initial time-domain feature sequence to obtain deep time-domain features, and the deep time-domain features are mapped to a single-channel time-domain waveform sequence. The mapping method for a single-channel time-domain waveform sequence is expressed as follows: ; In the formula, It is a single-channel time-domain waveform sequence. t For time indexing, To output the mapping matrix, It represents deep temporal domain features.

[0016] The beneficial effects of this preferred technical solution are as follows: The feature vector is extended into an initial time-domain feature sequence through linear mapping. Using the linear mapping matrix and bias vector, the feature vector is transformed into an initial time-domain feature sequence matching the time dimension. This allows the information contained in the feature vector to correspond to the sequence structure in the time domain. Then, this initial time-domain feature sequence undergoes multi-layer convolution processing to progressively deepen the expression of the time-domain features, resulting in deep time-domain features. This makes the time-domain characteristics of the feature sequence more closely match the ideal signal structure corresponding to distorted data. Finally, the deep time-domain features are mapped to a single-channel time-domain feature sequence through the output mapping matrix. The time-domain waveform sequence transforms deep time-domain features into a single-channel sequence with the same structure as the ideal reference data. The entire process, through the synergistic effect of linear mapping, multi-layer convolution processing, and output mapping, gradually restores the information in the feature vector to the time-domain waveform form, thereby reconstructing the signal. The reconstruction process corresponds to the signal presentation form of non-contact current sensing, directly repairing distorted data and keeping the deviation between the reconstructed time-domain waveform sequence and the ideal reference data within a reasonable range. This achieves non-linear distortion correction for non-contact current sensing, ensuring that the signal reconstruction result meets the signal output requirements of the actual measurement device.

[0017] As a preferred embodiment of the nonlinear distortion correction method for non-contact current sensing described in this invention, the step of sequentially performing multi-layer convolution processing on the initial time-domain feature sequence to obtain deep time-domain features includes: performing L-layer recursive convolution processing on the initial time-domain feature sequence, wherein the recursive relationship of each layer is expressed as: ; In the formula, For the feature sequence of the next layer, For pointwise convolution operation, For the first The intermediate feature sequence of the layer, The value range of is [0, L-1]; The calculation formula for the intermediate feature sequence is expressed as follows: ; In the formula, For the first The input feature sequence of the layer, For depthwise separable convolution operations, For normalization operations, For non-linear activation functions, when hour, For the initial time-domain feature sequence, when hour, This refers to the deep temporal domain features.

[0018] The beneficial effects of this preferred technical solution are as follows: The initial temporal feature sequence is processed through L layers of recursive convolution. Each layer first processes the input feature sequence through depthwise separable convolution, then adjusts the feature distribution through normalization, and transforms it with a nonlinear activation function to obtain an intermediate feature sequence. This process refines the expression of features at each layer. Then, pointwise convolution is used to generate the next layer's feature sequence, allowing the feature sequence to gradually deepen the information expression of the temporal dimension in the recursive process. The deep temporal features obtained after L layers of processing retain the basic structure of the initial temporal feature sequence. Furthermore, by combining depthwise separable convolution and pointwise convolution, the temporal correlation and detail presentation of features are optimized, making the deep temporal features more consistent with the temporal characteristics of the ideal signal. The entire multi-layer convolution process connects the operations of each layer through a recursive relationship, making the feature transformation process coherently correspond to the progression from the initial temporal features to the deep temporal features, directly realizing the deepening and optimization of temporal features, making the single-channel temporal waveform sequence obtained by subsequent mapping more consistent with the ideal reference data, completing the nonlinear distortion correction of non-contact current sensing, and meeting the signal output requirements of actual measurement devices.

[0019] Secondly, the present invention provides a non-contact current sensing nonlinear distortion correction system, comprising: Data acquisition module: used to place the input probe at the current input end of the transmission line to acquire ideal parameter data, and at the same time place the output probe at the current output end of the transmission line to acquire distortion data. During the current detection stage, the output probe is used to detect the current of the wire under test to obtain the distortion signal. Preprocessing module: used to preprocess the ideal parameter data and distortion data to obtain the training dataset, and to preprocess the distorted signal; Signal repair module: includes encoding and decoding networks, used to correct the preprocessed distorted signal to obtain corrected current signal data; Training module: Used to train the signal repair module using the training dataset.

[0020] Thirdly, the present invention provides an electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of a nonlinear distortion correction method for non-contact current sensing.

[0021] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the nonlinear distortion correction method for non-contact current sensing.

