A method and system for detecting faults in a generator core lamination press

Abstract

The application relates to the technical field of industrial equipment fault detection, and discloses a generator iron core stacking equipment fault detection method and system. The method comprises the following steps: acquiring an original vibration signal, synchronously calibrating the original vibration signal, and extracting a first vibration data set; performing time-frequency analysis and decomposition to obtain a second vibration data set; constructing a model to enhance a periodic weak impact signal, acquiring an enhanced signal feature set, and matching and quantifying a preset reference frequency template library to determine a potential fault signal distribution range; when the threshold value is exceeded, a fault feature data set is obtained through deep feature extraction, an hyperplane decision model is constructed based on the fault feature data set, a feature local evolution trend is analyzed, and a warning result is output. The method can accurately capture early fault signals, reduce the missed detection and misdiagnosis rate, and improve the equipment operation reliability.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault detection technology, and in particular to a fault detection method and system for generator core stacking equipment. Background Technology

[0002] Currently, generator core lamination equipment is the core equipment in generator manufacturing. Its operating status directly affects product accuracy and production efficiency. Fault detection technology has become a key support for ensuring the stability of the production line, while the application of sensor network chips provides basic communication support for multi-node data acquisition.

[0003] In existing technologies, fault detection for generator core stacking equipment mainly relies on a single sensor to collect vibration signals and judge the equipment status through simple time-domain analysis. However, this cannot comprehensively capture the operating information of multiple parts of the equipment, and is prone to misjudgment of faults due to incomplete signals. Alternatively, the timing of the acquisition of different sensors is not synchronized, resulting in phase deviations in the data of each channel. At the same time, using only basic noise reduction methods such as low-pass filtering makes it difficult to separate the high-frequency aliasing signals formed by the superposition of mechanical impact, electromagnetic interference, and environmental vibration during equipment operation. As a result, the periodic weak impact characteristics related to the fault are masked by background noise, making it impossible to accurately extract effective fault information.

[0004] Existing technologies lack enhancement mechanisms for weak features. Even if some abnormal signals can be captured, it is difficult to quantify their correlation with faults, resulting in insufficient reliability and accuracy of fault detection results. Summary of the Invention

[0005] This invention provides a method and system for fault detection of generator core stacking equipment, in order to solve the problem of insufficient reliability and accuracy of fault detection results in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a fault detection method for generator core stacking equipment, comprising:

[0007] The original vibration signal of the generator core stacking equipment during operation is acquired, and the original vibration signal is synchronously calibrated and feature extracted to obtain the first vibration dataset.

[0008] The first vibration dataset is decomposed and denoised by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components and the low-frequency effective components, and reconstructing the second vibration dataset.

[0009] For the second vibration dataset, a bandpass filter is constructed using the spectral kurtosis map for filtering, and the enhanced signal is reconstructed and enhanced using a trained complete dictionary matrix to obtain an enhanced signal feature set;

[0010] The enhanced signal feature set is matched with a preset reference frequency template library for pattern matching, and the matching results are quantitatively analyzed to determine the distribution range of potential fault signals.

[0011] Calculate the statistical dispersion of the distribution range of the potential fault signal. If it exceeds the preset dispersion threshold, extract the corresponding non-stationary signal segment and perform deep feature extraction through multi-scale wavelet packet decomposition and singular value decomposition to obtain the fault feature dataset.

[0012] A hyperplane decision model is constructed based on the fault feature dataset. By comparing and analyzing the local evolution trend of the features, a fault warning result is output.

[0013] The first vibration dataset is decomposed and denoised by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components from the low-frequency effective components, and reconstructing the second vibration dataset, including:

[0014] The time-frequency distribution matrix of the first vibration dataset is calculated using the Wignerville distribution algorithm;

[0015] Based on the energy ridge line of the time-frequency distribution matrix, the instantaneous frequency and instantaneous amplitude are extracted, and the first vibration dataset is decomposed into a sequence of intrinsic mode components;

[0016] For the intrinsic mode component sequence, calculate the permutation entropy value. If the permutation entropy value exceeds the preset noise tolerance threshold, then the corresponding intrinsic mode component sequence is determined to be a high-frequency aliasing component; otherwise, it is determined to be a low-frequency effective component.

