Current sensor error analysis system for multi-dimensional signal anomaly detection

Abstract

The present application relates to the technical field of current measurement, in particular to a multi-dimensional signal anomaly detection current sensor error analysis system, comprising four deeply coupled closed-loop modules of a multi-dimensional signal topology feature extraction module, a multi-channel time sequence correlation anomaly detection module, a working condition adaptive error characteristic online learning module and an error component blind source separation module, the system extracts the Betti number and the persistent barcode topology invariant feature from each channel sampling signal to distinguish the error type by using the time delay embedding combined with the persistent coherence method, combines the online Granger causality matrix with the graph difference to locate the error source candidate channel, executes the working condition adaptive compensation by using the variational autoencoder hidden space nearest neighbor search, and decomposes the residual error into three components of bias, noise and interference for independent compensation by using the constrained independent component analysis, and then updates the baseline, thereby realizing the joint online identification of three elements of the error type, the error source channel and the independent error component amplitude.

Description

Technical Field

[0001] This invention relates to the field of current measurement technology, specifically to a current sensor error analysis system for multi-dimensional signal anomaly detection, which is particularly suitable for online error analysis and compensation of multi-channel current sensor arrays in smart charging piles, electric drive systems for new energy vehicles, distributed photovoltaic grid-connected devices, energy storage systems, and high-voltage direct current transmission projects. Background Technology

[0002] Current sensors are key sensing components in electric drive systems for new energy vehicles, smart charging piles, distributed photovoltaic grid-connected devices, energy storage systems, and high-voltage direct current transmission projects. Their measurement accuracy directly determines the dynamic response quality and system stability of the upper-level control system. In actual operation, current sensors typically experience multiple disturbances simultaneously, including temperature drift, external magnetic field interference, electromagnetic interference coupling, and core aging due to long-term service. This results in composite errors caused by the superposition of bias error, gain error, and nonlinearity error. Identifying the type of composite error, locating its source, and estimating the amplitude of each independent component are crucial prerequisites for ensuring the long-term reliable operation of a high-precision current measurement system.

[0003] Chinese invention patent application CN119224675A discloses a current calibration measurement method and related device for a current transformer. This method preprocesses maintenance information, power data, and temperature time-series data; generates calibration parameters based on harmonic analysis and temperature influence change analysis; and constructs an error analysis model by analyzing magnetic field interference using a finite element simulation model and nonlinear parameterized functions. Finally, the calibration measurement of the current transformer is jointly controlled by the calibration parameters and the target error compensation value. However, this scheme uses single-channel harmonic amplitude statistics as the basis for error analysis and relies solely on offline calibration mapping relationships pre-stored in the error compensation database to generate compensation values. In actual operating conditions where bias error, gain error, and nonlinear error are mixed and superimposed with arbitrary amplitudes, the three types of errors mask each other at the amplitude statistics level. Reliable judgment of error types cannot be performed based solely on the two scalar characteristics of harmonic amplitude and dielectric loss. Furthermore, it is impossible to pinpoint the specific sensor channel in the multi-sensor array to be attributable to the error source based solely on the signal itself, and it is also impossible to perform targeted separation and independent compensation for the independent components within the composite error. The fundamental problem lies in the low feature dimension of error analysis—remaining only on the single dimension of amplitude statistics, lacking the systematic perception of the inherent geometric structure of multi-channel signals, the collaborative timing correlation between channels, and the nonlinear mapping of operating parameters and error characteristics.

[0004] Chinese invention patent application CN112183477A discloses a method for extracting continuous features of EEG signals based on continuous coherence. This method uses continuous coherence tools to extract topological invariants such as Betti number and continuous barcode to characterize the signal structure of a one-dimensional time series. However, this scheme only uses continuous coherence as a single-channel feature extractor for EEG signals. The mapping from topological features to EEG task labels is learned by a downstream classifier in a black box. It does not reveal the geometric discrimination mechanism between specific physical perturbation patterns and phase space topological invariants, nor does it establish a synergistic relationship between multi-channel topological features and channel-level anomaly localization, nor does it integrate topological feature extraction into a multi-module closed-loop processing link. Chinese invention patent application CN119004059A discloses a data fusion method and system based on multiple sensors. The solution explicitly acknowledges that existing multi-sensor data fusion methods face three challenges: individual sensor manufacturing deviations, time-varying drift caused by temperature and humidity aging, and lack of future performance prediction capabilities. However, its approach remains at the level of data preprocessing and comprehensive evaluation, without providing a systematic closed-loop processing scheme for performing blind source separation and independent compensation for each independent component within the composite error.

[0005] In summary, existing current sensor error analysis schemes face a core technical problem when dealing with the mixed superposition of bias, gain, and nonlinear errors under wide temperature range and strong electromagnetic interference conditions: they fail to effectively address the joint online identification of the three key elements—error type, error source channel, and amplitude of each independent error component—based solely on the signal itself, rather than offline calibration data. The key to solving this problem lies in shifting the feature perception dimension of error analysis from the amplitude statistics domain to the phase space topology domain, and establishing a deeply coupled closed-loop processing architecture that encompasses multi-channel topology perception, channel-level causal anomaly localization, adaptive compensation under operating conditions, and blind source separation of error components. Summary of the Invention

[0006] To address the core bottleneck in existing current sensor error analysis technologies, which rely solely on the signal itself and cannot jointly identify the three elements of error type, error source channel, and independent error component amplitude online, this invention provides a current sensor error analysis system for multi-dimensional signal anomaly detection. By coupling four types of sensing capabilities—phase space topological invariant characteristics of each channel's sampled signal, anomalies in the multi-channel time-series causal relationship matrix, implicit space mapping of operating parameters and error characteristics, and blind source separation of error components—into a closed-loop processing link, this system achieves joint online identification and independent compensation of the three elements of composite error type, channel, and component without relying on offline calibration data, based on the multi-dimensional signal geometry and multi-module collaborative mechanism.

[0007] The technical solution of the present invention is: a current sensor error analysis system for multidimensional signal anomaly detection, including a multidimensional signal topology feature extraction module, a multi-channel time-series correlation anomaly detection module, an online learning module for adaptive error characteristics under operating conditions, and a blind source separation module for error components.

[0008] The multidimensional signal topology feature extraction module performs time-delay embedding reconstruction on the sampled signals of each channel of the current sensor array to obtain the phase space trajectory point cloud of each channel. Then, it uses topological data analysis methods to extract topological invariant features, including Betti number and continuous barcode, from the phase space trajectory point cloud of each channel. Error type discrimination is performed based on the deviation pattern of the topological invariant features relative to the pre-stored baseline topological features. Among them, bias error is manifested as a shift in the centroid of the phase space trajectory point cloud while the topological complexity index maintains the baseline value; gain error is manifested as a constant overall scaling of the continuous barcode length distribution while the topological complexity index maintains the baseline value; nonlinear error is manifested as an integer-level abrupt change in the one-dimensional Betti number accompanied by a change in the continuous barcode length distribution from single-mode to dual-mode or multi-mode.

