A multi-state detection method for equipment based on electromagnetic field energy analysis

Abstract

The application discloses a kind of equipment multi-state detection methods based on electromagnetic field energy analysis, it is related to computer processing technical field, comprising the following steps: S01, the magnetic field energy near the power cord of target equipment is obtained, and the magnetic field energy is converted into voltage time sequence signal;S02, the voltage time sequence signal is carried out wavelet packet decomposition, and frequency domain fingerprint feature is extracted after calculation;S03, the correlation coefficient between the calculated frequency band energy ratio and each frequency band is used as constraint weight, and dynamics feature is extracted;S04, in the reconstructed phase space, mutation point detection is carried out, and fault early warning signal is judged and generated.The application acquires magnetic field energy by non-contact, and fuses frequency domain fingerprint feature, chaotic dynamics feature and mutation point joint determination mechanism, combines adaptive standard atlas dynamic updating based on incremental distillation, realizes the accurate identification of equipment multi-state, early fault early warning and long-term robust monitoring.

Description

Technical Field

[0001] This invention relates to the field of computer processing technology, and in particular to a method for multi-state detection of equipment based on electromagnetic field energy analysis. Background Technology

[0002] Real-time monitoring and fault early warning of equipment operating status are crucial for ensuring the safe and stable operation of industrial systems. Traditional equipment condition monitoring methods typically rely on installing various sensors on the equipment itself or directly connecting them to its electrical circuits. However, these contact-based monitoring methods have many inherent limitations: First, the installation process of sensors and cables is complex and may damage the original structure of the equipment, affecting its sealing or electromagnetic compatibility; second, for many old equipment already in operation or equipment in confined, high-risk environments, there is often a lack of reserved detection interfaces or installation space, making traditional methods difficult to implement. In recent years, with the advancement of electromagnetic induction technology, analyzing the state of equipment by collecting the spatial magnetic field energy radiated around its power lines or casing during operation has become a promising non-contact, low-cost alternative. However, most existing magnetic field energy-based solutions focus on improving energy harvesting efficiency to achieve self-powering, while neglecting to deeply explore the rich state characteristics (such as frequency band distribution and nonlinear dynamic characteristics) contained in the magnetic field signal. Therefore, these methods typically struggle to accurately distinguish subtle differences in electromagnetic radiation under various operating conditions (such as normal operation, low-power standby, and complete shutdown), and are even less effective at identifying transient characteristic changes caused by sudden faults (such as partial discharge and power fluctuations). In summary, how to extract stable, sensitive, and comprehensive characteristic information that fully characterizes the device's state from the weak magnetic field energy collected non-contactly, thereby achieving accurate identification of multiple device states and early fault warning, is a pressing technical problem in this field. Summary of the Invention

[0003] To address the aforementioned technical problems, the present invention provides a multi-state detection method for equipment based on electromagnetic field energy analysis, comprising the following steps: S01. Obtain the magnetic field energy near the power line of the target device and convert the magnetic field energy into a voltage timing signal; S02. Perform wavelet packet decomposition on the voltage time-series signal, calculate the energy ratio in each frequency band after decomposition, obtain the frequency band energy ratio, analyze the correlation coefficient between different frequency bands, and extract frequency domain fingerprint features. S03. Using the calculated frequency band energy ratio and correlation coefficient between each frequency band as constraint weights, apply frequency domain constraint embedding optimization to the embedding dimension determination process of phase space reconstruction, then perform phase space reconstruction on the voltage time series signal, map the one-dimensional voltage time series signal to a high-dimensional phase space, and calculate the maximum Lyapunov exponent and correlation exponent in the reconstructed phase space to extract dynamic features. S04. In the reconstructed phase space, abrupt change point detection is performed. When the fluctuation characteristics of the phase space trajectory deviate from the historical pattern or a discontinuous jump occurs, it is determined that equipment fault characteristics have occurred and a fault warning signal is generated.

[0004] 2. The method for multi-state detection of equipment based on electromagnetic field energy analysis according to claim 1, characterized in that, in step S02, wavelet packet decomposition is performed, the energy proportion in each frequency band after decomposition is calculated to obtain the frequency band energy ratio, and the correlation coefficient between different frequency bands is analyzed, including: Select wavelet basis functions and determine the number of decomposition levels, perform wavelet packet transform on the voltage time series signal, and obtain wavelet packet coefficients for multiple frequency bands; Calculate the energy value of each frequency band based on the wavelet packet coefficients of each frequency band, use the total energy of all frequency bands as the denominator, calculate the energy ratio of each frequency band, and form the frequency band energy ratio. Calculate the Pearson correlation coefficient between any two energy sequences from different frequency bands, and use it as the correlation coefficient; Furthermore, the total energy of the frequency band energy ratio is used as the representative energy feature of the voltage time sequence signal, and its dimensions are normalized and aligned with the feature sequence after phase space reconstruction.

[0005] Preferably, the phase space reconstruction of the voltage timing signal in step S03 includes: The mutual information method is used to determine the time delay of phase space reconstruction, and the time point corresponding to the first decrease of the mutual information function to the global minimum is taken as the delay time. The pseudo nearest neighbor method is used to determine the embedding dimension of the phase space reconstruction, and the minimum embedding dimension that makes the pseudo nearest neighbor rate fall below the threshold for the first time is selected. Based on the time delay and the embedding dimension, the one-dimensional voltage timing signal is reconstructed into a phase point trajectory in a high-dimensional phase space.

