Coupling separation and state collaborative prediction system and method based on power supply characteristic spectrum

Abstract

The invention belongs to the technical field of equipment state detection, and discloses a coupling separation and state collaborative prediction system and method based on a power supply characteristic spectrum. Comprising the following steps: collecting current and voltage signals of all equipment in the shared power supply topology and extracting a spectrum feature vector; carrying out random Fourier feature mapping on the spectrum feature vectors of other devices except the target device, and training a group sparse regression model to obtain a group regression coefficient set; calculating an external interference component, and obtaining a pure feature vector of the target equipment based on the external interference component; performing significance index calculation on the pure feature vector to obtain a high-dominance window set; obtaining a window abnormal value of the high-dominance window, averaging to obtain an abnormal evaluation index, comparing the abnormal evaluation index with an alarm threshold, and judging whether the state of the current target equipment is normal or not; constructing an anomaly evaluation sequence, establishing a trend function, and predicting residual safe operation time; accurate judgment and prediction of the state of the power supply equipment are realized through external coupling separation.

Description

Technical Field

[0001] This invention relates to the field of equipment condition monitoring technology, and more specifically, to a system and method for power supply characteristic spectrum coupling separation and condition co-prediction. Background Technology

[0002] In existing technologies, power equipment condition monitoring mainly relies on single-device analysis. However, in complex industrial scenarios with multiple devices and shared power supply topologies, the electrical connectivity of the shared power supply link easily leads to coupling interference between devices' power signals. This results in the monitoring signal of the target device being mixed with interference components from other devices. Because existing single-device analysis methods cannot effectively separate external coupling interference from the true characteristics of the device itself, the spectral characteristics of the target device are distorted. This distortion directly causes a significant decrease in the accuracy of equipment condition prediction. Long-term inaccurate condition predictions prevent timely identification of equipment faults, ultimately leading to a surge in the risk of unexpected equipment downtime, severely disrupting the continuity of industrial production and reducing production efficiency.

[0003] In view of this, the present invention proposes a system and method based on power source characteristic spectrum coupling separation and state cooperative prediction to solve the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a method based on power source characteristic spectrum coupling separation and state collaborative prediction, comprising:

[0005] Collect current and voltage signals of all devices in the shared power supply topology and extract spectral feature vectors;

[0006] Random Fourier feature mapping is performed on the spectral feature vectors of devices other than the target device, and a group sparse regression model is trained to obtain a set of group regression coefficients;

[0007] The external interference component is calculated based on the set of group regression coefficients, and the external interference component is extracted from the spectral feature vector of the target device to obtain the pure feature vector of the target device.

[0008] The significance index of the pure feature vector of each window of the target device is calculated to obtain the set of highly explicit windows;

[0009] Obtain the window outlier values ​​of each high-visibility window in the high-visibility window set, average the window outlier values ​​to obtain the anomaly assessment index, compare the anomaly assessment index with the alarm threshold, and determine whether the current target device status is normal.

[0010] After determining that the current target device is in normal condition, an anomaly assessment sequence consisting of the most recent L consecutive window anomaly assessment indicators is constructed. A trend function is established using the anomaly assessment sequence as the fitting object. Based on the trend function, the remaining safe operating time of the target device when it reaches the alarm threshold is predicted.

[0011] Furthermore, the spectral eigenvectors include:

[0012] The root mean square amplitude of the fundamental current and the root mean square amplitude of the fundamental voltage; the root mean square amplitudes of the low and middle order odd harmonic currents and voltages within the preset harmonic order range; and the root mean square amplitudes of the current and voltage in the B frequency bands divided by order within the preset high harmonic order range.

[0013] Furthermore, methods for constructing group sparse regression models include:

[0014] H sets of training data are collected in advance, where H is a positive integer greater than 1;

[0015] Each set of training data corresponds to a time window, including the spectral feature vector of the target device in that time window, the spectral feature vectors of other devices besides the target device in the same time window, and the window time identifier used for alignment;

[0016] The input sample is the spectral feature vectors of the devices other than the target device in the time window, concatenated in device index order.

[0017] Perform data preprocessing on the training sample set;

[0018] Perform random Fourier feature mapping on each device subvector in the input vector to obtain the corresponding explicit finite-dimensional random feature vector; then concatenate the explicit finite-dimensional random feature vectors corresponding to each device according to the device index order to obtain the mapped input feature vector, and use the mapped input feature vector as the input feature of the group sparse regression model;

[0019] In an explicit finite-dimensional feature space, a group sparse regression model is constructed. The group sparse regression model takes the target device spectral feature vector as the fitting target, groups the input vector according to the device source, and each device group corresponds to a non-target device. Sparse constraints are applied on a per-device basis.

[0020] The training sample set is divided into a training set, a validation set, and a test set according to a preset ratio, and training termination conditions are set.

[0021] The regression coefficients of each equipment group are initialized, and iterative solutions are performed to obtain the set of group regression coefficients.

[0022] Furthermore, methods for performing iterative solutions to obtain the set of group regression coefficients include:

[0023] Initialize the regression coefficients of each equipment group to a zero vector;

[0024] In each iteration, the regression coefficients of the remaining device groups except the current device group to be updated are fixed, and only the regression coefficients of the current device group are updated, so that the fitting error of the fitting target is reduced and the regression coefficients meet the group-level sparsity constraint.