[0009] The multi-channel temporal correlation anomaly detection module receives the error type output by the multi-dimensional signal topology feature extraction module, performs online estimation of the temporal causal relationship matrix for each channel's sampled signal within a sliding time window, uses the change in the spectral norm of the current sliding time window causal relationship matrix relative to the baseline causal relationship matrix as the multi-channel collaborative anomaly detection index, and triggers a graph difference algorithm to perform edge-by-edge comparison of the two causal relationship matrices when the change in the spectral norm exceeds a preset anomaly threshold, and locates nodes where the causal relationship weights undergo directional changes or the absolute value of the weights changes beyond the edge change threshold as candidate error source channels, and the candidate error source channels carry the error type as the error type-channel joint identification result.

[0010] The online learning module for adaptive error characteristics uses an encoder-decoder network to perform latent space representation learning on the error characteristics of the sensor under different combinations of temperature, frequency, and DC bias parameters. The module encodes the parameter vector into a latent space vector and then decodes it back into an error characteristic parameter vector. During online operation, the module performs nearest neighbor retrieval in the latent space using the current parameters and calls the local error characteristic model associated with the nearest neighbor region. It then performs adaptive compensation for the error data of the candidate error source channels.

[0011] The error component blind source separation module uses a blind source separation algorithm to separate the residual error after adaptive compensation for the operating conditions, obtaining three independent error components: systematic bias component, random noise component, and interference coupling component. Kalman smoothing is used to perform optimal estimation and inverse compensation for the systematic bias component, adaptive Wiener filtering is used to suppress the random noise component, and electromagnetic interference spectrum feature matching is used to perform source tracing and labeling for the interference coupling component. The cleaned signal after compensation of the three types of components is fed back to the multi-dimensional signal topology feature extraction module to perform incremental updates on the baseline topology features, forming a closed-loop processing link from error type discrimination to error component compensation and then to baseline self-updating.

[0012] Compared with the prior art, the present invention achieves the following beneficial effects:

[0013] First, the feature dimension for error type discrimination shifts from single-channel amplitude statistics to phase space topology, enabling discriminative discrimination of three types of errors even in scenarios with superimposed composite errors. The mechanism lies in the fact that bias error, after time delay embedding, corresponds to the overall translation of the phase space trajectory point cloud; gain error corresponds to the proportional scaling of the phase space trajectory point cloud; and nonlinear error corresponds to the folding and self-intersection of the phase space trajectory point cloud. These three geometric deformation modes exert independent characteristic influences on topological invariants such as the Betti number and persistent barcodes, exhibiting geometric non-masking properties. Existing technologies perform single-scalar discrimination using single-channel harmonic amplitudes. This scheme's geometric discrimination mapping of the three types of error topological features remains independent and separable even when any two or three of the bias, gain, and nonlinearity are superimposed, improving the error type discrimination accuracy to over 95%.

[0014] Secondly, the error source channel-level localization capability represents a leap from overall alarm of the current sensor array to precise identification of specific channels. The mechanism lies in the fact that when a single sensor malfunctions, the temporal causal relationship between that sensor and the other channels undergoes a directional change. By using the change in spectral norm as an early detection indicator of multi-channel coordinated anomalies and combining it with a graph differencing algorithm for edge-by-edge comparison, the node where the causal relationship has abnormally changed can be automatically located. Compared to existing solutions that rely on overall threshold alarms of the current sensor array and cannot pinpoint specific channels, this solution improves the accuracy of specific channel identification to 96% in error source localization experiments with a sixteen-channel sensor array, providing a precise target for subsequent targeted compensation.

[0015] Third, it possesses continuous adaptive capability to dynamic drift of operating condition-dependent error characteristics. The mechanism lies in using a variational autoencoder to perform latent space manifold encoding on the nonlinear mapping relationship between operating condition parameters and error characteristics. During online operation, it performs nearest neighbor search in the latent space using the current operating condition parameters and calls the corresponding local model for compensation. This discretizes the continuous value space of the operating condition parameters into a set of local models on the manifold, avoiding the interpolation distortion problem of fixed offline calibration tables when operating condition parameters change continuously. Compared to fixed offline calibration models, this scheme reduces the overall error by approximately 63% over a wide temperature range from -40℃ to 120℃.

[0016] Fourth, the ability to specifically separate and independently compensate for the independent components within the composite error. The mechanism lies in using kurtosis and negative entropy indices as joint independence constraints to perform constrained independent component analysis. This decomposes the composite error into three independent components: systematic bias components, random noise components, and interference coupling components. Each component is processed by its most suitable compensation branch, forming a precise compensation chain targeting the differences in the physical origins of the error. Compared to existing schemes that perform overall filtering on the composite error, this scheme reduces the composite error margin by approximately 70%.

[0017] Fifth, the deep coupling and closed-loop synergistic effect of the four modules—topological features, causal matrix, operating condition latent space, and error components—is achieved. The mechanism lies in the fact that the output of the preceding module serves as the key input to the following module, while the results of the subsequent module update the baseline topological feature library of the preceding module through a clean signal feedback loop. This enables the system to adapt to the slow characteristic drift caused by long-term sensor aging. The coupling of the four modules generates a nonlinear synergistic gain of 1+1+1+1>4, giving this scheme a significant performance advantage over any single-module scheme in terms of composite error identification accuracy. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall architecture of the current sensor error analysis system for multidimensional signal anomaly detection according to the present invention.

[0019] Figure 2 This is a detailed structural diagram of the multidimensional signal topology feature extraction module of the present invention.

[0020] Figure 3 This is a detailed structural diagram of the multi-channel timing correlation anomaly detection module of the present invention.

[0021] Figure 4 This is a detailed structural diagram of the online learning module for adaptive error characteristics under working conditions of the present invention.

[0022] Figure 5 This is a detailed structural diagram of the error component blind source separation module of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Those skilled in the art should understand that the following embodiments are for illustrative purposes only and should not be considered as limiting the scope of protection of the present invention.

[0024] See Figure 1 This invention provides a current sensor error analysis system for multi-dimensional signal anomaly detection, comprising a multi-dimensional signal topology feature extraction module 1, a multi-channel time-series correlation anomaly detection module 2, an online learning module for adaptive error characteristics under operating conditions 3, and an error component blind source separation module 4, all connected in a closed loop. The current sensor array consists of N sensors, where N is a positive integer ranging from 2 to 64. In typical vehicle electric drive system applications, N is 3; in typical smart charging pile applications, N is 16; and in typical energy storage system applications, N is 32. Each current sensor outputs a channel sampling signal after synchronous sampling by a synchronous sampling circuit. The channel sampling signal is connected to the system's internal data acquisition unit via a PCI Express bus or a Gigabit Ethernet bus. After A / D conversion, timestamp alignment, and channel number labeling in the data acquisition unit, the signal is sent to the multi-dimensional signal topology feature extraction module 1 for further processing.