[0017] After performing soft threshold shrinkage on the high-frequency aliasing components, they are linearly superimposed and reconstructed with the low-frequency effective components to obtain the second vibration dataset.

[0018] In one implementation, the step of acquiring the original vibration signal during the operation of the generator core stacking equipment, and performing synchronous calibration and feature extraction processing on the original vibration signal to obtain a first vibration dataset includes:

[0019] The vibration signal of the generator core stacking equipment during operation is collected synchronously at multiple points by a sensor array to obtain the original vibration signal containing multiple frequency components and instantaneous change characteristics.

[0020] The original vibration signal is processed by analog-to-digital conversion and phase calibration to eliminate channel transmission delay, thereby obtaining time-synchronized vibration data;

[0021] The time-synchronized vibration data is decomposed using wavelet packet decomposition to separate steady-state background noise and transient impact components, and multi-frequency component feature vectors are extracted.

[0022] The spatial coordinates of the sensor array are obtained and associated with the feature vector of the multi-frequency components to obtain the first vibration dataset.

[0023] In one implementation, for the second vibration dataset, a bandpass filter is constructed using a spectral kurtosis map for filtering, and reconstruction and enhancement are performed using a trained complete dictionary matrix to obtain an enhanced signal feature set, including:

[0024] The spectral kurtosis plot of the second vibration dataset is obtained to determine the resonance demodulation frequency band. A bandpass filter is constructed to filter the second vibration dataset to obtain a narrowband filtered signal.

[0025] Based on the overcomplete dictionary matrix trained by the narrowband filtered signal, the sparse coefficient vector of the narrowband filtered signal under the overcomplete dictionary matrix is ​​calculated.

[0026] If the sparse coefficient vector satisfies a preset sparsity threshold, it is retained and linearly combined with the overcomplete dictionary matrix for reconstruction. A Hilbert transform is then performed on the reconstructed signal to obtain an enhanced signal feature set.

[0027] In one implementation, the step of performing pattern matching between the enhanced signal feature set and a preset reference frequency template library, and then performing quantitative analysis on the matching results to determine the distribution range of potential fault signals includes:

[0028] Calculate the matching degree between the enhanced signal feature set and each template in the preset reference frequency template library, and generate a multidimensional similarity matrix;

[0029] Filter out local maxima points in the multidimensional similarity matrix whose amplitude exceeds a preset matching threshold, and identify the corresponding specific frequency patterns;

[0030] Calculate the energy entropy value of the signal segment corresponding to the specific frequency mode to obtain the quantization evaluation sequence;

[0031] A fault probability distribution histogram is constructed based on the quantitative evaluation sequence to determine the distribution range of potential fault signals.

[0032] In one implementation, the statistical dispersion of the distribution range of the potential fault signal is calculated. If it exceeds a preset dispersion threshold, the corresponding non-stationary signal segment is extracted and deep feature extraction is performed through multi-scale wavelet packet decomposition and singular value decomposition to obtain a fault feature dataset, including:

[0033] Calculate the statistical dispersion of the potential fault signal distribution range. If the statistical dispersion exceeds a preset dispersion threshold, then extract the corresponding non-stationary signal segment.

[0034] Multi-scale wavelet packet decomposition is performed on the non-stationary signal segment to reconstruct a high-resolution time-frequency energy spectrum;

[0035] Singular value decomposition is performed on the high-resolution time-frequency energy spectrum to extract the non-zero singular value sequence and construct the singular value feature vector.

[0036] The singular value eigenvectors are dimensionality reduced by kernel principal component analysis to generate independent eigenvectors, which are then structured and arranged to obtain a fault feature dataset.

[0037] In one implementation, the step of constructing a hyperplane decision model based on the fault feature dataset and outputting a fault warning result by comparing and analyzing the local evolution trend of the features includes:

[0038] Based on the independent feature vectors in the fault feature dataset, a hyperplane decision model is constructed.

[0039] The feature deviation of the independent feature vector in the hyperplane decision model is calculated in real time, and an evolution time series is constructed based on the feature deviation at consecutive time points;

[0040] The local evolution trend of the evolution time series is extracted, the rate of change and the anomaly confidence are calculated, and the early failure risk index is obtained by mapping.

[0041] If the early failure risk index exceeds the preset warning threshold, a failure warning result containing a risk level identifier will be output.