[0006] Preferably, step S04, which involves detecting abrupt change points in the reconstructed phase space, includes: The local tangential distortion value of each phase point in the reconstructed phase space is calculated using a preset mutation point detection formula, which is constructed based on the Euclidean distance and the angle between the evolution direction of the phase point and its neighboring historical phase points. The local tangential distortion value is compared with a preset mutation threshold. If the local tangential distortion value exceeds the mutation threshold, the corresponding time point is determined to be a candidate mutation point. Further, the slope of the change in the frequency band energy ratio sequence within several time windows before and after the candidate mutation point is called. When the local tangential distortion value exceeds the mutation threshold and the slope of the change in the frequency band energy ratio exceeds the preset slope threshold, it is determined that a device fault characteristic has occurred at the corresponding time of the time series.

[0007] Preferably, after determining the occurrence of equipment failure characteristics in step S04, the following steps are included: The frequency band energy ratio sequence within a time window before and after the mutation point is dynamically time-warped and matched with the standard frequency band energy change template in a preset library of multiple fault modes, and the matching distance is calculated. Simultaneously, the local tangential distortion evolution curve of the reconstructed phase space trajectory within this time window is compared with the standard distortion template in the fault mode library in terms of shape similarity. Based on the weighted comprehensive score of matching distance and shape similarity, the specific fault type corresponding to the fault characteristics of the equipment is output. The specific fault type includes at least partial discharge fault, power ripple abnormal fault, and mechanical loosening-induced magnetic field disturbance fault.

[0008] As a preferred option, it also includes: S05. Based on the extracted frequency band energy ratio, correlation coefficient, maximum Lyapunov index, correlation index, and mutation point detection results, construct a multimodal state feature map; S06. The maximum Lyapunov index, correlation index, frequency band energy ratio and correlation coefficient are jointly mapped to a normalized continuous dimension vector. The current state of the target device is determined by segmenting the Euclidean distance of the continuous dimension vector in high-dimensional space with the ratio between the pre-stored multiple standard feature map vectors.

[0009] Preferably, the dynamic update method for the multimodal state feature map includes: During equipment operation, when the equipment is determined to be in a stable operating condition, the newly acquired and calculated multimodal state feature map is used to incrementally distill the standard feature map of the corresponding state in order to smoothly update the center value of the standard feature map.

[0010] Preferably, in step S05, when constructing the multimodal state feature map, a multi-granularity interaction strategy is adopted, including: The frequency band energy ratio and the correlation coefficient are used as the first granularity features to characterize the frequency domain energy distribution pattern of the device status; The maximum Lyapunov index and the correlation index are used as the second granularity features to characterize the nonlinear dynamic mode of the equipment state. The results of the mutation point detection are used as the third granularity feature to characterize the temporal transient pattern of the device state. The first granular feature, the second granular feature, and the third granular feature are fused using a weighted fusion or feature splicing method to obtain the multimodal state feature map.

[0011] The present invention has at least the following beneficial effects: 1. By embedding the frequency band energy ratio and correlation coefficient as constraint weights into the process of determining the embedding dimension of phase space reconstruction, the traditional deficiency of the separation between frequency domain features and time domain chaotic features is fundamentally solved. This significantly amplifies the differences in phase point trajectory distribution under different device states in the high-dimensional phase space, forming a joint feature space with strong discriminative power.

[0012] 2. When a discontinuous jump occurs in the phase space trajectory, this invention does not directly identify it as a fault. Instead, it further verifies it by calling the slope of the frequency band energy ratio sequence change within multiple time windows before and after that moment. This effectively avoids misjudging electromagnetic field jumps generated during equipment state switching as equipment faults, while ensuring high sensitivity in capturing sudden early faults (such as partial discharge and instantaneous power fluctuations). Furthermore, by jointly mapping the maximum Lyapunov exponent, correlation index, frequency band energy ratio, and correlation coefficient to a normalized continuous dimension vector, this invention achieves highly interpretable and computationally inefficient multi-state decision-making using the Euclidean distance ratio in high-dimensional space, avoiding the problems of existing deep learning methods that over-rely on large-scale training data and lack physical interpretability. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of a device multi-state detection method based on electromagnetic field energy analysis provided in Embodiment 1 of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0017] Example 1

[0018] This embodiment provides a multi-state detection method for devices based on electromagnetic field energy analysis. The method includes the following steps: Figure 1 As shown: S01. Obtain the magnetic field energy near the power line of the target device and convert the magnetic field energy into a voltage timing signal; Specifically, the aforementioned magnetic field energy is obtained by using one or more miniature hollow coils (also known as Rogowski coils) or induction coils with magnetic cores, which are non-contactly fitted or placed close to the power cord sheath of the target device. When the device is powered on, the alternating current flowing through the power cord generates an alternating magnetic field in the surrounding space. According to Faraday's law of electromagnetic induction, the magnetic flux through the induction coil changes with time, and an induced electromotive force (EMF) is generated at the ends of the coil. The instantaneous value of this induced EMF is proportional to the rate of change of the current in the power cord, and its waveform envelope reflects the dynamic change of the device's load current. To accommodate the differences in magnetic field strength between devices of different power levels, the output of the induction coil is usually connected to an adaptive gain conditioning circuit. This circuit consists of an instrumentation amplifier and a programmable gain amplifier, capable of amplifying the weak induced voltage at the millivolt level to a standard signal range of 0~5V or 0~10V. Subsequently, the conditioned analog voltage signal is digitized using a high-precision analog-to-digital converter (ADC, with a bit depth of at least 12 bits and a sampling rate set to 1kHz~100kHz depending on the device status change rate), ultimately generating a one-dimensional voltage time-series signal sequence for subsequent algorithm processing, denoted as V(t)={v1,v2,…,v…}. NThis method utilizes the equipment's power supply line as the primary conductor and the induction coil as the secondary winding, with the two coupled only through magnetic field, without any electrical connection or physical contact. When the equipment is in different operating states (e.g., full load operation, light load standby, complete shutdown), the amplitude, frequency components, and harmonic distribution of the current in the power supply line will show significant differences. Under normal operating conditions, the current waveform typically contains abundant fundamental and higher harmonics, resulting in drastic changes in the magnetic field, and the induced voltage has a large time-domain waveform amplitude and abundant high-frequency components. In standby mode, the current decreases significantly and the waveform tends to be sinusoidal, the induced voltage amplitude decreases, and high-frequency noise is significantly reduced. In shutdown mode, the current is zero, and the induced voltage is theoretically zero (in practice, there may be weak power frequency interference in the environment, which can be eliminated through threshold judgment). Therefore, the converted voltage timing signal naturally carries the fingerprint information of the equipment's state.

