Electric power and magnetic power dynamic joint detection method and system based on multivariable fusion

Abstract

The invention provides an electric power and magnetic power dynamic joint detection method and system based on multivariable fusion, and relates to the technical field of power electronic equipment detection. According to the method, voltage, current and magnetic field sensors are arranged in a to-be-tested electromagnetic system, original physical quantity data are collected, and a multivariable time sequence data set is constructed; through preprocessing and feature extraction, a dynamic feature matrix related to electric power and magnetic power is generated; constructing a joint estimation model based on the matrix, and comparing the joint estimation model with a reference value to obtain a detection residual error; generating a state label according to the residual size and trend, and evaluating the estimation credibility in combination with a confidence interval; and a feedback mechanism is further constructed, model parameters and feature structures are optimized by adopting an adaptive algorithm, and online iteration and self-learning are realized. The method is high in detection precision, high in state recognition capability, good in adaptability and robustness, and suitable for energy flow monitoring and intelligent diagnosis of a complex electromagnetic system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power electronic device detection, in particular to a dynamic joint detection method and system for electric power and magnetic power based on multivariate fusion. BACKGROUND

[0002] With the rapid development of high-frequency power electronic devices, wireless energy transmission systems and smart grids, the coupling of electromagnetic energy in multiple physical fields is increasingly significant. The traditional electric power detection method (such as the instantaneous power method based on voltage and current sampling) cannot fully reflect the actual energy flow state in the system. Especially under the conditions of complex carrier frequency, multi-frequency excitation and multi-load nonlinearity, the electric power and magnetic power in the system present dynamic interaction and non-stationary variation characteristics. Some studies propose to use magnetic field probes or magnetometers to monitor the change of magnetic flux, but their monitoring accuracy and dynamic response capability are still limited.

[0003] The current research trend is shifting from single energy form detection to multi-physical field fusion detection. Multivariate fusion information such as voltage, current, magnetic flux density, power factor, frequency drift, etc. is used to establish a dynamic energy flow modeling mechanism for the system. At the algorithm level, machine learning, Kalman filtering, adaptive signal fusion and other methods are introduced to improve the real-time performance and robustness of dynamic detection. At the hardware level, a composite detection platform integrating electromagnetic sensors, vector network analyzers and synchronous acquisition modules is gradually formed, but there is still a lack of high-robustness and low-latency dynamic joint detection technology path in complex electromagnetic environments.

[0004] The existing methods generally have the following shortcomings:

[0005] (1) The electric power and magnetic power detection mechanisms are independent, and a linkage model cannot be established, resulting in a lack of complementary and verification mechanisms between the detection results;

[0006] (2) Multivariate dynamic correlation modeling is not introduced, making it difficult to deal with energy fluctuations and mutations under non-stationary conditions;

[0007] (3) The data acquisition path and fusion strategy are disconnected, resulting in large system delay and poor robustness;

[0008] (4) The existing systems lack adaptive parameter adjustment mechanisms and cannot adjust feature weights or detection parameters for different application scenarios.

[0009] Therefore, there is an urgent need for a dynamic detection method based on multivariate fusion modeling and joint consideration of electric power and magnetic power variation rules. SUMMARY

[0010] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide a method and system for dynamic joint detection of electric power and magnetic power based on multivariable fusion, which realizes dynamic collaborative modeling and intelligent feedback optimization of electric power and magnetic power, and improves the accuracy and adaptability of energy detection of complex electromagnetic systems.

[0011] To achieve the above object, the present invention provides the following solutions:

[0012] A dynamic joint detection method of electric power and magnetic power based on multivariable fusion, comprising:

[0013] Multiple voltage, current, and magnetic field sensors are deployed in the electromagnetic system to be tested to collect raw physical quantity data including voltage signals, current signals, power factor, frequency, magnetic flux density, and magnetic induction intensity. The signals of the raw physical quantity data are time-series aligned through a multi-channel synchronous acquisition device to construct a unified multivariate time series data set;

[0014] Preprocessing the multivariate time series data set to obtain preprocessed data, and extracting electric power-related features and magnetic power-related features from the preprocessed data based on a time domain, frequency domain, or time-frequency domain analysis method to form a multi-source dynamic feature matrix;

[0015] constructing a joint estimation model of electric power and magnetic power based on the dynamic characteristic matrix, and comparing the electric power and magnetic power values ​​output by the joint estimation model with reference values ​​independently derived from the original physical quantity data to calculate a detection residual;

[0016] Based on the size and evolution trend of the detection residual, a state discrimination label of the current power state is generated. At the same time, a confidence interval is calculated based on the distribution range of the detection residual to quantify the uncertainty of the current power estimation result and assist in determining the credibility level of the state discrimination;

[0017] A feedback mechanism is constructed based on the detection residual, the state discrimination label and the historical estimation accuracy, and an adaptive algorithm is used to iteratively optimize the fusion parameters, weight distribution and feature dimensions of the joint estimation model to achieve self-learning and robustness enhancement of the dynamic detection system.

[0018] A dynamic joint detection system of electric power and magnetic power based on multivariable fusion, comprising:

[0019] A multi-source signal acquisition module is used to deploy multiple voltage, current, and magnetic field sensors in the electromagnetic system to be tested, respectively collecting raw physical quantity data including voltage signals, current signals, power factor, frequency, magnetic flux density, and magnetic induction intensity, and to perform time-series alignment on the signals of the raw physical quantity data through a multi-channel synchronous acquisition device to construct a unified multivariate time series data set;

[0020] a feature preprocessing and extraction module, configured to preprocess the multivariate time series dataset to obtain preprocessed data, and extract electric power related features and magnetic power related features in the preprocessed data based on time domain, frequency domain or time-frequency domain analysis methods, to form a multi-source dynamic feature matrix;

[0021] a joint power modeling module, configured to construct a joint estimation model of electric power and magnetic power based on the dynamic feature matrix, and compare electric power and magnetic power values output by the joint estimation model with reference values independently derived according to the original physical quantity data, to calculate detection residuals;

[0022] a reference power calculation and residual evaluation module, configured to generate a state discrimination label of a current power state based on a size and evolution trend of the detection residuals, and calculate a confidence interval based on a distribution range of the detection residuals, to quantify uncertainty of a current power estimation result and assist in judging a confidence level of state discrimination;

[0023] a feedback optimization and model adaptation module, configured to construct a feedback mechanism based on the detection residuals, the state discrimination label and historical estimation accuracy, and iteratively optimize fusion parameters, weight distribution and feature dimensions of the joint estimation model using an adaptive algorithm, to realize self-learning and robustness enhancement of a dynamic detection system.

