State monitoring system for compression-shear testing machine

Abstract

The invention belongs to the technical field of structural component monitoring, and discloses a pressure-shear testing machine state monitoring system, which comprises an acquisition module for performing signal preprocessing on an acquired signal sequence to obtain signal data; the feature extraction module is used for acquiring a feature vector of each piece of signal data based on the signal data; marking the feature vector obtained by the testing machine in the healthy state as a healthy feature vector; the training module is used for training an isolation forest model and a mechanical coding model based on the health feature vectors; the detection module is used for calculating an abnormal score by an isolation forest model based on a feature vector obtained in real time for each dynamic loading event; marking the dynamic loading events of which the abnormal scores are greater than or equal to a preset suspicious threshold value as suspicious events; the abnormity confirmation module is used for calculating a reconstruction error by a mechanical coding model for a suspicious event; and marking the suspicious event of which the reconstruction error is greater than or equal to a preset abnormal threshold value as an abnormal event. The abnormal detection rate is effectively improved, and the false alarm rate is remarkably reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of structural component monitoring, and more particularly to a state monitoring system for a compression shear testing machine. Background Art

[0002] In the fields of structural strength assessment and material performance testing, large-scale compression-shear testing machines are widely used to simulate the behavior of components under complex load conditions. They play an irreplaceable role in extreme dynamic testing, especially in impact loading or high-rate cyclic loading. However, extreme loading conditions also place higher demands on the structure of the testing machine and the health of key structural components. In particular, initial signs of failure, such as loosening of connecting structural components or hydraulic system anomalies, that may occur during testing need to be identified promptly and reliably without disrupting the test process.

[0003] Traditional anomaly detection methods often rely on fixed thresholds, manual judgment, or single statistical models. These methods often suffer from insufficient detection sensitivity and high false alarm rates when faced with signal non-stationarity, transient anomaly characteristics, and complex background noise interference encountered during dynamic testing. Furthermore, existing methods generally lack in-depth analysis and diagnostic support for anomaly causes, making it difficult for maintenance personnel to provide effective fault location information.

[0004] Therefore, in order to solve the problem of identifying the initial fault characteristics of the test machine structure under extreme dynamic test conditions, there is an urgent need for an identification method that can extract weak transient abnormal features in a strong interference background and has high detection accuracy and interpretability to assist in predictive maintenance and system-level risk control.

[0005] In view of this, the present invention proposes a compression shear testing machine state monitoring system to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned purpose, the present invention provides the following technical solution: a compression shear testing machine state monitoring system, comprising:

[0007] The acquisition module performs signal preprocessing on the acquired signal sequence to obtain signal data;

[0008] The feature extraction module performs continuous wavelet transform on each signal data to obtain the wavelet coefficient of each signal data; based on the wavelet coefficient, the signal analysis feature of each signal data is obtained; based on the signal analysis feature, the feature vector of each signal data is constructed; the feature vector obtained when the test machine is in a healthy state is marked as a healthy feature vector;

[0009] The training module trains the isolation forest model and the mechanical encoding model based on the health feature vector;

[0010] The detection module calculates an anomaly score for each dynamic loading event based on the feature vector obtained in real time using the isolation forest model; dynamic loading events with an anomaly score greater than or equal to a preset suspicious threshold are marked as suspicious events;

[0011] The abnormality confirmation module calculates the reconstruction error of suspicious events based on the feature vectors obtained in real time using the mechanical coding model; suspicious events with a reconstruction error greater than or equal to a preset abnormality threshold are marked as abnormal events;

[0012] The alarm module calculates the comprehensive damage score of each abnormal event based on the anomaly score and reconstruction error; obtains the damage level based on the comprehensive damage score; and generates an alarm message when the comprehensive damage score is higher than the preset alarm level.

[0013] Furthermore, the method for performing continuous wavelet transform includes: performing continuous wavelet transform on each signal data based on a preset mother wavelet function, a preset scale parameter and a time parameter to obtain wavelet coefficients of a combination of each signal data, the preset scale parameter and the time parameter;

[0014] The method for selecting the preset scale parameter includes: calculating the corresponding minimum scale and maximum scale according to the upper and lower frequency limits of the frequency range of the initial fault characteristic component in the signal data and the center frequency of the preset mother wavelet function to obtain the scale range; the minimum scale is smaller than the ratio of the center frequency to the upper frequency limit; the maximum scale is larger than the ratio of the center frequency to the lower frequency limit; within the determined scale range, extracting an arithmetic sequence with the minimum scale as the first item, the maximum scale as the last item and the total number of preset scales, and using the scales in the arithmetic sequence as the preset scale parameter.

[0015] Furthermore, the signal analysis features include global time-frequency entropy, regional concentration, and local time-frequency entropy.

[0016] Furthermore, the method for obtaining the global time-frequency entropy includes: calculating a wavelet spectrum based on wavelet coefficients; calculating a global probability distribution of each signal data based on the wavelet spectrum; and calculating a global time-frequency entropy of each signal data based on the global probability distribution.

[0017] The method for obtaining regional concentration includes: calculating the regional energy of each signal data in each preset region of interest based on the wavelet spectrum; calculating the regional concentration of each signal data in each preset region of interest based on the regional energy;

[0018] The method for obtaining local time-frequency entropy includes: calculating the local probability distribution of each signal data in each preset region of interest based on the wavelet spectrum; and calculating the local time-frequency entropy of each signal data in each preset region of interest based on the local probability distribution and applying the definition of Shannon entropy.

[0019] Furthermore, the wavelet spectrum calculation method includes: calculating the square modulus of the wavelet coefficient to obtain the wavelet energy; for each two-dimensional wavelet coefficient matrix corresponding to each signal data, calculating the wavelet energy of each wavelet coefficient in the two-dimensional wavelet coefficient matrix to obtain the wavelet spectrum of each signal data;

[0020] The calculation method of the local probability distribution includes: for each signal data and its corresponding wavelet spectrum, and for each preset region of interest, calculating the regional energy of the signal data in the preset region of interest; for each scale-time unit in the preset region of interest, dividing the wavelet energy by the regional energy to obtain the normalized energy of the wavelet energy of each scale-time unit relative to the regional energy, that is, the local probability of the scale-time unit in the preset region of interest; all local probabilities of a signal data in a preset region of interest constitute the local probability distribution of the signal data in the preset region of interest.

[0021] Furthermore, the method for constructing the feature vector includes: for each dynamic loading event, orderly arranging all signal analysis features extracted from each corresponding signal data to form a feature vector for each signal data;

[0022] The healthy state is defined as the state in which the test machine operates under clear conditions without known structural damage or functional failure.

[0023] Furthermore, the calculation method of the anomaly score includes:

[0024] The feature vector obtained in real time is input into the trained isolation forest model; each isolation tree in the isolation forest model is traversed. In each isolation tree, starting from the root node, the feature vector is directed to the corresponding child node according to the current feature dimension and split point of each non-leaf node in the isolation tree until it reaches a leaf node. The number of edges that the feature vector passes through from the root node to the leaf node where the feature vector is finally located is recorded to obtain the path length of the feature vector in each isolation tree; the expected value of the path length of the feature vector on all isolation trees in the isolation forest model is calculated to obtain the average path length; the reciprocal of the ratio of the average path length to the normalization factor is calculated to obtain the normalized length; and an exponential function with the normalized length as the exponent and base 2 as the base is calculated to obtain the anomaly score.

[0025] Furthermore, the method for calculating the reconstruction error includes: marking the feature vector of the suspicious event as a suspicious feature vector; inputting the suspicious feature vector into a trained mechanical coding model; the mechanical coding model reconstructs the original suspicious feature vector to obtain a reconstructed feature vector; calculating the reconstruction loss between the suspicious feature vector and the corresponding reconstructed feature vector to obtain the reconstruction error.

[0026] Furthermore, the calculation method of the comprehensive damage score includes: normalizing the reconstruction error to obtain a normalized error; and performing a weighted linear combination of the abnormality score and the normalized error according to a preset weight to obtain a comprehensive damage score.

[0027] Furthermore, the method for calculating the normalized error includes: determining a preset upper limit of the reconstruction error under a healthy state to obtain the upper limit of the reconstruction error; determining a preset baseline value of the reconstruction error under a healthy state to obtain the lower limit of the reconstruction error; calculating the difference between the input reconstruction error and the lower limit of the reconstruction error to obtain the relative error; selecting the larger of the relative error and 0 to obtain the relative reconstruction error; calculating the difference between the upper limit of the reconstruction error and the lower limit of the reconstruction error to obtain the error interval; calculating the ratio of the relative reconstruction error to the error interval to obtain the error ratio; selecting the smaller of the error ratio and 1 to obtain the normalized error.

[0028] Furthermore, the alarm information includes: a unique identifier of the abnormal event, the date and time of the abnormal event, the damage level of the abnormal event, key abnormal characteristics, a comprehensive damage score, an abnormality score, and a reconstruction error;

[0029] The identification method of key abnormal features includes: extracting the feature vector corresponding to the abnormal event; in the healthy feature vector set, respectively counting the mean and standard deviation of each feature dimension to obtain the healthy mean and healthy standard deviation; for each feature dimension, first calculating the difference between the corresponding component in the feature vector of the abnormal event and the healthy mean, calculating the absolute value of the ratio of the difference to the healthy standard deviation, and obtaining the feature deviation; sorting the feature deviations corresponding to each feature dimension in descending order, and selecting the top feature dimension. The signal analysis features corresponding to the characteristic deviations are used as the key abnormal features of abnormal events.