[0019] S02. Perform wavelet packet decomposition on the voltage time-series signal, calculate the energy ratio in each frequency band after decomposition, obtain the frequency band energy ratio, analyze the correlation coefficient between different frequency bands, and extract frequency domain fingerprint features.

[0020] Furthermore, in the above embodiments, wavelet packet decomposition is performed, the energy proportion within each frequency band after decomposition is calculated to obtain the frequency band energy ratio, and the correlation coefficient between different frequency bands is analyzed, including: By selecting wavelet basis functions and determining the number of decomposition levels, wavelet packet transform is performed on the voltage time series signal to obtain wavelet packet coefficients for multiple frequency bands. Calculate the energy value of each frequency band based on the wavelet packet coefficients of each frequency band, use the total energy of all frequency bands as the denominator, calculate the energy ratio of each frequency band, and form the frequency band energy ratio. Calculate the Pearson correlation coefficient between any two energy sequences from different frequency bands, and use it as the correlation coefficient. Furthermore, the total energy of the frequency band energy ratio is used as the representative energy feature of the voltage time sequence signal, and its dimensions are normalized and aligned with the feature sequence after phase space reconstruction.

[0021] Specifically, the voltage time-series signal is first decomposed using wavelet packet decomposition. Wavelet packet decomposition is an extension of wavelet transform, which decomposes not only the low-frequency part of the signal but also the high-frequency part with equal precision, thus achieving multi-level division of the entire frequency band. In practical implementation, it is necessary to pre-select appropriate wavelet basis functions (such as the db4 or db8 wavelets in the Daubechies series, which are suitable for handling transient signals due to their good regularity and tight support characteristics) and determine the number of decomposition levels. The choice of the number of decomposition levels depends on the sampling frequency of the signal and the target frequency band resolution: for magnetic field energy signals generated by equipment near the power frequency, 3 to 5 levels of decomposition are usually used to divide the signal into 8 to 32 equal-width frequency bands. Through wavelet packet transform, the original voltage time-series signal is decomposed into wavelet packet coefficients of multiple frequency bands, each frequency band corresponding to a specific frequency interval. These coefficients retain the time-frequency local information of the signal within that frequency band.

[0022] After obtaining the wavelet packet coefficients, the energy percentage within each frequency band is calculated, specifically including: The energy value of a frequency band is obtained by squaring and summing the wavelet packet coefficients of each frequency band. The total energy of the signal is obtained by summing the energy values ​​of all frequency bands. Then, divide the energy value of each frequency band by the total energy to obtain the energy percentage of that frequency band (band energy ratio).

[0023] Furthermore, in the above embodiments, the correlation coefficient between energy sequences of different frequency bands is calculated, and the energy change sequences of any two different frequency bands on the time axis are extracted, and the Pearson correlation coefficient between them is calculated. The Pearson correlation coefficient ranges from [-1, 1], and the larger its absolute value, the stronger the same or opposite trend of energy change between the two frequency bands.

[0024] Specifically, to organically integrate frequency domain features with subsequent phase space reconstruction (time-domain chaotic features), this embodiment also uses the total energy value of the frequency band energy ratio (i.e., the total energy of the original voltage time-series signal) as a representative energy feature. Before performing phase space reconstruction, this representative energy feature is normalized and aligned with the phase space trajectory feature sequence to be extracted. The purpose of normalizing and aligning the dimensions is to eliminate the inconsistency in feature scale caused by differences in the absolute amplitude of magnetic field strength between different devices and different acquisition times. In specific implementation, Z-score normalization or Min-Max normalization methods can be used: mapping the total energy value of the frequency band energy ratio to the [0,1] interval, and mapping the Lyapunov exponent and correlation index obtained after phase space reconstruction to the same numerical scale, so that the two have comparable weights in subsequent weighted fusion or Euclidean distance calculation.

[0025] S03. Using the calculated frequency band energy ratio and correlation coefficient between each frequency band as constraint weights, apply frequency domain constraint embedding optimization to the process of determining the embedding dimension of phase space reconstruction. Then, reconstruct the voltage time series signal in phase space, map the one-dimensional voltage time series signal to a high-dimensional phase space, and calculate the maximum Lyapunov exponent and correlation exponent in the reconstructed phase space to extract dynamic features.

[0026] Furthermore, in the above embodiments, phase space reconstruction of the voltage timing signal includes: The mutual information method is used to determine the time delay of phase space reconstruction, and the time point corresponding to the first decrease of the mutual information function to the global minimum is taken as the delay time. The pseudo nearest neighbor method is used to determine the embedding dimension of the phase space reconstruction, and the minimum embedding dimension that makes the pseudo nearest neighbor rate fall below the threshold for the first time is selected. Based on time delay and embedding dimension, a one-dimensional voltage timing signal is reconstructed into a phase point trajectory in a high-dimensional phase space.