[0024] According to the specific embodiments provided by the present application, the following technical effects are disclosed:

[0025] The present application provides a multi-variable fusion based electric power and magnetic power dynamic joint detection method and system. The present application can break through the limitations of traditional methods of independent observation and static evaluation of electromagnetic power by constructing a joint detection process of multi-source fusion of electric power and magnetic power. It first constructs a unified multivariate time series dataset through synchronous acquisition of multiple physical quantities, extracts electromagnetic features with time-frequency information on this basis, and realizes comprehensive modeling of system power behavior; secondly, through residual calculation and state discrimination, it can not only identify abnormal operating states in real time, but also quantify estimation uncertainty; finally, combined with a feedback mechanism and an adaptive optimization algorithm, the model is given dynamic learning and robustness enhancement capability, which can continuously adapt to energy flow changes in complex electromagnetic environments, and improves the accuracy, stability and intelligent level of the detection system. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0027] Figure 1 A method flowchart is provided for the embodiments of the present application.

[0028] Figure 2 A system structure schematic diagram is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0030] The purpose of the present application is to provide a multi-variable fusion-based dynamic joint detection method and system for electric power and magnetic power, which realizes dynamic collaborative modeling and intelligent feedback optimization for electric power and magnetic power, and improves the accuracy and adaptive ability of energy detection of complex electromagnetic systems.

[0031] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0032] Figure 1 A method flowchart is provided for the embodiments of the present application, as shown in Figure 1 The present application provides a multi-variable fusion-based dynamic joint detection method for electric power and magnetic power, which comprises:

[0033] Step 100: A plurality of voltage, current and magnetic field sensors are arranged in the electromagnetic system to be detected, and original physical quantity data including voltage signals, current signals, power factors, frequencies, magnetic flux densities and magnetic induction intensities are collected, and the signals of the original physical quantity data are time-aligned through a multi-channel synchronous acquisition device to construct a unified multi-variable time series data set;

[0034] Step 200: The multi-variable time series data set is preprocessed to obtain preprocessed data, and based on a time domain, frequency domain or time-frequency domain analysis method, electric power-related features and magnetic power-related features in the preprocessed data are extracted to form a multi-source dynamic feature matrix;

[0035] Step 300: Based on the dynamic feature matrix, a joint estimation model for electric power and magnetic power is constructed, and the electric power and magnetic power values output by the joint estimation model are compared with reference values independently derived according to the original physical quantity data to calculate detection residuals;

[0036] Step 400: based on the size and evolution trend of the detection residual, a state discrimination label of the current power state is generated, and a confidence interval is calculated based on the distribution range of the detection residual to quantify the uncertainty of the current power estimation result and assist in judging the credibility level of state discrimination;

[0037] Step 500: based on the detection residual, state discrimination label and historical estimation accuracy, a feedback mechanism is constructed, and an adaptive algorithm is used to iteratively optimize the fusion parameters, weight distribution and feature dimension of the joint estimation model to realize self-learning and robustness enhancement of the dynamic detection system.

[0038] Preferably, a plurality of voltage sensors, current sensors and magnetic field sensors are arranged in the electromagnetic system to be measured to collect raw physical quantity data, and a multi-channel synchronous acquisition device is used to time-align the signals of the raw physical quantity data, and a unified multivariate time series data set is constructed, including:

[0039] Voltage sensors, current sensors and magnetic field sensors are arranged at the input end, output end and key energy flow path of the electromagnetic system to be measured;

[0040] The voltage sensors and current sensors are used to collect three-phase or single-phase voltage and current waveform data, and to calculate power factor and operating frequency parameters in real time;

[0041] The magnetic field sensors are used to continuously collect magnetic induction intensity values and magnetic flux density values at designated monitoring points in the electromagnetic system to be measured within a preset sampling period to obtain measurement samples;

[0042] The measurement samples at different times are organized in time sequence as magnetic power related time series data for representing the dynamic changes of the magnetic field during system operation;

[0043] The three-phase or single-phase voltage, current waveform data, power factor, operating frequency parameters and time series data are determined as the raw physical quantity data;

[0044] The raw physical quantity data is input into a multi-channel synchronous acquisition device for time stamp labeling, inter-channel time alignment and sampling frequency uniformization processing to obtain preprocessed data;

[0045] The preprocessed data is organized into a multi-dimensional multivariate time series data set according to a unified time reference.

[0046] In this embodiment, preferably, a number of voltage sensors, current sensors and magnetic field sensors are respectively arranged at the input end, output end and key energy flow path position of the electromagnetic system to be measured, for real-time collection of raw physical quantity data such as voltage, current and magnetic field. The voltage sensor and current sensor can be selected from Hall or photoelectric sensors with high sampling accuracy, which can respectively collect three-phase or single-phase voltage and current waveform data, and calculate the real-time power factor and operating frequency parameters based on the synchronous data stream. At the same time, the magnetic field sensor preferably adopts a high-sensitivity fluxgate or Hall type sensor that supports vector measurement, which is set in the sensitive area of ​​the system magnetic flux change, and continuously collects the magnetic induction intensity value and magnetic flux density value at the target monitoring point within the set sampling period (such as every 1 to 5 ms) to constitute the original measurement sample related to magnetic power.

[0047] Subsequently, the three-phase or single-phase voltage waveform data, current waveform data, power factor, frequency parameters, and time series samples of magnetic induction intensity and magnetic flux density collected above are input into a multi-channel synchronous acquisition device as raw physical quantity data. The multi-channel synchronous acquisition device has a high-precision clock synchronization mechanism and a unified sampling frequency control logic, which can perform timestamp marking, channel alignment, and unified sampling frequency processing on the data of all channels, thereby ensuring the consistency of the registration of different physical quantities in the time domain. Finally, the pre-processed data stream is reorganized according to a unified time base to construct a multivariate time series data set with consistent time series attributes and feature dimensions, which serves as the input basis for the subsequent feature extraction and joint modeling modules.

[0048] Preferably, the multivariate time series data set is preprocessed to obtain preprocessed data, and based on a time domain, frequency domain or time-frequency domain analysis method, electric power-related features and magnetic power-related features in the preprocessed data are extracted to form a multi-source dynamic feature matrix, including:

[0049] Normalizing each physical quantity channel in the multivariate time series data set so that different physical quantities have comparable numerical scales, thereby obtaining normalized data;

[0050] The normalized data is subjected to denoising by using a median filtering method to eliminate high-frequency interference and transient mutation anomalies in the signal to obtain filtered data;

[0051] According to the time series structure of the filtered data, the signal is respectively extracted using time domain analysis, frequency domain analysis or time-frequency domain analysis methods to obtain the electric power-related features and the magnetic power-related features; the electric power-related features include: instantaneous electric power, effective electric power, power factor and its fluctuation range, voltage-current phase difference change rate, harmonic distortion rate and main frequency power spectrum peak; the magnetic power-related features include: magnetic induction intensity change rate with time, unit area magnetic flux change, hysteresis loop area, magnetic flux density offset and fluctuation coefficient in a typical sampling interval, and local magnetic field energy spectrum distribution;

[0052] A multi-source dynamic feature matrix is ​​constructed based on the electric power-related features and the magnetic power-related features; each column of the multi-source dynamic feature matrix corresponds to a physical quantity feature, and each row corresponds to a multi-variable state at the same time point.