[0025] Updated cyclically according to the order of equipment groups;

[0026] Block coordinate descent is used as the iterative update method, and an early stopping mechanism is set. Training stops when any of the following conditions are met: the maximum number of iterations is reached; or the decrease in fitting error between two adjacent iterations is less than the convergence threshold; or the decrease in validation set error for T consecutive iterations is less than the preset threshold ε.

[0027] Once the stopping condition is met, output the set of group regression coefficients obtained in the last iteration as the training result.

[0028] Furthermore, the method for obtaining the pure feature vector is as follows:

[0029] S111: During the real-time operation phase, random Fourier feature mapping is performed on the spectral feature vectors of each non-target device except the target device to obtain the corresponding explicit finite-dimensional random feature vectors.

[0030] S112: Renumber all devices except target device i from index 1 to N-1, and denote the g-th renumbered device as device g. For each feature dimension... , , The external disturbance component is calculated using the set of group regression coefficients obtained by training the group sparse regression model, which is the feature dimension of the spectral feature vector.

[0031] S113: Extract the external interference components of each feature dimension from the spectral feature vector of the target device to obtain the pure feature vector of the target device.

[0032] Furthermore, the method for obtaining the set of highly explicit windows is as follows:

[0033] Calculate the spectral sum and fluctuation gradient of the pure feature vectors for each window of the target device;

[0034] A significance index is obtained by weighted summation of the spectrum and the fluctuation gradient;

[0035] Using the current window as the baseline, obtain the significance index of the most recent L consecutive windows, sort the L windows in descending order according to the significance index, and take the top Q% of windows to form a set of highly significant windows.

[0036] Furthermore, the methods for obtaining the alarm threshold include:

[0037] The feature dimension of any single pure feature vector is defined as the degree of freedom for statistical analysis. ;

[0038] Preset significance level α;

[0039] Calculate the cumulative probability P of health data falling into the safe zone;

[0040] Based on the squared Mahalanobis distance, under the health assumption, it approximately follows a degree of freedom of 1. The statistical properties of the chi-square distribution are determined using statistical calculation tools. The inverse cumulative distribution function is then calculated, with the cumulative probability P and degrees of freedom input. The value returned by the inverse cumulative distribution function is defined as the alarm threshold.

[0041] Furthermore, methods for determining whether the current target device is in normal condition include:

[0042] Acquire historical operating data of the target device under healthy operating conditions, divide it into multiple healthy windows according to a fixed sliding window, calculate the corresponding pure feature vector for each healthy window, and collect the pure feature vectors to form a baseline dataset;

[0043] Calculate the health mean vector and covariance matrix of the target device based on the baseline dataset;

[0044] Based on the health mean vector and covariance matrix, the squared Mahalanobis distance is calculated for the pure feature vectors of each high-dominance window in the set of high-dominance windows to obtain the window outliers of each high-dominance window.

[0045] The outlier values ​​of the window are averaged based on the number of high-explicit windows in the set of high-explicit windows to obtain the outlier assessment index.

[0046] Compare the anomaly assessment metrics with the alarm thresholds:

[0047] If the abnormal assessment index is greater than or equal to the alarm threshold, the current state of the target device is determined to be abnormal, and an abnormal warning is immediately triggered.

[0048] If the abnormal assessment index is less than the alarm threshold, the target device is determined to be in a normal state.

[0049] Furthermore, the method for obtaining the remaining safe operating time is as follows:

[0050] S121: Calculate the window outliers for each of the L most recent consecutive windows. Arrange the L outliers in window time order to construct an anomaly assessment sequence;

[0051] S122: Based on the fitted trend function of the anomaly assessment sequence, the trend of the window outlier value with the window number is characterized. A quadratic polynomial regression model is used as the trend function, and the window outlier value in the anomaly assessment sequence is used as the fitted sample value. The regression coefficient is determined by the least squares method.

[0052] S123: After obtaining the fitted trend function, solve the equation for the trend function to reach the alarm threshold to obtain one or more candidate window numbers. When there is a real solution among the candidate window numbers that is greater than the current window number, select the smallest real solution as the candidate number for the threshold to be reached. When the candidate number is a decimal, round it up to an integer window number as the final window number. Calculate the difference between the final window number and the current window number to obtain the predicted number of remaining windows. Multiply the predicted number of remaining windows by a fixed sliding time to obtain the predicted remaining safe running time and output it.

[0053] Based on the power source characteristic spectrum coupling separation and state co-prediction system, an implementation method based on power source characteristic spectrum coupling separation and state co-prediction is carried out, including:

[0054] Data acquisition module: Collects current and voltage signals of all devices in the shared power supply topology and extracts spectral feature vectors;

[0055] Model training module: Perform random Fourier feature mapping on the spectral feature vectors of devices other than the target device and train a group sparse regression model to obtain a set of group regression coefficients;

[0056] Coupling stripping module: Calculates external interference components based on the set of group regression coefficients, and strips the external interference components from the spectral feature vector of the target device to obtain the pure feature vector of the target device;

[0057] Explicit screening module: Calculates the significance index of the pure feature vector of each window of the target device to obtain a set of highly explicit windows;

[0058] Status judgment module: Obtain the window anomaly value of each high-visibility window in the high-visibility window set, calculate the average of the window anomaly values ​​to obtain the anomaly evaluation index, compare the anomaly evaluation index with the alarm threshold, and determine whether the current target device status is normal.