[0027] Specifically, the frequency band energy ratio and correlation coefficients between frequency bands obtained through wavelet packet decomposition are essentially frequency domain fingerprints of the electromagnetic radiation energy of the device under different operating states. These fingerprint features can reflect the essential differences in physical processes such as internal circuit topology, load characteristics, and switching actions. By using this frequency domain fingerprint as constraint weights, the process of determining the embedding dimension in phase space reconstruction is guided. Traditional phase space reconstruction usually uses a fixed embedding dimension or determines the dimension based solely on the statistical characteristics of the time series itself (such as the pseudo nearest neighbor method). This approach ignores the differences in energy contribution of the signal in different frequency bands. However, this invention uses the frequency band energy ratio as the importance weight of each frequency band to adjust the embedding dimension search process. That is, the frequency band with a high energy proportion is given a higher phase space dimension contribution factor during reconstruction, so that the reconstructed phase point trajectory can preferentially retain the dynamic information of key frequency bands. At the same time, the correlation coefficients between different frequency bands are used to constrain the redundancy of the phase space trajectory: if two frequency bands are highly correlated, their corresponding virtual dimensions are merged in the embedding optimization to avoid phase space dimension expansion due to multicollinearity; if the correlation is low, more independent dimensions are retained to capture the decoupling characteristics of different physical processes. This frequency-domain constrained embedding optimization process essentially introduces prior frequency-domain physical knowledge into the phase space dimension selection, so that the final high-dimensional phase space not only faithfully restores the chaotic dynamic characteristics of the original time-series signal, but also highlights the frequency-domain sensitive region related to the device state in terms of structure, thereby amplifying the distribution differences of phase point trajectories under different states (working, standby, power off, etc.).

[0028] After completing the frequency domain constraints, the voltage time series signal is reconstructed in phase space. The theoretical basis of phase space reconstruction is Takens' embedding theorem. In short, for a nonlinear dynamic system, a high-dimensional space is constructed using the time-delay sequence of a single observed variable. The topological structure of this space is differentially homeomorphic to the state space of the original system, thus allowing the extraction of the system's nonlinear invariants from the reconstructed phase space. In this invention, the voltage time series signal originates from non-contact magnetic field energy acquisition. This signal is characterized by nonlinearity, non-stationarity, and low signal-to-noise ratio, making it highly suitable for revealing its intrinsic dynamic laws using phase space reconstruction. Specifically, two key parameters need to be determined: the time delay τ and the embedding dimension m. This invention uses the mutual information method to determine τ. The mutual information method estimates the optimal delay by calculating the information redundancy between the original time series x(t) and its delayed sequence x(t+τ). The specific process is as follows: the time series is delayed at equal intervals, and the mutual information function I(τ) is calculated for each delay. A larger mutual information value indicates a stronger nonlinear correlation between the two sequences, while a smaller value indicates a weaker correlation. The optimal delay is usually taken as the value of τ when the mutual information function first drops to the global minimum, because at this point the two sequences are neither excessively redundant (τ is too small) nor completely independent (τ is too large).

[0029] After determining τ, the pseudo-nearest neighbor method is used to determine the embedding dimension m. The principle of the pseudo-nearest neighbor method is that in a low-dimensional space, two points that are not originally adjacent may appear to be adjacent due to projection compression (i.e., pseudo-nearest neighbor points). As the embedding dimension gradually increases, these pseudo-adjacency relationships will gradually be eliminated. When the proportion of pseudo-nearest neighbor points first falls below a set threshold (preferably 1% in this invention), it is considered that a sufficiently high dimension has been obtained to fully unfold the attractor. During the operation, starting from m=1, for each phase point x(i), its nearest neighbor x(j) is found in the m-dimensional space, and then the rate of change of distance between the two points is calculated as the dimension increases to m+1. If the rate of change exceeds an empirical threshold (usually 10-15), the nearest neighbor is determined to be a pseudo-nearest neighbor. The proportion of pseudo-nearest neighbors for all phase points is statistically analyzed, and the m corresponding to the first drop below 1% is the optimal embedding dimension. For example, for the time series data with τ=12 above, the calculated proportion of pseudo nearest neighbors is 8% when m=5, 3.2% when m=6, and 0.8% when m=7. Therefore, m=7 is taken as the embedding dimension.

[0030] After obtaining τ and m, the one-dimensional voltage time-series signal {x1, x2, …, xN} can be reconstructed into phase point trajectories in a high-dimensional phase space. Each phase point is represented as... These phase points form a trajectory in m-dimensional space, and its geometric structure reflects the attractor morphology of the system. Based on this, this invention calculates the maximum Lyapunov exponent and the correlation exponent. The maximum Lyapunov exponent quantitatively describes the exponential divergence rate of adjacent trajectories in phase space. A positive value indicates that the system has chaotic characteristics and is highly sensitive to initial conditions, and the magnitude of the value reflects the reciprocal of the predictable time scale of the system. The calculation employs a small data method: the nearest neighbor of each phase point is found in the reconstructed phase space, the evolution distance between them over time is tracked, and a linear fit is performed on the logarithm of the distance; the slope of this fit is the maximum Lyapunov exponent. The correlation exponent (also known as the correlation dimension) is used to characterize the complexity and self-similarity of the system's attractor. It is obtained by calculating the correlation integral where the distance between phase points is less than a given radius at different scales, followed by log-log linear regression.

[0031] S04. In the reconstructed phase space, abrupt change point detection is performed. When the fluctuation characteristics of the phase space trajectory deviate from the historical pattern or a discontinuous jump occurs, it is determined that equipment fault characteristics have occurred and a fault warning signal is generated.

[0032] Furthermore, in the above embodiments, mutation point detection is performed in the reconstructed phase space, including: The local tangential distortion value of each phase point in the reconstructed phase space is calculated using a preset mutation point detection formula. The mutation point detection formula is constructed based on the Euclidean distance and the angle between the evolution direction and the phase point and its neighboring historical phase points. The local tangential distortion value is compared with a preset mutation threshold. If the local tangential distortion value exceeds the mutation threshold, the corresponding time point is determined to be a candidate mutation point. Further, the slope of the change in the frequency band energy ratio sequence within several time windows before and after the candidate mutation point is called. When the local tangential distortion value exceeds the mutation threshold and the slope of the change in the frequency band energy ratio exceeds the preset slope threshold, it is determined that the equipment failure feature has occurred at the corresponding time point.

