Edge computing-based fault detection and management system and method for electromagnetic vibration table

Abstract

This invention discloses an edge computing-based fault detection and management system and method for electromagnetic vibration tables, relating to the field of electromagnetic vibration table condition detection technology. The system includes a transmission and acquisition module, a fault analysis module, and a management decision module. Its key technical points are: sending a target excitation sequence to the electromagnetic vibration table and acquiring multiple response sequences output by the electromagnetic vibration table; wherein the electromagnetic vibration table is associated with multiple test task types; extracting features from multiple response sequences through edge nodes, identifying coupling deviations of each feature coefficient, updating evaluation weights using the MapReduce framework, and employing ART neural network execution mode clustering to determine the health index of the electromagnetic vibration table; when acquiring test tasks to be executed, introducing a rule engine, and triggering a graded fault management response in conjunction with the health index; this invention improves the timeliness of fault identification for different test task types and enhances management efficiency.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic vibration table condition detection technology, specifically to an electromagnetic vibration table fault detection and management system and method based on edge computing. Background Technology

[0002] Electromagnetic vibration tables are widely used in aviation, aerospace, automotive, and steel industries. By simulating environmental vibration, they can be used to diagnose faults in various structures and equipment. They need to be compatible with various test task types such as sinusoidal vibration, random vibration, and impact tests. Their health status directly determines the accuracy of test data, the safety of the test process, and the service life of the equipment itself.

[0003] Traditional electromagnetic vibration table fault detection and management technologies have the following limitations: Firstly, they typically rely on a central server for global calculations. Given the multi-dimensional response sequences generated during electromagnetic vibration table testing, high transmission delays are common. While neural network-based health assessment models have been developed to extract features from multi-dimensional response sequences and obtain multi-dimensional feature vectors, the central node struggles to dynamically update large-scale weights in real time due to the coupling of these multi-dimensional feature vectors, resulting in significant lag in health status monitoring. Secondly, when performing current test tasks, such as sinusoidal vibration, the system struggles to capture potential fault characteristics in acceleration signals, driving current signals, and moving coil displacement signals. This leads to a single assessment dimension, failing to accurately identify the correlations and deviation patterns between potential fault characteristics under electromechanical coupling. Furthermore, the lack of task-adaptive hierarchical responses during electromagnetic vibration table fault assessment significantly reduces management efficiency. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an edge computing-based electromagnetic vibration table fault detection and management system and method. By sending a target excitation sequence to the electromagnetic vibration table and collecting multiple response sequences output by the table, features are extracted from these response sequences through edge nodes to identify coupling deviations in each feature coefficient. A MapReduce framework is introduced to dynamically update the evaluation weights in parallel. An ART neural network is used to perform pattern clustering, which eliminates the need for pre-setting the number of clusters or labels, ensuring the accuracy and timeliness of health index generation. A rule engine is introduced to trigger graded fault responses based on the health index, thus solving the problems mentioned in the background technology.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: Firstly, this application provides an edge computing-based fault detection and management system for electromagnetic vibration tables. The system includes: a transmission and acquisition module that sends a target excitation sequence to the electromagnetic vibration table and acquires multiple response sequences output by the electromagnetic vibration table; wherein the electromagnetic vibration table is associated with multiple test task types; a fault analysis module that extracts features from multiple response sequences through edge nodes, identifies the coupling deviation of each feature coefficient, updates the evaluation weights using the MapReduce framework, and uses ART neural network execution mode clustering to determine the health index of the electromagnetic vibration table; and a management decision module that, when acquiring test tasks to be executed, introduces a rule engine and triggers a graded fault management response in conjunction with the health index.

[0006] Furthermore, the target excitation sequence is a set of drive control waveforms containing a preset frequency band distribution, the response sequence includes at least acceleration signals, drive current signals and moving coil displacement signals, and the test task types include at least sinusoidal vibration, random vibration and shock tests.

[0007] Furthermore, the characteristic coefficients are identified, including: Based on any response sequence, extract and calculate the mean coordinates and mean features of the sampling points in the response sequence, and neutralize them to obtain centered coordinates and centered features; construct a design matrix with the centered coordinates as row vectors, the total number of sampling points as the number of rows, and the corresponding feature space dimension as the number of columns; construct observation vectors with the centered features according to the sampling order; establish the correlation between the design matrix and the observation vectors, and determine the corresponding feature coefficients.

[0008] Furthermore, the characteristic coefficients also include the first characteristic coefficient, and the following is performed: The displacement signal of the moving coil is retrieved to determine the first design matrix; the driving current signal is retrieved to determine the first observation vector; the first correlation relationship between the first design matrix and the first observation vector is established, and the odd-numbered coefficients are solved by least squares fitting as the first characteristic coefficients; the nonlinear components in the first characteristic coefficients are extracted and inverted, and normalized by the L2 norm to obtain the main direction of stiffness attenuation.