[0059] Status prediction module: After determining that the current target device is in normal status, it constructs an anomaly assessment sequence consisting of the most recent L consecutive window anomaly assessment indicators, and uses the anomaly assessment sequence as the fitting object to establish a trend function. Based on the trend function, it predicts the remaining safe operating time of the target device to reach the alarm threshold.

[0060] The technical effects and advantages of this invention based on the power source characteristic spectrum coupling separation and state cooperative prediction system and method are as follows:

[0061] First, this invention utilizes random Fourier features to introduce nonlinear fitting capabilities, and combines this with health condition data to learn and quantify the interference contributions of other devices to the target device's feature vector. It accurately calculates external interference components and separates them from the target device's spectral feature vector, obtaining a pure feature vector. This coupling separation strategy solves the problem of overlapping coupling interference and ontological degradation features, ensuring that the feature data used for subsequent condition assessment is pure and improving the accuracy of condition monitoring.

[0062] Then, based on the obtained pure feature vector, the present invention calculates the significance index and filters the set of high-dominance windows, so that the state judgment focuses on the key windows with significant feature changes; based on the health mean vector and covariance matrix obtained from the health condition baseline dataset, the squared Mahalanobis distance of the set of high-dominance windows is calculated and averaged to obtain the anomaly assessment index. The anomaly assessment index is compared with the alarm threshold determined based on the chi-square distribution to achieve high-confidence identification of the degree of deviation of multi-dimensional features from the health state and improve the reliability of the state judgment.

[0063] Finally, after determining that the target equipment is in a normal state, this invention constructs an anomaly assessment sequence based on anomaly assessment indicators within a continuous time window, and fits the changing trends of the anomaly assessment indicators over time to predict the remaining time before the anomaly assessment indicators reach the alarm threshold, which is then used as the remaining safe operating time of the target equipment. By combining threshold determination with trend extrapolation, this invention not only enables timely judgment of the current state of the target equipment but also provides quantifiable predictions of the remaining safe operating time. This provides maintenance personnel with a basis for decision-making regarding advance maintenance, spare parts preparation, and scheduling optimization, thereby reducing the risk of unplanned downtime caused by sudden failures and improving the continuity and safety of industrial production.

[0064] In summary, coupling and separation solve the problem of data purity for target equipment, high-exposure window sets ensure accurate and efficient status judgment, and remaining safe operating time prediction further realizes a closed loop from "anomaly identification" to "risk warning and early intervention." The synergy of these three aspects enables accurate judgment and precise prediction of the target equipment status, avoiding the risk of unexpected equipment downtime due to long-term inaccurate status assessments, and ensuring the continuity and efficiency of industrial production. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the process of the power source characteristic spectrum coupling separation and state cooperative prediction method in Embodiment 1 of the present invention;

[0066] Figure 2 This is a flowchart of the target device status determination process in Embodiment 1 of the present invention;

[0067] Figure 3This is a schematic diagram of the power source characteristic spectrum coupling separation and state collaborative prediction system in Embodiment 2 of the present invention. Detailed Implementation

[0068] 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.

[0069] Example 1

[0070] This embodiment provides a method based on power source characteristic spectrum coupling separation and state co-prediction, including:

[0071] Collect current and voltage signals of all devices in the shared power supply topology and extract spectral feature vectors.

[0072] The current and voltage signals of all N devices in the shared power supply topology are collected. The continuous signal is divided into windows with a fixed sliding time, and each window is used as a sample. For each window w, a fast Fourier transform is performed to obtain the spectrum, and the spectral feature vector of each device j is extracted. .in, j represents the device index, and N represents the total number of devices in the shared power supply topology that participate in spectral feature extraction and subsequent coupling and decoupling. w represents the sample window index, and W represents the total number of sample windows obtained by dividing the continuous signal into fixed sliding time windows within an observation period.

[0073] The spectral feature vector includes: the root mean square amplitude of the fundamental current and the root mean square amplitude of the fundamental voltage; the root mean square amplitudes of the low and middle odd harmonic currents and voltages within a preset harmonic order range; and the root mean square amplitudes of the current and voltage in B frequency bands divided according to the order within a preset high harmonic order range.

[0074] Specifically, the low and medium odd harmonics within the preset range can be selected as the 3rd, 5th, 7th, up to the 25th harmonics. The preset high harmonic frequency band is divided into B sub-bands, which can be selected as the 26th to 100th harmonic frequency band, and further divided into three sub-bands: 26th to 40th, 41st to 60th, and 61st to 100th.

[0075] It should be noted that the spectral feature vector is used to characterize the power supply side operating characteristics of the device within this time window. Specifically, the root mean square amplitude of the fundamental current and the root mean square amplitude of the fundamental voltage characterize the basic load level and power supply amplitude characteristics of the device within this time window; the root mean square amplitudes of the low- and mid-order odd harmonic currents and voltages within a preset harmonic order range characterize the main harmonic distortion components caused by nonlinear loads; and the comprehensive amplitude index of current and voltage in several frequency bands divided according to order within a preset high-order harmonic order range reflects the changing trend of high-frequency switching ripple and electromagnetic disturbances on the frequency band scale while maintaining controllable feature dimensions.