[0053] In this embodiment, preferably, the multivariate time series data set must first undergo standardized preprocessing and feature extraction processing before being input into the joint power modeling module. Specifically, normalization processing is performed on each type of original physical quantity channel in the data set (including voltage, current, power factor, frequency, magnetic induction intensity and magnetic flux density) to uniformly map different physical quantities to a comparable numerical range, usually normalized to the [0,1] interval or the standard normal distribution range to avoid the interference of the numerical scale on the modeling accuracy. Subsequently, the normalized data is denoised using the median filtering method, which suppresses the extreme values ​​and mutation values ​​of the signal within the time window, thereby effectively eliminating transient noise caused by high-frequency interference, measurement fluctuations or pulse anomalies, and obtaining stable and continuous filtered data.

[0054] Based on the time series structure of the filtered data, time domain analysis, frequency domain analysis, and time-frequency domain analysis methods are used to extract multiple characteristic variables reflecting the energy change behavior of the system. Among them, electric power-related features include: instantaneous electric power (obtained by multiplying the instantaneous voltage and current), effective electric power (calculated based on sliding window average or root mean square), power factor and its fluctuation range (reflecting energy utilization efficiency), the phase difference change rate between voltage and current, total harmonic distortion (THD), and the power spectrum peak at the main frequency (obtained by FFT or PSD calculation). Magnetic power-related features include: the time derivative of magnetic induction intensity (dB / dt), the change in magnetic flux per unit area (calculated based on the area of ​​the sensor layout region), the area of ​​the hysteresis loop (representing magnetic energy loss), the offset value and standard deviation of the magnetic flux density within a typical sampling interval (reflecting local fluctuations), and the local magnetic field energy spectrum distribution characteristics obtained by wavelet packet decomposition, Hilbert transform, or energy spectrum density analysis. By aligning all the characteristic variables in rows according to a unified timestamp, a multi-source dynamic characteristic matrix can be constructed. Each column in the matrix corresponds to a physical quantity feature, and each row represents the system state at the same time point, which is used for subsequent joint estimation modeling.

[0055] Preferably, the expression of the multi-source dynamic feature matrix is:

[0056]

[0057] Wherein, F is a multi-source dynamic feature matrix, which represents the multivariate feature sequence of the electric power related features and the magnetic power related features extracted at K consecutive sampling moments; f e (t k ) is at time point t k The extracted electric power related feature vector, f m (t k ) is the magnetic power related feature vector extracted at the same time point; n e is the number of electric power related features, n m is the number of magnetic power related features; K is the total number of sampling time points; the electric power related feature subvector f e (t k ) includes the following components: instantaneous electric power, effective electric power, power factor, rate of change of phase difference between voltage and current, harmonic distortion rate and peak power spectrum density at main frequency; instantaneous electric power is the point-by-point product of voltage and current; effective electric power represents the average active power within a certain time window; power factor reflects the efficiency of electric energy utilization; rate of change of phase difference between voltage and current is used to reflect the fluctuation of load properties; harmonic distortion rate is used to measure the proportion of non-fundamental components in the signal; peak power spectrum density at main frequency represents the main energy frequency component; the magnetic power related characteristic subvector fm (t k ) includes the following components: the rate of change of magnetic induction intensity with time, the change of magnetic flux per unit area, the area of ​​hysteresis loop, the offset value and fluctuation coefficient of magnetic flux density in a typical sampling interval, and the characteristic value of the local magnetic field energy spectrum; the rate of change of magnetic induction intensity with time is used to measure the dynamic response intensity of the magnetic field; the change of magnetic flux per unit area is the increment of magnetic flux per unit area over time; the area of ​​hysteresis loop represents the magnetic energy consumed in the unit magnetic flux loop; the offset value and fluctuation coefficient of magnetic flux density in a typical sampling interval reflect the fluctuation characteristics of the magnetic field; the characteristic value of the local magnetic field energy spectrum is used to capture the energy focusing situation in the local frequency domain.

[0058] Specifically, in this embodiment, to construct a multi-source dynamic feature matrix that can comprehensively characterize the system's electrical and magnetic power states, the preprocessed data of different physical quantities is first synchronized and aligned based on a unified time series, ensuring that all features have corresponding values ​​at each sampling moment. At each time point, the extracted power-related features include instantaneous power, effective power, power factor, the rate of change of the phase difference between voltage and current, harmonic distortion, and the peak power spectrum density of the main frequency. Instantaneous power is calculated by point-by-point multiplication of the collected voltage and current signals; effective power is calculated based on the mean within a sliding time window, reflecting the average energy flow; power factor is calculated from the cosine of the phase angle between voltage and current and serves as a measure of energy utilization efficiency; the rate of change of the phase difference is used to capture dynamic fluctuations in electrical load properties; harmonic distortion is calculated based on the harmonic content of the signal using a fast Fourier transform; and the peak power spectrum density of the main frequency is identified using power spectrum density analysis to identify the frequency point with the highest energy concentration.

[0059] In terms of constructing magnetic power features, the time rate of change of magnetic induction intensity, the change in magnetic flux per unit area, the area of ​​the hysteresis loop, the offset and fluctuation coefficient of magnetic flux density within a typical sampling interval, and the eigenvalues ​​of the local magnetic field energy spectrum are similarly extracted at each sampling time point. The rate of change of magnetic induction intensity is estimated by the difference between the sampled values ​​of the magnetic induction sensor at consecutive time points and is used to characterize the magnetic field response speed. The change in magnetic flux per unit area is obtained by multiplying the area of ​​the sensor monitoring area by the magnetic induction increment. The area of ​​the hysteresis loop is obtained by integrating the magnetic field-flux relationship curve and is used to describe energy loss. The offset and fluctuation coefficient of magnetic flux density are calculated based on the mean and standard deviation within a statistical window to reflect magnetic field stability. The eigenvalues ​​of the local magnetic field energy spectrum are extracted using wavelet packet decomposition, Hilbert transform, and other methods to extract characteristic parameters of energy-concentrated regions in the frequency domain. All of these features are structurally organized as "feature types" and "behaviors at unified time points," ultimately forming a two-dimensional multi-source dynamic feature matrix that serves as the input for the subsequent joint estimation model.