[0009] Furthermore, the characteristic coefficients also include a second characteristic coefficient, and the following is performed: The acceleration signal is retrieved, and the fundamental frequency component and harmonic components are extracted through Fourier transform. The energy distortion ratio is calculated to determine the second design matrix. The acceleration signal is scalar projected onto the principal direction of stiffness attenuation to obtain the projected acceleration signal to determine the second observation vector. By establishing the second correlation between the second design matrix and the second observation vector, the mapping weights corresponding to each harmonic are determined as the second characteristic coefficients.

[0010] Furthermore, the characteristic coefficients also include a third characteristic coefficient, and the following is performed: All response sequences are input into a pre-defined neural network inverse model. The third observation vector is determined by the residual sequence between the output predicted excitation waveform and the actual input drive control waveform. The frequency distribution of the drive control waveform is extracted to determine the third design moment. Establish the third correlation between the third design matrix and the third observation vector, and determine the weighted sum of the optimal weight vector as the third characteristic coefficient with the goal of minimizing the sum of squared residuals.

[0011] Furthermore, the health index of the electromagnetic vibration table is determined, including: The MapReduce framework is used to perform distributed statistics on the coupling deviation of different feature coefficients, and the triggering frequency of each feature coupling deviation within a preset time window is calculated. The evaluation weight of each feature coefficient is dynamically updated based on the triggering frequency. In the ART neural network, a training layer is introduced, and a competitive learning dataset is formed by combining coupling bias with a pre-stored health template. Randomly selected sample data from the response sequences collected in real time at the edge nodes are provided to perform incremental learning on the training layer. Fault indicators are introduced to perform similarity significance evaluation until the training converges and a health recognition model is obtained. Among them, the fault indicators include standard warning thresholds and initial weights of neurons. The obtained coupling deviations are stratified into test tasks according to time series, and any time series segment is marked as a monitoring step. The real-time coupling deviations under each monitoring step are substituted into the health identification model to determine whether each monitoring step belongs to steady-state deviation or abnormal fault, and the corresponding calculation method is selected. The health index is obtained by combining different feature coefficients.

[0012] Furthermore, if the health type is a steady-state deviation, weighted calculation is selected; if the health type is an abnormal fault, exponential weighted calculation is selected.

[0013] Furthermore, a rules engine is introduced, combined with health index triggering tiered fault management responses, including: Divide the time series into equal time steps, label each response sequence corresponding to each time step with the corresponding health type, and form a feature source; collect all feature sources in the preset monitoring period and aggregate them, and extract the proportion of feature sources whose health type is abnormal fault from the feature source set. Extract the maximum and minimum values ​​of the coupling bias corresponding to each response sequence, and construct multiple task vectors by combining the proportion of feature sources and the current experimental task to be executed, including a first task vector and a second task vector; establish multiple task vectors and a health index. The correlation between the numbers is analyzed and the corresponding correlation degree is extracted. Any response sequence with a correlation degree greater than the corresponding correlation degree threshold is selected. Through cluster analysis, it is identified into three categories: low, medium and high, and a graded fault management response is triggered accordingly.

[0014] Secondly, this application provides a fault detection and management method for an electromagnetic vibration table based on edge computing. The method includes: sending a target excitation sequence to the electromagnetic vibration table and collecting multiple response sequences output by the electromagnetic vibration table; wherein the electromagnetic vibration table is associated with multiple test task types. By extracting features from multiple response sequences through edge nodes, the coupling deviation of each feature coefficient is identified. The evaluation weights are updated using the MapReduce framework, and the ART neural network is used to perform pattern clustering to determine the health index of the electromagnetic vibration table. When acquiring test tasks to be executed, a rules engine is introduced, which, combined with health indices, triggers a graded fault management response.

[0015] (III) Beneficial Effects This invention provides an electromagnetic vibration table fault detection and management system and method based on edge computing, which has the following beneficial effects: 1. This invention establishes an electromechanical coupling mechanism for an electromagnetic vibration table with precise frequency band distribution matching for different test task types. It analyzes the response sequence composed of acceleration signal, driving current signal and moving coil displacement signal to form differentiated and identifiable features, which facilitates subsequent fault analysis. 2. This invention performs feature extraction on multiple response sequences through edge nodes, identifies the coupling deviation of the first to third feature coefficients, and improves the extraction accuracy of fault characteristics in complex environments by using mathematical modeling. It introduces the MapReduce framework to perform parallel dynamic updates of evaluation weights, ensuring the accuracy and timeliness of health index calculation. At the same time, it only updates the weights of the ART neural network training layer, without the need for full retraining, and realizes parallel statistical analysis of the coupling deviation of multiple feature coefficients, thereby improving the robustness of the system throughout its entire life cycle. 3. This invention introduces a rule engine, constructs a first task vector and a second task vector, and establishes a correlation model in conjunction with a health index to trigger graded fault response management, thereby improving the timeliness of fault identification and the efficiency of fault detection for different types of test tasks. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the modules of the present invention. Detailed Implementation

[0017] 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.