[0076] The feature dimension of the spectral eigenvector is denoted as K, where K is a positive integer greater than 1. The feature dimension is determined by the number of selected fundamental components, the number of mid-to-low-order odd harmonic orders, and the number of high-order harmonic sub-bands. Specifically, the root-mean-square (RMS) amplitude of the fundamental current and the fundamental voltage each correspond to one feature dimension, for a total of two feature dimensions. In the mid-to-low-order odd harmonic components, each selected harmonic order corresponds to two feature dimensions: the RMS amplitude of the current and the RMS amplitude of the voltage. In the high-order harmonic components, each sub-band corresponds to two feature dimensions: the RMS amplitude of the current and the RMS amplitude of the voltage. The feature dimension K of the spectral eigenvector can be represented as the total dimension obtained by adding the above three feature dimensions.

[0077] For example, when the low-order odd harmonics are selected as odd harmonics of the 3rd to 25th orders, there are a total of 12 harmonic orders; when the high-order harmonic frequency band is divided into three sub-bands of the 26th to 40th, 41st to 60th, and 61st to 100th orders, the number of sub-bands is 3; in this example, the feature dimension K of the spectral feature vector is 2 + 2 × 12 + 2 × 3, that is, K = 32.

[0078] The spectral feature vectors are mapped to an explicit finite-dimensional feature space using stochastic Fourier eigenmaps to characterize the nonlinear coupling relationship between different devices through a shared power supply topology. In the explicit finite-dimensional feature space, a group sparse regression model is pre-constructed to group and model the feature contributions from different devices, and the training yields a set of group regression coefficients to characterize the degree of interference effect of each device on the target device.

[0079] The method for constructing the group sparse regression model includes:

[0080] H sets of training data are collected in advance, where H is a positive integer greater than 1;

[0081] The training data is selected from the historical operating data of devices in a healthy operating condition in the shared power supply topology and obtained by dividing it by a sliding window. The healthy operating condition is used to characterize that the device itself is fault-free and the operating conditions are within the normal range, so as to reduce the interference of fault conditions on the learning of coupling relationship.

[0082] Each set of training data corresponds to a time window, including the spectral feature vector of the target device in that time window, the spectral feature vectors of other devices besides the target device in the same time window, and the window time identifier used for alignment.

[0083] For example, the window length can be set to 5s, the sliding step can be set to 1s, and the number of training windows H can be set to no less than 2000 groups;

[0084] The input sample is the spectral feature vectors of the devices other than the target device within the time window, concatenated in device index order.

[0085] The training sample set undergoes data preprocessing, including window time alignment, outlier removal, and feature dimension unification. For example, z-score standardization can be applied to each spectral feature dimension to ensure zero mean and unit variance on the training set. The mean and variance parameters required for standardization are saved for use in the online phase when applying the same transformation to new window features.

[0086] To characterize the nonlinear coupling relationships in a shared power supply topology, random Fourier feature mapping is performed on each device sub-vector in the input vector to obtain the corresponding explicit finite-dimensional random feature vector. Then, the explicit finite-dimensional random feature vectors corresponding to each device are concatenated according to the device index order to obtain the mapped input feature vector, which is used as the input feature of the group sparse regression model. The grouping unit of the group sparse regression model is a device group. Each device group corresponds to all the dimensional features obtained by random Fourier feature mapping of a non-target device. Sparse constraints are applied on a device group basis to learn the coupling contribution of each non-target device to the spectral features of the target device. For example, the random feature dimension D can be set to 256–1024, preferably 512; the kernel bandwidth or scale parameter can be set to 0.8–1.5 times the median distance of the input vectors in the training set, preferably 1.0 times; random frequencies and phase offsets are generated according to a preset random method and are fixed and saved together with the model parameters after training to ensure consistent mapping during the online phase.

[0087] A group sparse regression model is constructed in an explicit finite-dimensional feature space. This model uses the target device's spectral feature vector as the fitting target, groups the input vectors according to device origin, and applies sparsity constraints on each device group. This allows the group sparse regression model to automatically highlight device groups with significant influence on the target device while suppressing those with weaker influence, thus obtaining coupling contribution parameters with device-level interpretability. For example, the regularization strength λ can be set to 10. -4 ~10 -1The logarithmic scaling candidate set is used, and the optimal λ is determined by cross-validation or Bayesian optimization.

[0088] The training sample set is divided into a training set, a validation set, and a test set according to a preset ratio, and training termination conditions are set. For example, the training set, validation set, and test set can be divided into 70%, 15%, and 15% respectively; the maximum number of iterations can be set to 100–300; and the convergence threshold can be set to 10. -5 ;

[0089] The regression coefficients of each equipment group are initialized, and iterative solutions are performed to obtain the set of group regression coefficients.

[0090] Specifically, the regression coefficients of each group are initialized to zero vectors; in each iteration, the coefficients of the remaining device groups except the current device group to be updated are fixed, and only the regression coefficients of the current device group are updated, so that the fitting error of the fitting target is reduced and the regression coefficients meet the group-level sparsity constraint; the updates are performed in a loop according to the device group order.