[0060] Preferably, based on the dynamic characteristic matrix, a joint estimation model of electric power and magnetic power is constructed, and the electric power and magnetic power values ​​output by the joint estimation model are compared with reference values ​​independently derived from the original physical quantity data to calculate the detection residual, including:

[0061] A joint power estimation model is constructed using the multi-source dynamic characteristic matrix as input; the joint power estimation model uses a multivariate weighted combination method to model each characteristic variable and outputs an estimated electric power value and an estimated magnetic power value corresponding to each moment;

[0062] Based on the same time point as the characteristic matrix, using the corresponding voltage, current, magnetic induction intensity and magnetic flux in the collected original physical quantity data, a reference electric power value and a reference magnetic power value are independently calculated through physical analytical expressions as actual observation benchmarks;

[0063] A moment-by-moment difference operation is performed on the estimated electric power value, the estimated magnetic power value, the reference electric power value, and the reference magnetic power value to obtain an electric power residual sequence and a magnetic power residual sequence to form the detection residual.

[0064] Preferably, the expression of the joint power estimation model is:

[0065]

[0066] in: At time t k Estimated electric power value; At time t k Estimated magnetic power value; [f e (t k )|f m (t k )] T is the column vector after feature concatenation, with dimension (n e +n m )×1; V is the weight matrix from input feature to hidden layer, with dimension d×(n e +n m );σ(·) is the nonlinear activation function; W is the weight matrix from hidden layer to output, with dimension 2×d; b1 is the hidden layer bias term, with dimension d×1; b2 is the output bias term, with dimension 2×1; d is the hidden layer dimension, which is set according to the model complexity.

[0067] Preferably, the calculation formula of the reference electric power value is:

[0068]

[0069] in: For the tth kReference electric power value at the moment; U(t k ) is the instantaneous voltage collected by the voltage sensor; I(t k ) is the instantaneous current collected by the current sensor; -φ(t k ) is the instantaneous phase difference between voltage and current;

[0070] The calculation formula of the reference magnetic power value is:

[0071]

[0072] in: For the tth k Reference magnetic power at the moment; is the time derivative of magnetic induction intensity, which comes from the magnetic field sensor; A eff is the effective projected area under the action of the magnetic field; μ is the magnetic permeability of the magnetic medium.

[0073] Specifically, in this embodiment, preferably, the multi-source dynamic feature matrix is ​​used as the input data for joint modeling, and the feature variables of each column in the matrix are first spliced ​​and combined to form a unified feature input vector sequence. At each sampling time point, the electric power-related features and magnetic power-related features extracted at the corresponding time point are spliced ​​into a set of ordered vectors by column, and sent as input samples to the joint power estimation model. The estimation model used in this embodiment is a multivariate weighted combination model with nonlinear mapping capabilities. The preferred implementation form is a feedforward neural network with a layer of implicit structure, whose input is a multi-source feature column vector, and the output is the estimated electric power and magnetic power values ​​at that time point.

[0074] In terms of model structure, the number of input layer nodes is equal to the total number of dimensions of the eigenvector, and the number of hidden layer nodes can be set according to the system complexity and modeling accuracy requirements, with a preferred range of 10 to 50 nodes. The input weight matrix is ​​used to project the eigenvector into the hidden layer space; after the hidden layer is processed by a nonlinear activation function (such as hyperbolic tangent, ReLU or Sigmoid function), it works together with the output weight matrix to finally output the estimated electric power and estimated magnetic power values ​​corresponding to the current moment. The network training process can use an adaptive gradient descent method with momentum, or an integrated Bayesian regularization strategy to optimize the parameters, with the goal of minimizing the sum of squared residuals between the output value and the true power value.

[0075] In order to obtain a true reference power value as a benchmark for joint estimation model training and residual evaluation, an independent derivation method based on the analytical relationship of physical quantities is used to calculate the electric power and magnetic power. Specifically, the reference electric power value is directly calculated from the instantaneous waveform data synchronously collected by the voltage sensor and the current sensor, where the amplitude of the voltage and current and their instantaneous phase difference serve as the core input parameters. In this embodiment, the power factor is estimated by the cosine value of the phase angle, and the voltage and current values ​​are converted into effective value inputs according to the RMS principle to ensure that the calculation results of the reference power are physically consistent. For three-phase systems, symmetrical component transformation can be used to convert the three-phase power into equivalent total power.

[0076] In the calculation of the reference value of magnetic power, it is preferred to calculate the change in magnetic energy flux per unit time based on the rate of change of the magnetic induction intensity collected by the magnetic field sensor at consecutive moments, combined with the effective magnetic receiving area of ​​the sensor installation location and the magnetic permeability of the material in which it is located. The magnetic field effective area can be measured by calibrating the device geometry or the sensor projection coverage range, and the magnetic permeability is set by looking up the material properties of the medium in which the sensor is located. In the calculation, the differential method is used to estimate the derivative of the magnetic induction intensity with respect to time. The sampling period is set by the data acquisition system, and the typical value range is 1 to 10ms. The calculated reference electric power and magnetic power values ​​are respectively subjected to a one-to-one moment-by-moment difference operation with the output of the joint model estimation to form an electric power residual sequence and a magnetic power residual sequence, which in turn constitute a complete detection residual for subsequent state evaluation and model optimization feedback.

[0077] Preferably, based on the size and evolution trend of the detection residual, a state discrimination label of the current power state is generated. At the same time, a confidence interval is calculated based on the distribution range of the detection residual to quantify the uncertainty of the current power estimation result and assist in determining the credibility level of the state discrimination, including:

[0078] Get the joint power estimation model at each sampling time t k The residual between the estimated output power value and the reference power value is recorded as the electric power residual Magnetic power residual and calculate the absolute value to assess the magnitude of the instantaneous estimation error;

[0079] Within a preset time sliding window, the residual absolute value sequence of each type of residual is counted, and the maximum value, average value, standard deviation and change trend slope are extracted to characterize the overall intensity and dynamic evolution characteristics of the power estimation deviation; the overall intensity is the residual amplitude;

[0080] According to a set multi-level state discrimination rule, the state discrimination label of the current power state is generated based on the residual amplitude, wherein when the mean value and the maximum value of the residual absolute value are lower than a first threshold value and the slope is close to zero, a "normal" state label is assigned; when the residual mean value or the maximum value is in an intermediate interval or shows a sustained upward trend, a "fluctuation" state label is assigned; when the residual maximum value exceeds a second threshold value or the slope abnormally increases, an "abnormal" state label is assigned.