[0022] Compared with existing technologies, the beneficial effects of this invention are as follows: By synchronously acquiring ideal reference data and distortion data, it provides a sample pair that perfectly matches the actual measurement scenario for the training of the signal restoration module, avoiding the problem of sample disconnection from the actual distortion scenario in traditional correction methods. On this basis, the encoding network in the signal restoration module first decomposes the distortion data into linear components and nonlinear residual components. The residual components are then processed by multi-scale convolution to extract fusion features and integrated by temporal memory to obtain feature vectors. The decoding network then reconstructs the time-domain waveform sequence through linear mapping, multi-layer convolution processing, and output mapping. This process can accurately capture complex nonlinear distortion information coupled with magnetic saturation, temperature drift, and external electromagnetic interference, and establish the corresponding feature-signal mapping relationship. This improves the problem that existing technologies rely on physical models or traditional signal processing methods and are difficult to cope with complex multi-factor coupled nonlinear distortion, which limits the correction effect. Meanwhile, the temporal memory integration of fused features in the signal restoration module involves inputting the fused features into a recurrent neural network in ascending chronological order, passing the current time output to the next time as an auxiliary input, accumulating and generating a feature sequence containing historical information, and then obtaining a feature vector through global weighting. This deeply mines the temporal dependencies of the signal, enabling the use of historical distortion information to assist in the correction at the current moment, improving the ability to identify and process nonlinear distortion patterns with long-term correlation, and making up for the shortcomings of existing technologies that ignore the temporal correlation of signals and cannot fully utilize historical information. In addition, this invention also clearly separates the distortion data into linear components and nonlinear residual components, and adopts targeted processing methods for the two types of components: the linear components are obtained by a linear dynamic system composed of parallel first-order state units, and the residual components are processed by multi-scale convolution and temporal integration. This avoids the coarse mode of traditional methods that mix linear background and nonlinear distortion components, which reduces unnecessary calculation processes and computational complexity, and can more accurately focus on the core features of nonlinear distortion, improving the accuracy and fidelity of the final reconstructed signal, and solving the problems of computational redundancy and limited signal accuracy caused by the failure to separate components in existing technologies. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the overall process of a non-linear distortion correction method for non-contact current sensing according to an embodiment of the present invention. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figure 1 This invention proposes a nonlinear distortion correction method to address signal distortion problems introduced by core saturation, temperature drift, and external interference during non-contact current measurement. It should be noted that this invention comprises two parts: an offline training phase and an online application phase. In the offline training phase, a training dataset is constructed by collecting distortion data and ideal reference data in a laboratory environment. This training dataset is then used to train a signal restoration module, enabling the module to learn the mapping relationship from distorted signals to ideal current. After training, the signal restoration module is deployed to the actual measurement device. In the online application phase, the deployed signal restoration module is used to correct the distorted signals collected in the actual measurement scenario in real time, obtaining the corrected current signal data. It is important to emphasize that the offline training phase is a means, not an end; the ultimate goal is to serve the distortion correction in the online application phase, enabling the signal restoration module to accurately eliminate nonlinear distortion in actual operating conditions and improve current measurement accuracy.

[0027] This invention provides a nonlinear distortion correction method for non-contact current sensing, comprising steps S100~S400: S100: Place the input probe at the current input end of the transmission line to collect ideal reference data, and at the same time place the output probe at the current output end of the transmission line to collect distortion data.

[0028] S200: Preprocess the ideal reference data and the distorted data to obtain the training dataset, train the signal restoration module using the training dataset, and deploy the trained signal restoration module in the production environment for application.

[0029] S300: The current detection of the wire under test is performed by the output probe to obtain the distortion signal, and the distortion signal is preprocessed.

[0030] S400: Input the preprocessed distorted signal to the signal repair module to obtain the corrected current signal data.

[0031] It should be noted that during non-contact current measurement, factors such as magnetic core saturation, temperature drift, and external electromagnetic interference can easily cause nonlinear distortion of the signal. Traditional correction methods mostly rely on hardware compensation, which has the disadvantages of high cost and poor adaptability to operating conditions. At the same time, the ideal reference data in the laboratory environment differs greatly from the operating condition data in the actual measurement scenario. If the signal is directly used to train the repair module, it is easy to cause insufficient correction accuracy of the module in actual applications. Moreover, the omission in the distortion signal preprocessing stage will further amplify the measurement error, making it difficult to meet the actual needs of high-precision current measurement. Therefore, accurate correction of nonlinear distortion in non-contact current sensing is crucial.

[0032] Therefore, to address the aforementioned limitations of hardware compensation, poor adaptability of modules to operating conditions, and insufficient correction accuracy, steps S100~S400 are used to collect ideal reference data and distortion data to construct a training set, train and deploy the signal repair module, and then perform correction on the distorted signal after preprocessing the actual measurement. This solves the problems of high cost and low adaptability of traditional hardware compensation; it enables the signal repair module to adapt to actual operating conditions, improves the effectiveness of distorted signal preprocessing, and thus completes real-time and accurate correction of nonlinear distortion in actual measurement scenarios, effectively improving the accuracy of non-contact current measurement and meeting the needs of high-precision current monitoring.

[0033] Example 2, refer to Figure 1 As an embodiment of the present invention, based on the above embodiment, a nonlinear distortion correction method for non-contact current sensing is provided.

[0034] In this embodiment of the application, step S200, which involves training the signal repair module using a training dataset and deploying the trained signal repair module in a production environment for application, includes steps A1 to A3: A1. Input the distorted data in the training dataset into the encoding network, and perform layer-by-layer transformation and feature extraction on the distorted data through the encoding network to obtain the feature vector; In this embodiment, the encoding network processes distorted data in three stages. First, the distorted data is input into a stable state-space absorption layer. The state-space absorption layer decomposes the distorted data into linear explanatory quantities and nonlinear residuals using a learnable state-space model. The linear explanatory quantities represent the background trend components in the distorted data that can be explained by linear laws, while the nonlinear residuals contain the true nonlinear distortion features introduced by factors such as core saturation and temperature drift. Subsequently, the linear explanatory quantities and nonlinear residuals are concatenated along the channel dimension and then input into a depthwise separable one-dimensional convolutional compression layer. The depthwise separable one-dimensional convolutional compression layer performs local morphological refinement on the concatenated data using a multi-scale convolutional structure. With channel expansion, local features of different granularities are extracted and fused while keeping the time length unchanged to obtain fused features. Finally, the fused features are input into the memory pooling layer, which recursively accumulates them in the time dimension through gated recurrent units to form a feature sequence containing historical information. The feature sequence is then globally weighted and integrated using an attention mechanism. The integrated representation is compressed into a low-dimensional feature vector through linear mapping. The low-dimensional feature vector is then further divided into a first subspace vector representing the instantaneous response and a second subspace vector representing historical dependence, which serve as the final output of the encoding network.