[0018] The core of this invention lies in sending a target excitation sequence to an electromagnetic vibration table, collecting multiple response sequences output by the electromagnetic vibration table, extracting features from multiple response sequences through edge nodes, identifying coupling deviations of each feature coefficient, introducing a MapReduce framework to dynamically update evaluation weights in parallel, and using an ART neural network to perform pattern clustering. This network does not require pre-setting the number of clusters or labels, ensuring the accuracy and timeliness of health index generation. A rule engine is introduced to trigger graded fault responses in conjunction with the health index, making it suitable for large-scale, high-frequency electromagnetic vibration table test scenarios.

[0019] Example 1: This invention provides an electromagnetic vibration table fault detection and management system and method based on edge computing; Figure 1 This is a schematic diagram of the modules of the present invention; please refer to it. Figure 1 The system includes: a data acquisition module, a fault analysis module, and a management decision module, and the data acquisition module, the fault analysis module, and the management decision module are connected in communication. The following is an explanation of each module involved: Sending and collecting modules: The target excitation sequence is sent to the electromagnetic vibration table, and multiple response sequences output by the electromagnetic vibration table are collected. The electromagnetic vibration table is associated with multiple test task types. The response sequences include the acceleration signal, driving current signal, and moving coil displacement signal of the electromagnetic vibration table. It should be noted that the electromagnetic vibration table mainly consists of a working platform, moving parts including an armature and coils, a support device, inner and outer magnetic yokes, and permanent magnets. By sending a target excitation sequence to the electromagnetic vibration table, a set of drive control waveforms containing a preset frequency band distribution can be sent to trigger the electromagnetic vibration table to generate a mechanical response characterizing the electromechanical coupling state, thereby enabling... The characteristics of electromagnetic vibration table failures are highlighted in the response sequence, such as the impedance change caused by short coils, which provides effective data support for subsequent edge node feature extraction and fault diagnosis. The test tasks include sinusoidal vibration, random vibration, shock tests, and mixed vibration. Different test task types send different drive control waveforms and have different preset frequency band coverage ranges. While outputting the corresponding target excitation sequence, each task simultaneously acquires the acceleration signal, drive current signal, and moving coil displacement signal of the electromagnetic vibration table as response sequences. Piezoelectric accelerometers, Hall effect closed-loop current sensors, and eddy current displacement sensors are used to acquire the corresponding acceleration response sequences, drive current response sequences, and moving coil displacement response sequences, respectively. None of the aforementioned sensors are shown in the figure; they are adaptively installed, and the obtained response sequences are directly transmitted to the edge computing nodes. For sinusoidal vibration, the drive control waveform table is shown below. The control waveform is characterized as a linear swept-frequency sine wave and a fixed-frequency sine wave. The preset frequency band distribution includes low-frequency, mid-frequency, and high-frequency bands, and each frequency band adopts a graded sweep step size and gradient amplitude setting. For random vibration, the drive control waveform is characterized as a broadband Gaussian random noise waveform. The preset frequency band distribution includes low-frequency, mid-frequency, and high-frequency bands, and the power spectral density is allocated according to different frequency bands. The mid-frequency band is set as the energy core distribution range. For impact tests, the drive control waveform is characterized as a half-sine impact pulse waveform. For mixed vibration, the drive control waveform is characterized as a composite waveform of synchronous superposition of sine sweep, random noise, and impact pulse. Thus, a unique mapping relationship is established between test task type, control waveform form, preset frequency band distribution, and response sequence acquisition.

[0020] For the above typical test task types, an electromechanical coupling mechanism for a frequency band distribution precisely matched electromagnetic vibration table is established, so that different types of fault characteristics form differentiated and identifiable features in the response sequence composed of acceleration signal, driving current signal and moving coil displacement signal, which facilitates subsequent fault analysis.