[0042] Secondly, the present invention provides a fault detection system for generator core stacking equipment, comprising:

[0043] The data acquisition module is used to acquire the original vibration signal during the operation of the generator core stacking equipment, and to perform synchronous calibration and feature extraction processing on the original vibration signal to obtain the first vibration dataset.

[0044] The decomposition and denoising module is used to decompose and denoise the first vibration dataset by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components and the low-frequency effective components, and reconstructing the second vibration dataset.

[0045] The feature enhancement module is used to filter the second vibration dataset by constructing a bandpass filter through the spectral kurtosis map, and to reconstruct and enhance it by training a complete dictionary matrix, so as to obtain an enhanced signal feature set.

[0046] The fault matching module is used to perform pattern matching between the enhanced signal feature set and the preset reference frequency template library, perform quantitative analysis on the matching results, and determine the distribution range of potential fault signals.

[0047] The feature extraction module is used to calculate the statistical dispersion of the distribution range of the potential fault signal. If it exceeds the preset dispersion threshold, the corresponding non-stationary signal segment is extracted and deep feature extraction is performed through multi-scale wavelet packet decomposition and singular value decomposition to obtain the fault feature dataset.

[0048] The judgment and early warning module is used to construct a hyperplane decision model based on the fault feature dataset, and output fault early warning results by comparing and analyzing the local evolution trend of the features.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] (1) This invention achieves multi-channel and multi-point synchronous acquisition through sensor array, eliminates channel transmission delay by combining analog-to-digital conversion and phase calibration, and accurately separates steady-state background noise and transient impact components by combining wavelet packet decomposition. This effectively solves the problems of one-sided signal acquisition and asynchronous timing in the prior art, and improves the comprehensiveness and accuracy of fault feature extraction.

[0051] (2) This invention decomposes the first vibration dataset by using the Wignerville distribution algorithm, filters high-frequency aliasing components based on permutation entropy and performs soft threshold shrinkage processing to achieve accurate separation of effective signals and interference components; then, it determines the resonance demodulation frequency band by using spectral kurtosis analysis, and combines overcomplete dictionary matrix training and Hilbert transform to amplify periodic weak impact signals, thus solving the technical pain point that weak fault characteristics are masked by background noise and significantly improving the sensitivity of early minor fault identification.

[0052] (3) This invention determines the distribution range of potential fault signals by multi-dimensional similarity matrix matching and energy entropy quantification assessment. For signals exceeding the threshold, deep fault features are extracted by multi-scale wavelet packet decomposition, singular value decomposition and kernel principal component analysis. Then, the risk index is obtained by analyzing the feature evolution trend through hyperplane decision model. This realizes the quantitative judgment and graded early warning of faults, reduces the fault missed detection rate and false detection rate, provides accurate basis for equipment operation and maintenance, and effectively improves the reliability of equipment operation. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the fault detection method for generator core stacking equipment provided in the first embodiment of the present invention;

[0054] Figure 2 This is a schematic diagram of the fault detection system for generator core stacking equipment provided in the second embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Reference Figure 1 The first embodiment of the present invention provides a fault detection method for generator core stacking equipment, comprising the following steps:

[0057] S11, acquire the original vibration signal of the generator core stacking equipment during operation, perform synchronous calibration and feature extraction processing on the original vibration signal to obtain the first vibration dataset;

[0058] S12, the first vibration dataset is decomposed and denoised by calculating the time-frequency distribution matrix, separating the high-frequency aliasing component and the low-frequency effective component, and reconstructing to obtain the second vibration dataset.

[0059] S13, For the second vibration dataset, a bandpass filter is constructed using the spectral kurtosis map for filtering, and the enhanced signal is reconstructed and enhanced using a trained complete dictionary matrix to obtain an enhanced signal feature set;

[0060] S14, perform pattern matching between the enhanced signal feature set and the preset reference frequency template library, perform quantitative analysis on the matching results, and determine the distribution range of potential fault signals;

[0061] S15, calculate the statistical dispersion of the distribution range of the potential fault signal. If it exceeds the preset dispersion threshold, extract the corresponding non-stationary signal segment and perform deep feature extraction through multi-scale wavelet packet decomposition and singular value decomposition to obtain the fault feature dataset.