[0035] For example, for a distorted data segment with 512 sampling points, after processing by a steady-state spatial absorption layer, two signals with the shape [1,512,2] are obtained (where 1 represents the batch size, 512 represents the time length, and 2 represents the two channels). After processing by a depthwise separable one-dimensional convolutional compression layer, the number of channels is expanded to 64, and the shape becomes [1,512,64]. After processing by a memory convergence layer, it is compressed into a low-dimensional feature vector with the shape [1,32], where the first 16 dimensions are the first subspace vector and the last 16 dimensions are the second subspace vector.

[0036] A2. Input the feature vector into the decoding network, and perform layer-by-layer inverse transformation and signal reconstruction on the feature vector through the decoding network to obtain the time-domain waveform sequence. Calculate the mean square error between the time-domain waveform sequence and the corresponding ideal reference data in the training dataset. In this embodiment, the decoding network first receives the feature vector output by the encoding network, and then expands the feature vector into an initial time-domain feature sequence through linear mapping. The length of the initial time-domain feature sequence in the time dimension is consistent with the input distorted data, but the number of channels is smaller. Subsequently, the initial time-domain feature sequence undergoes multiple convolutional processing layers. Each convolutional processing layer includes depthwise separable convolution, normalization operation, and nonlinear activation function to gradually restore the time-domain characteristics of the ideal reference data. During the multi-layer convolutional processing, the decoding network refines the local structure and dynamic characteristics of the signal layer by layer, while keeping the time length strictly unchanged. After multi-layer convolutional processing, the decoding network compresses the processed multi-channel features into a single-channel time-domain waveform sequence through the final linear mapping. The time-domain waveform sequence is the estimation result of the ideal current at the input end by the decoding network. Finally, the time-domain waveform sequence is compared point by point with the corresponding ideal reference data in the training dataset, and the mean square error between the two is calculated. The mean square error reflects the current reconstruction accuracy of the signal restoration module.

[0037] For example, for the feature vector with shape [1,32] in step A1, the decoding network first expands it into an initial temporal feature sequence with shape [1,512,16] through linear mapping (where 512 is the time length and 16 is the initial number of channels). After three layers of convolution processing, the number of channels changes to 32, 48, and 64 respectively. Finally, it is compressed into a single-channel temporal waveform sequence with shape [1,512,1] through linear mapping. Then, this temporal waveform sequence is compared with the ideal reference data with the same shape [1,512,1], and the mean square error is calculated to be 0.0042.

[0038] A3. Optimize the encoding and decoding networks based on the mean square error and backpropagation algorithm and iterate until the mean square error converges to the preset threshold range to complete the training of the signal restoration module. Then, deploy the trained signal restoration module to the actual measurement device for distortion correction. In this embodiment, iterative optimization enables the encoding and decoding networks to accurately map distorted signals to ideal currents, thus ensuring accurate distortion correction in practical applications. In each training iteration, several distorted data points and corresponding ideal reference data are randomly selected from the training dataset to form a training batch. The distorted data are then sequentially processed through the encoding and decoding networks to obtain a time-domain waveform sequence, and the mean square error (MSE) is calculated. Subsequently, the gradient of the MSE relative to all learnable parameters in the encoding and decoding networks is calculated using the backpropagation algorithm. This includes the state transition coefficients in the steady-state space absorption layer, the kernel weights in the depthwise separable one-dimensional convolutional compression layer, the gated recurrent unit parameters and attention mechanism parameters in the memory-converging layer, and the convolutional parameters and linear mapping matrices of each layer in the decoding network. Based on the calculated gradient, an optimization algorithm (such as the Adam optimizer) is used to synchronously update all parameters of the encoding and decoding networks, gradually reducing the MSE. During training, the performance of the signal restoration module is evaluated periodically on the validation set (divided from the training dataset). Training is considered complete when the mean square error has converged to a preset threshold range.

[0039] For example, in this embodiment, the training batch size is set to 16, the learning rate is 0.0001, and the Adam optimizer is used for parameter updates. The preset threshold range is a mean squared error of less than 0.001. After approximately 5000 training iterations, the mean squared error on the validation set decreases from the initial 0.152 to 0.00087, meeting the preset threshold range. The training process terminates, completing the training of the signal restoration module. At this point, all parameters of the encoding and decoding networks have converged to their optimal state, enabling accurate reconstruction of a time-domain waveform sequence close to the ideal reference data from distorted data.

[0040] After training, all parameters of the encoding and decoding networks in the signal restoration module are fixed and no longer updated. The trained signal restoration module is then deployed to an actual non-contact current measurement device for real-time correction of distorted signals acquired on-site. In practical applications, only the distorted signal of the wire under test needs to be acquired through the output probe. The distorted signal undergoes the same preprocessing as in the training phase, and then the preprocessed distorted signal is input into the trained signal restoration module. The distorted signal then propagates forward through the encoding and decoding networks sequentially. During forward propagation, the encoding network first performs the same processing as in the training phase, transforming and extracting features from the distorted signal layer by layer through a steady-state spatial absorption layer, a depthwise separable one-dimensional convolutional compression layer, and a memory convergence layer to obtain a feature vector. The decoding network then receives this feature vector and expands it into an initial time-domain feature sequence through linear mapping. After multiple convolutional processing layers, it is mapped into a single-channel time-domain waveform sequence, which is the corrected current signal data. Since the parameters of the signal restoration module have been optimized to their optimal state through a large number of samples during the training phase, it can accurately reconstruct a waveform close to the ideal current from the distorted signal, thereby achieving effective correction of nonlinear distortion. It is worth noting that in practical applications, the distortion correction process only requires one forward propagation, without the need for ideal reference data as a supervision label, nor for calculating the mean square error or performing backpropagation. Therefore, the computational complexity is much lower than that during the training phase, meeting the requirements of real-time measurement.