[0030] Beneficial effects of the compression shear testing machine state monitoring system of the present invention:

[0031] By extracting multi-dimensional time-frequency features through refined preprocessing and continuous wavelet transform, the system's ability to capture and characterize weak, transient initial fault characteristics is significantly improved, and the signal-to-noise ratio in complex backgrounds is improved.

[0032] By using the isolation forest model for rapid initial screening and deep confirmation through the mechanical coding model, a two-stage joint discrimination is achieved, which effectively improves the detection rate of real anomalies and significantly reduces the false alarm rate.

[0033] Construct a comprehensive damage score that integrates the anomaly score and normalized reconstruction error to support multi-level classification of abnormal events and achieve more detailed health status assessment.

[0034] Automatically identify key abnormal characteristics and their signal sources, assist in locating possible fault sites, provide clear decision-making references for operation and maintenance personnel, and support rapid response and targeted maintenance.

[0035] The overall system is suitable for online monitoring of compression and shear testing machines under extreme loading conditions. It has good real-time performance and adaptability, which helps to achieve predictive maintenance and ensure the safe operation of testing equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a system module diagram of the compression shear testing machine state monitoring system of the present invention;

[0037] Figure 2 This is a schematic diagram of a scenario of a compression shear testing machine state monitoring system of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] See Figure 1 The state monitoring system of the compression shear testing machine described in this embodiment includes: an acquisition module, a feature extraction module, a training module, a detection module, an abnormality confirmation module and an alarm module.

[0040] This embodiment provides a system for identifying the initial fault characteristics of a compression shear testing machine under extreme dynamic testing. Initial fault characteristics refer to the initial signs that the compression shear testing machine has not completely failed in function, but has already shown performance degradation or indicates that a future failure may occur; these signs are usually manifested as weak, transient physical signals generated by the equipment structure or system. In the present invention, initial fault characteristics specifically include but are not limited to the following situations: early fatigue damage or loosening of the compression shear testing machine structure, early loosening of actuator connection parts, and initial abnormalities of the hydraulic system, such as internal leakage and unstable pressure control. The compression shear testing machine is prone to initial fault characteristics under extreme dynamic testing conditions including impact or high-rate cyclic loading. The abnormal signals generated by the initial fault characteristics often appear in a specific frequency range. For the frequency range of the abnormal signal corresponding to the target initial fault characteristic, this system is uniformly referred to as the abnormal frequency range. For the convenience of description, the compression shear testing machine will be referred to as the testing machine; dynamic loading events include impact testing or high-rate cyclic loading testing.

[0041] The acquisition module performs signal preprocessing on the acquired signal sequence to obtain signal data.

[0042] Signal sequences include vibration sequences and pressure sequences. Signal data includes vibration data and pressure data. Vibration sequences are obtained by signal preprocessing, while pressure sequences are obtained by signal preprocessing, respectively.

[0043] The vibration sequence acquisition method includes: determining the vibration monitoring position and arranging an acceleration sensor at each vibration monitoring position; synchronously acquiring the vibration voltage signal of each acceleration sensor at a preset sampling frequency during the actual process of the dynamic loading event of the testing machine; converting the vibration voltage signal into digital acceleration time series data through signal conditioning (such as amplification and filtering) and analog-to-digital conversion (ADC, for example, using a 24-bit ADC to ensure sufficient dynamic range and accuracy), thereby obtaining the vibration sequence of each vibration monitoring position.

[0044] In this embodiment, a piezoelectric accelerometer, such as one that complies with the IEPE standard, is used. Due to its high sensitivity, wide frequency response range, and excellent signal-to-noise ratio, this type of sensor can effectively capture weak, high-frequency vibration signals generated by early-stage structural damage. The specific sensor model and sensitivity range should be selected based on the specific structure of the testing machine, the amplitude of the damage signal to be monitored, and the input range of the data acquisition system.

[0045] The method for determining the vibration monitoring locations includes: performing finite element analysis (FEA) on the design drawings of the testing machine to identify key monitoring areas where stress concentration, large strain, or fatigue damage under extreme dynamic loads are common; combining the experience and knowledge of field experts with the historical failure data of the testing machine to further optimize and confirm these key monitoring areas; accelerometers are placed within these identified key monitoring areas, or at locations near them that can effectively reflect the status of the area, such as key connection nodes of the testing machine frame, and the surfaces of the connecting parts between the actuator and the frame. The reason for selecting these locations is that they can effectively capture the characteristic signals of initial failures.

[0046] The vibration monitoring location is crucial to effectively capture early damage signals. A rigid connection between the accelerometer and the mounting surface, such as bolt fastening or using a specialized adhesive, should be used to ensure effective transmission of high-frequency signals.

[0047] The principles for determining the preset sampling frequency include: based on the Nyquist-Shannon sampling theorem, the preset sampling frequency is set to at least 2 times the expected highest frequency component in the signal of interest (vibration signal or pressure signal). Usually, in order to better reconstruct the waveform and provide margin for subsequent signal processing (such as filtering), 2.5 times to 10 times or even higher will be selected. When selecting the specific value, it is first necessary to analyze the characteristic frequency range of the initial fault characteristic signal that may be generated by the testing machine under extreme dynamic testing. For example, early fatigue damage or loosening of the structure may produce high-frequency vibration or acoustic emission signals, the frequency of which can reach tens of kilohertz; pressure pulsations caused by initial abnormalities of the hydraulic system (such as internal leakage, unstable pressure control) may also contain high-frequency components. Therefore, it is necessary to conduct preliminary dynamic characteristic tests on the testing machine or refer to research data of similar equipment to estimate the upper limit of the maximum frequency of these abnormal signals, denoted as Secondly, considering that the data acquisition system (DAQ) is usually equipped with an anti-aliasing filter, the cutoff frequency (roll-off point) of the filter should be reasonably set according to the preset sampling frequency, for example, set to 0.4 to 0.8 times the preset sampling frequency, so as to effectively filter out useless high-frequency noise higher than the Nyquist frequency (i.e. 0.5 times the preset sampling frequency) and prevent signal aliasing. Therefore, the preset sampling frequency The choice should satisfy and ensure that the full All useful signal components within . In this embodiment, considering the need to capture high-frequency acoustic emission signals or high-frequency vibrations generated by the expansion of small cracks in the structure, as well as transient pressure changes in the hydraulic system, the preset sampling frequency can be set to a specific value in the range of 50 kHz to 200 kHz, such as 100 kHz. The setting of this value is intended to ensure that signal details of sufficiently high frequencies can be captured while taking into account data volume and processing efficiency. Selecting a higher preset sampling frequency, such as 100 kHz, can more accurately capture the transient characteristics of the signal, which is crucial for identifying early, weak abnormal signals.

[0048] The pressure sequence acquisition method includes: determining the pressure monitoring position and arranging a pressure sensor at each pressure monitoring position; synchronously acquiring the pressure voltage signal of each pressure sensor at a preset sampling frequency during the actual process of a dynamic loading event on the testing machine; converting the pressure voltage signal into digital pressure time series data through signal conditioning and analog-to-digital conversion to obtain the pressure sequence at each pressure monitoring position.

[0049] In this embodiment, a high-frequency dynamic pressure sensor is used. For example, a pressure sensor installed near the inlet and outlet ports of the testing machine's hydraulic system actuator or on key hydraulic lines should be used. This type of sensor should have a fast response time and a sufficiently high upper operating frequency limit to accurately capture transient pressure changes during loading. Its measurement range should cover the maximum dynamic pressure that the testing machine's actuator can generate.

[0050] Pressure monitoring locations include: key pressure measurement points in the hydraulic system, such as the master cylinder, accumulator, or the oil inlet and return ports of the actuator cylinder. These locations are chosen because they can directly reflect the operating status of the hydraulic system and the dynamic changes in loading force.

[0051] The signal preprocessing method for a signal sequence includes: segmenting the original signal segment corresponding to the dynamic loading event in the signal sequence; removing the DC offset or low-frequency trend drift in the original signal segment to obtain a corrected signal segment; performing noise suppression on the corrected signal segment to obtain a noise-reduced signal segment; and normalizing the noise-reduced signal segment to obtain signal data. Signal preprocessing aims to convert the acquired signal sequence into clean and standardized signal data suitable for subsequent advanced feature extraction.

[0052] The method for acquiring the original signal segment includes: selecting one or more reference signal channels (such as the pressure signal channel of the main loading actuator or the acceleration signal channel of the key position) and setting the starting amplitude threshold and the ending amplitude threshold for them; when the amplitude of the reference signal channel exceeds the preset starting amplitude threshold for the first time in time, it is marked as the start time of the dynamic loading event; when the amplitude of the reference signal falls below the preset ending amplitude threshold and persists for a period of time, or when a loading cycle is judged to have ended according to the loading instruction, it is marked as the end time of the dynamic loading event. Based on these marked start and end times, the signal sequence of the corresponding time range is cut out from the signal sequence of all acquisition channels to form an independent original signal segment. This step is used to separate the original signal segment corresponding to each independent dynamic loading event executed by the testing machine from the continuous signal sequence.