[0021] Fault Analysis Module: By extracting features from multiple response sequences through edge nodes, identifying coupling deviations of each feature coefficient, updating evaluation weights using the MapReduce framework, and performing pattern clustering using the ART neural network, the health index of the electromagnetic vibration table is determined. Identify the characteristic coefficients, including: Any response sequence is retrieved from the edge nodes. Through frame-based processing, several data frames are divided. For any response sequence within each data frame, the mean coordinates and mean features of the sampling points are extracted and calculated. The range of sampling point numbers (1 to N) is determined by marking the sampling point numbers of N consecutive sampling points in the data frame. The mean coordinates are obtained by summing all sampling point numbers and dividing by the total number of sampling points N. The sampling points are then numbered sequentially. Signal acquisition values ​​from any response sequence within the data frame are collected, and outlier filtering is performed. The mean features are obtained by summing all filtered signal acquisition values ​​and dividing by the total number of sampling points N, where N is a positive integer. Neutralization processing is used to obtain centered coordinates and centered features. For example, using the mean coordinates as the center unifies the temporal reference of data from different frames, avoiding coordinate dimension deviations caused by different sampling start times. Using the mean feature as the center eliminates systematic errors such as DC offset and temperature drift in the corresponding signals, ensuring the accuracy of subsequent coupling relationship modeling and improving the representation accuracy of feature coefficients. For the centered coordinates of each sampling point, feature expansion is performed according to the definition of the feature space dimension, generating a 1×M dimensional row vector. The total number of sampling points is the number of rows, and the corresponding feature space dimension is the number of columns, constructing an N×M dimensional design matrix. Here, the feature space dimension is a predefined setting, M is a positive integer, and each row vector element corresponds one-to-one with the feature space dimension, with each element representing the projection value of the centered coordinate onto the corresponding feature dimension. The centered features are arranged according to the sampling order, i.e., the sampling point indices from 1 to N, and the sorted centered features are converted into an N×1 dimensional column vector, i.e., the observation vector. The relationship between the design matrix and the observation vector is established by mapping the observation vector to the feature space defined by the design matrix, determining the feature coefficients corresponding to the data frame, including the first feature coefficient, second feature coefficient, third feature coefficient, and fourth feature coefficient. The relationship is represented as: the observation vector is the dot product of the design matrix and the feature coefficients. It should be noted that the total number of sampling points is represented by N in the following examples. The characteristic coefficients also include the first characteristic coefficient, and the following steps are performed: the moving coil displacement signal X(t) is retrieved as the first reference term, and a centering process is performed to obtain the centered displacement feature. This feature is then expanded into a row vector according to the odd-order terms, such as by first-order, third-order, and fifth-order terms. The vectors are then arranged vertically according to the sampling order to form an N×3-dimensional first design matrix. The driving current signal I(t), which is completely synchronized with the moving coil displacement signal in time, is retrieved as the first characteristic term, and a centering process is performed to obtain the centered current feature. This current feature is then arranged vertically according to the sampling order, strictly corresponding to the row numbers of the first design matrix, to form an N×1-dimensional first observation vector. Here, t represents the timestamp. Establish the first correlation between the first design matrix and the first observation vector: the first observation vector is equal to the dot product of the first design matrix and the first characteristic coefficient; through least squares fitting, determine the first-order linear coefficient β1, the third-order nonlinear coefficient β3, and the fifth-order nonlinear coefficient β5, construct the corresponding matrix by concatenation, and form the first characteristic coefficient β by transposition; extract the nonlinear component from the first characteristic coefficient, invert it, and normalize it to obtain the main direction of stiffness attenuation: extract the third-order nonlinear coefficient β3 and the fifth-order nonlinear coefficient β5 from the first characteristic coefficient β to form the nonlinear component vector; it should be noted that the first-order term represents the ideal linear coupling relationship between the moving coil displacement and the driving current; the third-order term represents mild nonlinear coupling distortion, such as slight air gap inhomogeneity; the fifth-order term represents severe nonlinear coupling distortion, such as severe moving coil eccentricity and local short circuit of the coil; in addition, the sign of the nonlinear coefficient directly reflects the direction and degree of coupling distortion: if the sign of the nonlinear coefficient is positive, it indicates that the nonlinear coupling of that order is enhanced; if the sign of the nonlinear coefficient is negative, it indicates This order of nonlinear coupling attenuation; the elements of the nonlinear component vector are inverted, and the inverted nonlinear component vector is normalized by L2 norm to eliminate the influence of magnitude, and finally the normalized stiffness attenuation principal direction vector is obtained. The characteristic coefficients also include a second characteristic coefficient, and the following steps are performed: The acceleration signal a(t) that is completely synchronized with the timing of the moving coil displacement signal is retrieved, and the fundamental frequency component and harmonic components are extracted after Fourier transform. The fundamental frequency component corresponds to the rated excitation frequency of the drive control waveform and is the reference component of the frequency domain energy. The harmonic components include multiple orders, such as selecting the 2nd, 3rd, 4th, and 5th harmonics. The amplitudes of the fundamental frequency component and each harmonic component are extracted to determine the energy of the extracted fundamental frequency component and each harmonic component. For each harmonic component, the ratio of the harmonic energy to the fundamental frequency energy is used as the energy distortion ratio of that harmonic to form an energy distortion ratio sequence, which serves as the second reference term. An N×4-dimensional second design matrix is ​​constructed in the order of the 2nd harmonic energy distortion ratio, the 3rd harmonic energy distortion ratio, the 4th harmonic energy distortion ratio, and the 5th harmonic energy distortion ratio. The acceleration signal is scalar projected onto the principal direction of stiffness attenuation, including: extracting the amplitude of each sampling point of the acceleration signal, multiplying it by the value of the principal direction vector of stiffness attenuation to obtain the projected amplitude of each sampling point, forming the projected acceleration signal according to the sampling order, and using the projected acceleration signal as the second feature term. The second observation vector is determined through centering processing and is 4×1 dimension. A second correlation relationship is established between the second design matrix and the second observation vector. The second correlation relationship is characterized as follows: the second observation vector is approximately equal to the dot product of the second design matrix and the second feature coefficient. The least squares method is used to fit the solution, and the theoretical fitting value obtained by matrix dot product operation with the second feature coefficient as the mapping weight minimizes the overall deviation from the second observation vector. If the mapping weight is larger, it indicates that even a small distortion of the energy of this harmonic can induce a significant vibration response in the principal direction of stiffness attenuation, that is, the higher the excitation sensitivity of this harmonic to stiffness attenuation and the stronger its contribution. If the mapping weight is closer to 0, it indicates that the energy distortion of this harmonic will hardly cause any change in the vibration in the stiffness direction and has no effective contribution to the principal direction of stiffness attenuation. The feature coefficients also include a third feature coefficient, and the following steps are performed: all response sequences of the single-frame dynamic displacement signal, driving current signal, and acceleration signal synchronously acquired through edge nodes are input into a preset neural network inverse model; wherein, the preset neural network inverse model includes an input layer, a first hidden layer, a second hidden layer, and an output layer: the input layer is set to have 3×N neurons, the three response sequences are concatenated into a one-dimensional vector, and after one dimensionality reduction processing by the second hidden layer, the number of neurons in the second hidden layer is set to 256, and the ReLU activation function is used; after a second dimensionality reduction processing by the second hidden layer, the number of neurons in the second hidden layer is set to 128, and the ReLU activation function is used. The activation function outputs a predicted excitation waveform through the output layer. A differential calculation is performed between the predicted excitation waveform and the actual input drive control waveform. The residual value is obtained by calculating the point-by-point deviation between the predicted excitation waveform and the actual drive control waveform. It should be noted that if the residual value is greater than 0: the amplitude