[0041] For example, for a segment of distorted signal actually acquired, after preprocessing, the shape is [1,512,1]. This signal is then input into the deployed signal repair module. The distorted signal is first processed by an encoding network, and decomposed into linear interpreters and nonlinear residuals through a steady-state spatial absorption layer. Then, it is processed by a depthwise separable one-dimensional convolutional compression layer to extract fusion features, and then compressed by a memory convergence layer to obtain a feature vector with a shape of [1,32]. This feature vector is then processed by a decoding network, and expanded into an initial time-domain feature sequence through linear mapping. After three layers of convolution processing, it is mapped into a single-channel time-domain waveform sequence with a shape of [1,512,1], which is the corrected current signal data.

[0042] In an optional implementation, step S200, which trains the signal restoration module using a training dataset and deploys the trained module in a production environment, can also employ a phased training and gradual deployment strategy. Specifically, the encoding and decoding networks are first pre-trained separately. First, the parameters of the decoding network are fixed, and the encoding network is trained using only the distorted data from the training dataset, enabling it to learn the ability to extract compact feature representations from the distorted data. Then, the parameters of the encoding network are fixed, and the decoding network is trained using the feature vectors output by the encoding network and the corresponding ideal reference data, enabling it to learn the ability to reconstruct the ideal current waveform from the feature vectors. After the pre-training phase, the encoding and decoding networks are jointly fine-tuned. At this stage, all parameters of both networks are optimized simultaneously, and a stability constraint on the feature vectors is added as an auxiliary loss term based on the mean square error. Through a multi-objective optimization strategy, the signal restoration module improves the robustness of the feature representation while ensuring reconstruction accuracy. During training, a learning rate decay strategy is adopted, using a larger learning rate initially to accelerate convergence and gradually reducing the learning rate later for fine-tuning. After training, the signal repair module was first deployed in a small-scale test environment for trial operation. Actual distorted signals were collected for correction, and the correction effect was evaluated. Once the corrected current signal data met the accuracy requirements, the signal repair module was then deployed to a large-scale production environment. During production operation, distorted signal samples with poor correction results were periodically collected and added to the training dataset for incremental training, continuously optimizing the performance of the signal repair module to adapt to different operating conditions and environmental changes, thus maintaining the long-term stability of distortion correction.

[0043] In this embodiment of the application, the step A1 of performing layer-by-layer transformation and feature extraction on the distorted data through an encoding network includes B1~B2: B1. Decompose the distorted data into linear components and nonlinear residual components, and extract local features of different granularities from the residual components through multi-scale convolution and fuse them to obtain fused features. In this embodiment, the steady-state spatial absorption layer decomposes the distortion data by connecting multiple first-order state units in parallel. Each state unit models the dynamic characteristics of the signal with a different time constant. After parallel processing, the units converge into a linear component, which characterizes the background trend in the distortion data that can be explained by linear dynamic laws. Subtracting the linear component from the distortion data yields the nonlinear residual component. The nonlinear residual component concentrates the nonlinear distortion information caused by core saturation, temperature drift, and external interference. To ensure training stability, the transition coefficients of the state units are constrained and parameterized. At the same time, the time constants of each state unit are uniformly distributed over a short- to long-term range, enabling the signal restoration module to have multi-scale time memory capabilities.

[0044] Linear and nonlinear residual components are concatenated and input into a depthwise separable one-dimensional convolutional compression layer. A multi-scale convolution strategy is used to extract and fuse local features of different granularities. The first convolutional block uses a medium-length kernel to capture the temporal morphology of the fundamental frequency band. The second layer employs a dual-branch parallel structure: the main branch uses a longer kernel to capture steady morphological changes, while the parallel branch uses a shorter kernel to capture fast, detailed features. The outputs of the two branches are fused during point convolution operations, achieving multi-scale temporal feature integration. Finally, a channel recalibration mechanism is introduced at the convolutional layer output. By evaluating the global response strength of each channel, the channel weights are adjusted to highlight feature channels related to nonlinear distortion and suppress redundant and noisy channels, achieving channel-dimensional fusion and ultimately obtaining the fused features.

[0045] B2. Perform temporal memory integration on the fused features and output the feature vector; In this embodiment, the memory aggregation layer recursively processes the fused features over time through a gated loop unit, accumulating historical dependency information. The gated loop unit dynamically adjusts information flow through update and reset gates; the update gate controls the retention of historical states, while the reset gate controls the forgetting of historical information. At each time step, the gated loop unit calculates the activation value of the control gate based on the current input and the hidden state of the previous time step. Combining the control gates, it selectively forgets and updates the historical state, generating a new hidden state that integrates the current input and historical information. Through this time-step recursion, the signal repair module accumulates global temporal dependencies from the initial time step to the current time step, forming a feature sequence containing historical memories.

[0046] To avoid the dilution of early information by the final hidden state, an attention mechanism is introduced for global weighted integration. The attention mechanism calculates an importance score for each time step, obtains the attention weight distribution through nonlinear mapping and normalization, and then performs a weighted summation of the feature sequence based on these attention weights to obtain a global context vector. This global context vector dynamically highlights the time segments most critical to the task. The global context vector and the final hidden state are combined using a gating fusion strategy. The fusion gate determines the weight allocation of the two information sources, generating a fused memory representation. Finally, the fused memory representation is mapped to a low-dimensional space through a linear transformation to obtain a feature vector. This feature vector retains both instantaneous nonlinear features and captures global historical dependencies across time, serving as the final output of the encoding network for the decoding network.