[0091] Block coordinate descent is used as the iterative update method, and an early stopping mechanism is set. Training stops when any of the following conditions are met: the maximum number of iterations is reached; or the decrease in fitting error between two adjacent iterations is less than the convergence threshold; or the decrease in validation set error for T consecutive iterations is less than a preset threshold ε, where T can be set to 10 and ε can be set to 10. -3 .

[0092] Once the stopping condition is met, output the set of group regression coefficients obtained in the last iteration as the training result.

[0093] The set of group regression coefficients is used to characterize the coupling contribution strength of each device group to the target device in each feature dimension. It is used in the online stage to estimate the interference component introduced by external devices through the shared power supply topology coupling in the spectral features of the target device based on the spectral features of the other devices, and to remove the interference component from the spectral feature vector of the target device to obtain a pure feature vector.

[0094] The external interference component is calculated based on the set of regression coefficient parameters obtained from the output of the group sparse regression model, and then the external interference component is extracted from the spectral feature vector of the target device to obtain the pure feature vector of the target device. The method includes:

[0095] S111: During the real-time operation phase, random Fourier feature mapping is performed on the spectral feature vectors of each non-target device except the target device to obtain the corresponding explicit finite-dimensional random feature vectors.

[0096] S112: Renumber the remaining devices except the target device i from index 1 to N-1, and denote the g-th device after renumbering as device g. For each feature dimension k, , Given the feature dimension of the spectral eigenvector, the external disturbance component is calculated using the set of group regression coefficients obtained from training a group sparse regression model:

[0097] ;

[0098] in, Let be the external interference component of the target device i in the kth feature dimension. Let g be the group regression coefficient of the k-th feature dimension and device g. Let represent the explicit finite-dimensional random eigenvector of the g-th device. This represents the transpose of a vector.

[0099] S113: Calculate the external disturbance components for each feature dimension. The pure feature vector of the target device is obtained by stripping it from the spectral feature vector of the target device. :

[0100] ;

[0101] in, Let i represent the spectral feature vector of device i, i represent the target device index, w=1,2,…,W represent the window index of continuous sampling, and the window index increases in the sliding sampling order, and there is no need to pre-limit a fixed sampling time period or a fixed total number of windows in real time operation.

[0102] In this embodiment, the spectral feature vector extracted by the target device within the sample window can be considered as being composed of two superimposed parts: one part reflects the feature components caused by changes in the target device's own load and operating status; the other part is the interference contribution component of other devices coupled to the target device's measurement quantities through the shared power supply topology. The set of group regression coefficient parameters obtained through the group sparse regression model is used to estimate the coupling contribution of other devices in each feature dimension, obtaining the external interference component. This external interference component is then stripped from the spectral feature vector of the target device, thus obtaining the pure feature vector.

[0103] Therefore, the role of pure feature vectors is to prioritize the use of decoupled feature representations in subsequent state recognition, state co-prediction, or anomaly detection, thereby reducing feature contamination caused by co-source power supply coupling, harmonic superposition, and measurement crosstalk, making the state changes of the target device more distinguishable and stable.

[0104] The significance index is calculated for the pure feature vector of each window of the target device to obtain the set of highly explicit windows.

[0105] After separating the coupling interference from the target device, the system obtained a clean feature vector. In actual condition prediction or health assessment, not all features at every point in time are equally important. Features in some time periods contribute significantly to equipment status information, while features in other time periods may mainly contain minor fluctuations or residual noise. To improve prediction accuracy, this embodiment introduces a high-exposure window selection step.

[0106] Pure feature vector for each window w of the target device Calculate the significance index S(w), which measures the importance of the feature at the current point in time.

[0107] The significance index includes three components: the first is the sum of the spectral frequencies, used to capture fluctuation segments with strong energy and significant amplitude changes; the calculation formula is as follows:

[0108] ;

[0109] in, The sum of the spectrum, Let i be the pure feature vector of the k-th feature dimension of the target device i within the window w.

[0110] Second is the fluctuation gradient, used to capture transient, rapidly changing features. Its calculation formula is:

[0111] ;

[0112] in, The gradient represents the fluctuation.

[0113] The significance index is the weighted sum of the two components above:

[0114] ;

[0115] in, As a significance indicator, and The weighting is set by those skilled in the art based on their experience. For example, it can be automatically calibrated based on baseline data of healthy operating conditions, and determined by making the sum of the spectra and the contribution of the fluctuation gradient to the significance index under healthy operating conditions of the same order of magnitude.

[0116] Using the current window as the benchmark, obtain the significance index S(w) of the most recent consecutive L windows, sort the L windows in descending order according to the significance index S(w), and take the top Q% of windows to form a set of highly significant windows.

[0117] Where L is the preset number of windows and Q is the preset ratio. L can be selected from 1000 to 1500 windows, and Q can be selected from 10% to 30%, with 20% being preferred.