[0025] See Figure 2 The multidimensional signal topology feature extraction module 1 includes a time delay embedding unit, a topology invariant calculation unit, and an error type discrimination unit. The time delay embedding unit receives the sampled signals from each channel and, based on the Takens embedding theorem, performs topology invariant calculation on the first-order signal. One-dimensional sampled signal sequence of the channel Perform a delay embedding transformation and map it to A set of trajectory points in a phase space. Embedding dimension. The false nearest neighbor ratio of each channel's sampled signal is adaptively selected, and the integer dimension determined when the false nearest neighbor ratio first drops below 5% is used as the final embedding dimension, with a typical value range of 6 to 10; delay parameter The time delay is adaptively selected based on the first minimum point of the mutual information function of the sampled signals from each channel, typically ranging from 2 to 8 sampling intervals. After the time delay embedding is completed, the... Channel phase space trajectory point cloud is shaped like of The vector is composed of dimensional vectors, where the superscripts are... This represents the vector transpose operator.

[0026] The topological invariant computation unit constructs simple complex sequences from the phase space trajectory point clouds of each channel using Vietoris-Rips filtering, and calculates the persistent homology group based on the GUDHI open-source topological data analysis library to extract the zero Viberty number. One-dimensional Betty number With continuous barcode length distribution Three types of topological invariant features. Among them The number of connected components in a phase space trajectory point cloud. Characterizing the number of loop structures in phase space trajectory point clouds, and the distribution of persistent barcode lengths. It represents the scale span from the appearance to the disappearance of each topological feature; the larger the length, the more stable the topological feature.

[0027] The error type discrimination unit discriminates the error type based on the deviation pattern of the topological invariant features relative to the pre-stored baseline topological features. A key finding of this invention is that the three typical errors of the current sensor exhibit distinguishable geometric deformation patterns on the phase space trajectory point cloud after time-delay embedding: the measurement model corresponding to the bias error is... After time-delay embedding, each point in the phase space trajectory point cloud is translated by the same vector. ,in for A vector of all 1s, therefore and All baseline values ​​are maintained, but the trajectory centroid shifts; the measurement model corresponding to the gain error is... After time-delay embedding, the phase space trajectory point cloud is scaled proportionally along each dimension. Therefore, and All maintain baseline values ​​while maintaining barcode length according to Overall scaling; the measurement model corresponding to nonlinear error is ,in For non-monotonic nonlinear functions or piecewise functions with saturation and dead-zone characteristics, the phase space trajectory folds and self-intersects after time-delay embedding, thus adding a one-dimensional hole. An integer-level abrupt change occurs, accompanied by a continuous shift in the barcode length distribution from unimodal to bimodal or multimodal. Based on this geometric deformation pattern, the error type discrimination unit employs a composite discrimination method to perform the discrimination.

[0028] The criteria for determining the centroid offset of the phase space trajectory point cloud are as follows:

[0029]

[0030] in: Let be the centroid offset of the current phase space trajectory point cloud, which is a scalar and has a value range of . The unit is A (ampere), which is calculated by this formula and is a geometric measure characterizing the bias error component; Let the current phase space trajectory point cloud centroid be, and let be A dimensional vector, whose dimensional vector is the first dimensional vector. The dimension value is the current phase space trajectory point cloud at the th position. The arithmetic mean over the dimension, with a range of values ​​of . The unit is A, which is obtained by cloud computing of the current phase space trajectory points and represents the geometric center of the current trajectory. Let the centroid of the point cloud be the baseline phase space trajectory point cloud, and let the centroid be... 1D vector, with the same range of values ​​as The unit is A, which is obtained by cloud computing from pre-stored baseline points and represents the geometric center of the trajectory under error-free conditions of the sensor. `L2 norm` is the operator that represents the Euclidean length of a vector. (Left side of the equation) The unit is A, on the right side of the equation Each component of the vector is in unit A. After L2 norm operation, the dimension of A is retained, and the dimensions on both sides are strictly consistent.

[0031] The scaling factor for continuous barcode length distribution is determined according to the following formula:

[0032]

[0033] in: The scaling factor is the total length of the continuous barcode, where is a scalar and its value range is [value range missing]. , dimensionless, is calculated by this formula, and is a geometric measure characterizing the gain error component; The first point cloud of the current phase space trajectory The length of a continuous barcode is a scalar, with a value ranging from 1 to 2. The unit is A, which is calculated from the persistent homology group and characterizes the scale persistence of the topological feature; The first point cloud of the baseline phase space trajectory The length of a continuous barcode is a scalar, and the unit is the same. It is obtained by cloud computing from pre-stored baseline points; The total number of continuous barcodes is a positive integer, with a value ranging from 1 to 2. , dimensionless, determined by the results of calculations from the continuous homology group; This is a continuous barcode number, a positive integer, with a value range of [value range missing]. Dimensionless, used to iterate over all continuous barcodes. Left side of the equation. Since A is a dimensionless quantity, the numerator and denominator on the right side of the equation are both sums of units A. After division, the units cancel each other out, resulting in a dimensionless quantity. The dimensions on both sides are strictly consistent.

[0034] The nonlinear error discrimination criterion is based on the following formula for the kurtosis index of the continuous barcode length distribution:

[0035]

[0036] The specific judgment rule is: when and and When neither of them changed by an integer order, the bias error was determined to be dominant. The centroid Euclidean distance threshold, with a value range of [value range missing]. , This is the rated current value of the current sensor. When... and and When no integer-level changes occur, the gain error is determined to be dominant. This is the threshold for the scaling ratio of the continuous barcode length, typically set to 0.05. Integer-level sudden increase or The time-based determination is dominated by nonlinear error, among which To ensure the kurtosis threshold for the continuous barcode length distribution, it is set to 1.5 times the baseline kurtosis value obtained by sampling each channel under zero current conditions.

[0037] When multiple discrimination conditions are met simultaneously, the error type discrimination unit uses the normalized distance of each topological index from the baseline as weights to perform a composite proportion estimation of multiple error types:

[0038]

[0039] in: For the first The composite weight of the error type is a scalar, and its value range is... , dimensionless, is calculated by this formula and characterizes the relative contribution of this type of error to the composite error; For the first The absolute distance of a topological index from the baseline is a scalar, and its value ranges from [value missing]. The unit is consistent with the original unit of the topological index, and is obtained by subtracting each topological index from the baseline value and taking the absolute value. For the first The standard deviation of the topological index on the baseline data is a scalar with a range of values ​​of 1. The units are consistent with the original units of the topological index, obtained from baseline data statistics, and are used to normalize indices of different dimensions to the same scale; subscript Traversing a collection These correspond to four types of topological indicators: centroid offset, barcode scaling, one-dimensional Betty number, and kurtosis; subscript For the traversal variable in normalized summation, the range of values ​​is the same as... The left side of the equation For dimensionless quantities, each term on the right side of the equation The dimensions of the quantity have been canceled out in the numerator and denominator, resulting in a dimensionless quantity, and the dimensions on both sides are strictly consistent.