[0047] In an optional implementation, step A1 can involve performing layer-by-layer transformation and feature extraction on the distorted data through an encoding network, or directly inputting the distorted data into a multi-layer convolutional neural network. By progressively increasing the receptive field and the number of channels, multi-level feature representations from local to global are extracted. Batch normalization and ReLU activation functions are introduced after each convolutional layer to accelerate convergence and enhance nonlinear expressive power. Then, a skip connection mechanism is introduced in the middle layer of the encoding network to fuse the detailed features extracted from the shallow layer with the semantic features extracted from the deep layer, avoiding gradient vanishing and preserving multi-scale information. Finally, global average pooling is used to compress the multi-channel features into a fixed-length feature vector.

[0048] In this embodiment of the application, the step B1 of decomposing the distorted data into linear components and nonlinear residual components includes C1~C2: C1. Construct a linear dynamic system consisting of K parallel first-order state units, input distortion data into the linear dynamic system, and use the calculated linear response as the linear component. In this embodiment, the linear dynamic system consists of K independent first-order state units connected in parallel. Each state unit maintains an internal state variable to memorize the dynamic characteristics of the signal at different time scales. The distortion data at time t... When input to a linear dynamic system, the k-th state unit (where k ranges from 1 to K) first determines its internal state. and the input distorted data The state is updated according to the following recursive relationship: In the formula, This is the state value at the next moment. This is the current state value. These are the state transition coefficients. The input weights are the state transition coefficients. The state transition coefficient controls the rate of memory decay in this state unit; the closer the absolute value is to 1, the longer the memory retention time. To ensure stability, the state transition coefficient... Generate through constraint parameterization, i.e. ,in These are learnable, unconstrained parameters.

[0049] After each state unit completes its state update, the linear dynamic system weights and aggregates the outputs of all state units to generate a linear component. Specifically, the linear component consists of two parts: the first part is distorted data. Direct feedback items The direct feedback coefficient dThe first part controls the proportion of distorted data directly transmitted to the output; the second part is the weighted sum of the outputs of K state units. ,in The output weight of the k-th state unit is given by the complete formula for calculating the linear component. This linear component characterizes the background trend portion of distorted data that can be explained by linear dynamics, such as the fundamental frequency component and slowly varying low-frequency components.

[0050] During the parameter initialization phase, in order to enable each state unit to have different time-scale memory capabilities, the fundamental period of the signal is used. Set K time constants In the interval The system is uniformly distributed within the timescale, thus covering multiple timescales from fast transients to long-term trends; subsequently, it is based on a time constant. and sampling period Calculate the initial state transition coefficients This exponential decay form simulates the discrete-time response characteristics of a first-order low-pass filter; input weights The initial value is set to This ensures that the output amplitude of the state unit matches the input amplitude under steady-state conditions. For example, when the fundamental frequency period... Sampling period When the number of state units K=8, the time constants of each unit are 10, 20, 30, ..., 160ms, and the corresponding initial state transition coefficients are uniformly distributed from 0.99 to 0.995.

[0051] C2. Subtract the linear component from the distorted data to obtain the nonlinear residual component; In this embodiment, linear components are obtained. Then, the distorted data was extracted by point-by-point subtraction. Remove the linear component, that is, calculate the nonlinear residual component. The residual component contains all the nonlinear distortion information in the distorted data that cannot be explained by the linear dynamic system. Specifically, this includes nonlinear compression or clipping of the signal amplitude caused by core saturation, baseline shift and gain fluctuations due to temperature drift, and high-frequency noise and abrupt pulses superimposed by external electromagnetic interference. Since the linear component has absorbed the background trend and periodic fundamental components of the distorted data, the nonlinear residual cost centrally reflects the true distortion characteristics, exhibiting higher information density and stronger discriminative power.

[0052] linear components and nonlinear residuals The decomposition is performed along the channel dimension to form a dual-channel decomposition result, which serves as the input to the subsequent depthwise separable one-dimensional convolutional compression layer. Through this decomposition strategy, the signal inpainting module can separately process the linear background and nonlinear distortion components of the distorted data, avoiding mutual interference and thus improving the targeting and accuracy of feature extraction. For example, for distorted data of shape [1,512,1], step C1 calculates the linear component of shape [1,512,1] and the nonlinear residual cost of shape [1,512,1], which are then concatenated to obtain a decomposition result of shape [1,512,2].

[0053] In an alternative implementation, the decomposition of the distorted data into linear and nonlinear residual components in step B1 can also be achieved using a frequency domain decomposition strategy combined with a filter bank. Specifically, the distortion data is first subjected to a Fast Fourier Transform (FFT) to convert the time-domain signal to the frequency domain. In the frequency domain, the fundamental frequency, its corresponding amplitude, and phase are identified. The fundamental frequency typically corresponds to the power frequency component of the current signal (e.g., 50Hz or 60Hz), which is represented by the most concentrated energy peak in the frequency domain. Next, a narrowband bandpass filter is designed, with its center frequency set to the fundamental frequency and its bandwidth covering a reasonable range near the fundamental frequency (e.g., ±5Hz). The narrowband bandpass filter is applied to the frequency domain signal to extract the fundamental frequency and nearby stationary frequency components. Then, the filtered frequency domain signal is converted back to the time domain using an inverse Fourier transform. The resulting time-domain signal is the linear component, which mainly contains the fundamental periodic component and low-frequency trend in the distortion data. Finally, the linear component is subtracted from the original distortion data to obtain the nonlinear residual component, which contains higher harmonics, interharmonics, and frequency shift components generated by nonlinear distortion.