[0081] Based on the distribution characteristics of the residual absolute value in the sliding window, an estimated confidence interval is calculated, which represents the reliable range of the power estimation result in a statistical sense, and the expression is:

[0082] CI(t k )=[μ r (t k )-α·σ r (t k ),μ r (t k )+α·σ r (t k )]

[0083] Wherein, μ r (t k ) is the mean value of the residual absolute value in the current window, σ r (t k ) is the standard deviation, and α is a confidence coefficient for setting the confidence level.

[0084] According to the state discrimination label and the width of the corresponding confidence interval, the confidence level of the current estimation result is jointly judged, wherein if the state discrimination label is "normal" and the confidence interval width is less than a preset threshold, it is determined that the current estimation result has high confidence; if the state discrimination label is "abnormal" and the confidence interval width is greater than or equal to the preset threshold, it is determined that the current estimation result has greater uncertainty and low confidence.

[0085] In this embodiment, preferably, the system obtains the estimated electric power value and the estimated magnetic power value output by the joint power estimation model at each sampling moment, and performs difference calculations with the reference electric power value and the reference magnetic power value respectively to obtain the electric power residual and the magnetic power residual. To facilitate subsequent state judgment and statistical analysis, the absolute value of the residual value at each moment is first taken as the instantaneous error amplitude of the estimated deviation at that moment. All residual absolute values ​​are arranged continuously in chronological order to form a residual absolute value sequence. Within each sliding statistical window (for example, every 20 consecutive sampling moments), the system extracts the statistical features of the residual absolute value sequence within the window, including the maximum value, mean value, standard deviation, and slope of the linear trend. Among them, the maximum value and mean value are used to characterize the overall intensity of the power estimation error, that is, the residual amplitude; the slope is used to reflect the evolution trend of the residual over time, thereby assisting in judging whether the system is stable or fluctuating.

[0086] Based on the residual amplitude and trend indicators extracted above, the system generates a state discrimination label for the current power state according to the preset state discrimination rules. Specifically, when the mean and maximum values ​​of the residuals are both lower than the first threshold, and the slope of the trend change is close to zero, it indicates that the power estimation result is stable and reliable, and the "normal" state label is assigned at this time; when the mean or maximum value of the residuals is in the middle threshold range, or the slope shows a gradual upward trend, it indicates that the fluctuation of the system power estimation is aggravated, and the "fluctuation" state label is assigned at this time; if the maximum value of the residual exceeds the second threshold, or the slope increases suddenly, it indicates that the system has significant abnormal behavior, and the estimation model is difficult to effectively track the true state, and the "abnormal" state label is assigned at this time. The above-mentioned first and second thresholds can be set by clustering or statistically analyzing multiple groups of historical power residual data in the offline training phase, and can also be dynamically and adaptively set based on the system usage scenario.

[0087] To further quantify the confidence level of the estimated value, the system performs statistical analysis on the residual absolute value sequence within the sliding window and constructs a confidence interval based on its mean and standard deviation, reflecting the statistically reliable bounds of the estimated result. The confidence level is determined by a confidence coefficient, which is configured based on the risk tolerance of the target application scenario. For example, a larger confidence coefficient can be set for a high-reliability scenario. Based on the upper and lower bounds of the confidence interval, the confidence level of the estimated result is determined in conjunction with the current state discrimination label. If the state is "normal" and the confidence interval width is less than the preset confidence tolerance, the estimate is considered highly reliable. If the state is "abnormal" and the confidence interval width exceeds the tolerance, the estimate is considered to have significant uncertainty and a low confidence level. This preset confidence tolerance is obtained through empirical statistical analysis of the confidence interval distribution characteristics in multiple typical system operation samples and can be dynamically updated during actual operation.

[0088] Preferably, a feedback mechanism is constructed based on the detection residual, the state discrimination label and the historical estimation accuracy, and an adaptive algorithm is used to iteratively optimize the fusion parameters, weight distribution and feature dimension of the joint estimation model to achieve self-learning and robustness enhancement of the dynamic detection system, including:

[0089] A feedback signal index set is constructed, and a feedback state vector is formed; the feedback signal index set includes: the mean, variance and slope of the current detected residual sequence, the corresponding power state label, and the historical estimation accuracy in the past N detection periods;

[0090] It is judged whether the feedback state vector satisfies the trigger condition, and when any of the following conditions is met, the model optimization mechanism is started: the current state label is "abnormal" and the residual mean is greater than the first adaptive threshold; the historical estimation accuracy continuously decreases by more than a set number of times; the current confidence interval width exceeds the upper limit of system stability;

[0091] In the optimization process of the model optimization mechanism, the following components of the joint estimation model are iteratively updated based on the feedback state vector:

[0092] The feature fusion weight in the joint estimation model is optimized; the feature fusion weight is used to adjust the contribution degree of the multi-source feature after input layer splicing to the estimation output, and the initial value of the feature fusion weight is obtained from the training process. Subsequently, it is dynamically adjusted according to the residual contribution rate to improve the response weight of important feature dimensions;

[0093] The parameter matrix inside the joint estimation model is updated, and adaptive gradient descent, Kalman filtering or Bayesian update method is used to minimize the residual function value in the current period, so as to improve the fitting ability of the model to nonlinear feature combination;

[0094] The feature dimension screening strategy of the joint estimation model input is adjusted, and the correlation or information gain degree between each feature and the residual is sorted to dynamically eliminate redundant features with low contribution to estimation accuracy or high collinearity, so as to reduce the model complexity and improve the generalization ability;

[0095] The optimized feature fusion weight, parameter matrix and feature dimension screening strategy are reloaded into the structure of the joint estimation model, and the power estimation and residual analysis of the next period are continued to realize online iterative optimization and long-term self-learning update of the detection model.

[0096] In the present embodiment, to realize online adaptive optimization of the joint power estimation model, the present embodiment first constructs a feedback signal index set in each detection period, collects the electrical power residual and magnetic power residual sequences in the current detection period, calculates the statistical features thereof, including the mean, variance, and change trend slope; simultaneously records the power state label corresponding to the current detection period, and the historical estimation accuracy record in the past preset period number (such as N=10), for example, the mean square error or residual mean. The above feedback signal indexes are organized into a structured feedback state vector, which is used as the core basis for judging whether to start the model optimization. To improve the response capability, the present embodiment sets three trigger conditions, and automatically starts the model adaptive optimization mechanism when any condition is met, specifically including: the current state label is "abnormal" and the residual mean exceeds the first adaptive threshold; the historical estimation accuracy continuously decreases by more than a preset number of times (such as 3 times); the confidence interval width of the current detection period exceeds the stability upper limit allowed by the present embodiment, indicating that the response capability of the model to the actual power change decreases.