[0062] S16. Based on the fault feature dataset, construct a hyperplane decision model, and output fault warning results by comparing and analyzing the local evolution trend of the features.

[0063] The first vibration dataset is decomposed and denoised by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components from the low-frequency effective components, and reconstructing the second vibration dataset, including:

[0064] The time-frequency distribution matrix of the first vibration dataset is calculated using the Wignerville distribution algorithm;

[0065] Based on the energy ridge line of the time-frequency distribution matrix, the instantaneous frequency and instantaneous amplitude are extracted, and the first vibration dataset is decomposed into a sequence of intrinsic mode components;

[0066] For the intrinsic mode component sequence, calculate the permutation entropy value. If the permutation entropy value exceeds the preset noise tolerance threshold, then the corresponding intrinsic mode component sequence is determined to be a high-frequency aliasing component; otherwise, it is determined to be a low-frequency effective component.

[0067] After performing soft threshold shrinkage on the high-frequency aliasing components, they are linearly superimposed and reconstructed with the low-frequency effective components to obtain the second vibration dataset.

[0068] In step S11, it is necessary to acquire the original vibration signal during the operation of the generator core stacking equipment, and perform synchronous calibration and feature extraction processing on the original vibration signal to obtain the first vibration dataset, including:

[0069] The vibration signal of the generator core stacking equipment during operation is collected synchronously at multiple points by a sensor array to obtain the original vibration signal containing multiple frequency components and instantaneous change characteristics.

[0070] The original vibration signal is processed by analog-to-digital conversion and phase calibration to eliminate channel transmission delay, thereby obtaining time-synchronized vibration data;

[0071] The time-synchronized vibration data is decomposed using wavelet packet decomposition to separate steady-state background noise and transient impact components, and multi-frequency component feature vectors are extracted.

[0072] The spatial coordinates of the sensor array are obtained and associated with the feature vector of the multi-frequency components to obtain the first vibration dataset.

[0073] It should be noted that the original vibration signal is a signal derived from the physical quantities of vibration generated by mechanical movement and structural vibration in key parts of the generator core lamination equipment during operation. The sensor array is a data acquisition unit formed by arranging multiple piezoelectric accelerometers. Multi-frequency components refer to the vibration components in different frequency ranges contained in the original vibration signal, covering low-frequency structural vibration, medium-frequency generator operating vibration, and high-frequency impact vibration, etc. Instantaneous change characteristics refer to the dynamic characteristics of the signal such as abrupt changes and fluctuations in the time dimension. Together, these two constitute the core information dimension reflecting the operating status of the equipment.

[0074] In this step, the sensor array is first arranged and multi-point synchronous data acquisition is performed. Based on the structural characteristics and fault-prone areas of the generator core lamination equipment, the sensor installation locations are determined, typically selecting key components such as the pressure head, frame, and base. The sensor spacing is set according to the equipment size and vibration propagation characteristics, generally 50-500 mm, ensuring comprehensive coverage of the main vibration areas of the equipment. Piezoelectric accelerometers are used, with a measurement range of 0-50g and a frequency response range of 1 Hz-10 kHz, meeting the data acquisition requirements for vibration signals of different frequencies during equipment operation. Synchronous acquisition is achieved through an external hardware trigger signal. The trigger signal uses a TTL level signal with a rising edge trigger mode and a trigger delay of less than 1 microsecond, ensuring that all sensor channels start acquiring data at the same time. The acquisition duration is set according to the equipment operating cycle, generally 10-60 seconds. The acquired signal is a sequence of analog vibration signals containing multiple frequency components and instantaneous changes.

[0075] Furthermore, the acquired multi-channel analog vibration signals are transmitted to a data acquisition card. The card's resolution is set to 24 bits, and the sampling rate is set to 50 kHz. This parameter setting is based on the fact that the frequency of fault impact signals during equipment operation can typically reach several kilohertz, and a sampling rate of 50 kHz satisfies the Nyquist sampling theorem, ensuring that the signal is not distorted. During analog-to-digital conversion, the acquisition card quantizes and encodes the analog signal, converting it into a digital signal. Due to differences in circuitry between hardware channels and variations in transmission line length, there will be microsecond-level transmission delays in the digital signals of each channel, resulting in phase deviation. To eliminate this deviation, a reference signal injection method is used for phase calibration. A sinusoidal reference signal with a known frequency (e.g., 1 kHz) and phase is injected into each sensor channel. The response output of each channel to this reference signal is acquired, and the phase difference between the response signal of each channel and the original reference signal is calculated to obtain the phase offset of each channel. Based on the phase offset, phase compensation is performed on the vibration data of each channel. The compensation method involves shifting the vibration data of each channel along the time axis. The shift duration is calculated based on the phase offset and the reference signal frequency, ensuring that the data of each channel remains consistent in the time dimension. Through this compensation process, time-synchronized vibration data is obtained, eliminating phase deviations between channels.