[0054] In this embodiment of the application, step B2, which involves temporal memory integration of the fused features and outputting the feature vector, includes steps D1 to D2: D1. Input the fused features into the recurrent neural network in ascending order of time for sequential processing, and pass the output of the current time to the next time as an auxiliary input, thereby accumulating and generating a feature sequence containing historical information. First, a gated loop unit is used to recursively process the fused features. In this embodiment, the shape of the fused features is... Where B is the batch size and T is the time duration. The number of channels; the hidden dimension H of the gated loop unit is set to 32, and the initial hidden state is set to zero vector.

[0055] During the recursive processing, the gated loop unit dynamically adjusts the information flow using update and reset gates. For the t-th time step, the update and reset gates are first calculated. The update gate controls the degree of retention of historical states, while the reset gate controls the degree of forgetting of historical information. Subsequently, candidate hidden states are generated based on the current input and the historical states adjusted by the reset gate. Finally, the hidden state at the current time step is obtained by weighted fusion of the historical states and candidate states using the update gate. By iteratively accumulating information step by step from the initial time to the final time T, a feature sequence containing historical memories is formed. , ,…, }

[0056] For example, for a fused feature of shape [1,512,64], after processing by a gated recurrent unit, a feature sequence of shape [1,512,32] is obtained, where 32 is the hidden dimension H. Each time step in this feature sequence contains accumulated memory information from the start time to the current time.

[0057] D2. Perform global importance weighting and summation on the feature sequences to obtain the feature vectors that represent the characteristics of the entire time period; To avoid the dilution of early key information by the hidden state at the final time step T, this embodiment introduces an attention mechanism to globally weighted integrate the feature sequence. Specifically, firstly, a nonlinear transformation is applied to the hidden state at each time step in the feature sequence generated in step D1 to generate an importance score. Then, softmax normalization is used to obtain the attention weight distribution. Next, the feature sequence is weighted and summed based on the attention weight distribution to obtain a global context vector. Then, the final time step hidden state is concatenated with the global context vector. A gating fusion strategy is used to combine the two information sources. The fusion gate assigns weights according to task requirements to generate a fused memory representation. Subsequently, a linear transformation is used to map the fused memory representation to a low-dimensional space, resulting in a shape... The feature vector is obtained by considering the length of the original sequence, where r is much smaller than the length of the original sequence. For example, for a feature sequence of shape [1, 512, 32], after attention-weighted integration and dimensionality reduction mapping, a feature vector of shape [1, 32] is obtained.

[0058] In an optional implementation, step B2 involves temporal memory integration of the fused features, and the output feature vector can also employ a parallel encoding method based on a self-attention mechanism. First, positional encoding is added to the fused features, injecting temporal positional information. The positional encoding is generated using sine and cosine functions and element-wise added to the fused features to obtain position-aware features. Subsequently, the position-aware features are input into a multi-layer Transformer encoder for global dependency modeling. Each encoder layer contains a multi-head self-attention module and a feedforward neural network module. The multi-head self-attention module calculates the correlation matrix between time steps using query, key, and value triples, enabling each time step to correlate information from all other time steps in parallel, thereby capturing long-distance temporal dependencies. Based on the feature sequence output by the encoder, a dual-pooling strategy is used to extract global representations: one path performs average pooling on the features of all time steps to obtain overall statistical properties, and the other path performs max pooling to capture significant peak features. The results of the two pooling paths are concatenated and mapped to the target dimension through a fully connected layer, resulting in a feature vector that integrates global and local temporal information.

[0059] In this embodiment of the application, step A2, which involves performing layer-by-layer inverse transformation and signal reconstruction of the feature vector using a decoding network, includes steps E1 to E2: E1. Extend the feature vector into an initial time-domain feature sequence through a linear mapping. The formula for calculating the initial time-domain feature sequence is as follows: ; In the formula, This is the initial time-domain feature sequence. t For time indexing, z For feature vectors, It is a linear mapping matrix. It is the bias vector; In this embodiment, a linear mapping layer is used to expand and transform the feature vector, restoring the compressed feature vector to its time-dimensional unfolded sequence form. Specifically, the feature vector... z The shape is Where B is the batch size. For the feature dimension; the weight matrix of the linear mapping layer and bias vector All of these are learnable parameters, obtained through training and optimization.

[0060] weight matrix The shape is ,in T For the target time length, The initial number of channels, Represents the total number of dimensions after flattening the time and channel dimensions; bias vector The shape is During the mapping process, the feature vector z With weight matrix Performing matrix multiplication yields a matrix of shape... The intermediate result is then added with the bias vector. After performing an affine transformation, the intermediate results are reshaped into shapes. The three-dimensional tensor, i.e., the initial temporal feature sequence .

[0061] For example, for a feature vector of shape [1, 32], when the target time length T = 512 and the initial number of channels... When =64, the weight matrix The shape is [32,32768], and after linear mapping and reshaping, the initial time-domain feature sequence with shape [1,512,64] is obtained.

[0062] E2. Perform multi-layer convolution processing on the initial time-domain feature sequence to obtain deep time-domain features, and map the deep time-domain features into a single-channel time-domain waveform sequence; The mapping method for a single-channel time-domain waveform sequence is expressed as follows: ; In the formula, It is a single-channel time-domain waveform sequence. t For time indexing, To output the mapping matrix, It represents deep temporal domain features; In this embodiment, each convolutional module includes operations such as depthwise separable convolution, normalization, and nonlinear activation. Through layer-by-layer processing, the feature sequence is gradually refined from coarse-grained representation to high-resolution deep temporal features. , where L is the total number of convolutional layers and t is the time index.