[0033] Specifically, firstly, for each phase point in the reconstructed high-dimensional phase space (i.e., the state point of the voltage timing signal under delayed coordinate mapping), its local tangential distortion value is calculated. The core design idea of ​​this value is: when the equipment is running smoothly, the phase space trajectory should be smooth and continuous, and the evolution direction (tangential) and distance changes between adjacent phase points should be relatively consistent; however, when the equipment experiences a sudden failure, the transient changes in magnetic field energy will cause the phase point trajectory to become distorted or discontinuous. The specific calculation formula is based on two factors: one is the rate of change of the Euclidean distance between the current phase point and its neighboring historical phase points, and the other is the difference in the angle between the evolution directions. For example, it can be defined as: taking the current phase point Xi and its k nearest historical phase points in time. The angular deviation between the vector from Xi to each historical phase point and the vector at the previous moment is calculated, and then multiplied by the distance normalization factor to obtain the local tangential distortion value Di. In implementation, a sliding window method is usually used to dynamically calculate this value for each phase point, and a mutation threshold obtained through trial operation or statistical analysis of historical normal data is set. .

[0034] When Di>D th Instead of immediately classifying it as a fault, this moment is marked as a candidate mutation point. This is because a simple phase space trajectory jump may originate from a device state switch (e.g., a motor transitioning from standby to operating state causes a step increase in its electromagnetic field strength, resulting in a jump in the phase point trajectory). To filter out false alarms caused by such normal state transitions, this method further utilizes the slope of the frequency band energy ratio sequence within several time windows (e.g., 5 sampling periods before and after) before and after the candidate mutation point. The frequency band energy ratio is a stable frequency domain feature obtained from wavelet packet decomposition: the energy proportion of each frequency band has a unique "fingerprint" under different device states. Normal state transitions (such as standby to operating) usually cause synchronous, monotonic changes in the energy ratio of multiple frequency bands, with relatively gentle slopes; while sudden faults (such as partial discharge) often cause drastic, non-monotonic fluctuations in the high-frequency band energy ratio, with the absolute value of the slope far exceeding the normal range. In implementation, a slope threshold is preset. Only when the local tangential distortion value of the candidate mutation point exceeds the mutation threshold and the absolute value of the slope of the frequency band energy ratio change exceeds the slope threshold, is it finally determined that a real equipment fault characteristic has occurred at that moment, and a fault warning signal is then generated.

[0035] Example 2

[0036] Based on the above embodiment one, this embodiment aims to include the following in step S04 after determining the occurrence of equipment failure characteristics: The frequency band energy ratio sequence within a time window before and after the mutation point is dynamically time-warped and matched with the standard frequency band energy change template in a preset library of multiple fault modes, and the matching distance is calculated. Simultaneously, the local tangential distortion evolution curve of the reconstructed phase space trajectory within this time window is compared with the standard distortion template in the fault mode library in terms of shape similarity. Based on the weighted comprehensive score of matching distance and shape similarity, the specific fault type corresponding to the fault characteristics of the output device is determined. The specific fault type includes at least partial discharge fault, power supply ripple abnormal fault, and mechanical loosening-induced magnetic field disturbance fault.

[0037] After determining that a device has malfunctioned, this method further performs attribution analysis to output the specific fault type. The core of this step is to perform template matching between the local signal characteristics before and after the abrupt change point and a pre-set fault mode library. The fault mode library is constructed by pre-collecting magnetic field energy data under various typical faults, and also by wavelet packet decomposition and phase space reconstruction. It contains standard frequency band energy change templates and standard distortion templates corresponding to each fault.

[0038] In practice, the frequency band energy ratio sequence within a time window before and after the mutation point (e.g., 0.1 seconds before to 0.2 seconds after the mutation point) is extracted to form a curve that changes over time. Since the duration, frequency band location of energy mutations, and waveform shape vary among different faults, directly using Euclidean distance matching is easily affected by time axis scaling. Therefore, the Dynamic Time Warping (DTW) algorithm is used for matching. DTW can find the optimal correspondence between the test sequence and the standard template through nonlinear alignment, calculating a matching distance (the smaller the distance, the more similar the frequency band energy change patterns of the two). Simultaneously, for the reconstructed phase space trajectory within the same time window, the curve of the local tangential distortion value evolving over time (i.e., the distortion evolution curve) is extracted. This curve is then compared with the standard distortion template in the fault mode library for shape similarity, for example, using the Pearson correlation coefficient or a method based on Frèchet distance to calculate a similarity score (the higher the score, the closer the fluctuation shape of the distortion curve is to the typical pattern of a certain fault). Finally, the DTW matching distance (converted to a similarity score after normalization) and shape similarity score are assigned weights (e.g., frequency domain weight 0.6, time domain distortion weight 0.4), and a weighted composite score is calculated. The fault type with the highest score is the output result.

[0039] It should be noted that the preset fault types in the above embodiments include at least: Partial discharge fault (manifested as instantaneous energy spikes in the high-frequency band plus isolated singularities in phase space); Power supply ripple abnormality fault (manifested as periodic fluctuation of power frequency octave band energy + small oscillation envelope of phase space trajectory); Mechanical loosening induces magnetic field disturbance faults (manifested as a slow drift in the low-frequency band energy ratio + a large-scale drift in the phase space trajectory rather than a local abrupt change). Preset information such as these.

[0040] Example 3

[0041] Based on the above embodiment one, this embodiment aims to provide the following steps in addition to the above steps: S05. Based on the extracted frequency band energy ratio, correlation coefficient, maximum Lyapunov index, correlation index, and mutation point detection results, construct a multimodal state feature map.