[0097] Once the feedback state vector meets any of the above trigger conditions, the present embodiment will activate the adaptive optimization process of the joint power estimation model. In this process, first, the feature fusion weight vector of the model is adjusted, which is used to control the contribution degree of different input feature variables to the final power estimation result. In the initial training stage, the weight is obtained by fitting historical data, while in the online optimization stage, the present embodiment updates the weight according to the correlation between each feature dimension and the residual change, so that the features that are highly sensitive to estimation error change obtain higher weight, and noise features are gradually suppressed, thereby improving the modeling accuracy and stability. The optimization process can be realized using the residual sensitivity sorting method or the feature contribution rate backtracking mechanism, which has the characteristics of low algorithm implementation threshold and high execution efficiency.

[0098] Subsequently, the present embodiment iteratively updates the internal parameter matrix of the joint estimation model. Preferably, the adaptive gradient descent method can be selected to quickly adjust the weight coefficients; when it is necessary to improve the robustness to time series fluctuations, the Kalman filter method can be selected to realize smoothing correction; if dynamic evaluation combined with model uncertainty is required, the Bayesian update method can be selected for prior parameter adjustment. At the same time, the present embodiment further optimizes the structure of the feature input dimension, sorts the correlation coefficients or information gain values between the features and the residuals from high to low, filters and retains the main feature dimensions, and dynamically eliminates redundant features that have no significant contribution to estimation or even introduce errors. After all optimization operations are completed, the new weight, parameter, and input structure configuration are loaded into the joint power estimation model, and are used for data input in the next detection period, realizing online iterative update and continuous self-learning of the present embodiment, and significantly improving the generalization ability and long-term robustness of the model in dynamic scenarios.

[0099] Corresponding to the above method, as Figure 2As shown, this embodiment also provides a dynamic joint detection system of electric power and magnetic power based on multivariable fusion, including:

[0100] A multi-source signal acquisition module is used to deploy multiple voltage, current, and magnetic field sensors in the electromagnetic system to be tested, respectively collecting raw physical quantity data including voltage signals, current signals, power factor, frequency, magnetic flux density, and magnetic induction intensity, and to perform time-series alignment on the signals of the raw physical quantity data through a multi-channel synchronous acquisition device to construct a unified multivariate time series data set;

[0101] a feature preprocessing and extraction module, configured to preprocess the multivariate time series data set to obtain preprocessed data, and extract electric power-related features and magnetic power-related features from the preprocessed data based on a time domain, frequency domain, or time-frequency domain analysis method to form a multi-source dynamic feature matrix;

[0102] a joint power modeling module, configured to construct a joint estimation model of electric power and magnetic power based on the dynamic characteristic matrix, and compare the electric power and magnetic power values ​​output by the joint estimation model with reference values ​​independently derived from the original physical quantity data to calculate a detection residual;

[0103] A reference power calculation and residual evaluation module is used to generate a state discrimination label for the current power state based on the size and evolution trend of the detection residual. At the same time, a confidence interval is calculated based on the distribution range of the detection residual to quantify the uncertainty of the current power estimation result and assist in determining the confidence level of the state discrimination;

[0104] The feedback optimization and model adaptation module is used to build a feedback mechanism based on the detection residual, the state discrimination label and the historical estimation accuracy, and adopt an adaptive algorithm to iteratively optimize the fusion parameters, weight distribution and feature dimensions of the joint estimation model to achieve self-learning and robustness enhancement of the dynamic detection system.

[0105] The beneficial effects of the present invention are as follows:

[0106] (1) By constructing a multi-source joint detection system for electric and magnetic power, this invention breaks through the limitations of traditional methods that only perform unidirectional power assessment based on voltage and current. By introducing magnetic power elements such as magnetic induction intensity and magnetic flux density, the system's comprehensive perception of complex electromagnetic energy flows is effectively enhanced. This system is particularly suitable for power electronics systems, wireless energy transmission devices, and electromagnetic drive equipment with strong coupling effects. By constructing a dynamic characteristic matrix, a unified expression of multi-physical quantity and multi-time scale energy characteristics is achieved, enhancing the input information integrity of the detection model.

[0107] (2) The application realizes the collaborative prediction of electric power and magnetic power by adopting a joint estimation model, fuses the nonlinear relationship of multi-source features in the modeling process, and effectively improves the adaptability of the model to dynamic changes and non-steady state conditions. Combined with the reference power value for residual analysis, the system state can be evaluated at each time, with high real-time and accuracy guarantee. At the same time, the power state label is established by the statistical trend of the residual, which provides a quantitative basis for the steady state identification and fluctuation warning of the system operation.

[0108] (3) The application further introduces a confidence interval calculation mechanism, realizes the dynamic evaluation of the detection confidence by measuring the uncertainty of the estimation result. Under the dual analysis of state label and confidence interval, the system can identify the stable interval and risk interval of the estimation result, and enhance the fault trend prediction ability. In addition, this method does not need to rely on a single threshold judgment, and has higher flexibility and robustness, which is suitable for deployment requirements and performance tolerance settings of various systems.

[0109] (4) Based on the residual feedback mechanism and the adaptive optimization algorithm driven by the state label, the application realizes the dynamic updating ability of the joint power estimation model, which can adjust the feature fusion weight, modeling parameter and input structure in real time according to the system running state, so as to realize continuous learning and precision enhancement. The optimization mechanism significantly improves the stability and generalization ability of the detection system in the long-term running process, has good online scalability and cross-system portability, and is suitable for deployment in intelligent power grid, industrial electric drive, electromagnetic manufacturing and other high dynamic electromagnetic environments.

[0110] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between the embodiments can be referred to each other.

[0111] In this paper, specific examples are used to illustrate the principles and implementation methods of the application. The above description of the embodiments is only used to help understand the method and its core idea; at the same time, for those skilled in the art, according to the idea of the application, the specific implementation method and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the application.

Claims

1. A dynamic joint detection method of electric power and magnetic power based on multivariable fusion, characterized in that: include: Multiple voltage, current, and magnetic field sensors are deployed in the electromagnetic system to be tested to collect raw physical quantity data including voltage signals, current signals, power factor, frequency, magnetic flux density, and magnetic induction intensity. The signals of the raw physical quantity data are time-series aligned through a multi-channel synchronous acquisition device to construct a unified multivariate time series data set; Preprocessing the multivariate time series data set to obtain preprocessed data, and extracting electric power-related features and magnetic power-related features from the preprocessed data based on a time domain, frequency domain, or time-frequency domain analysis method to form a multi-source dynamic feature matrix; constructing a joint estimation model of electric power and magnetic power based on the dynamic characteristic matrix, and comparing the electric power and magnetic power values ​​output by the joint estimation model with reference values ​​independently derived from the original physical quantity data to calculate a detection residual; Based on the size and evolution trend of the detection residual, a state discrimination label of the current power state is generated. At the same time, a confidence interval is calculated based on the distribution range of the detection residual to quantify the uncertainty of the current power estimation result and assist in determining the credibility level of the state discrimination; A feedback mechanism is constructed based on the detection residual, the state discrimination label and the historical estimation accuracy, and an adaptive algorithm is used to iteratively optimize the fusion parameters, weight distribution and feature dimensions of the joint estimation model to achieve self-learning and robustness enhancement of the dynamic detection system.