[0053] The method for obtaining a corrected signal segment includes: selecting an appropriate detrending method for each raw signal segment captured from a dynamic loading event (whether a raw signal segment of a vibration sequence or a raw signal segment of a pressure sequence); applying the selected detrending method to each raw signal segment to obtain a corrected signal segment. For example, if high-pass filtering is used, the raw signal segment is passed through a designed high-pass filter; if polynomial detrending is used, a polynomial fit is performed on the raw signal segment, and the fitted trend is subtracted.

[0054] Detrending methods include high-pass filtering, polynomial detrending, and baseline correction. The choice depends on the characteristics of the original signal segment (e.g., whether the drift resembles a DC offset or a slowly varying trend) and the sensitivity of subsequent analysis to the signal's low-frequency components. For example, if the signal drift manifests as slow fluctuations in the baseline, methods such as segmented averaging or moving average can be used to estimate the baseline, which can then be subtracted from the original signal segment.

[0055] The necessity of removing DC offset or low-frequency trend drift is that these components are not directly generated by the dynamic response of the structure or hydraulic system of the testing machine under dynamic loading events, but may come from the zero drift of the sensor itself, the impact of ambient temperature changes on the sensor, slow quasi-static load changes, or slow pressure baseline changes in the hydraulic system that are unrelated to dynamic loading events, such as slow fluctuations in system pressure caused by oil temperature changes. If these components are not removed, they may interfere with subsequent feature extraction, such as continuous wavelet transform, mask or distort the true high-frequency or transient features related to initial fault characteristics, or affect the accuracy of amplitude-based statistical features. Therefore, removing DC offset or low-frequency trend drift aims to separate the dynamic components of the signal and improve the signal-to-noise ratio and accuracy of subsequent analysis.

[0056] The method for obtaining the noise reduction signal segment includes: for each corrected signal segment, analyzing the noise type and noise characteristics that may exist in the corrected signal segment and selecting an appropriate noise reduction method; carefully determining the key parameters of the noise reduction method based on the selected noise reduction method; and applying the configured noise reduction method to each corrected signal segment to obtain the noise reduction signal segment.

[0057] Noise reduction techniques include: bandpass filtering when the useful signal and the main noise are known to be in different frequency bands, low-pass filtering when high-frequency noise is more significant, notch filtering when there is narrow-band interference of a specific frequency, wavelet denoising is particularly suitable for processing noise in non-stationary signals and transient impact signals, and adaptive filtering is used when the noise characteristics are unknown or change over time.

[0058] The key parameters of the noise reduction method can be determined based on prior knowledge (such as known noise frequency), analysis of the signal spectrum, or experimental optimization to achieve the best signal-to-noise ratio improvement effect.

[0059] Noise suppression during signal preprocessing is crucial because the raw sampled signal is inevitably contaminated by various noises. Left untreated, this noise can be mistakenly identified as signal features during subsequent feature extraction, or it can drown out weak early-stage fault signatures, reducing the sensitivity and reliability of anomaly detection. The goal of noise suppression is to improve the signal-to-noise ratio (SNR), providing a high-quality data foundation for accurate extraction of features sensitive to early-stage faults.

[0060] The signal data acquisition method includes: for each noise reduction signal segment, selecting a suitable normalization method, adjusting the amplitude of the noise reduction signal segments from different sources, different dynamic loading events or different sensor channels to a unified and comparable numerical range, and obtaining the signal data.

[0061] Normalization methods include: min-max normalization, which linearly scales the data to a preset interval, usually [0, 1] or [-1, 1]. In this embodiment, the signal amplitude is scaled to [0, 1]. For a noise reduction signal segment ,in is the total number of data points in the denoised signal segment, The formula for normalizing the Vth data point in the noise reduction signal segment to the interval [0, 1] is:

[0062]

[0063] Where, is a data point in the noise-reduced signal segment, is the normalized data point, and are the minimum and maximum values ​​of the data points in the noise reduction signal segment respectively. (For example, the signal segment is a constant), then all Can be set to 0 or 0.5.

[0064] Normalization is very important in the final stage of signal preprocessing. The signal amplitude ranges of different sensors (even sensors of the same type due to calibration or sensitivity differences) and different physical quantities may vary greatly. Normalization can unify them to a similar numerical scale to avoid certain signals or features with larger amplitudes from disproportionately dominating the analysis results in subsequent feature extraction or machine learning model training. Normalization allows signal data from different dynamic loading events or different working conditions to be compared and analyzed on a relatively unified benchmark. Although the basic normalization methods (such as minimum-maximum normalization or Z-score normalization) are universal for the noise reduction signal segments of vibration sequences and the noise reduction signal segments of pressure sequences, the key lies in the calculation range of the normalization parameters. As mentioned above, this embodiment recommends normalizing each independent noise reduction signal segment. This means that the noise-reduced signal segment for a particular dynamic loading event from one vibration sensor is normalized using its own minimum / maximum values ​​(or mean / standard deviation), while the noise-reduced signal segment for the same dynamic loading event from another pressure sensor is normalized using the pressure signal's own minimum / maximum values ​​(or mean / standard deviation). This independent normalization strategy is appropriate because it preserves the relative dynamic variation within each signal segment while aligning the signal of each sensor within each dynamic loading event to a standard range. This avoids normalization bias caused by some events having significantly greater overall energy than others, ensuring that subsequent feature extraction is sensitive to the signal patterns of each event.

[0065] The feature extraction module performs continuous wavelet transform (CWT) on each signal data to obtain the wavelet coefficient of each signal data; based on the wavelet coefficient, the signal analysis feature of each signal data is obtained; based on the signal analysis feature, the feature vector of each signal data is constructed; and the feature vector obtained when the testing machine is in a healthy state is marked as a healthy feature vector.

[0066] The healthy state is defined as the state in which the test machine operates under clear conditions without known structural damage or functional failure. More specifically, it can be operationally defined based on one or more of the following aspects:

[0067] Initial baseline: The testing machine is considered to be in a healthy state during its initial operation after passing factory acceptance or undergoing comprehensive maintenance and confirmation that there are no known faults or abnormalities.

[0068] Performance indicator compliance: In this state, when the testing machine performs standard test conditions, its key performance indicators, such as loading accuracy, response speed, noise level, vibration amplitude, etc., should be within the normal fluctuation range allowed by the design specifications or acceptance standards.

[0069] Signal pattern stability: In this state, the vibration signals and pressure signals collected and processed by the system described in the present invention should exhibit stable and predictable patterns after feature extraction, and should not contain known abnormal signal characteristics related to specific damage or faults.

[0070] The reason for defining a healthy state is to obtain reliable baseline data for subsequent anomaly detection model training and threshold setting. The feature vectors collected and processed in this healthy state are labeled as healthy feature vectors and used for subsequent model training.

[0071] The continuous wavelet transform method includes: performing continuous wavelet transform on each signal data based on a preset mother wavelet function, preset scale parameter and time parameter to obtain wavelet coefficients of a combination of each signal data, preset scale parameter and time parameter.

[0072] The method for selecting the preset mother wavelet function includes: preliminarily screening several candidate mother wavelet functions, such as Morlet wavelet, Mexican wavelet, etc., according to the characteristics of the signal to be analyzed and the expected abnormal signal characteristics. Hat wavelet, Daubechies series wavelet (such as db4, db6), Symlet series wavelet, etc., these candidate mother wavelet functions have their own characteristics in time-frequency localization ability, symmetry, regularity, etc., and are suitable for analyzing different types of signals, such as impact signals, oscillation signals, and signals with specific discontinuities; collect a part of the baseline signal data processed by the acquisition module when the test machine is running in a clear healthy state; for each candidate mother wavelet function, these baseline signal data are analyzed by continuous wavelet transform, and their performance in representing healthy signals is evaluated. The evaluation indicators may include: the concentration of time-frequency energy (the energy of the expected healthy signal is relatively concentrated in the time-frequency plane and has a low entropy value), the ability to express the dominant response frequency of the known structure, and the sensitivity to the introduced simulated weak abnormal signals (that is, which mother wavelet can more significantly highlight the time-frequency feature changes caused by the abnormality); based on the above evaluation results and combined with the experience and knowledge of experts in the field, one or a group of optimal mother wavelet functions are finally selected as the preset mother wavelet functions, denoted as , which is used for the continuous wavelet transform calculation of all subsequent signal data. For example, in this embodiment, if the initial fault characteristic signals are mostly manifested as transient shocks or high-frequency oscillations, a complex Morlet wavelet or a Mexican Hat wavelet with good time-frequency localization characteristics may be selected. In order to ensure the consistency and comparability of the analysis of vibration data and pressure data, the same preset mother wavelet function is usually selected for both. The selection of the preset mother wavelet function directly affects the effect of the continuous wavelet transform analysis. The appropriate preset mother wavelet function can better match and reveal specific features in the signal, thereby improving the effectiveness of subsequent feature extraction and the accuracy of anomaly detection.

[0073] The method for selecting the preset scale parameter includes: selecting the upper limit frequency of the frequency range of the signal component of the initial fault characteristic in the signal data; and lower frequency , and the center frequency of the preset mother wavelet function , estimate the corresponding minimum scale and maximum scale ; Minimum scale Should be less than the ratio of the center frequency to the upper limit frequency, denoted as (Due to the preset sampling frequency , ); the maximum scale should be greater than the ratio of the center frequency to the lower limit frequency, denoted as ; within a certain scale range Extract the smallest scale as the first item, the largest scale as the last item, and the total number of preset scales. The arithmetic progression, the scale of the arithmetic progression as a preset scale parameter.