[0118] In this embodiment, the introduction of a highly explicit window selection mechanism enables an on-demand evaluation strategy. This effectively filters out weak signal changes caused by residual noise, communication jitter, or slight environmental fluctuations, ensuring that only abnormal signals occurring during critical time periods with sufficient physical information and significant feature fluctuations are sent to the next stage of anomaly detection. This reduces sensitivity to background noise. By focusing the analysis on these time periods with the most significant degradation and highest information gain, the datasets used for training and evaluation are optimized, avoiding dilution by redundant, stable data. This improves the ability to capture subtle degradation patterns and the accuracy of subsequent state judgments.

[0119] See Figure 2 As shown, the pure feature vectors of each high-dominance window in the high-dominance window set are respectively compared with the health mean vector and covariance matrix determined by the coupled and separated health condition baseline dataset to calculate the squared Mahalanobis distance, thereby obtaining the window outlier value of each high-dominance window. The outlier value is averaged to obtain the anomaly assessment index. The anomaly assessment index is compared with the alarm threshold to determine whether the current target device status is normal.

[0120] The historical operating data of the target device under healthy operating conditions is obtained, and multiple healthy windows are obtained by dividing the data into a fixed sliding window. The corresponding pure feature vector is calculated for each healthy window, and the pure feature vectors are aggregated to form a baseline dataset.

[0121] The health mean vector of the target device is calculated based on the baseline dataset, using the following formula:

[0122] ;

[0123] in, This is a vector of health mean values. This represents the total number of clean feature vectors in the baseline dataset. Let represent the pure feature vector of the u-th window.

[0124] And the covariance matrix, calculated using the following formula:

[0125] ;

[0126] in, Let be the covariance matrix.

[0127] Based on the healthy mean vector and covariance matrix, the squared Mahalanobis distance is calculated for the pure eigenvectors of each high-dominance window in the set of high-dominance windows to obtain the outlier value of each high-dominance window. The calculation formula is as follows:

[0128] ;

[0129] in, This is an abnormal value for the window. For the pure feature vectors of the highly explicit window, This is a vector of health mean values. Let be the covariance matrix.

[0130] The maximum value of the window outlier is used as the anomaly assessment indicator. .

[0131] The method for obtaining the alarm threshold includes:

[0132] The feature dimension of any single pure feature vector is defined as the degree of freedom for statistical analysis. ;

[0133] The significance level α is preset, for example, α=0.003, which represents the allowable probability of misclassifying health data as abnormal;

[0134] Calculate the cumulative probability P that the health data falls into the safe zone: P = 1 - α;

[0135] Based on the squared Mahalanobis distance, under the health assumption, it approximately follows a degree of freedom of 1. The statistical properties of the chi-square distribution can be determined using statistical computing tools, such as Python's SciPy statistics module, by performing the inverse cumulative distribution function calculation on the chi-square distribution, with the cumulative probability P and degrees of freedom as input. The value returned by the inverse cumulative distribution function is defined as the alarm threshold. .

[0136] In this embodiment, the alarm threshold calculation relies on the mean vector and covariance matrix of the pure feature vector under healthy operating conditions, and uses the dimension of the pure feature vector as the statistical degree of freedom, which can accurately match the real health status of the target device and avoid judgment deviation caused by the mismatch between the general threshold and the actual health characteristics of the device.

[0137] Anomaly assessment indicators With alarm threshold Comparison:

[0138] If abnormal assessment indicators Greater than or equal to the alarm threshold If the target device is determined to be in an abnormal state, an anomaly warning will be triggered immediately.

[0139] If abnormal assessment indicators Less than the alarm threshold The target device is determined to be in a normal state.

[0140] After determining that the current target device is in normal condition, an anomaly assessment sequence is constructed and a trend function is fitted to predict the remaining safe operating time. The specific method is as follows:

[0141] S121: Calculate the window outliers for each of the L most recent consecutive windows. The L outliers in each window are arranged in chronological order to construct an anomaly assessment sequence.

[0142] S122: Based on the aforementioned anomaly assessment sequence, a trend function is fitted to characterize the trend of window outliers as a function of window number. Considering the potential for nonlinear acceleration during equipment degradation, a quadratic polynomial regression model is used as the trend function, and the window outliers in the anomaly assessment sequence are used as the fitted sample values. The regression coefficients are determined using the least squares method.

[0143] ;

[0144] in, Let z be the fitting function, z be the window number, and b, c, and d be the regression coefficients to be determined.

[0145] The window numbering is based on the earliest window among L consecutive windows, with the starting window number being z=1.

[0146] The regression coefficients are measured using the least squares method, which ensures that the selected regression coefficients minimize the sum of squared residuals between the values ​​of the fitted function and the values ​​of the anomaly assessment index.

[0147] S123: Obtain the well-fitted trend function Then, by solving the trend function Reaching alarm threshold The equation yields one or more candidate window numbers. When there is a real solution among the candidate window numbers that is greater than the current window number, the smallest real solution is selected as the candidate number to reach the threshold. When the candidate number is a decimal, it is rounded up to an integer window number, which is then used as the final window number. The difference between the final window number and the current window number is calculated to obtain the predicted number of remaining windows. The predicted number of remaining windows is multiplied by a fixed sliding time to obtain the predicted remaining safe running time and is then output.

[0148] It should be noted that if the equation has no solution or the real solutions are all less than the current window number, zero will be output directly as an invalid prediction value, indicating that the prediction failed to obtain a usable remaining safe running time result.