of the actual drive control waveform is lower than the amplitude of the ideal predicted excitation waveform, corresponding to attenuation of the drive control signal and insufficient response of the actuator; if the residual value is less than 0: the amplitude of the actual drive control waveform is higher than the amplitude of the ideal predicted excitation waveform, corresponding to overdrive of the control drive signal and command distortion; if the residual value is approximately equal to 0: there is no significant deviation in the drive control link, and the execution accuracy is high. All residual values ​​are arranged in sampling order to form an N×1 dimensional residual sequence. The residual mean, residual standard deviation, residual peak value, residual root mean square value, residual fundamental frequency amplitude, and residual total harmonic distortion rate are extracted using feature extraction. A 6×1 dimensional third observation vector is then formed in the order of residual mean, residual standard deviation, residual peak value, residual root mean square value, residual fundamental frequency amplitude, and residual total harmonic distortion rate. The frequency distribution of the drive control waveform is extracted. Fourier transform is used to extract and calculate the amplitude proportions of the fundamental frequency component and each harmonic component, including the fundamental frequency amplitude proportion, the 2nd harmonic amplitude proportion, the 3rd harmonic amplitude proportion, the 4th harmonic amplitude proportion, the 5th harmonic amplitude proportion, and the frequency interval coefficient between the fundamental frequency and the 2nd harmonic. The frequency interval coefficient is calculated by dividing the fundamental frequency by (fundamental frequency minus the 2nd harmonic). The absolute value of the harmonic frequency is used to represent the frequency distribution. The eigenvectors of the 6-dimensional frequency distribution are used as row vectors and arranged vertically in the order of fundamental frequency amplitude proportion, second harmonic amplitude proportion, third harmonic amplitude proportion, fourth harmonic amplitude proportion, fifth harmonic amplitude proportion, and frequency interval coefficient between the fundamental frequency and the second harmonic, forming an N×6-dimensional third design matrix. A third correlation relationship is established between the third design matrix and the third observation vector, which is represented as follows: the third observation vector is approximately equal to the dot product of the third design matrix and the third eigenvalue coefficient. The least squares method is used to determine the weighted sum of the optimal weight vector as the third eigenvalue coefficient with the goal of minimizing the sum of squared residuals. In the optimal weight vector, each weight component corresponds to the contribution weight of the eigenvector of a frequency distribution. The health index of the electromagnetic vibration table is determined by: using the MapReduce framework to perform distributed statistics on the coupling deviations of the first, second, and third characteristic coefficients, and the third characteristic coefficient; calculating the trigger frequency of each characteristic coupling deviation within a preset time window; and dynamically updating the evaluation weights of each characteristic coefficient based on the trigger frequency. The coupling deviation is calculated by subtracting the health template benchmark value from the characteristic coefficient, and then standardizing the coupling deviation to eliminate the influence of differences in the order of magnitude between different characteristic coefficients. For scalar characteristic coefficients, the coupling deviation is the difference between the actual characteristic coefficient and the scalar benchmark value in the health template. For vector characteristic coefficients, the health template pre-stores a benchmark vector of the same dimension, and the coupling deviation is the same-dimensional deviation vector obtained by subtracting the two vectors point-by-point according to their component order. For matrix characteristic coefficients, the health template pre-stores a benchmark matrix with the same number of rows, columns, element physical meaning, and arrangement order as the actual characteristic coefficient matrix. All rows and columns of the matrix are traversed, and the operation of subtracting the health template benchmark element from the actual characteristic coefficient element is performed point-by-point for the elements corresponding to the same row and column number to obtain a coupling deviation matrix of the same dimension as the original characteristic coefficient matrix. In addition, for the matrix-form coupling deviation, in order to adapt to the subsequent MapReduce distributed statistics, dynamic weight update and health index calculation requirements, post-processing is performed on the matrix-form coupling deviation for subsequent fault analysis and classification: extract the global maximum value, minimum value, element mean and Frobenius norm of the coupling deviation matrix, select the norm as the scalar coupling deviation to characterize the overall deviation of the matrix, or expand the matrix into a one-dimensional vector by rows. The MapReduce framework is divided into Map and Reduce phases. The evaluation weights of each feature coefficient are dynamically updated based on the trigger frequency. This includes: Map phase: The coupling deviation processed by edge nodes is split into multiple independent data shards according to different feature coefficient types, such as the first to third shards. Each shard is assigned to a Map node, and trigger judgment is performed: cases where the coupling deviation exceeds the standard warning threshold are filtered out, and a deviation event is marked. Intermediate key-value pairs are generated by statistically analyzing the correlation between the frequency of deviation events and the timestamps within the time window for each type of feature coefficient. Standard warning threshold: Scalar coupling deviation data from past operation of the electromagnetic vibration table is analyzed. By collecting and statistically analyzing scalar coupling deviations under different test task types, the average and standard deviation of the scalar coupling deviations under different test task types are determined. A standard threshold is set based on the statistical results, which is the average scalar coupling deviation plus twice the standard deviation. It should be noted that the multiplier setting is just an example; the specific setting should be based on the actual situation. Reduce phase: Aggregate key-value pairs corresponding to the same feature coefficient, and aggregate local trigger frequencies according to different feature coefficient types. The local trigger frequency is the ratio of the occurrence frequency to the total number of frames in the time window to obtain the global trigger frequency. This is obtained by summing the ratios of multiple local trigger frequencies to the number of Map nodes and then normalizing the results. The evaluation weight is positively correlated with the trigger frequency of the feature coefficient, that is, the higher the trigger frequency, the greater the contribution of the feature to the health assessment. The evaluation weight of each feature coefficient is dynamically updated in this way, and the sum of the evaluation weights of different feature coefficients is guaranteed to be 1. An ART neural network was used to cluster all coupling deviations. A training layer was introduced into the ART neural network, combining the coupling deviations with a pre-stored health template. The health template was a baseline dataset of the health status of an offline calibrated electromagnetic vibration table, containing the baseline values, deviation thresholds, and distribution characteristics of the first to third feature coefficients under typical test tasks such as sinusoidal, random, and shock tests. A competitive learning dataset was formed by identifying the similarity between the coupling deviations and the health template under different typical test tasks, used for the activation and weight update of neurons in the competitive layer of the ART neural network. Randomly selected sample data from the response sequences collected in real time at edge nodes were provided to perform incremental learning on the training layer. Fault indicators were introduced to evaluate the similarity significance between the coupling deviations and the target winning neurons until training converged, obtaining a health identification model. The fault indicators included standard warning thresholds and neuron weights. The obtained coupling deviations were stratified according to the test tasks based on time series, and any time series segment was marked as a monitoring step. The real-time coupling deviations under each monitoring step were substituted into the health identification model to determine whether each monitoring step belonged to a steady-state deviation or an abnormal fault. If the health type is steady-state deviation, choose weighted calculation: retrieve the current evaluation weights, and sum the first, second, and third characteristic coefficients using weighted methods. Specifically, calculate the product of the evaluation weight of the first characteristic coefficient and the first characteristic coefficient, the product of the evaluation weight of the second characteristic coefficient and the second characteristic coefficient, the product of the evaluation weight of the third characteristic coefficient and the first characteristic coefficient, and sum all the products to obtain the health index. If the health type is abnormal fault, choose multiplicative calculation: retrieve the current evaluation weights, and sum the first, second, and third characteristic coefficients using weighted methods. The weighted summation of the first and third characteristic coefficients is achieved by multiplying the first characteristic coefficient as the base and using the evaluation weight corresponding to the first characteristic coefficient as the exponent. This yields the weighted summation of the first characteristic coefficient's components. The same process is repeated to obtain the weighted summation of the second and third characteristic coefficients' components. Finally, these weighted summations are summed to obtain the health index.