[0074] Scale range It should ensure that the target frequency information related to the dynamic response of the test machine structure and the characteristics of potential initial failures can be fully covered. The choice of scales (or the step size between adjacent scales) requires a trade-off between frequency resolution and computational efficiency. More scales (smaller step size) can provide higher frequency resolution, but also means more computation and longer processing time. Usually, the number of scales is The selection should ensure that there is sufficient resolution to distinguish different frequency components and capture energy changes within the "abnormal frequency range" and the subsequently defined "preset region of interest". For example, you can set A value between 64 and 256, depending on signal complexity and analysis needs.

[0075] The method for obtaining the time parameter includes: for each signal data obtained after signal preprocessing, it is composed of a series of discrete sampling data points in time; the time information of these discrete sampling points constitutes the time parameter required for continuous wavelet transform. Specifically, if a signal data contains data points, and the sampling time interval is , which is the reciprocal of the preset sampling frequency, then the time point sequence Each time point in is a time parameter, where Represents the number of times the signal is counted from the start time. When performing continuous wavelet transform, these discrete time parameters are To calculate the wavelet coefficients, we can get the representation of the signal data at different times and different scales (frequencies). Therefore, the time parameters are directly obtained from the temporal structure of the input signal data without the need for additional selection or presetting.

[0076] The calculation method of continuous wavelet transform includes: for each input signal data, it is recorded as And each preset mother wavelet function , each preset scale parameter And each time parameter , apply the following definition of continuous wavelet transform to calculate the corresponding wavelet coefficients :

[0077]

[0078] Where, is the time variable, is the preset mother wavelet function The meaning of this formula is that the original signal With a translation (by ) and scaling (determined by By traversing all the preset scale parameters and all time parameters , we can get a two-dimensional complex matrix, each element in the matrix At a specific scale parameter and specific time parameters Therefore, for each input signal data, a set of wavelet coefficients will be generated, which completely describes the energy distribution and phase information of the signal data in the time-frequency plane.

[0079] In actual calculations, since the original signal data is discrete, the above-mentioned integral operation is usually efficiently calculated in the frequency domain through discrete convolution or the use of fast Fourier transform (FFT). Many scientific computing software libraries provide implementations of continuous wavelet transforms. For example, in the Python language environment, you can use the pywt.cwt function in the PyWavelets library or related functions in the SciPy library; in the MATLAB environment, you can use the cwt function in the Wavelet Toolbox™. These libraries typically allow users to specify the input signal, a preset mother wavelet function, and a preset scale parameter sequence, and then automatically calculate the corresponding wavelet coefficient matrix.

[0080] For each input signal data, a wavelet coefficient matrix in the form of a two-dimensional complex matrix is ​​generated. For example, the rows correspond to the preset scale parameters and the columns correspond to the time parameters. Therefore, each signal data has a corresponding wavelet coefficient under each specific combination of the preset scale parameter and each time parameter. The wavelet coefficient matrix is ​​recorded as .

[0081] Signal analysis features include: global time-frequency entropy, regional concentration, and local time-frequency entropy.

[0082] The method for obtaining the global time-frequency entropy includes: calculating a wavelet spectrum based on wavelet coefficients; calculating a global probability distribution of each signal data based on the wavelet spectrum; and calculating a global time-frequency entropy of each signal data based on the global probability distribution.

[0083] The calculation method of wavelet spectrum includes: calculating the square modulus of wavelet coefficients to obtain wavelet energy, which is recorded as ; For each signal data corresponding to the two-dimensional wavelet coefficient matrix , calculate the wavelet energy of each wavelet coefficient in the two-dimensional wavelet coefficient matrix, and obtain the wavelet spectrum of each signal data (also called energy spectrum density map or scale map), which is recorded as The specific calculation formula of wavelet energy includes:

[0084]

[0085] in, Representative Preset scale parameters, Representative time parameters.

[0086] Wavelet Energy Indicates the signal data in time parameter Nearby, preset scale parameters The wavelet spectrogram visually displays the distribution of signal energy on the time-frequency plane. The wavelet spectrogram is the basis for extracting various time-frequency features. It converts complex wavelet coefficients into real energy representations, making it easier to analyze changes in signal energy over time and frequency.

[0087] The calculation method of global probability distribution includes: for each signal data corresponding to the wavelet spectrum , first of all, the wavelet spectrum is calculated at all preset scale parameters ( ) and all time parameters ( ) and sum up to get the total energy ; Then, each wavelet energy in the wavelet spectrum Divide by the total energy , get the value at a specific scale-time unit The normalized energy on , that is, the probability of the corresponding scale-time unit; all normalized energies It constitutes the global probability distribution of the signal data.

[0088] The calculation formula for total energy includes:

[0089]

[0090] The calculation formula for normalized energy includes:

[0091]

[0092] The global probability distribution describes how the signal energy is distributed across the entire time-frequency plane and satisfies Therefore, each signal data has a corresponding global probability distribution. The global probability distribution is the basis for calculating information entropy (such as time-frequency entropy). It converts the energy spectrum into a probability distribution, making it possible to quantify the complexity or uncertainty of the signal energy distribution using information theory methods.

[0093] The calculation method of global time-frequency entropy includes: for each signal data and its corresponding global probability distribution, the definition of Shannon entropy is applied to calculate the global time-frequency entropy of the signal data, which is recorded as The calculation formula of global time-frequency entropy includes:

[0094]

[0095] Where, is the number of preset scale parameters, is the number of time parameters (i.e. the length of the signal data segment). hour, .

[0096] Global time-frequency entropy The higher the value, the more diffuse, disordered, or complex the distribution of signal energy on the time-frequency plane. Conversely, if the energy is concentrated in a few time-frequency regions, the global time-frequency entropy value is low. The emergence of incipient fault characteristics often causes the signal energy to become dispersed from its normal concentrated distribution, or the emergence of new energy concentration regions. These changes can cause changes in the global time-frequency entropy value. Therefore, global time-frequency entropy can serve as a sensitive indicator for measuring changes in the overall time-frequency characteristics of a signal. A corresponding global time-frequency entropy value is calculated for each signal data point (vibration data or pressure data). Global time-frequency entropy can quantify the complexity or uncertainty of the signal's time-frequency representation as a whole. It is sensitive to changes in the distribution pattern of signal energy on the entire time-frequency plane (such as dispersion or the emergence of new components), and is an important indicator of incipient fault characteristics.

[0097] The method for obtaining regional concentration includes: calculating the regional energy of each signal data in each preset region of interest based on the wavelet spectrum; and calculating the regional concentration of each signal data in each preset region of interest based on the regional energy.

[0098] The method for selecting the preset region of interest includes: according to the structural characteristics of the test machine, the known natural frequencies of the components, the natural frequencies obtained through modal analysis or theoretical calculation, the frequency band (or scale band) and time period sensitive to the initial fault characteristics in the historical fault data, or through preliminary signal analysis (such as observing the difference in wavelet spectra of signals under healthy state and known fault state) as well as engineering experience and expert knowledge, preliminarily determining one or more regions on the time-frequency plane that may be closely related to specific abnormal modes or structural response changes; these preliminarily determined regions are accurately defined on the time-frequency (or scale-time) plane, and each preset region of interest is recorded as ,in is the index of a region defined by its boundaries in the scale (or frequency) dimension (e.g., scale range [ ],in For the region The starting scale, is the end scale; or the corresponding frequency range [ ],in is the starting frequency, is the end frequency) and the boundaries in the time dimension (e.g., the time range [ ],in is the starting time, These boundaries can be fixed or relative to the total length of the signal segment or a specific event (such as an impulse peak). A preset ROI can be rectangular or irregularly shaped as needed. All preset ROIs are grouped together to form a preset ROI library.

[0099] The purpose of defining preset ROIs is to focus on those time-frequency subspaces most likely to reflect specific incipient fault patterns or critical structural response changes, enabling more refined feature extraction. By pre-setting these ROIs, energy changes within these areas can be monitored specifically. This helps identify characteristic signals associated with specific incipient fault characteristics or locations, improving detection specificity and sensitivity.

[0100] The calculation method of regional energy includes: for each signal data and its corresponding wavelet spectrum , and for each preset region of interest , for all the wavelet spectrograms that fall within the preset region of interest Wavelet energy in the range Sum up and get the regional energy, which is recorded as .if By scale index range[ ] and time index range [ ] definition, then:

[0101]

[0102] The range of the summation is all the Therefore, each signal data in each preset region of interest A corresponding regional energy will be calculated .

[0103] Regional energy directly reflects the energy intensity of the signal in a specific time-frequency region. Its changes can indicate whether the physical process corresponding to the region (such as vibration at a specific frequency, impulse response in a specific time period) has undergone abnormal changes.

[0104] The calculation method of regional concentration includes: for each signal data, the regional energy and total energy The ratio is used as signal data in the preset region of interest The regional concentration within is denoted as The calculation formula for regional concentration includes:

[0105]

[0106] Regional concentration The value of ranges from 0 to 1 (assuming the energy is non-negative). Regional concentration represents the proportion of the energy of a specific preset region of interest in the total signal energy. If the regional concentration of a preset region of interest changes significantly, for example, the energy concentration in a preset region of interest corresponding to a frequency band that represents the main response in a healthy state decreases, or the energy concentration in a preset region of interest corresponding to a frequency band that originally has less energy but is related to a certain abnormality increases, it may indicate that the distribution of signal energy in the time-frequency plane has changed, which may be related to the initial fault characteristics of the structure or system. A corresponding regional concentration will be calculated for each signal data in each preset region of interest. Regional concentration can quantify the energy proportion of a specific time-frequency region and is not affected by fluctuations in the overall energy level of the signal. It is very effective for monitoring changes in the energy distribution of known frequency bands or the emergence of new frequency bands, thereby being able to more robustly reflect changes in the state of the structure or system.