[0076] Furthermore, wavelet packet decomposition was performed on the time-synchronized vibration data using the db8 wavelet basis. The db8 wavelet basis was chosen because of its good temporal localization characteristics and frequency domain resolution, which can effectively separate different frequency components in the signal. The number of decomposition layers was set to four. This number was determined because the main frequency range of the equipment vibration signal is 1 Hz to 10 kHz. Four layers of decomposition can evenly divide the frequency band into 16 sub-bands. The frequency range of each sub-band is evenly distributed according to the sampling rate and the number of decomposition layers, which can comprehensively cover the main frequency components of the signal while avoiding increased computational complexity due to too many decomposition layers. During the decomposition process, the signal was iteratively decomposed according to the Mallat algorithm for wavelet packet decomposition to obtain the wavelet packet coefficients of each decomposition node. The wavelet packet coefficients of each decomposition node were reconstructed to obtain the signal of each sub-band. The energy of each sub-band signal was calculated using either the sum of squares of the signal amplitude (for discrete signals) or the integral of the square of the amplitude (for continuous signals) to obtain the energy value of each sub-band. The energy values ​​of each sub-band are normalized by dividing the energy value of each sub-band by the sum of the energies of all sub-bands, resulting in the normalized energy percentage of each sub-band. This normalized energy percentage constitutes a multi-frequency component feature vector, reflecting the proportion of each frequency component in the signal. This achieves the separation of steady-state background noise from transient impact components. Specifically, the energy of the low-frequency sub-bands primarily reflects steady-state background noise and high-energy structural vibrations, while the energy of the high-frequency sub-bands primarily reflects transient impact components.

[0077] Finally, spatial coordinate association and encapsulation are performed. The three-dimensional spatial coordinates of each sensor on the device are pre-measured using a laser rangefinder with an accuracy of ±0.1 mm. The origin of the coordinates is set to the geometric center of the device or a fixed reference point, such as the lower left corner of the base. The three-dimensional coordinates of each sensor are bound to the multi-frequency component feature vector of the corresponding channel. This binding method involves constructing a structured data structure, with each data unit containing a coordinate field and a feature vector field. All bound structured data are integrated and encapsulated to form the first vibration dataset. The dataset is stored in JSON format for easy subsequent data reading and processing. This dataset not only contains the time and frequency domain information of the signal but also incorporates spatial distribution characteristics, providing data support for subsequent vibration source localization.

[0078] It is worth noting that the arrangement of the sensor array can be adjusted according to the structural differences of the specific equipment and the fault statistics. For areas with high failure rates, the sensor density can be increased. The sampling rate and the number of decomposition layers can be optimized according to the actual conditions such as the operating speed and vibration frequency range of the equipment. If the operating speed of the equipment is high and the vibration frequency range is wide, the sampling rate and the number of decomposition layers can be appropriately increased. During phase calibration, the frequency of the reference signal can be selected from the main characteristic frequencies of the equipment during operation to improve the targeting of the calibration. Other energy indicators, such as wavelet packet entropy and peak factor, can also be used as supplementary features to improve the comprehensiveness of the features when constructing multi-frequency component feature vectors.

[0079] The sensor spacing is primarily determined by the vibration transmission characteristics of the equipment structure and the accuracy requirements for fault location. For compact structures with complex vibration modes, a smaller spacing of 50-150 mm can be used to improve spatial resolution. For large racks and other parts with long vibration propagation distances, a larger spacing of 200-500 mm can be used to achieve economical coverage. The acquisition duration must cover at least several complete typical operating cycles of the equipment (such as multiple stacking actions) to ensure that the signal samples can fully reflect the periodic and transient behavior of the equipment. Typically, a duration of 10-60 seconds can cover most operating conditions. In specific applications, the minimum number of complete cycles can be calculated based on the equipment's operating cycle time (such as the number of actions per minute) and appropriately extended to ensure data redundancy.