[0022] Feature extraction is performed on multiple response sequences through edge nodes to identify the coupling deviation of the first to third feature coefficients. During this process, the principal direction of stiffness attenuation is determined by extracting the odd-numbered coefficients of the first feature coefficient, and precise identification is achieved. The harmonic order that dominates stiffness attenuation is not determined by the harmonic order. The contribution of each harmonic order is quantified to obtain the second characteristic coefficient. The mathematical modeling method greatly improves the extraction accuracy and physical interpretability of fault features in complex nonlinear environments. The MapReduce framework is introduced to dynamically update the evaluation weights in parallel, and the distributed parallel processing capability is used to improve the efficiency of the weight iteration of feature coefficients. This solves the problems of long processing time and poor real-time performance of traditional centralized processing, ensuring that the health index can be updated in real time and adapting to the dynamic monitoring needs of the electromagnetic vibration table test process. At the same time, only the weights of the ART neural network training layer are updated, without the need for full retraining, and the data parallel statistics of the coupling deviation of multiple feature coefficients are realized, which improves the robustness of the system throughout its entire life cycle.

[0023] Management Decision Module: When acquiring test tasks to be executed, a rule engine is introduced, which is combined with a health index to trigger a graded fault management response; A rule engine is introduced, and a health index is used to trigger a graded fault management response, including: dividing the time series into equal time steps, marking each response sequence corresponding to each time step with a corresponding health type to form a feature source; collecting and aggregating all feature sources in a preset monitoring period, and extracting the proportion of feature sources whose health type is abnormal fault from the feature source set. The maximum and minimum values ​​of the coupling bias corresponding to the response sequence within each time step are extracted. Several task vectors are constructed by combining the feature source proportion and the test task type, including a first task vector and a second task vector. The first task vector is: maximum value plus minimum value, feature source proportion, and test task type. The second task vector is: maximum value minus minimum value, feature source proportion, and test task type. A correlation between multiple task vectors and the health index is established: the health index is retrieved, and based on the first task vector, a first time label is created for the health index according to the time series, forming a correlation model between the first task vector and the health index; the health index is retrieved, and based on the second task vector, a first time label is created for the health index according to the time series, forming a correlation model between the second task vector and the health index. Based on the correlation model, the correlation between the first task vector and the health index, and the correlation between the second task vector and the health index are calculated using the Pearson correlation coefficient. Response sequences with correlation values ​​greater than a corresponding threshold are selected and used as the dataset for analysis. K-means clustering is employed to perform clustering analysis on the feature matrix. Three cluster centers are initialized, and Euclidean distance is used as the sample similarity metric. The cluster centers are iteratively updated until the maximum convergence is achieved, thus identifying three classification ranges: low, medium, and high. The low classification range represents the lowest fault characteristic intensity, corresponding to a minor fault risk, triggering a Level 1 fault response: the test task is not interrupted. The medium classification range represents a mild to moderate fault characteristic, corresponding to a moderate fault risk, triggering a Level 2 fault response: graded control is implemented according to the test task priority, with priority increasing in the order of sinusoidal vibration, random vibration, and impact tests. The high classification range represents the highest fault characteristic intensity, corresponding to a severe fault risk, triggering a graded fault response: a red warning and continuous beeping are activated, and equipment drive control is immediately cut off. The system controls the output, performs safety braking and mechanical interlocking, terminates the current test task, and prevents the malfunction from escalating and causing equipment damage.