[0107] The method for obtaining local time-frequency entropy includes: calculating the local probability distribution of each signal data in each preset region of interest based on the wavelet spectrum; and calculating the local time-frequency entropy of each signal data in each preset region of interest based on the local probability distribution.

[0108] The calculation method of local probability distribution includes: for each signal data and its corresponding wavelet spectrum , and for each preset region of interest First, calculate the regional energy of the signal data in the preset region of interest (already obtained in the regional energy calculation method); then, for each scale-time unit in the preset region of interest , the wavelet energy Divided by the regional energy , we get the normalized energy of the wavelet energy of each scale-time unit relative to the regional energy, that is, the scale-time unit The local probability within the preset region of interest is denoted as ,in, All local probabilities of a signal data within a preset region of interest constitute the local probability distribution of the signal data within the preset region of interest.

[0109] The calculation formula for local probability includes:

[0110]

[0111] The local probability distribution describes how the signal energy is distributed within a specific preset region of interest and satisfies Therefore, each signal data in each preset region of interest There is a corresponding local probability distribution.

[0112] The local probability distribution is the basis for calculating the information entropy within a specific region. It converts the regional energy spectrum (that is, all wavelet energies within the region) into the probability distribution form within the region, making it possible to use information theory methods to quantify the complexity or uncertainty of the energy distribution within a specific time-frequency region.

[0113] The calculation method of local time-frequency entropy includes: for each signal data and its location in each preset region of interest The local probability distribution within , the definition of Shannon entropy is applied to calculate the signal data in the preset region of interest The local time-frequency entropy within is denoted as The calculation formula of local time-frequency entropy is as follows:

[0114]

[0115] The range of the summation is all the Scale-time unit In the calculation, it is also agreed that hour, .

[0116] Local time-frequency entropy A higher value indicates a more diffuse, disordered, or complex distribution of signal energy within a specific preset region of interest. Conversely, if the energy is concentrated at a few time-frequency points within the region, the local time-frequency entropy value is lower. Changes in local time-frequency entropy can capture subtler and more localized changes in energy distribution patterns than global time-frequency entropy, making it potentially more sensitive for identifying early-stage fault signatures that affect only specific time-frequency regions. A corresponding local time-frequency entropy value is calculated for each signal data point in each preset region of interest.

[0117] Changes in local time-frequency entropy can capture more subtle, localized changes in energy distribution. Local time-frequency entropy focuses on the complexity of energy distribution within a specific region, revealing subtle changes in signal characteristics in that region and compensating for the shortcomings of global features in locating anomalies.

[0118] The method for constructing the feature vector includes: for each independent dynamic loading event, all the signal analysis features extracted from each corresponding signal data (i.e., vibration data or pressure data from a specific sensor) are arranged in order to form a comprehensive feature vector to fully characterize the time-frequency domain dynamic characteristics of the signal data in the dynamic loading event. Specifically, if a signal data extracts the global time-frequency entropy , and in Regional concentration on a preset region of interest and local time-frequency entropy , then it is the feature vector constructed for the signal data is an ordered arrangement of all signal analysis features, namely:

[0119]

[0120] If a dynamic loading event collects The vibration sequence of each vibration monitoring location and The pressure sequence of each pressure monitoring position will be obtained Vibration data and Pressure data. Each of the signal data is constructed with its own feature vector according to the above method. Then, the feature vectors of all the signal data from the same dynamic loading event can be further spliced ​​together to form a higher-dimensional total feature vector representing the overall state of the dynamic loading event. Alternatively, you can also use the eigenvector of each signal data As a sample input for the subsequent model, it is necessary to record the sensor source information corresponding to the feature vector at the same time. In this embodiment, consider constructing an independent feature vector for each signal data. , and uses it as the basic unit for subsequent processing. Therefore, each signal data (for example, vibration data from an accelerometer, or pressure data from a pressure sensor) will have a feature vector composed of its global time-frequency entropy, the regional concentration of all preset regions of interest, and the local time-frequency entropy of all preset regions of interest.

[0121] Fusion of multiple scalar features describing signal characteristics from different dimensions into a feature vector provides a standardized and more informative input for subsequent machine learning models, helping the models learn complex anomaly patterns.

[0122] The training module trains the isolation forest model and the mechanical encoding model based on the health feature vector.

[0123] The Isolation Forest (iForest) model is an anomaly detection algorithm based on ensemble learning. Its core principle is that outliers are typically few and distinct, making them easier to isolate in feature space than normal points. The Isolation Forest model achieves this by constructing an ensemble of multiple Isolation Trees (iTrees).

[0124] The training method of the isolation forest model includes: using the healthy feature vector as training data to obtain a training data set; randomly generating initial subsets with a preset total number of trees, where each initial subset is a sample set with a preset number of samples randomly selected independently from the complete training data set; constructing a corresponding isolation tree for each initial subset to obtain a preset total number of isolation trees; all constructed isolation trees together constitute the isolation forest model.

[0125] The isolation tree construction methods include:

[0126] Step F1F1: Create a root node of the isolation tree and set the root node as the current node; set the current node subset of the current node to be equal to the initial subset, and set the node depth of the current node to 0.

[0127] Steps F1 and F2: Determine whether the current node meets the stopping condition; if so, mark the current node as a leaf node and terminate the splitting of the current node.

[0128] Steps F1 and F3: If the stopping condition is not met, a feature dimension of a healthy feature vector (for example, a local time-frequency entropy of a vibration data) is randomly selected from the current node subset of the current node as the current feature dimension used for splitting.

[0129] Steps F1 and F4: Randomly select a value between the minimum and maximum values ​​of the current node subset on the current feature dimension to obtain a split point.

[0130] Steps F1F5: All training data in the current node subset whose values ​​on the current feature dimension are less than the split point enter the left subset; all training data in the current node subset whose values ​​on the current feature dimension are greater than or equal to the split point enter the right subset.

[0131] Step F1F6: Create a left child node and a right child node for the current node. For the left child node, use the left subset as the current node subset, add 1 to the node depth of the current node as the current node depth, and return to step F1F2 to continue the construction. For the right child node, use the right subset as the current node subset, add 1 to the node depth of the current node as the current node depth, and return to step F1F2 to continue the construction.

[0132] Steps F1F7: If the current node is a non-leaf node, save the current feature dimension and split point used to split the current node; if the current node is a leaf node, record the number of training data in the current node subset and the final depth (i.e., the current node depth of the leaf node).

[0133] Stop conditions include:

[0134] The current node subset contains only one training data, or all training data are the same;

[0135] The depth of the tree has reached the preset maximum depth;

[0136] All training data has been isolated.

[0137] It is worth noting that it is necessary to ensure that the total number of training data in the training dataset (for example, hundreds to thousands of samples, covering various normal operating conditions when the test machine is running healthy) is sufficient to represent the characteristic distribution of the test machine when it is running healthy. The preset number of samples is determined by the hyperparameter "Number of samples used for training when building each tree" (max_samples). For example, it can be set to an absolute number such as 256, or set to a fixed proportion of the total number of training data. Using subsamples helps to increase the diversity between trees and improve the generalization ability of the model. If the preset number of samples is not specified or the preset number of samples is set to be equal to the total number of samples, all training data are used to build each isolation tree. A smaller preset number of samples can increase the diversity of the tree, reduce the risk of overfitting, and reduce training time, but it may cause the discrimination ability of a single tree to decrease. The preset maximum depth is related to the preset number of samples. For example, the preset maximum depth can be set to , where is the logarithmic function with base 2, The symbol for rounding up (indicates rounding to the smallest integer not less than its internal value). The total number of trees is specified by the hyperparameter "number of isolated trees" (n_estimators), which can be set to an integer between 100 and 500, for example. A larger number of trees generally improves model stability and anomaly detection accuracy, but also increases computational overhead and memory usage. Hyperparameters can be selected based on empirical values ​​or tuned on a validation set through cross-validation to achieve optimal detection performance and training efficiency.

[0138] The isolation tree is a random binary tree structure that recursively divides the current node subset by performing random feature selection and random split point selection on the training data, with the goal of isolating each training data to a leaf node.

[0139] For a given input feature vector, the number of edges from the root node to the leaf node where the input feature vector is located in an isolation tree is the path length of the input feature vector in the isolation tree. Outliers are expected to be isolated with shorter path lengths because they are "few and different."

[0140] The isolation forest model is suitable for the task of identifying the initial fault features of the compression and shear testing machine of the present invention, mainly based on the following advantages: it does not need to assume that the data obeys a specific probability distribution (such as a Gaussian distribution) and has good adaptability to various forms of healthy feature vector distributions; due to its random partitioning characteristics and usually shallow tree depth, the isolation forest has high computational efficiency in both the training and prediction stages, and is suitable for processing large-scale data sets and real-time detection needs; it can effectively process high-dimensional feature vectors, and the feature vectors constructed in the present invention may contain multiple time-frequency features; because it only focuses on how to "isolate" samples, it has the ability to detect global outliers far away from dense data areas and local anomalies that form small clusters.