[0040] In its specific implementation, the time delay embedding unit, topological invariant calculation unit, and error type discrimination unit of the multidimensional signal topology feature extraction module 1 share the same sliding time window data buffer area, and the sliding time window length is... In the application scenario of vehicle electric drive system, the typical sampling points are 1024; in the application scenario of smart charging pile, the typical sampling points are 2048; and in the application scenario of energy storage system, the typical sampling points are 4096, to match the balance requirements of error analysis response speed and statistical stability in different application scenarios. Upper bound of the scale parameter of Vietoris-Rips filtration. The filtering resolution is set to 100 scale levels by default, based on the 95th percentile of the inter-point distance distribution of the current phase space trajectory point cloud, to achieve a balance between computational complexity and topological feature resolution. The calculation of the continuous homology group utilizes the SimplexTree data structure and continuous homology algorithm from the GUDHI open-source topology data analysis library. On the NVIDIA Jetson AGX Orin edge computing platform, the continuous homology calculation for a single-channel 2048-point phase space trajectory point cloud takes approximately 50ms to 150ms. For the 16-channel parallel continuous homology calculation, the overall time can be controlled within 300ms after acceleration via the CUDA pipeline, meeting the real-time requirements of online processing.

[0041] The multidimensional signal topology feature extraction module 1 outputs the error type discrimination result and the corresponding composite weight vector, which are then fed into the multi-channel time-series correlation anomaly detection module 2 as prior constraints. The baseline topology features of the multidimensional signal topology feature extraction module 1 are obtained during the initial system deployment by loading an error-free signal onto a current sensor array using a calibration-grade current source. The baseline topology feature data is stored in the system's built-in non-volatile memory in a structured binary format. The geometric discrimination mapping relationship of the three types of errors in the phase space topology feature space can be intuitively understood as follows: the bias error causes the phase space trajectory point cloud to move along... The 2D space diagonal is translated but the trajectory shape remains unchanged; the gain error causes the phase space trajectory point cloud to move along... The topological connectivity of the trajectory remains unchanged when each dimension of the phase space is scaled up or down proportionally; nonlinear errors cause the phase space trajectory point cloud to fold, self-intersect, or branch, which manifests topologically as the addition of a one-dimensional hole or splitting into multiple connected components.

[0042] See Figure 3The multi-channel time-series correlation anomaly detection module 2 includes a causal relationship matrix estimation unit, a spectral norm anomaly detection unit, and a graph difference node localization unit. The causal relationship matrix estimation unit receives the error type output by the multi-dimensional signal topology feature extraction module 1 and performs online estimation of the time-series causal relationship matrix for each channel's sampled signal within a sliding time window. The online estimation of the time-series causal relationship matrix uses the Granger causality estimation algorithm, specifically implemented as follows: taking the first... The passage is Before the moment The historical sampling points are used as historical data for the predicted variable, with the first historical sampling point as the first historical data. The passage is Before the moment Using historical sampling points as candidate explanatory variables, a vector autoregressive model is used to perform a least-squares fit on both variables. The significance of the fit residuals is determined by comparing the variance using only the historical data of the predicted variable with the variance using the historical data of the candidate explanatory variables. The F-statistic is then used to determine the significance of the first historical sampling point. Channel to the first Does the channel exhibit Granger causality? The value of the F-statistic is used as a basis for determining the time-series causality matrix. The absolute value of the element's weight is recorded as a positive weight when the statistic is significant and as zero when it is not significant. The sliding time window length typically ranges from 256 to 4096 sampling points, depending on the sensor sampling rate and the application's dynamic response requirements.

[0043] During the system power-on initialization phase or after baseline update, the causal relationship matrix estimation unit continuously collects causal relationship matrices for several sliding time windows under normal operating conditions and takes their arithmetic mean as the baseline causal relationship matrix, which is stored in the system baseline library, denoted as . During online operation, the causal relationship matrix estimation unit outputs the current causal relationship matrix for each sliding time window. .

[0044] The spectral norm anomaly detection unit uses the change in the spectral norm of the current sliding time window causality matrix relative to the baseline causality matrix as the multi-channel collaborative anomaly detection index, calculated according to the following formula:

[0045]

[0046] in: Let be the change in spectral norm, and be a scalar with a range of values ​​of . , dimensionless, is calculated by this formula, and characterizes the overall degree of abnormality in the multi-channel causal structure; Let be the causal relationship matrix of the current sliding time window. Square formation, its first element For the first Channel to the first The significance weights of the channel Granger causality F-statistic, with values ​​ranging from [value range missing]. Dimensionless, output online by the causal relationship matrix estimation unit, representing the current causal structure; The baseline causal relationship matrix has the same dimensions as the baseline causal relationship matrix. The element values ​​and units are the same. The data is read from the system baseline library and characterizes the multi-channel causal structure of the sensor under error-free operating conditions. In a matrix, it is defined as the maximum singular value of the matrix, i.e. ,in Operator for the largest eigenvalue of a matrix; The total number of channels in the current sensor array is a positive integer, typically ranging from 2 to 64. Both sides of the equation are dimensionless quantities with strictly consistent dimensions.

[0047] when The time-triggered edge-by-edge comparison operation of the graph differential node localization unit is performed, where The spectral norm anomaly threshold is set at three times the statistical standard deviation of the spectral norm during baseline acquisition by the sensor array.

[0048] The graph differential node localization unit performs an edge-by-edge comparison between the current causal relationship matrix and the baseline causal relationship matrix, calculating the edge change for each edge:

[0049]

[0050] in: For the first The change in the edge value is a scalar, and its range is [value range missing]. , dimensionless, is calculated by this formula, and represents the magnitude of change in the causal weight of the edge. If a directional change also occurs, an additional directional penalty term is added. and The definition is the same as the formula mentioned above; Here, is the penalty coefficient for directional change, and is a scalar with a range of values. Dimensionless, with a default value of 1.0. The larger the value, the more sensitive it is to changes in the direction of cause and effect. This is an indicator function operator; it takes a value of 1 when the condition is met and a value of 0 when the condition is not met. For sign function operators, positive numbers output +1, negative numbers output -1, and zero outputs 0; This is the scalar absolute value operator. Both sides of the equation are dimensionless quantities with strictly consistent dimensions.

[0051] Subsequently, the graph differential node localization unit defines the first... The node anomaly of a channel is the sum of the changes in all edges that use that channel as their start and end point:

[0052]

[0053] in: For the first The node anomaly of the channel is a scalar, and its value ranges from 1 to 2. , dimensionless, is calculated by this formula and characterizes the overall intensity of the changes in causal relationships that this channel participates in in a multi-channel causal graph; and The definition is the same as the formula mentioned above; The target channel number has a value range of [value range missing]. ; The channel numbers are those other than the target channel, and the traversal range is... ; Let be the total number of channels in the current sensor array. Both sides of the equation are dimensionless quantities with strictly consistent dimensions.