[0080] For example, during the operation of the generator core stacking equipment, eight piezoelectric accelerometers are evenly arranged on the pressure head, frame, and base of the equipment. The sensor coordinates at the pressure head are (0,0,300) mm and (0,200,300) mm, the sensor coordinates on the frame side are (500,200,150) mm, (500,0,150) mm, and (500,-200,150) mm, and the sensor coordinates at the base are (0,0,0) mm, (200,0,0) mm, and (-200,0,0) mm. The acquisition of all channels is achieved at the same time through an external hardware trigger signal, capturing a multi-frequency component analog signal sequence containing low-frequency structural vibration (1-10 Hz), medium-frequency generator operation (50-100 Hz), and high-frequency impact (1-10 kHz). After analog-to-digital conversion at 24-bit resolution and a 50 kHz sampling rate, the data acquisition card injects a 1 kHz reference sinusoidal signal into each channel. The phase offset of each channel is measured to be between 0.01π and 0.05π. Based on this offset, the time delay compensation is calculated, and software compensation is used to achieve time synchronization of data from all channels. The original vibration data from time synchronization is decomposed into 4-level wavelet packets using the db8 wavelet basis, dividing the frequency band evenly into 16 sub-bands. The energy proportion of each sub-band node is calculated to form a multi-frequency component feature vector. The energy proportion of the high-frequency sub-band (2.5-10 kHz) is significantly higher in the fault state than in the normal state. The three-dimensional coordinates of each sensor are bound to the feature vector of the corresponding channel and encapsulated into a structured data packet to obtain the first vibration dataset. In subsequent analysis, this dataset can be used to locate the vibration source. When the high-frequency impact energy is concentrated in the feature vector corresponding to the sensor at the indenter, it can be quickly identified as localized overlapping loosening.

[0081] In step S12, the first vibration dataset needs to be decomposed and denoised by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components from the low-frequency effective components, and reconstructing the second vibration dataset, including:

[0082] The time-frequency distribution matrix of the first vibration dataset is calculated using the Wignerville distribution algorithm;

[0083] Based on the energy ridge line of the time-frequency distribution matrix, the instantaneous frequency and instantaneous amplitude are extracted, and the first vibration dataset is decomposed into a sequence of intrinsic mode components;

[0084] For the intrinsic mode component sequence, calculate the permutation entropy value. If the permutation entropy value exceeds the preset noise tolerance threshold, then the corresponding intrinsic mode component sequence is determined to be a high-frequency aliasing component; otherwise, it is determined to be a low-frequency effective component.

[0085] After performing soft threshold shrinkage on the high-frequency aliasing components, they are linearly superimposed and reconstructed with the low-frequency effective components to obtain the second vibration dataset.

[0086] It should be noted that the time-frequency distribution matrix is ​​the output of the Wignerville distribution algorithm. It is a two-dimensional matrix where rows represent time and columns represent frequency. The matrix element values ​​represent the energy density of the signal at the corresponding time and frequency points. The energy ridge is the path with the most concentrated energy in the time-frequency distribution matrix. The energy ridge reflects the main frequency variation trend of the signal, and the core time-frequency parameters of the signal can be extracted by tracing the energy ridge. Permutation entropy is an indicator used to measure the complexity and regularity of a signal. High-frequency aliasing components (noise) have poor regularity and high permutation entropy values, while low-frequency effective components (useful signals) have strong regularity and low permutation entropy values. Permutation entropy can be used to distinguish between effective signals and noise.

[0087] The preset noise tolerance threshold is a critical value used to determine the high-frequency aliasing component and the low-frequency effective component. Its value is based on the statistical results of the arrangement entropy of a large number of equipment vibration signals under normal operating conditions. By collecting multiple sets of normal operating equipment data, the arrangement entropy value is calculated, and 1.2-1.5 times the statistical average value is taken as the threshold. In this embodiment, it is set to 0.8. This threshold can be adjusted according to the actual situation such as equipment type and operating conditions.