[0063] First, a mapping function P is introduced to perform dimensionality reduction projection on the feature sequence, converting multi-channel deep temporal features into single-channel waveforms. The mapping function P is usually implemented using pointwise convolution, i.e., a one-dimensional convolution with a kernel size of 1, which reduces the number of channels of deep features from... Compressed to 1. Specifically, the weight matrix of the pointwise convolution has the following shape: And by weighting the feature vector at each time step A weighted combination is performed, and the output scalar value is used as the predicted amplitude at that time step. By applying the same mapping function P at all time steps, a single-channel time-domain waveform sequence with shape [B,T,1] is obtained. .

[0064] For example, for a deep temporal feature of shape [1,512,1], a single-channel temporal waveform sequence of shape [1,512] is obtained after pointwise convolution mapping. Then, the temporal waveform sequence is compared with the corresponding ideal reference data in the training dataset to calculate the reconstruction error.

[0065] In this embodiment of the application, step E2, which involves performing layer-by-layer inverse transformation and signal reconstruction of the feature vectors using a decoding network, includes F1: F1. Perform L layers of recursive convolution on the initial temporal feature sequence, where the recursive relationship of each layer is expressed as: ; In the formula, For the feature sequence of the next layer, For pointwise convolution operation, For the first The intermediate feature sequence of the layer, The value range of is [0, L-1]; The formula for calculating the intermediate feature sequence is as follows: ; In the formula, For the first The input feature sequence of the layer, For depthwise separable convolution operations, For normalization operations, For non-linear activation functions, when hour, Given the initial time-domain feature sequence, when hour, It represents deep temporal domain features; In this embodiment, a depthwise separable convolutional architecture is used for recursive processing to achieve multi-layer feature extraction and transformation. Specifically, the L-layer convolutional network is arranged in sequence. The process is recursively applied layer by layer, with the input of each layer being the output of the previous layer. After processing by the convolution module, the input for the next layer is generated. .

[0066] Each layer's convolutional processing consists of two sub-steps: first, processing the input feature sequence... Applying depthwise convolution The depthwise convolution performs independent spatial convolution operations on each channel, with a kernel size of 3 or 5 and a stride of 1. After the depthwise convolution output, it undergoes layer normalization and non-linear activation function processing. Layer normalization standardizes the features along the channel dimension to stabilize the training process, while the non-linear activation function uses GELU or ReLU to introduce non-linear expressive power, resulting in an intermediate feature sequence. Subsequently, pointwise convolution is applied to this intermediate feature sequence, with a kernel size of 1, to fuse information across channels and adjust the number of channels, finally outputting the feature sequence for the next layer. .

[0067] For example, for an input feature sequence of shape [1, 512, 64] After depthwise convolution with a kernel size of 3, layer normalization, GELU activation, and pointwise convolution with the number of channels adjusted from 64 to 32, a shape of [1, 512, 32] is obtained. Through L-layer recursion, the number of feature channels decreases layer by layer while the expressive power increases layer by layer, eventually converging to the single-channel current signal reconstruction result.

[0068] In summary, by synchronously acquiring ideal reference data and distortion data, a sample pair that perfectly matches the actual measurement scenario is provided for the training of the signal restoration module. This avoids the problem of sample disconnect from the actual distortion scenario in traditional correction methods. Based on this, the encoding network in the signal restoration module first decomposes the distortion data into linear components and nonlinear residual components. The residual components are then processed by multi-scale convolution to extract fusion features and integrated through temporal memory to obtain feature vectors. The decoding network then reconstructs the time-domain waveform sequence through linear mapping, multi-layer convolution processing, and output mapping. This process can accurately capture complex nonlinear distortion information coupled with magnetic saturation, temperature drift, and external electromagnetic interference, and establish the corresponding feature-signal mapping relationship. This improves the problem that existing technologies rely on physical models or traditional signal processing methods and are difficult to cope with complex multi-factor coupled nonlinear distortion, which limits the correction effect. Meanwhile, the temporal memory integration of fused features in the signal restoration module involves inputting the fused features into a recurrent neural network in ascending chronological order, passing the current time output to the next time as an auxiliary input, accumulating and generating a feature sequence containing historical information, and then obtaining a feature vector through global weighting. This deeply mines the temporal dependencies of the signal, enabling the use of historical distortion information to assist in the correction at the current moment, improving the ability to identify and process nonlinear distortion patterns with long-term correlation, and making up for the shortcomings of existing technologies that ignore the temporal correlation of signals and cannot fully utilize historical information. In addition, this invention also clearly separates the distortion data into linear components and nonlinear residual components, and adopts targeted processing methods for the two types of components: the linear components are obtained by a linear dynamic system composed of parallel first-order state units, and the residual components are processed by multi-scale convolution and temporal integration. This avoids the coarse mode of traditional methods that mix linear background and nonlinear distortion components, which reduces unnecessary calculation processes and computational complexity, and can more accurately focus on the core features of nonlinear distortion, improving the accuracy and fidelity of the final reconstructed signal, and solving the problems of computational redundancy and limited signal accuracy caused by the failure to separate components in existing technologies.

[0069] Example 3 illustrates a schematic scheme for a non-contact current sensing nonlinear distortion correction method. It should be noted that the technical solution of this non-contact current sensing nonlinear distortion correction system belongs to the same concept as the technical solution of the aforementioned non-contact current sensing nonlinear distortion correction method. Details not described in detail in this embodiment can be found in the description of the aforementioned non-contact current sensing nonlinear distortion correction method.

[0070] This embodiment also provides a non-contact current sensing nonlinear distortion correction system, including: Data acquisition module: used to place the input probe at the current input end of the transmission line to acquire ideal parameter data, and at the same time place the output probe at the current output end of the transmission line to acquire distortion data. During the current detection stage, the output probe is used to detect the current of the wire under test to obtain the distortion signal. Preprocessing module: Used to preprocess ideal parameter data and distortion data to obtain training dataset, and to preprocess distorted signals; Signal repair module: includes encoding and decoding networks, used to correct the preprocessed distorted signal to obtain corrected current signal data; Training module: Used to train the signal repair module using the training dataset.