[0141] After training is complete, a trained isolation forest model is obtained. This model learns the distribution pattern of healthy feature vectors in the feature space, specifically the average path length characteristic required for effective isolation of healthy feature vectors. This trained isolation forest model is then used by the subsequent detection module to calculate anomaly scores for feature vectors generated by new dynamic loading events, providing a preliminary assessment of whether they deviate from a healthy state.

[0142] The mechanical encoding model is a mechanically aware autoencoder model. In this example, an autoencoder (AE) neural network architecture is employed. Its core objective is to learn an effective representation of healthy feature vectors and identify anomalies that deviate from healthy patterns by minimizing reconstruction loss. The so-called "mechanical encoding" does not mean that the model directly solves mechanical equations. Instead, it refers to indirectly learning the underlying patterns and constraints associated with normal mechanical behavior embedded in this health data by training on a large number of feature vectors representing the healthy mechanical behavior of the testing machine.

[0143] The training method of the mechanical coding model includes: using the same healthy feature vectors as those used in the isolation forest model training or from the same batch as the training data; dividing the training data into a training set and a test set with a ratio of 8:2; then using the training set to learn the model and evaluating its performance through the test set; when the model can accurately reconstruct the input training data, the training is considered complete and the mechanical coding model is obtained.

[0144] The mechanical coding model compresses the original high-dimensional health feature vector input through the encoder and maps it to a latent space representation as a low-dimensional vector. The decoder reconstructs the latent space representation output by the encoder back to the original health feature vector. The goal of the mechanical coding model training is to make the feature vector reconstructed by the decoder The healthy feature vector with the original input The reconstruction loss between them should be as small as possible.

[0145] The calculation method of reconstruction loss includes: calculating the mean square error between the input healthy feature vector and the reconstructed feature vector, and obtaining the reconstruction loss, which is recorded as The calculation formula of reconstruction loss includes:

[0146]

[0147] Where, is the dimension of the feature vector, is the first health feature vector of the input A quantity, is the reconstructed eigenvector A portion.

[0148] Because the mechanical encoding model is trained only on healthy feature vectors to reconstruct them as accurately as possible, it learns effective compressed representations and high-fidelity reconstruction capabilities for healthy feature vectors. However, when the input feature vector comes from an event that actually exhibits incipient fault characteristics, its pattern deviates from the learned healthy pattern, preventing the mechanical encoding model from accurately reconstructing it, resulting in a large reconstruction error. This large reconstruction error becomes the basis for determining an anomaly.

[0149] The detection module calculates the anomaly score of each dynamic loading event based on the feature vector obtained in real time using the isolation forest model; dynamic loading events with an anomaly score greater than or equal to the preset suspicious threshold are marked as suspicious events.

[0150] The calculation method of the abnormal score includes: taking the feature vector obtained in real time Input into the trained isolation forest model; traverse each isolation tree in the isolation forest model, and in each isolation tree Starting from the root node, according to the random splitting rules of each non-leaf node in the isolation tree (that is, in which feature dimension and at which cut point to divide), the feature vector Guide the corresponding child node until it reaches a leaf node and records the feature vector The number of edges from the root node to the leaf node where the eigenvector finally passes through, and the eigenvector in each isolation tree is obtained The path length in is denoted as ; Calculate the eigenvector The expected value of the path length on all isolation trees in the isolation forest model is used to obtain the average path length. ; Calculate the inverse of the ratio of the average path length to the normalization factor to obtain the normalized length; calculate the exponential function with the normalized length as the exponent and the base 2 to obtain the anomaly score, which is recorded as .

[0151] The calculation formula for average path length includes:

[0152]

[0153] Where, is the total number of isolated trees in the isolation forest model, is the eigenvector In the The length of the path in the isolation tree.

[0154] The calculation formula for the anomaly score includes:

[0155]

[0156] Where, is the preset number of samples. In this embodiment, ; Is a normalization factor used to standardize the average path length so that the average path length is comparable. The calculation formula of the anomaly score is to convert the average path length Converted to an anomaly score normalized between 0 and 1.

[0157] The calculation formula of the normalization factor is:

[0158]

[0159] Where, is a harmonic number and can be approximated as ,and , is the Euler–Mascheroni constant.

[0160] When the anomaly score is close to 1, it indicates that the average path length Relative to is small, which means that the eigenvector It is very easy to be isolated in the isolation tree (that is, only a few splits are needed to separate it from other samples). Therefore, the probability of this feature vector being considered an anomaly is high. When the anomaly score is much less than 0.5 (that is, close to 0), it means that the average path length Relatively long, meaning that the eigenvector It is hidden deep in the leaf nodes of the isolation tree and requires many splits to be isolated. Its behavior is similar to the healthy samples used in training. Therefore, the probability of this feature vector being considered a normal point is high. When the anomaly score is close to 0.5, it means that the average path length Approximately equal to At this time, the eigenvector has no obvious abnormal or normal tendency, and its judgment is relatively vague.

[0161] The core concept of Isolation Forest is to ensure that outliers are "few but distinct." Because their feature values ​​often deviate from the dense areas of normal data points, they can be isolated using random features and random split points without requiring many partitions. Consequently, their paths in the isolation tree are relatively short. In contrast, normal data points, located in dense areas, require more partitions to be isolated, resulting in relatively long paths.

[0162] When constructing each isolation tree, the Isolation Forest model randomly selects a feature dimension and a split point within that feature dimension to partition the data. If a feature vector is anomalous, some of its eigenvalues ​​may be significantly different from those of most normal samples. Therefore, when a feature dimension and split point that can distinguish anomalies are randomly selected, the anomalous feature vector is quickly divided into a separate branch and isolated. Because the Isolation Forest model contains multiple random isolation trees, this randomness ensures that even if some isolation trees fail to effectively isolate the anomalous feature vector, others have a chance to do so.

[0163] The method for determining the value of the preset suspicious threshold includes: inputting all the healthy feature vector sets used in training the isolation forest model (or an independent set of healthy feature vectors that did not participate in the training) into the trained isolation forest model, calculating the abnormality score of each healthy feature vector, and obtaining the healthy score; marking the distribution of the healthy score as the healthy score distribution; selecting a percentile of the healthy score distribution (for example, the 95th percentile or the 99th percentile) as the preset suspicious threshold, which is recorded as For example, if the 99th percentile is selected, it means that in a healthy state, only 1% of samples will be misclassified as suspicious. Domain experts can initially set a preset suspicious threshold based on the historical data of the test machine, the sensitivity requirements of the specific test, and operational experience. Then, through actual operation or simulation testing using historical data, the effectiveness of the preset suspicious threshold in screening known or simulated suspicious events (such as whether it can effectively detect while avoiding excessive and unnecessary "suspicious" marks) is observed, and iterative adjustments and optimizations are performed.

[0164] The purpose of selecting a preset suspicious threshold is to minimize the probability of misclassifying normal events as suspicious events while ensuring a sufficiently high detection rate for true anomalies, thereby reducing the burden on the subsequent anomaly confirmation module. Factors influencing the selection of the preset suspicious threshold include the required sensitivity for detecting early fault features, the acceptable false alarm cost for the system, and the representativeness and inherent fluctuation range of the healthy feature vector. The preset suspicious threshold is used in the detection module to initially mark test events whose anomaly scores output by the isolation forest model are greater than or equal to the preset suspicious threshold as suspicious events for subsequent further confirmation.

[0165] The abnormality confirmation module calculates the reconstruction error of suspicious events based on the feature vectors obtained in real time by the mechanical coding model; suspicious events with a reconstruction error greater than or equal to the preset abnormality threshold are marked as abnormal events.

[0166] The calculation method of reconstruction error includes:

[0167] Step 1: Mark the feature vector of the suspicious event as a suspicious feature vector, denoted as ; Input the suspicious feature vector into the trained mechanical encoding model.

[0168] Step 2: The mechanical coding model reconstructs the original suspicious feature vector and obtains the reconstructed feature vector, which is recorded as .

[0169] Step 3: Calculate suspicious eigenvectors The corresponding reconstructed eigenvector The reconstruction loss between them is obtained, and the reconstruction error is recorded as .

[0170] The calculation formula for the reconstruction error includes:

[0171]

[0172] Where, is the dimension of the feature vector, and are the suspicious eigenvectors and the reconstructed eigenvectors, respectively. A portion.

[0173] The mechanical encoding model learns the intrinsic structure and effective representation of the health data pattern by training on a large number of health feature vectors, so it can reconstruct the health input with a smaller error. When it comes from an event with actual early abnormality, its pattern will deviate from the learned healthy pattern, causing the mechanical encoding model to be unable to accurately reconstruct it, resulting in a large reconstruction error.

[0174] The method for determining the value of the preset abnormality threshold includes: inputting all the healthy feature vector sets used in training the mechanical coding model (or an independent set of healthy feature vectors that did not participate in the training) into the trained mechanical coding model, calculating the reconstruction error of each healthy feature vector, and obtaining the healthy error; marking the distribution of the healthy error as the healthy error distribution; selecting a high percentile of the healthy error distribution (for example, the 95th percentile or the 99th percentile) as the preset abnormality threshold For example, selecting the 99th percentile means that under healthy conditions, only 1% of samples will have a reconstruction error exceeding this threshold. Domain experts can initially set a preset anomaly threshold based on the historical data of the test machine, the sensitivity requirements of the specific test, and operational experience. This threshold can then be iteratively adjusted and optimized by observing its effectiveness in screening known or simulated anomalies during actual operation or through simulations using historical data.