[0088] Soft threshold shrinkage is a noise suppression method. The amplitude of high-frequency aliasing components is usually small. By setting a soft threshold, signal components with amplitudes below the threshold are shrunk to zero, and signal components with amplitudes above the threshold are appropriately adjusted to achieve noise filtering.

[0089] In this step, the first vibration dataset is input into the Wignerville distribution algorithm. The algorithm performs sliding window processing on the signal, calculating the spectral distribution of the signal within each time window, thus forming a time-frequency distribution matrix. Specifically, the length of the sliding window is first set, determined based on the signal's sampling rate and main frequency components. The specific length can be determined experimentally based on the duration and frequency bandwidth of the components of interest in the signal to be analyzed, typically a power of 2, generally between 1024 and 4096 sampling points, to ensure that the local frequency characteristics of the signal can be captured. For each signal segment within a time window, the integral of its product with its own delayed signal is calculated to obtain the energy density value at each frequency point corresponding to that time window. The calculation results for all time windows are arranged in chronological order to form a time-frequency distribution matrix. The rows of the matrix correspond to the time window number, the columns correspond to the frequency points, and the element values ​​are the energy density of that time window and frequency point.

[0090] Furthermore, based on the time-frequency distribution matrix, peak detection is used to extract energy ridges. Specifically, the frequency point with the highest energy density is found on the frequency axis corresponding to each time window, and the maximum energy frequency points of all time windows are connected to form an energy ridge. Based on the energy ridge, the instantaneous frequency and instantaneous amplitude corresponding to each time window are extracted. The instantaneous frequency is the frequency value of the energy ridge corresponding to that time window, and the instantaneous amplitude is calculated from the energy density value of that frequency point. Based on the extracted instantaneous frequency and instantaneous amplitude, empirical mode decomposition is used to decompose the first vibration dataset into a sequence of intrinsic modal components. The decomposition process follows these principles: each intrinsic modal component must satisfy the following condition: the number of extrema is equal to or differs from the number of zero-crossings by at most one over the entire signal length; at any given time, the average value of the upper envelope formed by local maxima and the lower envelope formed by local minima is zero. Through an iterative selection process, the first vibration dataset is decomposed into multiple intrinsic modal components that satisfy the above principles. These components are arranged in descending order of frequency, forming a sequence of intrinsic modal components.

[0091] Furthermore, the permutation entropy value is calculated for each component in the intrinsic mode component sequence. First, each intrinsic mode component is divided into multiple non-overlapping time windows. The window length is determined based on the frequency characteristics of the component, typically 20-100 sampling points, with no overlap between windows. For each signal segment within a time window, its data points are sorted in ascending order, and the sorted sequence numbers are recorded, forming a permutation pattern. The probability of the permutation pattern corresponding to each time window appearing in all windows is calculated. Then, the permutation entropy value is calculated based on the probability distribution of the permutation patterns. The magnitude of the permutation entropy reflects the regularity of the signal segments; the more concentrated the probability distribution, the smaller the permutation entropy, and the stronger the signal regularity; the more dispersed the probability distribution, the larger the permutation entropy, and the weaker the signal regularity. The permutation entropy value of each intrinsic mode component is compared with a preset noise tolerance threshold. If the permutation entropy value exceeds the threshold, the component is determined to be a high-frequency aliasing component; if the permutation entropy value does not exceed the threshold, the component is determined to be a low-frequency effective component.

[0092] Finally, for the intrinsic mode components identified as high-frequency aliasing components, soft thresholding is performed. First, the soft threshold is determined, based on the amplitude statistics of the high-frequency aliasing components, generally set to 1-2 times the standard deviation of the component's amplitude. Specifically, the standard deviation of the amplitude of all data points in the high-frequency aliasing component is calculated, and then multiplied by 1.5 to obtain the soft threshold. For data points in the high-frequency aliasing component with amplitudes less than the soft threshold, their amplitudes are reduced to zero; for data points with amplitudes greater than or equal to the soft threshold, their amplitudes are subtracted from the soft threshold (keeping the sign unchanged) to achieve noise suppression. The high-frequency aliasing components after soft thresholding are then linearly superimposed with all low-frequency effective components for reconstruction. The reconstruction method involves directly adding the time point data corresponding to each component to obtain the reconstructed signal, which is the second vibration dataset.