2. The method for dynamic joint detection of electric power and magnetic power based on multivariable fusion according to claim 1 is characterized in that: Multiple voltage sensors, current sensors, and magnetic field sensors are deployed in the electromagnetic system to collect raw physical quantity data. The signals of the raw physical quantity data are time-series aligned through a multi-channel synchronous acquisition device to construct a unified multivariate time series data set, including: Arrange voltage sensors, current sensors and magnetic field sensors on the input end, output end and key energy flow path of the electromagnetic system to be tested; The voltage sensor and current sensor are used to collect three-phase or single-phase voltage and current waveform data, and calculate the power factor and operating frequency parameters in real time; Utilizing the magnetic field sensor to continuously collect magnetic induction intensity values ​​and magnetic flux density values ​​at designated monitoring points in the electromagnetic system to be measured within a preset sampling period to obtain measurement samples; Organizing the measurement samples at different times into time series data related to magnetic power in chronological order, so as to characterize the dynamic changes of the magnetic field during the operation of the system; Determine the three-phase or single-phase voltage, the current waveform data, the power factor, the operating frequency parameter, and the time series data as the original physical quantity data; Inputting the raw physical quantity data into a multi-channel synchronous acquisition device for timestamp marking, inter-channel timing alignment and sampling frequency unification processing to obtain pre-processed data; The preprocessed data is organized into the multi-dimensional multivariate time series data set according to a unified time base.

3. The method for dynamic joint detection of electric power and magnetic power based on multivariable fusion according to claim 1 is characterized in that: Preprocessing the multivariate time series data set to obtain preprocessed data, and extracting electric power-related features and magnetic power-related features from the preprocessed data based on a time domain, frequency domain, or time-frequency domain analysis method to form a multi-source dynamic feature matrix, including: Normalizing each physical quantity channel in the multivariate time series data set so that different physical quantities have comparable numerical scales, thereby obtaining normalized data; The normalized data is subjected to denoising by using a median filtering method to eliminate high-frequency interference and transient mutation anomalies in the signal to obtain filtered data; According to the time series structure of the filtered data, the signal is respectively extracted using time domain analysis, frequency domain analysis or time-frequency domain analysis methods to obtain the electric power-related features and the magnetic power-related features; the electric power-related features include: instantaneous electric power, effective electric power, power factor and its fluctuation range, voltage-current phase difference change rate, harmonic distortion rate and main frequency power spectrum peak; the magnetic power-related features include: magnetic induction intensity change rate with time, unit area magnetic flux change, hysteresis loop area, magnetic flux density offset and fluctuation coefficient in a typical sampling interval, and local magnetic field energy spectrum distribution; A multi-source dynamic feature matrix is ​​constructed based on the electric power-related features and the magnetic power-related features; each column of the multi-source dynamic feature matrix corresponds to a physical quantity feature, and each row corresponds to a multi-variable state at the same time point.

4. The method for dynamic joint detection of electric power and magnetic power based on multivariable fusion according to claim 3 is characterized in that: The expression of the multi-source dynamic feature matrix is: Wherein, F is a multi-source dynamic feature matrix, which represents the multivariate feature sequence of the electric power related features and the magnetic power related features extracted at K consecutive sampling moments; f e (t k ) is at time point t k The extracted electric power related feature vector, f m (t k ) is the magnetic power related feature vector extracted at the same time point; n e is the number of electric power related features, n m is the number of magnetic power related features; K is the total number of sampling time points; the electric power related feature subvector f e (t k ) includes the following components: instantaneous electric power, effective electric power, power factor, rate of change of phase difference between voltage and current, harmonic distortion rate and peak power spectrum density at main frequency; instantaneous electric power is the point-by-point product of voltage and current; effective electric power represents the average active power within a certain time window; power factor reflects the efficiency of electric energy utilization; rate of change of phase difference between voltage and current is used to reflect the fluctuation of load properties; harmonic distortion rate is used to measure the proportion of non-fundamental components in the signal; peak power spectrum density at main frequency represents the main energy frequency component; the magnetic power related characteristic subvector f m (t k ) includes the following components: the rate of change of magnetic induction intensity with time, the change of magnetic flux per unit area, the area of ​​hysteresis loop, the offset value and fluctuation coefficient of magnetic flux density in a typical sampling interval, and the characteristic value of the local magnetic field energy spectrum; the rate of change of magnetic induction intensity with time is used to measure the dynamic response intensity of the magnetic field; the change of magnetic flux per unit area is the increment of magnetic flux per unit area over time; the area of ​​hysteresis loop represents the magnetic energy consumed in the unit magnetic flux loop; the offset value and fluctuation coefficient of magnetic flux density in a typical sampling interval reflect the fluctuation characteristics of the magnetic field; the characteristic value of the local magnetic field energy spectrum is used to capture the energy focusing situation in the local frequency domain.

5. The method for dynamic joint detection of electric power and magnetic power based on multivariable fusion according to claim 1 is characterized in that: Based on the dynamic characteristic matrix, a joint estimation model of electric power and magnetic power is constructed, and the electric power and magnetic power values ​​output by the joint estimation model are compared with reference values ​​independently derived based on the original physical quantity data to calculate the detection residual, including: A joint power estimation model is constructed using the multi-source dynamic characteristic matrix as input; the joint power estimation model uses a multivariate weighted combination method to model each characteristic variable and outputs an estimated electric power value and an estimated magnetic power value corresponding to each moment; Based on the same time point as the characteristic matrix, using the corresponding voltage, current, magnetic induction intensity and magnetic flux in the collected original physical quantity data, a reference electric power value and a reference magnetic power value are independently calculated through physical analytical expressions as actual observation benchmarks; A moment-by-moment difference operation is performed on the estimated electric power value, the estimated magnetic power value, the reference electric power value, and the reference magnetic power value to obtain an electric power residual sequence and a magnetic power residual sequence to form the detection residual.

6. The method for dynamic joint detection of electric power and magnetic power based on multivariable fusion according to claim 3 is characterized in that: The expression of the joint power estimation model is: in: At time t k Estimated electric power value; At time t k Estimated magnetic power value; [f e (t k )|f m (t k )] T is the column vector after feature concatenation, with dimension (n e +n m )×1; V is the weight matrix from input feature to hidden layer, with dimension d×(n e +n m );σ(·) is the nonlinear activation function; W is the weight matrix from hidden layer to output, with dimension 2×d; b1 is the hidden layer bias term, with dimension d×1; b2 is the output bias term, with dimension 2×1; d is the hidden layer dimension, which is set according to the model complexity.