[0175] The core purpose of selecting a preset anomaly threshold is to ensure a high confirmation rate for true anomalies while effectively filtering out false positives generated by the isolation forest model, thereby improving the accuracy of the entire anomaly identification system. Factors influencing the selection of the preset anomaly threshold include the confidence level required for confirming anomalies, the system's tolerance for missed positives, and the average reconstruction accuracy of the mechanical encoding model for healthy data. The preset anomaly threshold is used in the anomaly confirmation module to confirm suspicious events with a reconstruction error greater than or equal to this threshold as anomalies.

[0176] The alarm module calculates the comprehensive damage score of each abnormal event based on the anomaly score and reconstruction error; obtains the damage level based on the comprehensive damage score; and generates an alarm message when the comprehensive damage score is higher than the preset alarm level.

[0177] The calculation method of the comprehensive damage score includes: normalizing the reconstruction error to obtain the normalized error; performing a weighted linear combination of the abnormality score and the normalized error according to the preset weight to obtain the comprehensive damage score, which is recorded as .

[0178] The calculation method of the normalized error includes: determining a preset upper limit of the reconstruction error in a healthy state (for example, it can be taken as a preset abnormal threshold , or the 99.9 percentile of the reconstruction error of the healthy feature vector set), and obtain the upper limit of the reconstruction error, which is recorded as ; Determine a preset baseline value of the reconstruction error in a healthy state (For example, the mean or median of the reconstruction error of the healthy feature vector set) to obtain the lower limit of the reconstruction error; calculate the difference between the input reconstruction error and the lower limit of the reconstruction error to obtain the relative error; select the larger of the relative error and 0 to obtain the relative reconstruction error; calculate the difference between the upper limit of the reconstruction error and the lower limit of the reconstruction error to obtain the error interval; calculate the ratio of the relative reconstruction error to the error interval to obtain the error ratio; select the smaller of the error ratio and 1 to obtain the normalized error, which is recorded as The calculation formula for normalized error includes:

[0179]

[0180] Where, is the function that takes the minimum value, is the function that takes the maximum value.

[0181] The calculation formula for the comprehensive injury score includes:

[0182]

[0183] Where, is the anomaly score calculated by the isolation forest model in the detection module for abnormal events, i.e. , and is the preset weight; , ,and .

[0184] Preset weights are used to balance the impact of anomaly scores and normalized error on the overall damage score. Preset weights can be determined based on expert experience, historical data analysis, or an equal weighting strategy to reflect the relative importance and discernment of each indicator in damage identification.

[0185] The comprehensive damage score is designed to combine the rapid anomaly detection capabilities of the isolation forest model with the reconstruction-based anomaly confirmation capabilities of the mechanical encoding model, resulting in a more comprehensive and robust quantitative assessment of the damage severity of incipient fault characteristics of the testing machine. Events with high anomaly scores and large reconstruction errors are associated with higher comprehensive damage scores.

[0186] The method of obtaining the injury level includes: converting the continuous comprehensive injury score (between 0 and 1) is discretized into several ordered damage levels by comparing with a set of preset damage thresholds.

[0187] Preset damage thresholds include: Level 1 threshold , Level 2 threshold and level three thresholds .

[0188] Injury levels include:

[0189] When the comprehensive injury score is greater than the level 1 threshold and less than or equal to the level 2 threshold ( ), the abnormal event is rated as minor concern (i.e., level 1). Minor concern indicates that there may be very early or very minor abnormal signs. It is recommended that the operator increase the frequency of monitoring the operating status of the testing machine or pay attention to the relevant components during the next planned maintenance.

[0190] When the comprehensive injury score is greater than the level 2 threshold and less than or equal to the level 3 threshold ( ), the abnormal event is rated as a moderate warning (i.e., Level 2). A moderate warning indicates that there are obvious early signs of a fault. Maintenance personnel are advised to arrange for an inspection of the testing machine to initially locate the problem and consider whether to adjust subsequent test parameters or reduce the loading frequency / amplitude.

[0191] When the comprehensive injury score is greater than the level 3 threshold ( ), the abnormal event is rated as a serious warning (i.e., level 3). A serious warning indicates that there is a significant abnormality or damage that may have affected the test safety or data accuracy. It is recommended to immediately suspend the test and conduct a detailed inspection of the testing machine and necessary repairs.

[0192] Methods for determining the preset damage thresholds include: the level 1 threshold is a value that should be greater than the maximum comprehensive damage score that can be achieved by the healthy feature vector; the level 2 and level 3 thresholds can be determined by combining the following methods:

[0193] 1. Utilize the comprehensive damage score distribution of existing fault cases to extract typical interval boundaries corresponding to different damage levels, and set preset damage thresholds that can effectively distinguish between minor concerns, moderate warnings, and severe alerts.

[0194] 2. The value of the composite damage score should be determined based on the understanding of the potential risk and maintenance urgency associated with different values ​​by experienced testing machine operators, maintenance engineers, or structural health monitoring experts. For example, a composite damage score exceeding 0.7 may indicate the need for emergency maintenance.

[0195] 3. Determine the corresponding score threshold based on the manufacturer's maintenance guidelines and fault level standards, and refer to the safety margin or fault boundary.

[0196] The number and specific values ​​of the preset damage thresholds should be determined based on the actual testing machine management requirements, maintenance resources, and the ability to tolerate different levels of risk. Typically, three to five damage levels are defined to facilitate management and decision-making.

[0197] The method for determining the value of the preset alarm level is mainly determined by the system's risk management strategy and maintenance response capabilities. The preset alarm level is usually directly set to one or more of the damage levels defined above. For example, moderate warning and above (i.e., level two and level three) can be set as the level to trigger the alarm. This means that when the damage level of an abnormal event is assessed to reach or exceed level two, the system should generate an alarm message. The selection of the preset alarm level requires a balance between detection sensitivity, maintenance cost and potential risks. In some cases, the preset alarm level can also be dynamically adjustable. For example, based on the importance of the task currently performed by the test machine, the availability of maintenance resources or the evolution of historical faults, the preset alarm level can be temporarily adjusted to adapt to different operating scenarios and risk preferences.

[0198] Alert information includes: unique identifier of the abnormal event (e.g., based on timestamp, sequence number of dynamically loaded events, or automatically generated by global event counter), date and time of abnormal event occurrence, damage level of abnormal event, key abnormal characteristics, and comprehensive damage score , anomaly score , and reconstruction error .

[0199] The identification method of key abnormal features includes: extracting the feature vector corresponding to the abnormal event; in the healthy feature vector set, respectively counting the mean and standard deviation of each feature dimension to obtain the healthy mean and healthy standard deviation; for each feature dimension, first calculating the difference between the corresponding component in the feature vector of the abnormal event and the healthy mean, and then calculating the absolute value of the ratio of the difference to the healthy standard deviation (using the Z-score method) to obtain the feature deviation; sorting the feature deviations corresponding to each feature dimension in descending order, and selecting the top feature dimension. The signal analysis features corresponding to the characteristic deviations are used as the key abnormal features of abnormal events.

[0200] This embodiment, through the sequential connection and coordinated operation of the aforementioned acquisition module, feature extraction module, training module, detection module, anomaly confirmation module, and alarm module, forms a complete and efficient system for identifying early-stage fault characteristics in compression shear testing machines under extreme dynamic testing. This system first acquires high-quality signal data from raw vibration and pressure sensor signals through a sophisticated signal acquisition and multi-step preprocessing process (including data segmentation, baseline correction and detrending, noise suppression, and normalization). The feature extraction module then performs a continuous wavelet transform on these signal data, extracting multi-dimensional signal analysis features including global time-frequency entropy, regional concentrations of multiple pre-defined regions of interest, and local time-frequency entropy. This constructs an information-rich feature vector for each signal data point. The training module uses a large number of these feature vectors, acquired when the testing machine is in a healthy state, to perform offline training on the isolation forest model and the mechanical encoding model, enabling them to learn patterns of healthy conditions. During the online monitoring phase, the detection module uses the trained isolation forest model to rapidly score the feature vectors acquired in real time for anomalies, initially screening for suspicious events. Next, the anomaly confirmation module uses a trained mechanical coding model to calculate the reconstruction error of the feature vectors of these suspicious events. By comparing it with a preset threshold, it further confirms the true anomaly event, effectively reducing false alarms. Finally, the alarm module calculates a comprehensive damage score based on the anomaly score of the isolation forest model and the normalized reconstruction error of the mechanical coding model. Based on this score, the anomaly event is classified into different damage levels. When the damage level reaches the preset alarm level, an alarm is generated containing event details and key anomaly feature information.

[0201] Compared with the existing technology, the technical problem that this embodiment aims to solve is: under the harsh conditions of extreme dynamic tests such as impact or high-rate cyclic loading in large-scale compression and shear testing machines, traditional methods are difficult to effectively and reliably identify subtle, complex and diverse early fault characteristic signals such as early fatigue damage of structures, early loosening of component connections or early abnormalities of hydraulic systems. These signals often have non-stationary and transient characteristics and are easily masked by strong background noise or normal operation variations, resulting in traditional methods based on fixed thresholds or single models having problems such as insufficient detection sensitivity, high false alarm rate or lack of diagnostic information.