[0149] Example 2

[0150] See Figure 3 As shown, this embodiment provides a power source characteristic spectrum coupling separation and state co-prediction system. Implementing the power source characteristic spectrum coupling separation and state co-prediction method includes:

[0151] Data acquisition module: Collects current and voltage signals of all devices in the shared power supply topology and extracts spectral feature vectors;

[0152] Model training module: Perform random Fourier feature mapping on the spectral feature vectors of devices other than the target device and train a group sparse regression model to obtain a set of group regression coefficients;

[0153] Coupling stripping module: Calculates external interference components based on the set of group regression coefficients, and strips the external interference components from the spectral feature vector of the target device to obtain the pure feature vector of the target device;

[0154] Explicit screening module: Calculates the significance index of the pure feature vector of each window of the target device to obtain a set of highly explicit windows;

[0155] Status judgment module: Obtain the window anomaly value of each high-visibility window in the high-visibility window set, calculate the average of the window anomaly values ​​to obtain the anomaly evaluation index, compare the anomaly evaluation index with the alarm threshold, and determine whether the current target device status is normal.

[0156] Status prediction module: After determining that the current target device is in normal status, it constructs an anomaly assessment sequence consisting of the most recent L consecutive window anomaly assessment indicators, and uses the anomaly assessment sequence as the fitting object to establish a trend function. Based on the trend function, it predicts the remaining safe operating time of the target device to reach the alarm threshold.

[0157] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0158] In conclusion, the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method based on power signature spectrum coupling separation and state collaborative prediction, characterized in that, The method comprises the following steps: Collecting current and voltage signals of all devices in the power supply sharing topology and extracting a spectrum feature vector; Performing random Fourier feature mapping on the spectrum feature vectors of devices other than the target device and training a group sparse regression model to obtain a group regression coefficient set; Calculating an external interference component based on the group regression coefficient set and removing the external interference component from the spectrum feature vector of the target device to obtain a pure feature vector of the target device; Calculating a significance index of the pure feature vector of each window of the target device to obtain a high significance window set; Obtaining window outliers of each high significance window in the high significance window set, averaging the window outliers to obtain an anomaly evaluation index, comparing the anomaly evaluation index with an alarm threshold, and determining whether the current state of the target device is normal; After determining that the current state of the target device is normal, constructing an anomaly evaluation sequence composed of the last L consecutive window anomaly evaluation indexes, establishing a trend function taking the anomaly evaluation sequence as a fitting object, and predicting the remaining safe operation time of the target device to reach the alarm threshold based on the trend function.

2. The power source feature spectrum-based coupling separation and state collaborative prediction method according to claim 1, characterized in that, The spectrum feature vector comprises: The root mean square amplitude of the fundamental current and the root mean square amplitude of the fundamental voltage; the root mean square amplitudes of the low-order odd harmonic currents and voltages within a preset harmonic order range; and the root mean square amplitudes of the current and voltage of B frequency band subbands divided by order within a preset high-order harmonic order range.

3. The power source feature spectrum-based coupling separation and state collaborative prediction method according to claim 1, characterized in that, The training method of the group sparse regression model comprises: Pre-collecting H sets of training data, where H is a positive integer greater than 1; Each set of training data corresponds to a time window and comprises a spectrum feature vector of the target device in the time window, spectrum feature vectors of devices other than the target device in the same time window, and a window time identifier for alignment; The input sample is an input vector obtained by concatenating the spectrum feature vectors of the remaining devices in the time window in the order of device index; Data preprocessing is performed on the training sample set; Random Fourier feature mapping is performed on each device sub-vector in the input vector to obtain corresponding explicit finite-dimensional random feature vectors; then the explicit finite-dimensional random feature vectors corresponding to each device are concatenated in the order of device index to obtain a mapped input feature vector, which is used as the input feature of the group sparse regression model; A group sparse regression model is constructed in the explicit finite-dimensional feature space, which takes the spectrum feature vector of the target device as a fitting target, groups the input vector by device source, and applies sparse constraints on each device group; The training sample set is divided into a training set, a validation set and a test set according to a preset proportion, and a training termination condition is set; The regression coefficients of each device group are initialized and iteratively solved to obtain a group regression coefficient set.

4. The power source feature spectrum-based coupling separation and state co- prediction method according to claim 3, characterized in that, The method of iteratively solving to obtain the group regression coefficient set comprises: Initializing the regression coefficients of each device group to zero vectors; In each iteration round, the regression coefficients of the rest of the device groups except the current device group to be updated are fixed, only the regression coefficients of the current device group are updated, the fitting error of the fitting target is reduced, and the regression coefficients meet the group-level sparse constraint; The updating is performed in sequence by cycling through the device groups; The block coordinate descent is adopted as the iterative updating method, and an early stop mechanism is set, and the training is stopped when any of the following conditions is met: the maximum number of iterations is reached; or the fitting error of the adjacent two iteration rounds decreases by less than the convergence threshold; or the validation set error decreases by less than the preset threshold ε for T consecutive rounds; When the stop condition is met, the group regression coefficient set obtained in the last iteration round is output as the training result.