[0042] Furthermore, in the above embodiments, when constructing the multimodal state feature map, a multi-granularity interaction strategy is adopted, including: Using the frequency band energy ratio and correlation coefficient as the first granularity feature, the frequency domain energy distribution pattern of the equipment status is characterized. The maximum Lyapunov exponent and the correlation exponent are used as the second granularity features to characterize the nonlinear dynamic mode of the equipment state. The results of mutation point detection are used as the third granularity feature to characterize the temporal transient pattern of the equipment state. By using weighted fusion or feature splicing, the first-granularity features, the second-granularity features, and the third-granularity features are fused in a knowledge-enhanced manner to obtain a multimodal state feature map.

[0043] Specifically, the first granularity feature vector is formed by combining the band energy ratio and the correlation coefficient between bands. This vector reflects the distribution ratio of electromagnetic field energy in each frequency band and the dependence between bands when the device is in normal operation, standby, or fault state. Essentially, it is the frequency domain energy fingerprint of the device state. For example, when a switching power supply is operating normally, the main component of its magnetic field energy is concentrated in the 100kHz~500kHz band, with a band energy ratio of over 0.6, and the correlation coefficient between this band and other bands is low. However, when the device enters standby mode, the high-frequency components attenuate significantly, and the energy ratio concentrates on the power frequency and its harmonics (50Hz, 100Hz), and the correlation coefficient changes accordingly. Secondly, the second granularity feature vector is formed by using the maximum Lyapunov exponent and the correlation index. This vector characterizes the chaotic characteristics and attractor complexity of the voltage time-series signal in the reconstructed phase space: a Lyapunov exponent greater than 0 indicates that the system is sensitive to initial conditions, and the larger the exponent value, the more complex and unpredictable the system. The correlation index quantitatively describes the geometric dimension of the attractor, reflecting the number of degrees of freedom of the system. Taking a switching power supply as an example, during normal operation, its voltage signal exhibits a quasi-periodic or weakly chaotic state, with the Lyapunov exponent ranging from 0.1 to 0.3. When a partial discharge fault occurs, random spike pulses appear in the magnetic field signal, the phase space trajectory diverges rapidly, and the Lyapunov exponent can jump to over 0.8, with the correlation index also increasing accordingly. The third granularity feature comes from the abrupt change detection results, including the timestamp of the abrupt change, the intensity of the abrupt change (the peak value of the local tangential distortion), and the duration of the abrupt change. This information characterizes the transient non-stationary behavior of the device state in the time dimension; for example, a sudden increase in power supply ripple can cause a millisecond-level spike jump in magnetic field energy.

[0044] After obtaining the feature vectors at the three granularities mentioned above, knowledge-enhanced fusion is achieved through weighted fusion or feature concatenation to generate the final multimodal state feature map. Weighted fusion is suitable for scenarios where the physical meanings of the features at each granularity differ significantly, their dimensions are different, but their importance can be assessed a priori. First, each granularity feature is normalized (e.g., Z-score normalization or Min-Max normalization). Then, weight coefficients are assigned to each granularity based on historical experimental data or expert knowledge; for example, setting the weight of the first granularity to 0.3, the second granularity to 0.4, and the third granularity to 0.3. Finally, the weighted sum is calculated to obtain the fused feature vector. Feature concatenation is more direct: the three granularity feature vectors are concatenated end-to-end in a fixed order to form a higher-dimensional comprehensive feature vector, without changing the original feature values ​​but retaining all information. Regardless of the method used, the final result is a multimodal feature map containing frequency domain energy distribution, nonlinear chaotic characteristics, and transient mutation modes. This feature map can serve as input for real-time state identification (e.g., semantic alignment or Euclidean distance comparison with pre-stored standard maps) and can also be used for incremental distillation and dynamic updates of the fault mode library. For example, in a real-world test case, when an industrial robot servo drive gradually transitioned from a normal operating state to an IGBT soft failure state, its first granularity feature showed that the high-frequency (>1MHz) energy ratio slowly increased from 0.1 to 0.35, the second granularity feature showed that the Lyapunov exponent increased from 0.2 to 0.55, and the third granularity feature detected 2-3 weak abrupt changes every 10 minutes. The feature map formed by stitching these three features exhibits a clear drift direction in high-dimensional space, thus enabling an early warning 30 minutes before the fault occurs and helping to distinguish this fault type from simple load fluctuations. This fully demonstrates the technical advantages of multi-granularity interaction strategies in multi-state equipment detection.

[0045] S06. The maximum Lyapunov exponent, correlation index, frequency band energy ratio and correlation coefficient are jointly mapped to a normalized continuous dimension vector. The current state of the target device is determined by segmenting the Euclidean distance of the continuous dimension vector in high-dimensional space with the ratio between the pre-stored multiple standard feature map vectors.

[0046] Specifically, for the real-time calculated maximum Lyapunov exponent (typically between 0 and 2, dimensionless), correlation index (between 0 and 1), frequency band energy ratio (the sum of the energy proportions of each frequency band is 1, belonging to a probability distribution between 0 and 1), and correlation coefficient (between -1 and 1), range normalization or Z-score normalization methods are used to map them to a unified numerical range, such as the [0,1] interval. For the Lyapunov exponent, since it may be close to 0 (steady state) or greater than 0 (chaotic state), a linear mapping is required based on the statistical extreme values ​​of the equipment's historical data; for the correlation coefficient, it is mapped to the [0,1] interval through a (ρ+1) / 2 transformation.

[0047] The four normalized feature values ​​are concatenated in a fixed order to form a four-dimensional vector F = [fL, fD, fE, fC], where fL is the normalized maximum Lyapunov exponent, fD is the normalized correlation index, fE is the normalized band energy ratio (in practical applications, the principal energy ratios from several of the most discriminative bands can be selected, or the principal components can be taken after dimensionality reduction of all band energy ratios), and fC is the normalized correlation coefficient. Each component of this vector is in the interval [0,1], achieving a unified dimension.