[0054] Will Exceeding the node anomaly threshold The channel output is a list of candidate channels for error sources, where The node anomaly threshold is set at 1.5 times the 95th percentile of the node anomaly distribution during the baseline acquisition period. The candidate channel list of error sources and the error types output by the multidimensional signal topology feature extraction module 1 together constitute the joint identification result of error type and channel, which is then fed into the online learning module 3 for adaptive error characteristics under operating conditions.

[0055] The Granger causality matrix estimation in the multi-channel time-series correlation anomaly detection module 2 employs an adaptive selection strategy for the order of the vector autoregressive model. The baseline causality matrix is ​​jointly determined by the Akaike information criterion and the Bayesian information criterion during each update, with a typical value range of 3 to 10. When the error type output by the multidimensional signal topology feature extraction module 1 is bias error dominant, the multi-channel time-series correlation anomaly detection module 2 will apply a directional change penalty coefficient. The value was dynamically adjusted from the default value of 1.0 to 0.5 because the bias error mainly changes the absolute value of the causal weights and less so the causal direction; when the error type is dominated by nonlinear error, The dynamic adjustment to 2.0 is due to the fact that nonlinear errors may induce new nonlinear coupling paths between multiple channels, easily leading to causal directional changes. This is based on prior knowledge of the error type. The dynamic adjustment mechanism enables the explicit consumption of the error type prior output by the multi-dimensional signal topology feature extraction module 1 in the multi-channel time-series correlation anomaly detection module 2, reflecting the deep coupling between the two modules.

[0056] See Figure 4 The online learning module 3 for adaptive error characteristics includes an encoder, latent space, decoder, operating condition model mapping library, and nearest neighbor retrieval compensation unit. The encoder, latent space, and decoder together constitute a variational autoencoder. The output of the operating condition parameter acquisition unit is connected to the input of the online learning module 3, and the output of the incremental learning update unit is connected to the nearest neighbor retrieval compensation unit of the online learning module 3. The operating condition parameter acquisition unit acquires operating condition parameters of the sensor's environment, including temperature parameters (obtained by a Pt1000 platinum resistance temperature sensor mounted near the sensor core, with a sampling rate of 10Hz and a measurement range of -50℃ to 150℃), frequency parameters (obtained by the zero-crossing detection circuit of the main circuit, with a measurement range of DC to 5000Hz), and DC bias parameters (obtained by the DC current component extraction unit, with a measurement range of 0 to 120% of the rated current). These three types of operating condition parameters are concatenated to form an operating condition parameter vector. .

[0057] The variational autoencoder employs a dual-network structure comprising an encoder and a decoder. The encoder consists of three fully connected layers. The first layer has an input dimension of 3 (corresponding to the three types of operating parameters), while the second and third layers have 64 and 32 hidden units, respectively. The ReLU activation function is used. The encoder ultimately outputs the mean vector. With log-variance vector The two are concatenated to define a latent space variable. The posterior distribution of the latent space variables. Sampling using reparameterization techniques:

[0058]

[0059] in: For implicit space variables, A dimensional vector, with values ​​ranging from 1 to 2. , dimensionless, obtained by sampling from this formula, characterizes the manifold mapping of the operating parameters in the latent space; Let be the mean vector of the posterior distribution in the latent space, and be... A 3D vector, output by the encoder neural network; Let be the standard deviation vector of the posterior distribution in the latent space, and be... The dimensional vector is obtained by taking the exponent of the logarithmic variance output by the encoder. For standard normal sampling noise, dimensional vector, following distributed; The element-wise multiplication operator for vectors; Let be the dimension of the latent space, a positive integer, typically 8 or 16. Both sides of the equation are dimensionless quantities with strictly consistent dimensions.

[0060] The decoder consists of three fully connected layers, with an input dimension of... The output dimension is an error characteristic parameter vector. The dimension of the variable autoencoder (typically 6, corresponding to the magnitude estimation and frequency feature estimation of the three types of errors, respectively) is symmetric to the encoder. The training objective of the variational autoencoder uses the ELBO (lower bound of evidence) loss function.

[0061]

[0062] in: is the ELBO loss function value, which is a scalar and dimensionless quantity, calculated by this formula, and represents the training loss of the variational autoencoder; The log-likelihood expectation term represents the accuracy of the decoder's reconstruction error characteristic parameter vector; The Kullback-Leibler divergence operator represents the posterior distribution. With prior distribution The distribution differences between (taking the standard normal distribution) are used to constrain the latent space manifold structure; Here, is the weight coefficient of the KL divergence regularization term, and is a scalar with a range of values. , dimensionless, with a default value of 1.0. A value that is too small results in a loose latent space structure, while a value that is too large will suppress reconstruction accuracy. Both sides of the equation are dimensionless quantities, and their dimensions are strictly consistent.

[0063] During the offline training phase, the variational autoencoder uses error characteristic data collected at Cartesian product grid points covering the three operating conditions (temperature, frequency, and DC bias) as training samples to complete the training. After training, the latent space vectors corresponding to each operating condition grid point are... Local compensation parameters for error characteristics They are stored in pairs in the working condition model mapping library.

[0064] The nearest neighbor search compensation unit uses current operating parameters when the system is running online. The current latent space vector is obtained through forward propagation of the encoder. Then, a nearest neighbor search is performed in the working condition model mapping library:

[0065]

[0066] in: The nearest neighbor grid point number is a positive integer, with a value range of [value range missing]. , dimensionless, is calculated by this formula, and represents the offline training condition most similar to the current working condition; The latent space vector obtained by encoding the current operating condition parameters is defined in the same way as the aforementioned formula; For the first The latent space vectors corresponding to each offline training grid point are stored during the offline training phase. The total number of entries in the working condition model mapping library, a positive integer, typically ranging from 200 to 2000; The grid points are numbered according to the operating conditions, and the range of values ​​is: Both sides of the equation are dimensionless quantities, and their dimensions are strictly consistent.

[0067] Nearest neighbor retrieval compensation unit call Corresponding error characteristic local compensation parameters The error data of the candidate error source channels are subjected to adaptive compensation based on operating conditions. The online learning module 3 for adaptive error characteristics performs incremental learning on newly acquired operating condition-error characteristic data with an update cycle of 1 hour, so as to achieve continuous adaptation to characteristic drift caused by sensor aging.