[0071] This embodiment also provides an electronic device suitable for non-linear distortion correction of non-contact current sensing, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the non-linear distortion correction method for non-contact current sensing proposed in the above embodiment.

[0072] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the nonlinear distortion correction method for non-contact current sensing as proposed in the above embodiments.

[0073] The storage medium proposed in this embodiment and the nonlinear distortion correction method for non-contact current sensing proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0074] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A nonlinear distortion correction method for non-contact current sensing, characterized in that, include: The input probe is placed at the current input end of the transmission line to collect ideal reference data, while the output probe is placed at the current output end of the transmission line to collect distortion data. The ideal reference data and the distortion data are preprocessed to obtain a training dataset, and the signal restoration module is trained using the training dataset. The trained signal restoration module is then deployed in the production environment for application. The current of the wire under test is detected by the output probe to obtain the distortion signal, and the distortion signal is preprocessed. The preprocessed distorted signal is input into the signal repair module to obtain the corrected current signal data.

2. The nonlinear distortion correction method for non-contact current sensing as described in claim 1, characterized in that, The steps of training the signal repair module using the training dataset and deploying the trained signal repair module in a production environment include: The distorted data in the training dataset is input into the encoding network, and the distorted data is transformed and features are extracted layer by layer through the encoding network to obtain the feature vector; The feature vector is input into the decoding network, and the feature vector is subjected to layer-by-layer inverse transformation and signal reconstruction through the decoding network to obtain a time-domain waveform sequence. The mean square error between the time-domain waveform sequence and the corresponding ideal reference data in the training dataset is calculated. The encoding and decoding networks are optimized and iterated according to the mean square error and backpropagation algorithm until the mean square error converges to a preset threshold range. The training of the signal restoration module is then completed, and the trained signal restoration module is deployed to the actual measurement device for distortion correction.

3. The nonlinear distortion correction method for non-contact current sensing as described in claim 2, characterized in that, The steps of performing layer-by-layer transformation and feature extraction on distorted data through the coding network include: The distorted data is decomposed into linear components and nonlinear residual components. Local features of different granularities are extracted from the residual components through multi-scale convolution and fused to obtain fused features. The fused features are integrated using temporal memory to output a feature vector.

4. The nonlinear distortion correction method for non-contact current sensing as described in claim 3, characterized in that, The steps of decomposing the distorted data into linear components and nonlinear residual components include: Construct a linear dynamic system consisting of K parallel first-order state units, input the distortion data into the linear dynamic system, and use the calculated linear response as the linear component; The nonlinear residual component is obtained by subtracting the linear component from the distorted data.

5. The nonlinear distortion correction method for non-contact current sensing as described in claim 4, characterized in that, The steps of performing temporal memory integration on the fused features and outputting a feature vector include: The fused features are input into a recurrent neural network in ascending order of time for sequential processing, and the output of the current time is passed to the next time as an auxiliary input, thereby accumulating and generating a feature sequence containing historical information. The feature sequence is weighted globally by weighting and summing to obtain a feature vector representing the characteristics of the entire time period.

6. The nonlinear distortion correction method for non-contact current sensing as described in claim 5, characterized in that, The steps of performing layer-by-layer inverse transformation and signal reconstruction on the feature vector through the decoding network include: The feature vector is extended into an initial time-domain feature sequence through a linear mapping. The formula for calculating the initial time-domain feature sequence is as follows: ； In the formula, This is the initial time-domain feature sequence. t For time indexing, z For feature vectors, It is a linear mapping matrix. It is the bias vector; The initial time-domain feature sequence is subjected to multiple convolutional processes to obtain deep time-domain features, and the deep time-domain features are mapped to a single-channel time-domain waveform sequence. The mapping method for a single-channel time-domain waveform sequence is expressed as follows: ； In the formula, It is a single-channel time-domain waveform sequence. t For time indexing, To output the mapping matrix, It represents deep temporal domain features.

7. The nonlinear distortion correction method for non-contact current sensing as described in claim 6, characterized in that, The steps for obtaining deep temporal features by sequentially performing multi-layer convolution on the initial temporal feature sequence include: The initial temporal feature sequence is processed by L layers of recursive convolution, where the recursive relationship of each layer is expressed as: ； In the formula, For the feature sequence of the next layer, For pointwise convolution operation, For the first The intermediate feature sequence of the layer, The value range of is [0, L-1]; The calculation formula for the intermediate feature sequence is expressed as follows: ； In the formula, For the first The input feature sequence of the layer, For depthwise separable convolution operations, For normalization operations, For non-linear activation functions, when hour, For the initial time-domain feature sequence, when hour, This refers to the deep temporal domain features.

8. A non-contact current sensing nonlinear distortion correction system, using the method described in any one of claims 1-7, characterized in that, include: Data acquisition module: used to place the input probe at the current input end of the transmission line to acquire ideal parameter data, and at the same time place the output probe at the current output end of the transmission line to acquire distortion data. During the current detection stage, the output probe is used to detect the current of the wire under test to obtain the distortion signal. Preprocessing module: used to preprocess the ideal parameter data and distortion data to obtain the training dataset, and to preprocess the distorted signal; Signal repair module: includes encoding and decoding networks, used to correct the preprocessed distorted signal to obtain corrected current signal data; Training module: Used to train the signal repair module using the training dataset.

9. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the nonlinear distortion correction method for non-contact current sensing according to any one of claims 1 to 7.

10. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the nonlinear distortion correction method for non-contact current sensing according to any one of claims 1 to 7.

Images