5. The power source feature spectrum-based coupling separation and state co- prediction method according to claim 1, characterized in that, The method for obtaining the pure feature vector comprises the following steps: S111: In the real-time running stage, the random Fourier feature mapping is performed on the spectrum feature vectors of each non-target device except the target device to obtain the corresponding explicit finite-dimensional random feature vector; S112: renumber the remaining devices except the target device i in index order as 1 to N-1, and the gth device after renumbering is recorded as device g, and the feature vector of each feature dimension of device g is recorded as , , is the feature dimension of the spectrum feature vector, and the external interference component is calculated using the group regression coefficient set obtained by training the group sparse regression model; S113: The external interference component of each feature dimension is calculated and removed from the spectrum feature vector of the target device to obtain the pure feature vector of the target device.

6. The power source feature spectrum-based coupling separation and state co- prediction method according to claim 1, characterized in that, The method for obtaining the high-visibility window set comprises the following steps: The spectrum sum and fluctuation gradient of the pure feature vector are calculated for each window of the target device; The spectrum sum and fluctuation gradient are weighted and summed to obtain the significance index; Taking the current window as a reference, the significance indexes of the last L consecutive windows are obtained, the L windows are arranged in descending order according to the significance indexes, and the top Q% windows are taken as the high-visibility window set.

7. The power source feature spectrum-based coupling separation and state co- prediction method according to claim 1, characterized in that, The method for obtaining the alarm threshold comprises the following steps: defining the feature dimension of any single pure eigenvector as the degree of freedom for the statistical analysis ; A preset significance level α is set; The cumulative probability P of the health data falling into the safe area is calculated; Based on the statistical property that the Mahalanobis distance square approximates to the chi-square distribution with the degree of freedom of under the healthy hypothesis, the inverse cumulative distribution function is calculated for the chi-square distribution using a statistical calculation tool, the cumulative probability P and the degree of freedom are input, and the value returned from the inverse cumulative distribution function is defined as the alarm threshold.

8. The power source feature spectrum-based coupling separation and state co- prediction method according to claim 1, characterized in that, The method for determining whether the current target device state is normal comprises the following steps: The historical running data of the target device in the healthy working condition are obtained, the data are divided into multiple healthy windows by a fixed sliding window, the pure feature vector of each healthy window is calculated, and the pure feature vectors are collected to form a baseline data set; The health mean vector and the covariance matrix of the target device are calculated based on the baseline data set; The Mahalanobis distance square of the pure feature vector of each high-visibility window in the high-visibility window set is calculated based on the health mean vector and the covariance matrix, to obtain the window abnormal value of each high-visibility window; The window abnormal value is averaged according to the number of high-visibility windows in the high-visibility window set to obtain an abnormal evaluation index; The abnormal evaluation index is compared with the alarm threshold: If the abnormal evaluation index is greater than or equal to the alarm threshold, it is determined that the current state of the target device is abnormal, and an abnormal early warning is triggered immediately; If the abnormal evaluation index is less than the alarm threshold, it is determined that the current state of the target device is normal.

9. The power source feature spectrum-based coupling separation and state co- prediction method according to claim 1, characterized in that, The method for obtaining the remaining safe running time comprises the following steps: S121: Calculate the window outliers of each window in the last L continuous windows Arrange the L window outliers in window time order to construct an anomaly evaluation sequence; S122: A trend function is fitted based on the abnormal evaluation sequence to depict the trend of the window abnormal value with the window sequence number, a quadratic polynomial regression model is adopted as the trend function, the window abnormal values in the abnormal evaluation sequence are taken as the fitting sample values, and the regression coefficients are determined by the least square method. S123: After obtaining the fitted trend function, one or more candidate window numbers are obtained by solving the equation of the trend function reaching the alarm threshold. When there is a real solution greater than the current window number in the candidate window number, the smallest real solution is selected as the candidate number of the threshold reaching. When the candidate number is a decimal, the candidate number is rounded up to an integer window number as the final window number. The difference between the final window number and the current window number is calculated to obtain the predicted remaining window number. The predicted remaining window number is multiplied by the fixed sliding time to obtain the predicted remaining safe operation time and output.

10. A system for power signature spectrum coupling separation and state collaborative prediction, implementing the method for power signature spectrum coupling separation and state collaborative prediction according to any one of claims 1-9, characterized in that, The system comprises: a data acquisition module that acquires current and voltage signals of all devices in the shared power supply topology and extracts a spectrum feature vector; a model training module that performs random Fourier feature mapping on the spectrum feature vectors of devices other than the target device and trains a group sparse regression model to obtain a group regression coefficient set; a coupling and stripping module that calculates an external interference component based on the group regression coefficient set and strips the external interference component from the spectrum feature vector of the target device to obtain a pure feature vector of the target device; a dominant screening module that calculates a significance index for the pure feature vector of each window of the target device to obtain a high significance window set; a state judgment module that obtains window outliers of each high significance window in the high significance window set, averages the window outliers to obtain an anomaly evaluation index, compares the anomaly evaluation index with an alarm threshold, and determines whether the current target device state is normal; a state prediction module that, after determining that the current target device state is normal, constructs an anomaly evaluation sequence composed of the latest L consecutive window anomaly evaluation indexes, establishes a trend function with the anomaly evaluation sequence as the fitting object, and predicts the remaining safe operation time of the target device reaching the alarm threshold.

Images