[0068] The incremental learning of the online learning module 3 for adaptive error characteristics under operating conditions employs an elastic weight consolidation strategy to avoid catastrophic forgetting. Specifically, during each hourly incremental learning iteration, a quadratic regularization term, weighted by the Fisher information matrix, is applied to all weight parameters of the encoder and decoder. This ensures that parameters critical to historical operating condition performance remain stable, while parameters insensitive to historical operating conditions can be freely updated with newly acquired data. The offline training dataset for the operating condition model mapping library is obtained by performing Cartesian product grid sampling on the following parameters: temperature operating conditions from -40°C to 120°C with a step size of 10°C; frequency operating conditions from 10Hz to 5000Hz with a logarithmic scale and 4 sampling points per octave; and DC bias operating conditions from 0 to 120% of the rated current value with a step size of 10%. For the extrapolation region where the operating parameters fall outside the grid, the nearest neighbor retrieval compensation unit queries the nearest neighbor grid point in the latent space, and then further linearly corrects the local compensation parameters based on the residual vector between the current operating parameters and the operating parameters of the nearest neighbor grid point through the decoder Jacobian matrix, thereby avoiding the systematic deviation generated by the nearest neighbor retrieval in the latent space in the extrapolation region.

[0069] See Figure 5 The error component blind source separation module 4 includes a constrained independent component analysis (CFI) unit, a Kalman smoothing compensation branch, an adaptive Wiener filtering branch, an electromagnetic interference spectrum matching branch, and a closed-loop feedback unit. The CFI unit takes the residual error after adaptive compensation output by the online learning module 3 as input. It assumes that the residual error is a linear mixture of three independent error sources, and its source signal separation model is as follows: ,in The three-dimensional independent component vectors correspond to the systematic bias component, the random noise component, and the interference coupling component, respectively. The residual error observation vector, This is the separation matrix.

[0070] The constrained independent component analysis unit uses the kurtosis and negative entropy indices as joint independence constraints to perform optimization of the separation matrix:

[0071]

[0072] in: Let be the value of the joint independence objective function, be a scalar, dimensionless, calculated by this formula. Maximizing this objective function is equivalent to maximizing the mutual independence between the independent components. Let be the separation matrix, and be The matrix is ​​dimensionless and is solved iteratively by an optimization algorithm. It represents a linear mapping from residual error observations to independent components. and are the weighting coefficients for the kurtosis and negative entropy terms, respectively, and is a scalar with a value range of . The sum of the two is equal to 1, which is dimensionless, and the default value is 0.5. For the first Each component is an independent scalar; For mathematical expectation operators; For the calculation of the approximate value of negative entropy, this invention adopts a non-quadratic function. ,in The negative entropy approximation parameter has a value range of [value range missing]. The default value is 1.0; To and Standard normal random variables with the same variance are used as the reference distribution for calculating negative entropy; For each independent component number, the value range is: Both sides of the equation are dimensionless quantities, and their dimensions are strictly consistent.

[0073] The separation matrix obtained by solving After applying the residual error, three independent components are obtained, which are then fed into three parallel compensation branches.

[0074] The Kalman smoothing compensation branch performs optimal estimation of the systematic bias components. Because the systematic bias components drift slowly over time, the Kalman smoothing compensation branch establishes a first-order random walk state-space model, with the state transition equation being: The observation equation is ,in for Systematic bias component state estimation at time t. and These are process noise and observation noise, respectively. The Kalman smoothing compensation branch uses the Rauch-Tung-Striebel smoother. Perform the optimal estimation, and its estimated value As a reverse compensation quantity applied to the original measurement signal:

[0075]

[0076] in: The measurement signal after systematic bias component compensation is given; is a scalar with a value range of . The unit is A, which is calculated by this formula and represents the measurement signal after the first stage of compensation. The original measurement signal is a scalar quantity in A, output by the current sensor array. for The Kalman smoothed estimate of the time-series systematic bias component is a scalar in unit A, output by the Rauch-Tung-Striebel smoother. Both sides of the equation are scalars in unit A, with strictly consistent dimensions.

[0077] The adaptive Wiener filter branch suppresses random noise components. It uses the LMS adaptive algorithm to update the tap coefficients of the Wiener filter online, minimizing the mean square error of the filter output and achieving optimal linear estimation and suppression of random noise components.

[0078] The electromagnetic interference (EMI) spectrum matching branch performs source identification for interference coupling components. This branch performs classification matching using three feature spectral libraries: power system frequency integer multiple harmonics spectrum, pulse width modulation carrier spectrum, and radio frequency electromagnetic interference spectrum. Specifically, a fast Fourier transform is performed on the interference coupling components to obtain their spectra. Then, the cosine similarity between the spectrum and each feature spectral library is calculated. The interference source category corresponding to the feature spectral library with the highest cosine similarity exceeding the matching threshold (typically 0.7) is output as the source identification result.

[0079] The closed-loop feedback unit will compensate the three types of components in the purified signal. The feedback is sent to the multidimensional signal topology feature extraction module 1 to perform incremental updates on the baseline topology feature library. The baseline topology feature update rule is executed according to the following formula:

[0080]

[0081] in: The updated baseline topology feature vector has the same dimensions. The updated baseline topology reference is calculated using this formula. The baseline topological feature vector before the update, whose components correspond to... , , , , as well as Equal baseline quantities, with units consistent with their respective topological quantities; For purification signal The topological feature vectors obtained through time-delay embedding reconstruction, Vietoris-Rips filtering, and continuous cohomology calculation have the same dimension and units. ; The learning rate is updated to the baseline, where is a scalar and its value ranges from 1 to 2. Dimensionless, with a default value of 0.01. Values ​​that are too large can cause baseline oscillation instability, while values ​​that are too small will result in a sluggish response to sensor aging and drift. The units of each component on both sides of the equation are consistent with the corresponding topological quantities, ensuring strict dimensionality.

[0082] Each time the closed-loop feedback unit completes a baseline update, it synchronously triggers the baseline causality matrix of the multi-channel time-series correlation anomaly detection module 2. Based on the new purification signal, the system baseline is re-statistically migrated adaptively as the sensor drifts over a long period of time, forming a complete closed-loop processing link from error type discrimination through channel-level positioning, adaptive compensation for operating conditions, blind source separation of error components, and baseline self-updating.

[0083] In a typical application scenario of a 16-channel current sensor array for smart charging piles, the current sensor array consists of 16 LEM LF 1005-S type closed-loop Hall current sensors with a rated current of 1000A and a primary-to-secondary turns ratio of 1000:1. The secondary output current is converted into a voltage signal by a 10Ω precision sampling resistor and then synchronously sampled by a 16-channel synchronous sampling ADC (typical model AD7606) at a sampling rate of 100kHz. The sampling signals of each channel are sent to the industrial control computer (typical model Siemens IPC847E) running the system via Gigabit Ethernet. The GPU accelerator card (typical model NVIDIA Jetson AGX Orin) embedded in the industrial control computer is responsible for the Vietoris-Rips filtering calculation in the multi-dimensional signal topology feature extraction module 1 and the variational autoencoder forward inference in the online learning module 3 for adaptive error characteristics. The CPU is responsible for Granger causality estimation and graph difference calculation in the multi-channel time-series correlation anomaly detection module 2 and constrained independent component analysis and Kalman smoothing in the error component blind source separation module 4.