[0202] The technical solution of this embodiment achieves the following beneficial technical effects through the coordinated work of its specific modules and detailed method steps:

[0203] (a) By performing meticulous preprocessing of vibration and pressure signals (e.g., detrending, targeted noise reduction, and independent normalization) and employing advanced continuous wavelet transform (CWT) to extract multidimensional time-frequency features, including global time-frequency entropy, regional concentration for multiple preset regions of interest, and local time-frequency entropy, the system significantly enhances the ability to capture and characterize weak, transient, incipient fault characteristic signals in complex backgrounds, thereby improving the input quality and sensitivity of subsequent anomaly detection.

[0204] (b) A two-stage anomaly detection strategy that uses the isolation forest model for preliminary rapid screening and combines it with the mechanical coding model for in-depth confirmation can ensure a high detection rate of true initial fault characteristics while effectively filtering out false alarms that may be generated in the first stage through the reconstruction and analysis of suspicious events in the second stage, thereby significantly improving the overall accuracy and credibility of anomaly warnings and reducing unnecessary downtime inspections.

[0205] (c) By calculating a comprehensive damage score that combines the anomaly score and the normalized reconstruction error and mapping it to multiple predefined damage levels, including minor concern, moderate warning, and severe alert, this system provides a more refined and instructive quantitative assessment of the test machine's health status, rather than a simple binary (normal / abnormal) judgment. Furthermore, by identifying the key anomaly features that contribute most to the anomaly determination and the signal data types that originate from them, it provides operators and maintenance personnel with preliminary fault diagnosis clues, helping to more quickly locate the root cause of the problem.

[0206] In general, the system provided in this embodiment, through its multi-module integrated design, refined signal processing and feature engineering, and collaborative decision-making of dual machine learning models, achieves efficient, accurate, and reliable identification and evaluation of initial fault characteristics of the compression and shear testing machine under extreme dynamic test conditions, providing strong technical support for achieving predictive maintenance, ensuring the safe operation of the testing machine, improving the quality of test data, and optimizing maintenance strategies. Figure 2 The application scenario diagram shown in the figure shows that this system can ensure the safe operation of the testing machine.

[0207] Anything not described in this application can be achieved by adopting or drawing on existing technologies.

[0208] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. The compression shear testing machine state monitoring system is characterized by: include: The acquisition module performs signal preprocessing on the acquired signal sequence to obtain signal data; The feature extraction module performs continuous wavelet transform on each signal data to obtain the wavelet coefficient of each signal data; Based on the wavelet coefficients, the signal analysis characteristics of each signal data are obtained; Based on the signal analysis characteristics, a feature vector of each signal data is constructed; the feature vector obtained when the test machine is in a healthy state is marked as a healthy feature vector; The training module trains the isolation forest model and the mechanical encoding model based on the health feature vector; The detection module calculates an anomaly score for each dynamic loading event based on the feature vector obtained in real time using the isolation forest model; dynamic loading events with an anomaly score greater than or equal to a preset suspicious threshold are marked as suspicious events; The abnormality confirmation module calculates the reconstruction error of suspicious events based on the feature vectors obtained in real time using the mechanical coding model; suspicious events with a reconstruction error greater than or equal to a preset abnormality threshold are marked as abnormal events; The alarm module calculates the comprehensive damage score of each abnormal event based on the anomaly score and reconstruction error; Based on the comprehensive injury score, the injury level is obtained; When the comprehensive damage score is higher than the preset alarm level, an alarm message is generated.

2. The compression shear testing machine state monitoring system according to claim 1, characterized in that: The method for performing continuous wavelet transform comprises: performing continuous wavelet transform on each signal data based on a preset mother wavelet function, a preset scale parameter and a time parameter, and obtaining wavelet coefficients of a combination of each signal data, the preset scale parameter and the time parameter; The method for selecting the preset scale parameter includes: calculating the corresponding minimum scale and maximum scale according to the upper and lower frequency limits of the frequency range of the initial fault characteristic component in the signal data and the center frequency of the preset mother wavelet function to obtain the scale range; the minimum scale is smaller than the ratio of the center frequency to the upper frequency limit; the maximum scale is larger than the ratio of the center frequency to the lower frequency limit; within the determined scale range, extracting an arithmetic sequence with the minimum scale as the first item, the maximum scale as the last item and the total number of preset scales, and using the scales in the arithmetic sequence as the preset scale parameter.

3. The compression shear testing machine state monitoring system according to claim 1, characterized in that: Signal analysis features include global time-frequency entropy, regional concentration, and local time-frequency entropy.

4. The compression shear testing machine state monitoring system according to claim 3, characterized in that: The method for obtaining the global time-frequency entropy includes: calculating a wavelet spectrum based on wavelet coefficients; calculating a global probability distribution of each signal data based on the wavelet spectrum; and calculating a global time-frequency entropy of each signal data based on the global probability distribution. The method for obtaining regional concentration includes: calculating the regional energy of each signal data in each preset region of interest based on the wavelet spectrum; calculating the regional concentration of each signal data in each preset region of interest based on the regional energy; The method for obtaining local time-frequency entropy includes: calculating the local probability distribution of each signal data in each preset region of interest based on the wavelet spectrum; and calculating the local time-frequency entropy of each signal data in each preset region of interest based on the local probability distribution and applying the definition of Shannon entropy.

5. The compression shear testing machine state monitoring system according to claim 4, characterized in that: The calculation method of the wavelet spectrum includes: calculating the square modulus of the wavelet coefficient to obtain the wavelet energy; for each two-dimensional wavelet coefficient matrix corresponding to each signal data, calculating the wavelet energy of each wavelet coefficient in the two-dimensional wavelet coefficient matrix to obtain the wavelet spectrum of each signal data; The calculation method of the local probability distribution includes: for each signal data and its corresponding wavelet spectrum, and for each preset region of interest, calculating the regional energy of the signal data in the preset region of interest; for each scale-time unit in the preset region of interest, dividing the wavelet energy by the regional energy to obtain the normalized energy of the wavelet energy of each scale-time unit relative to the regional energy, that is, the local probability of the scale-time unit in the preset region of interest; all local probabilities of a signal data in a preset region of interest constitute the local probability distribution of the signal data in the preset region of interest.

6. The compression shear testing machine state monitoring system according to claim 1, characterized in that: The method for constructing the feature vector includes: for each dynamic loading event, sequentially arranging the signal analysis features extracted from each corresponding signal data to form a feature vector for each signal data.

7. The compression shear testing machine state monitoring system according to claim 1, characterized in that: The calculation method of anomaly score includes: The feature vector obtained in real time is input into the trained isolation forest model; each isolation tree in the isolation forest model is traversed. In each isolation tree, starting from the root node, the feature vector is directed to the corresponding child node according to the current feature dimension and split point of each non-leaf node in the isolation tree until it reaches a leaf node. The number of edges that the feature vector passes through from the root node to the leaf node where the feature vector is finally located is recorded to obtain the path length of the feature vector in each isolation tree; the average path length of the feature vector on all isolation trees in the isolation forest model is calculated to obtain the average path length; the negative number of the ratio of the average path length to the normalization factor is calculated to obtain the normalized length; and an exponential function with the normalized length as the exponent and base 2 as the base is calculated to obtain the anomaly score.

8. The compression shear testing machine state monitoring system according to claim 1, characterized in that: The calculation method of reconstruction error includes: The feature vector of the suspicious event is marked as a suspicious feature vector; the suspicious feature vector is input into the trained mechanical coding model; the mechanical coding model reconstructs the original suspicious feature vector to obtain a reconstructed feature vector; the reconstruction loss between the suspicious feature vector and the corresponding reconstructed feature vector is calculated to obtain the reconstruction error.

9. The compression shear testing machine state monitoring system according to claim 1, characterized in that: The calculation method of the comprehensive injury score includes: The reconstruction error is normalized to obtain the normalized error; the anomaly score and the normalized error are weighted linearly combined according to the preset weight to obtain the comprehensive damage score.

10. The compression shear testing machine state monitoring system according to claim 9, characterized in that: The calculation method of normalized error includes: Determine a preset upper limit of the reconstruction error in a healthy state to obtain the upper limit of the reconstruction error; determine a preset baseline value of the reconstruction error in a healthy state to obtain the lower limit of the reconstruction error; calculate the difference between the input reconstruction error and the lower limit of the reconstruction error to obtain the relative error; select the maximum value between the relative error and 0 to obtain the relative reconstruction error; calculate the difference between the upper limit of the reconstruction error and the lower limit of the reconstruction error to obtain the error interval; calculate the ratio of the relative reconstruction error to the error interval to obtain the error ratio; select the minimum value between the error ratio and 1 to obtain the normalized error.

11. The compression shear testing machine state monitoring system according to claim 1, characterized in that: Alarm information includes: unique identifier of the abnormal event, date and time of the abnormal event, damage level of the abnormal event, key abnormal characteristics, comprehensive damage score, abnormality score, and reconstruction error; The identification method of key abnormal features includes: extracting the feature vector corresponding to the abnormal event; in the healthy feature vector set, respectively counting the mean and standard deviation of each feature dimension to obtain the healthy mean and healthy standard deviation; for each feature dimension, calculating the difference between the corresponding component in the feature vector of the abnormal event and the healthy mean, calculating the absolute value of the ratio of the difference to the healthy standard deviation, and obtaining the feature deviation; sorting the feature deviations corresponding to each feature dimension in descending order, and selecting the top feature deviation. The signal analysis features corresponding to the characteristic deviations are used as the key abnormal features of abnormal events.

Images