[0048] When the device is in a known state (such as normal operation, standby, power off, fault warning, etc.), the average value of the above four-dimensional vector is calculated through multiple samplings and used as the standard center point for that state. For example, the standard vector corresponding to the normal operation state is Fnormal = [0.72, 0.45, 0.68, 0.51], the standard vector corresponding to the standby state is Fstandby = [0.12, 0.21, 0.25, 0.33], and the standard vector corresponding to the power off state is off = [0.03, 0.08, 0.05, 0.11]. These standard spectral vectors are stored in the system and can be dynamically updated through incremental distillation.

[0049] For the current feature vector Fcurr obtained in real time, calculate its Euclidean distance with each standard map vector: , And so on. Then, the distance ratio is calculated, for example, by defining the relative distance ratio Ri = di / (d1+d2+d3+…). This ratio reflects the probability weight of the current state belonging to each of the standard states.

[0050] There are preset segmentation thresholds. For example, if the distance ratio between the current vector and the normal state, Rnormal, is less than 0.4, it is determined to be a normal state; if Rnormal is between 0.4 and 0.6 and Rstandby is less than 0.3, it is determined to be a transitional state; if Rfault (a fault standard vector in the fault mode library) is the smallest and its absolute distance is less than the preset absolute threshold, it is determined to be a specific fault type.

[0051] Example 4

[0052] Based on the above embodiment three, the dynamic update method of the multimodal state feature map in the embodiment includes: During equipment operation, when the equipment is determined to be in a stable operating condition, the newly acquired and calculated multimodal state feature map is used to incrementally distill the standard feature map of the corresponding state in order to smoothly update the center value of the standard feature map.

[0053] Specifically, the system continuously collects the time-series signal of magnetic field energy during equipment operation and calculates the multimodal state feature spectrum at the current moment in real time according to the method described in claims 1 to 7. This spectrum is a high-dimensional vector, whose components include, but are not limited to, the energy ratio of each frequency band, the correlation coefficient between different frequency bands, the maximum Lyapunov exponent, the correlation index, and the local tangential distortion peak in the most recent abrupt change detection result. At the same time, the system uses a lightweight steady-state discriminator to determine whether the equipment is currently in a stable operating condition. For example, when the standard deviation of the Euclidean distance between the state vectors calculated within N consecutive time windows (N can be 3 to 5, and the length of each window corresponds to 1 / 4 of a complete operating cycle of the equipment) is less than a preset threshold (such as less than 5% of the mean of each component), it is considered that the equipment is currently in a steady state, excluding the interference of data under transient switching or fault conditions on the standard spectrum. Once a stable operating condition is confirmed, the system triggers incremental distillation updates: the multimodal feature map obtained in real time is used as a "sample", and the pre-stored standard feature map corresponding to the current state of the equipment is used as the update target. A weighted moving average is performed according to the preset distillation rate α (α is between 0.1 and 0.3, preferably 0.15), that is, the new standard map = (1-α) × the original standard map + α × the current measured map, thereby achieving a smooth update of the center value of the standard map.

[0054] Example 5

[0055] This invention provides a non-transitory computer-readable storage medium storing at least one instruction or at least one program segment, which is loaded and executed by a processor to implement the following steps: Acquire the magnetic field energy near the power line of the target device and convert the magnetic field energy into a voltage timing signal; Wavelet packet decomposition is performed on the voltage time-series signal, the energy ratio in each frequency band after decomposition is calculated, the frequency band energy ratio is obtained, the correlation coefficient between different frequency bands is analyzed, and frequency domain fingerprint features are extracted. The calculated frequency band energy ratio and correlation coefficient between each frequency band are used as constraint weights to apply frequency domain constraint embedding optimization to the process of determining the embedding dimension of phase space reconstruction. Then, the voltage time series signal is reconstructed into phase space, mapping the one-dimensional voltage time series signal to a high-dimensional phase space. The maximum Lyapunov exponent and correlation exponent are calculated in the reconstructed phase space to extract dynamic features. In the reconstructed phase space, abrupt change point detection is performed. When the fluctuation characteristics of the phase space trajectory deviate from the historical pattern or a discontinuous jump occurs, it is determined that equipment fault characteristics have occurred and a fault warning signal is generated.

[0056] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0057] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0058] Example 6

[0059] This invention provides an electronic device, including a processor and a memory, wherein the memory stores at least one instruction or at least one program segment, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the following steps: Acquire the magnetic field energy near the power line of the target device and convert the magnetic field energy into a voltage timing signal; Wavelet packet decomposition is performed on the voltage time-series signal, the energy ratio in each frequency band after decomposition is calculated, the frequency band energy ratio is obtained, the correlation coefficient between different frequency bands is analyzed, and frequency domain fingerprint features are extracted. The calculated frequency band energy ratio and correlation coefficient between each frequency band are used as constraint weights to apply frequency domain constraint embedding optimization to the process of determining the embedding dimension of phase space reconstruction. Then, the voltage time series signal is reconstructed into phase space, mapping the one-dimensional voltage time series signal to a high-dimensional phase space. The maximum Lyapunov exponent and correlation exponent are calculated in the reconstructed phase space to extract dynamic features. In the reconstructed phase space, abrupt change point detection is performed. When the fluctuation characteristics of the phase space trajectory deviate from the historical pattern or a discontinuous jump occurs, it is determined that equipment fault characteristics have occurred and a fault warning signal is generated.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for multi-state detection of equipment based on electromagnetic field energy analysis, characterized in that, Includes the following steps: S01. Obtain the magnetic field energy near the power line of the target device and convert the magnetic field energy into a voltage timing signal; S02. Perform wavelet packet decomposition on the voltage time-series signal, calculate the energy ratio in each frequency band after decomposition, obtain the frequency band energy ratio, analyze the correlation coefficient between different frequency bands, and extract frequency domain fingerprint features. S03. Using the calculated frequency band energy ratio and correlation coefficient between each frequency band as constraint weights, apply frequency domain constraint embedding optimization to the embedding dimension determination process of phase space reconstruction, then perform phase space reconstruction on the voltage time series signal, map the one-dimensional voltage time series signal to a high-dimensional phase space, and calculate the maximum Lyapunov exponent and correlation exponent in the reconstructed phase space to extract dynamic features. S04. In the reconstructed phase space, abrupt change point detection is performed. When the fluctuation characteristics of the phase space trajectory deviate from the historical pattern or a discontinuous jump occurs, it is determined that equipment fault characteristics have occurred and a fault warning signal is generated.