7. The method for dynamic joint detection of electric power and magnetic power based on multivariable fusion according to claim 6 is characterized in that: The calculation formula of the reference electric power value is: in: For the tth k Reference electric power value at the moment; U(t k ) is the instantaneous voltage collected by the voltage sensor; I(t k ) is the instantaneous current collected by the current sensor; -φ(t k ) is the instantaneous phase difference between voltage and current; The calculation formula of the reference magnetic power value is: in: For the tth k Reference magnetic power at the moment; is the time derivative of magnetic induction intensity, which comes from the magnetic field sensor; A eff is the effective projected area under the action of the magnetic field; μ is the magnetic permeability of the magnetic medium.

8. The method for dynamic joint detection of electric power and magnetic power based on multivariable fusion according to claim 7 is characterized in that: Based on the size and evolution trend of the detection residual, a state discrimination label for the current power state is generated. At the same time, a confidence interval is calculated based on the distribution range of the detection residual to quantify the uncertainty of the current power estimation result and assist in determining the credibility level of the state discrimination, including: Get the joint power estimation model at each sampling time t k The residual between the estimated output power value and the reference power value is recorded as the electric power residual Magnetic power residual and calculate the absolute value to assess the magnitude of the instantaneous estimation error; Within a preset time sliding window, the residual absolute value sequence of each type of residual is counted, and the maximum value, average value, standard deviation and change trend slope are extracted to characterize the overall intensity and dynamic evolution characteristics of the power estimation deviation; the overall intensity is the residual amplitude; Based on the residual amplitude, the state discrimination label of the current power state is generated according to the set multi-level state discrimination rules; wherein, when the mean and maximum values ​​of the residual absolute values ​​are lower than a first threshold value and the slope is close to zero, a "normal" state label is assigned; when the mean or maximum value of the residual is in the middle range or shows a continuous upward trend, a "fluctuation" state label is assigned; when the maximum value of the residual exceeds a second threshold value or the slope increases abnormally, an "abnormal" state label is assigned; Based on the distribution characteristics of the absolute value of the residual in the sliding window, the estimated confidence interval is calculated. The confidence interval represents the statistically credible range of the power estimation result, and the expression is: CI(t k )=[μ r (t k )-a·s r (t k ),m r (t k )+a·s r (t k )] Among them, μ r (t k ) is the mean of the absolute values ​​of the residuals in the current window, σ r (t k ) is the standard deviation, and α is the confidence coefficient, which is used to set the confidence level; Based on the state discrimination label and the width of the corresponding confidence interval, the credibility level of the current estimation result is jointly judged. If the state discrimination label is "normal" and the confidence interval width is less than the preset threshold, the current estimation result is judged to have high credibility; if the state discrimination label is "abnormal" and the confidence interval width is greater than or equal to the preset threshold, the current estimation result is judged to have large uncertainty and low credibility.

9. The method for dynamic joint detection of electric power and magnetic power based on multivariable fusion according to claim 7 is characterized in that: A feedback mechanism is established based on the detection residual, the state discrimination label, and the historical estimation accuracy, and an adaptive algorithm is used to iteratively optimize the fusion parameters, weight distribution, and feature dimensions of the joint estimation model to achieve self-learning and robustness enhancement of the dynamic detection system, including: Constructing a set of feedback signal indicators and forming a feedback state vector; the feedback signal indicator set includes: the mean, variance and slope of the currently detected residual sequence, the corresponding power state label, and the historical estimation accuracy over the past N detection cycles; Determine whether the feedback state vector meets the trigger conditions, and start the model optimization mechanism when any of the following conditions is met: the current state label is "abnormal" and the residual mean is greater than the first adaptive threshold; the historical estimation accuracy decreases continuously for more than a set number of times; the current confidence interval width exceeds the upper limit of system stability; During the optimization process of the model optimization mechanism, the following components of the joint estimation model are iteratively updated based on the feedback state vector: Optimize the feature fusion weights in the joint estimation model; the feature fusion weights are used to adjust the contribution of multi-source features to the estimated output after splicing at the input layer. The initial values ​​of the feature fusion weights are obtained during the training process and are subsequently dynamically adjusted based on the residual contribution rate to improve the response weights of important feature dimensions. Updating the parameter matrix within the joint estimation model, using methods such as adaptive gradient descent, Kalman filtering, or Bayesian updating to minimize the residual function value within the current cycle, thereby improving the model's ability to fit nonlinear feature combinations; Adjust the feature dimension screening strategy for the joint estimation model input, sort the features based on the correlation or information gain between the features and the residuals, and dynamically remove redundant features that have low contribution to estimation accuracy or high collinearity, so as to reduce model complexity and improve generalization ability; The optimized feature fusion weights, parameter matrix and feature dimension screening strategy are reloaded into the structure of the joint estimation model, and the power estimation and residual analysis of the next cycle are continued to realize the online iterative optimization and long-term self-learning update of the detection model.

10. A dynamic joint detection system of electric power and magnetic power based on multivariable fusion, characterized in that: include: A multi-source signal acquisition module is used to deploy multiple voltage, current, and magnetic field sensors in the electromagnetic system to be tested, respectively collecting raw physical quantity data including voltage signals, current signals, power factor, frequency, magnetic flux density, and magnetic induction intensity, and to perform time-series alignment on the signals of the raw physical quantity data through a multi-channel synchronous acquisition device to construct a unified multivariate time series data set; a feature preprocessing and extraction module, configured to preprocess the multivariate time series data set to obtain preprocessed data, and extract electric power-related features and magnetic power-related features from the preprocessed data based on a time domain, frequency domain, or time-frequency domain analysis method to form a multi-source dynamic feature matrix; a joint power modeling module, configured to construct a joint estimation model of electric power and magnetic power based on the dynamic characteristic matrix, and compare the electric power and magnetic power values ​​output by the joint estimation model with reference values ​​independently derived from the original physical quantity data to calculate a detection residual; A reference power calculation and residual evaluation module is used to generate a state discrimination label for the current power state based on the size and evolution trend of the detection residual. At the same time, a confidence interval is calculated based on the distribution range of the detection residual to quantify the uncertainty of the current power estimation result and assist in determining the confidence level of the state discrimination; The feedback optimization and model adaptation module is used to build a feedback mechanism based on the detection residual, the state discrimination label and the historical estimation accuracy, and adopt an adaptive algorithm to iteratively optimize the fusion parameters, weight distribution and feature dimensions of the joint estimation model to achieve self-learning and robustness enhancement of the dynamic detection system.

Images