[0084] After artificially injecting typical composite errors (0.8% gain error plus 0.3% nonlinear error into the 3rd channel and 0.5% bias error into the 11th channel) into the sixteen-channel current sensor array, the multi-dimensional signal topology feature extraction module 1 of this system identified that the scaling ratio of the total length of the continuous barcode in the 3rd channel deviated from the baseline by 0.008 and the kurtosis index reached 1.6 times the baseline value, and the centroid offset of the phase space trajectory point cloud in the 11th channel reached 8.2A. Based on this, it was determined that the 3rd channel was a gain-nonlinear composite error and the 11th channel was a bias error. The multi-channel time-series correlation anomaly detection module 2 identified that the two channels with the highest node anomaly were the 3rd and 11th channels, and the channel-level positioning was accurate. The working condition adaptive error characteristic online learning module 3 called the local model of the corresponding hidden space region under the 25℃ working condition to perform working condition adaptive compensation on the above two channels. The error component blind source separation module 4 further separated the residual into three components: bias, noise and interference, and compensated them independently. Finally, the baseline was updated by the feedback of the purified signal. The error type discrimination accuracy reached 96.3% in 1000 independent injection experiments, the channel-level positioning accuracy reached 96.8%, the wide temperature range comprehensive error was reduced by 62.4% compared with the fixed offline calibration scheme, and the composite error margin was reduced by 69.8% compared with the overall filtering scheme. All indicators are consistent with the expected results of this invention.

[0085] To verify the performance advantages of this invention over existing technologies, the inventors deployed the system of this invention (Scheme A), a traditional error compensation scheme based on single-channel harmonic amplitude statistics (Scheme B, implemented according to the technical route of CN119224675A), and an error correction scheme based on multi-sensor data fusion prediction (Scheme C, implemented according to the technical route of CN119004059A) on the same sixteen-channel current sensor array hardware platform, respectively, and compared the error identification and compensation performance of the three schemes under the same test conditions.

[0086] The test conditions were set as follows: the temperature range was from -40℃ to 120℃ with a scanning step of 10℃; the fundamental frequency of the measured current covered the range from DC to 3000Hz; and the injection error scenarios covered seven combinations: single bias error, single gain error, single nonlinear error, bias-gain dual composite error, bias-nonlinear dual composite error, gain-nonlinear dual composite error, and three types of error triple composite. Each combination was injected independently 50 times at each temperature point, and the identification and compensation results were statistically analyzed.

[0087] Regarding the accuracy of error type discrimination, Scheme A of this invention achieves an average discrimination accuracy of 96.3% across 7 error combinations. Scheme B achieves 78.4% accuracy in a single error scenario, but its accuracy drops to 42.7% and 23.1% in dual and triple composite scenarios, respectively. Scheme C achieves 85.2% accuracy in a single error scenario, but its accuracy decreases to 58.3% in dual composite scenarios and 31.6% in triple composite scenarios. The fundamental difference lies in the fact that Scheme B uses single-channel harmonic amplitude statistics as the discrimination criterion, which leads to mutual masking of amplitude statistics under composite errors, causing discrimination failure. Scheme C uses comprehensive statistics after multi-sensor data fusion as the discrimination criterion, which, although introducing multi-channel information, fails to perceive the inherent geometric structure of the signal. Under strong nonlinear errors, the statistical distribution deviates from the Gaussian assumption, resulting in decreased discrimination accuracy. Scheme A uses phase space topological invariants as the discrimination criterion, where the geometric deformation patterns of the three types of errors in the topological feature space are independent and geometrically unmasked, maintaining a discrimination accuracy above 95% in composite error scenarios.

[0088] Regarding the error source channel-level positioning performance, Scheme A of this invention achieves a single-channel error source positioning accuracy of 96.8% and a dual-channel composite positioning accuracy of 93.2% in a 16-channel array. Scheme B, lacking channel-level positioning capability, can only output an overall alarm for the current sensor array. Scheme C can identify abnormal channels among multiple channels, but its accuracy in dual-channel composite positioning scenarios is only 64.5%. The fundamental difference lies in the fact that Scheme A uses differential edge-by-edge comparison of the spectral norm anomaly triggering graph of the Granger causality matrix, enabling precise positioning of specific channels even if errors occur simultaneously in two channels. Scheme C estimates abnormal channels based on the comprehensive deviation after data fusion, limiting its ability to distinguish between errors occurring simultaneously in multiple channels.

[0089] In terms of comprehensive error index over a wide temperature range, the comprehensive relative error of Scheme A in this invention is 0.18% in the range of -40℃ to 120℃, the comprehensive relative error of Scheme B is 0.48%, and the comprehensive relative error of Scheme C is 0.34%. Scheme A reduces the error by 62.5% compared to Scheme B and by 47.1% compared to Scheme C. The fundamental difference lies in the fact that Scheme A replaces the fixed offline calibration table with a variational autoencoder operating condition latent space mapping, which can continuously call the nearest neighbor local model without interpolation distortion when the operating condition parameters change continuously; Scheme B performs lookup compensation using a pre-stored temperature coefficient table, which produces a significant step when the temperature operating condition crosses the table grid boundary; Scheme C, although possessing a certain dynamic adjustment capability, does not manifoldize the nonlinear mapping relationship between operating condition parameters and error characteristics.

[0090] Regarding the composite error margin index, the error margin of Scheme A in this invention is 0.09% under the three composite error scenarios, Scheme B is 0.31%, and Scheme C is 0.22%. Scheme A reduces the margin by 71.0% compared to Scheme B and by 59.1% compared to Scheme C. The fundamental difference lies in the fact that Scheme A uses constrained independent component analysis to separate the composite error into independent components, which are then processed by three branches: Kalman smoothing, adaptive Wiener filtering, and electromagnetic interference spectrum matching. Each branch performs optimal compensation for its respective error physical origin. Schemes B and C perform overall filtering or overall regression compensation on the composite error, which cannot distinguish the independent components within the error, leading to conflicts between compensation actions.

[0091] The above comparative embodiments fully verify the systematic performance advantages of the present invention over the prior art in four dimensions: error type identification, channel-level positioning, wide temperature range accuracy, and composite error component compensation. They also experimentally confirm the nonlinear synergistic gain of 1+1+1+1>4 generated by the deep coupling closed loop of the four modules: topology perception, multi-channel causal positioning, working condition hidden space mapping, and error component blind source separation.