[0024] By introducing a rule engine, a correlation model is established by constructing a first task vector and a second task vector, combined with a health index, to trigger hierarchical fault response management and improve the timeliness of fault identification and the efficiency of fault detection for different types of test tasks.

[0025] Example 2: This invention provides a method for fault detection and management of electromagnetic vibration tables based on edge computing; the method includes the following steps: The target excitation sequence is sent to the electromagnetic vibration table, and multiple response sequences output by the electromagnetic vibration table are collected; the electromagnetic vibration table is associated with multiple test task types. By extracting features from multiple response sequences through edge nodes, the coupling deviation of each feature coefficient is identified. The evaluation weights are updated using the MapReduce framework, and the ART neural network is used to perform pattern clustering to determine the health index of the electromagnetic vibration table. When acquiring test tasks to be executed, a rules engine is introduced, which, combined with health indices, triggers a graded fault management response.

[0026] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from the most recent real-world situation by collecting a large amount of data and conducting software simulations. The formulas are set by those skilled in the art according to the actual situation.

[0027] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0028] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0029] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A fault detection and management system for an electromagnetic vibration table based on edge computing, characterized in that, The system includes: The acquisition module sends the target excitation sequence to the electromagnetic vibration table and acquires multiple response sequences output by the electromagnetic vibration table; the electromagnetic vibration table is associated with multiple test task types. The fault analysis module extracts features from multiple response sequences through edge nodes, identifies the coupling deviation of each feature coefficient, updates the evaluation weights using the MapReduce framework, and uses ART neural network to perform clustering to determine the health index of the electromagnetic vibration table. The management decision-making module introduces a rule engine when acquiring test tasks to be executed, and triggers a graded fault management response in combination with health index.

2. The electromagnetic vibration table fault detection and management system based on edge computing according to claim 1, characterized in that, The target excitation sequence is a set of drive control waveforms containing a preset frequency band distribution. The response sequence includes at least acceleration signals, drive current signals, and moving coil displacement signals. The test task types include at least sinusoidal vibration, random vibration, and shock tests.