2. The method for multi-state detection of equipment based on electromagnetic field energy analysis according to claim 1, characterized in that, Step S02 involves performing wavelet packet decomposition, calculating the energy percentage within each frequency band after decomposition to obtain the frequency band energy ratio, and analyzing the correlation coefficients between different frequency bands, including: Select wavelet basis functions and determine the number of decomposition levels, perform wavelet packet transform on the voltage time series signal, and obtain wavelet packet coefficients for multiple frequency bands; Calculate the energy value of each frequency band based on the wavelet packet coefficients of each frequency band, use the total energy of all frequency bands as the denominator, calculate the energy ratio of each frequency band, and form the frequency band energy ratio. Calculate the Pearson correlation coefficient between any two energy sequences from different frequency bands, and use it as the correlation coefficient; Furthermore, the total energy of the frequency band energy ratio is used as the representative energy feature of the voltage time sequence signal, and its dimensions are normalized and aligned with the feature sequence after phase space reconstruction.

3. The method for multi-state detection of equipment based on electromagnetic field energy analysis according to claim 1, characterized in that, The phase space reconstruction of the voltage timing signal in step S03 includes: The mutual information method is used to determine the time delay of phase space reconstruction, and the time point corresponding to the first decrease of the mutual information function to the global minimum is taken as the delay time. The pseudo nearest neighbor method is used to determine the embedding dimension of the phase space reconstruction, and the minimum embedding dimension that makes the pseudo nearest neighbor rate fall below the threshold for the first time is selected. Based on the time delay and the embedding dimension, the one-dimensional voltage timing signal is reconstructed into a phase point trajectory in a high-dimensional phase space.

4. The method for multi-state detection of equipment based on electromagnetic field energy analysis according to claim 1, characterized in that, The step S04, which involves detecting abrupt change points in the reconstructed phase space, includes: The local tangential distortion value of each phase point in the reconstructed phase space is calculated using a preset mutation point detection formula, which is constructed based on the Euclidean distance and the angle between the evolution direction of the phase point and its neighboring historical phase points. The local tangential distortion value is compared with a preset mutation threshold. If the local tangential distortion value exceeds the mutation threshold, the corresponding time point is determined to be a candidate mutation point. Further, the slope of the change in the frequency band energy ratio sequence within several time windows before and after the candidate mutation point is called. When the local tangential distortion value exceeds the mutation threshold and the slope of the change in the frequency band energy ratio exceeds the preset slope threshold, it is determined that a device fault characteristic has occurred at the corresponding time of the time series.

5. The method for multi-state detection of equipment based on electromagnetic field energy analysis according to claim 1, characterized in that, After determining the equipment fault characteristics in step S04, the following steps are included: The frequency band energy ratio sequence within a time window before and after the mutation point is dynamically time-warped and matched with the standard frequency band energy change template in a preset library of multiple fault modes, and the matching distance is calculated. Simultaneously, the local tangential distortion evolution curve of the reconstructed phase space trajectory within this time window is compared with the standard distortion template in the fault mode library in terms of shape similarity. Based on the weighted comprehensive score of matching distance and shape similarity, the specific fault type corresponding to the fault characteristics of the equipment is output. The specific fault type includes at least partial discharge fault, power ripple abnormal fault, and mechanical loosening-induced magnetic field disturbance fault.

6. The method for multi-state detection of equipment based on electromagnetic field energy analysis according to claim 1, characterized in that, Also includes: S05. Based on the extracted frequency band energy ratio, correlation coefficient, maximum Lyapunov index, correlation index, and mutation point detection results, construct a multimodal state feature map; S06. The maximum Lyapunov index, correlation index, frequency band energy ratio and correlation coefficient are jointly mapped to a normalized continuous dimension vector. The current state of the target device is determined by segmenting the Euclidean distance of the continuous dimension vector in high-dimensional space with the ratio between the pre-stored multiple standard feature map vectors.

7. The method for multi-state detection of equipment based on electromagnetic field energy analysis according to claim 6, characterized in that, The dynamic update method for the multimodal state feature map includes: During equipment operation, when the equipment is determined to be in a stable operating condition, the newly acquired and calculated multimodal state feature map is used to incrementally distill the standard feature map of the corresponding state in order to smoothly update the center value of the standard feature map.

8. The method for multi-state detection of equipment based on electromagnetic field energy analysis according to claim 1, characterized in that, In step S05, when constructing the multimodal state feature map, a multi-granularity interaction strategy is adopted, including: The frequency band energy ratio and the correlation coefficient are used as the first granularity features to characterize the frequency domain energy distribution pattern of the device status; The maximum Lyapunov index and the correlation index are used as the second granularity features to characterize the nonlinear dynamic mode of the equipment state. The results of the mutation point detection are used as the third granularity feature to characterize the temporal transient pattern of the device state. The first granular feature, the second granular feature, and the third granular feature are fused using a weighted fusion or feature splicing method to obtain the multimodal state feature map.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the device multi-state detection method based on electromagnetic field energy analysis as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the device multi-state detection method based on electromagnetic field energy analysis as described in any one of claims 1 to 8.

Images