[0092] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A current sensor error analysis system for multidimensional signal anomaly detection, characterized in that, The system includes: The multidimensional signal topology feature extraction module is used to perform time delay embedding reconstruction on the sampled signals of each channel of the current sensor array to obtain the phase space trajectory point cloud of each channel. The topology invariant features containing Betti number and continuous barcode are extracted from the phase space trajectory point cloud of each channel using the topology data analysis method. The error type is determined by the deviation pattern of the topology invariant features relative to the pre-stored baseline topology features. A multi-channel temporal correlation anomaly detection module is used to receive the error type output by the multi-dimensional signal topology feature extraction module, perform online estimation of the temporal causal relationship matrix for each channel sampled signal within a sliding time window, use the change in spectral norm of the current sliding time window causal relationship matrix relative to the baseline causal relationship matrix as the multi-channel collaborative anomaly detection index, and perform edge-by-edge comparison of the two causal relationship matrices using a graph difference algorithm to locate nodes where the causal relationship weights change abnormally as candidate error source channels. The candidate error source channels carry the error type as the error type-channel joint identification result. The online learning module for adaptive error characteristics under operating conditions is used to perform latent space representation learning on the error characteristics of the sensor under different combinations of operating condition parameters using an encoder-decoder network, and to perform nearest neighbor retrieval in the latent space using the current operating condition parameters, and to call the local model of error characteristics of the corresponding operating condition region to perform adaptive compensation for the error data of the candidate channels of the error source. The error component blind source separation module is used to perform error component separation on the residual error after adaptive compensation of the working condition using the blind source separation algorithm to obtain three independent error components: systematic bias component, random noise component and interference coupling component. The original measurement signal is then independently compensated by the compensation branch corresponding to each component. The compensated and purified signal is fed back to the multi-dimensional signal topology feature extraction module to perform baseline topology feature update. The multidimensional signal topology feature extraction module constructs a simple complex by using Vietoris-Rips filtering for the phase space trajectory point cloud of each channel and calculates the persistent homology group. The zero-dimensional Betti number represents the number of connected components in the phase space trajectory point cloud, the one-dimensional Betti number represents the number of loops in the phase space trajectory point cloud, and the persistent barcode length distribution represents the scale persistence of each topology feature. In the error type discrimination, the bias error discrimination condition is that the centroid of the phase space trajectory point cloud shifts while both the zero-dimensional Betti number and the one-dimensional Betti number maintain the baseline value; the gain error discrimination condition is that the persistent barcode length distribution is scaled by a constant ratio while both the zero-dimensional Betti number and the one-dimensional Betti number maintain the baseline value; and the nonlinear error discrimination condition is that the one-dimensional Betti number undergoes an integer-level abrupt change and the persistent barcode length distribution changes from a single mode to a dual mode or multimode.

2. The current sensor error analysis system for multidimensional signal anomaly detection according to claim 1, characterized in that, The multidimensional signal topology feature extraction module adopts a composite discrimination method in the error type discrimination: when the centroid offset of the phase space trajectory point cloud exceeds the baseline centroid Euclidean distance threshold and neither the zero-dimensional Betti number nor the one-dimensional Betti number has changed by an integer order, the bias error is determined to be dominant; when the ratio of the change in the total length of the continuous barcode relative to the baseline exceeds a preset threshold and neither the zero-dimensional Betti number nor the one-dimensional Betti number has changed by an integer order, the gain error is determined to be dominant; when the one-dimensional Betti number shows an integer-order sudden increase or the kurtosis index of the continuous barcode length distribution exceeds the baseline kurtosis threshold, the nonlinear error is determined to be dominant. When multiple discrimination conditions are met simultaneously, the normalized distance of each topological index from the baseline is used as the weight to perform composite proportion estimation on multiple types of errors.

3. The current sensor error analysis system for multidimensional signal anomaly detection according to claim 2, characterized in that, The online estimation of the temporal causality matrix in the multi-channel temporal correlation anomaly detection module adopts the Granger causality estimation algorithm. The system also stores a baseline causality matrix pre-sampled under normal operating conditions. The multi-channel temporal correlation anomaly detection module uses the change in the spectral norm of the current sliding time pane Granger causality matrix relative to the baseline causality matrix as an early detection indicator for multi-channel collaborative anomalies. When the change in the spectral norm exceeds a preset anomaly threshold, the graph difference algorithm is triggered to compare edge by edge. Nodes whose causal weights have undergone directional changes or whose absolute weight changes exceed the edge change threshold are output as a candidate channel list of error sources.

4. The current sensor error analysis system for multidimensional signal anomaly detection according to claim 3, characterized in that, The encoder-decoder network of the online learning module for adaptive error characteristics under operating conditions adopts a variational autoencoder. The encoder of the variational autoencoder maps the operating condition parameter vector, which includes temperature operating condition parameters, frequency operating condition parameters, and DC bias operating condition parameters, into a latent space vector. The decoder of the variational autoencoder restores the latent space vector into an error characteristic parameter vector. The latent space is constrained by a Kullback-Leibler divergence regularization term and is approximately a multivariate Gaussian distribution. In the adaptive compensation under operating conditions, the online learning module for adaptive error characteristics under operating conditions performs nearest neighbor search in the latent space with the current operating condition parameters and calls the local error characteristic model associated with the nearest neighbor operating condition region to perform compensation on the error data of the candidate channels of the error source.

5. The current sensor error analysis system for multidimensional signal anomaly detection according to claim 4, characterized in that, The blind source separation algorithm of the error component blind source separation module adopts the constrained independent component analysis algorithm. The error component blind source separation module uses the kurtosis index and the negative entropy index as a joint independence constraint to decompose the residual error into three independent components: systematic bias component, random noise component, and interference coupling component. In the compensation branch corresponding to each component, the systematic bias component performs optimal estimation and back-compensation to the original measurement signal using Kalman smoothing, the random noise component performs suppression using adaptive Wiener filtering, and the interference coupling component performs source identification using electromagnetic interference spectrum feature matching. The error component blind source separation module feeds back the cleaned signal after compensation of the three types of components to the multi-dimensional signal topology feature extraction module to perform incremental update of the baseline topology features, forming a closed-loop processing link from error type discrimination to error component compensation and then to baseline self-updating.

6. The current sensor error analysis system for multidimensional signal anomaly detection according to claim 1, characterized in that, In the time delay embedding reconstruction of the multidimensional signal topology feature extraction module, the embedding dimension is adaptively selected based on the false nearest neighbor rate index of the sampled signals of each channel, and the time delay parameter is adaptively selected based on the first minimum point of the mutual information function of the sampled signals of each channel.

7. The current sensor error analysis system for multidimensional signal anomaly detection according to claim 1, characterized in that, The multidimensional signal topology feature extraction module sets the kurtosis threshold of the continuous barcode length distribution used for nonlinear error discrimination to be 1.5 times the baseline kurtosis value obtained by sampling each channel under zero current conditions.

8. The current sensor error analysis system for multidimensional signal anomaly detection according to claim 5, characterized in that, The error component blind source separation module performs electromagnetic interference spectrum feature matching on the interference coupling component using three feature spectrum libraries: power system power frequency integer multiple harmonic spectrum, pulse width modulation carrier spectrum, and radio frequency electromagnetic interference spectrum, and marks the interference source category.

9. The current sensor error analysis system for multidimensional signal anomaly detection according to claim 1, characterized in that, The encoding-decoding network of the online learning module for adaptive error characteristics of operating conditions performs incremental learning on newly collected operating condition-error characteristic data with an update cycle of 1 hour.

Images