3. The electromagnetic vibration table fault detection and management system based on edge computing according to claim 1, characterized in that, Identify the characteristic coefficients, including: Based on any response sequence, extract and calculate the mean coordinates and mean features of the sampling points in the response sequence, and neutralize them to obtain centered coordinates and centered features; construct a design matrix with the centered coordinates as row vectors, the total number of sampling points as the number of rows, and the corresponding feature space dimension as the number of columns; construct observation vectors with the centered features according to the sampling order; establish the correlation between the design matrix and the observation vectors, and determine the corresponding feature coefficients.

4. The electromagnetic vibration table fault detection and management system based on edge computing according to claim 3, characterized in that, The characteristic coefficients also include the first characteristic coefficient, and the following is performed: The displacement signal of the moving coil is retrieved to determine the first design matrix; the driving current signal is retrieved to determine the first observation vector; the first correlation relationship between the first design matrix and the first observation vector is established, and the odd-numbered coefficients are solved by least squares fitting as the first characteristic coefficients; the nonlinear components in the first characteristic coefficients are extracted and inverted, and normalized by the L2 norm to obtain the main direction of stiffness attenuation.

5. The electromagnetic vibration table fault detection and management system based on edge computing according to claim 4, characterized in that, The eigenvalues ​​also include a second eigenvalue, and the following is performed: The acceleration signal is retrieved, and the fundamental frequency component and harmonic components are extracted through Fourier transform. The energy distortion ratio is calculated to determine the second design matrix. The acceleration signal is then subjected to scalar projection along the main stiffness attenuation direction to obtain the projected acceleration signal and determine the second observation vector. By establishing a second correlation between the second design matrix and the second observation vector, the mapping weights corresponding to each harmonic are determined as the second characteristic coefficients.

6. The electromagnetic vibration table fault detection and management system based on edge computing according to claim 3, characterized in that, The eigenvalues ​​also include a third eigenvalue, and the following is performed: All response sequences are input into a preset neural network inverse model. The third observation vector is determined by the residual sequence of the output predicted excitation waveform and the actual input driving control waveform. The frequency distribution of the driving control waveform is extracted to determine the third design matrix. A third correlation relationship is established between the third design matrix and the third observation vector. With the goal of minimizing the sum of squared residuals, the weighted sum of the optimal weight vectors is determined as the third characteristic coefficient.

7. The electromagnetic vibration table fault detection and management system based on edge computing according to claim 1, characterized in that, Determine the health index of the electromagnetic vibration table, including: The MapReduce framework is used to perform distributed statistics on the coupling deviation of different feature coefficients, and the triggering frequency of each feature coupling deviation within a preset time window is calculated. The evaluation weight of each feature coefficient is dynamically updated based on the triggering frequency. In the ART neural network, a training layer is introduced, and a competitive learning dataset is formed by combining coupling bias with pre-stored health templates. Randomly selected sample data from response sequences collected in real time at edge nodes are provided to perform incremental learning on the training layer. Fault indicators are introduced to perform similarity significance evaluation until training converges and a health recognition model is obtained. Among them, the fault indicators include standard warning thresholds and initial weights of neurons. The obtained coupling deviations are stratified into test tasks according to time series, and any time series segment is marked as a monitoring step. The real-time coupling deviations under each monitoring step are substituted into the health identification model to determine whether each monitoring step belongs to steady-state deviation or abnormal fault, and the corresponding calculation method is selected. The health index is obtained by combining different feature coefficients.

8. The electromagnetic vibration table fault detection and management system based on edge computing according to claim 7, characterized in that, If the health type is a steady-state deviation, choose weighted calculation; if the health type is an abnormal fault, choose exponential weighted calculation.

9. The electromagnetic vibration table fault detection and management system based on edge computing according to claim 8, characterized in that, Introducing a rules engine, combined with health indices, to trigger tiered fault management responses, including: Divide the time series into equal time steps, label each response sequence corresponding to each time step with the corresponding health type, and form a feature source; collect all feature sources in the preset monitoring period and aggregate them, and extract the proportion of feature sources whose health type is abnormal fault from the feature source set. Extract the maximum and minimum values ​​of the coupling deviation corresponding to each response sequence, and construct multiple task vectors by combining the proportion of feature sources and the current test task to be executed, including the first task vector and the second task vector; establish the correlation between multiple task vectors and health index and extract the corresponding correlation degree; screen any response sequence with a correlation degree greater than the corresponding correlation degree threshold, and identify it into three classification ranges of low, medium and high through cluster analysis, and trigger the corresponding graded fault management response.

10. A method for fault detection and management of an electromagnetic vibration table based on edge computing, characterized in that, The method includes: The target excitation sequence is sent to the electromagnetic vibration table, and multiple response sequences output by the electromagnetic vibration table are collected; the electromagnetic vibration table is associated with multiple test task types. By extracting features from multiple response sequences through edge nodes, the coupling deviation of each feature coefficient is identified. The evaluation weights are updated using the MapReduce framework, and the ART neural network is used to perform pattern clustering to determine the health index of the electromagnetic vibration table. When acquiring test tasks to be executed, a rules engine is introduced, which, combined with health indices, triggers a graded fault management response.

Images