[0093] For example, for the first vibration dataset, the Wignerville distribution algorithm was used to calculate the time-frequency distribution matrix. With a sliding window length of 2048 sampling points, the resulting time-frequency distribution matrix clearly showed the frequency changes of the signal at different times. In the early stages of equipment operation, the signal was dominated by low-frequency vibrations around 10 Hz. During frequent stacking actions, the frequency briefly jumped to above 500 Hz, forming an energy concentration region. Based on the energy ridges of the time-frequency distribution matrix, instantaneous frequency and instantaneous amplitude were extracted. In one stacking impact action, the instantaneous frequency rapidly increased from 200 Hz to 800 Hz, and the amplitude also increased accordingly. The first vibration dataset was decomposed into five intrinsic modal component sequences using empirical mode decomposition. The fourth component's frequency was concentrated above 1 kHz, clearly distinguishing it from the other components. The permutation entropy value of each intrinsic modal component was calculated. With a preset noise tolerance threshold of 0.8, a certain modal component with a permutation entropy value of 0.9, exceeding the threshold, was identified as a high-frequency aliasing component; the permutation entropy values ​​of the remaining components were between 0.5 and 0.7, and were identified as low-frequency effective components. The high-frequency aliasing component is subjected to soft threshold shrinkage. The standard deviation of the amplitude of this component is calculated to be 0.02. The soft threshold is set to 0.03. Data points with amplitudes less than 0.03 are shrunk to zero, and data points with amplitudes greater than or equal to 0.03 are subtracted by 0.03. Then, it is linearly superimposed with the low-frequency effective component to reconstruct the second vibration dataset. This dataset filters out most of the background noise and retains the main features of the signal.

[0094] In step S13, for the second vibration dataset, a bandpass filter needs to be constructed using the spectral kurtosis map for filtering, and then reconstructed and enhanced using a trained complete dictionary matrix to obtain an enhanced signal feature set, including:

[0095] The spectral kurtosis plot of the second vibration dataset is obtained to determine the resonance demodulation frequency band. A bandpass filter is constructed to filter the second vibration dataset to obtain a narrowband filtered signal.

[0096] Based on the overcomplete dictionary matrix trained by the narrowband filtered signal, the sparse coefficient vector of the narrowband filtered signal under the overcomplete dictionary matrix is ​​calculated.

[0097] If the sparse coefficient vector satisfies a preset sparsity threshold, it is retained and linearly combined with the overcomplete dictionary matrix for reconstruction. A Hilbert transform is then performed on the reconstructed signal to obtain an enhanced signal feature set.

[0098] It should be noted that a spectral kurtosis plot is a graph reflecting the kurtosis values ​​of a signal at different frequencies. Kurtosis is a statistical measure of the steepness of the signal amplitude distribution. Fault impulse signals often cause a significant increase in kurtosis values ​​at the corresponding frequencies. The spectral kurtosis plot can be used to quickly locate the resonant demodulation band containing fault characteristics. The resonant demodulation band refers to the frequency range with high kurtosis values ​​in the spectral kurtosis plot. This band contains the periodic impulse components generated by equipment faults. Filtering this band can extract fault-related signals. An overcomplete dictionary matrix is ​​a matrix with a much larger number of columns than rows. Each column of the matrix is ​​called an atom, and each atom represents a basic signal pattern. A fault impulse signal can be represented as a linear combination of several atoms in the overcomplete dictionary matrix. By training the overcomplete dictionary matrix, an atom library that highly matches the fault signal can be obtained.

[0099] The preset sparsity threshold is a critical value used to determine whether a sparse coefficient vector is effective. Sparsity refers to the proportion of non-zero elements in a sparse coefficient vector. Fault impact signals exhibit sparsity, meaning the proportion of non-zero elements in their corresponding sparse coefficient vectors is extremely low. By collecting multiple sets of fault signal samples, the sparsity of their sparse coefficient vectors is calculated, and the statistical minimum value is taken as the threshold, typically set between 0.05 and 0.2. This threshold can be adjusted according to the sparsity of the fault signal. The Hilbert transform is a signal processing transform that can extract the envelope information of a signal. The envelope spectrum clearly displays the fault characteristic frequencies and their harmonics, thereby enhancing the fault characteristics.

[0100] In this step, spectral kurtosis is first calculated on the second vibration dataset. The calculation involves dividing the dataset into multiple time windows, the window length of which is determined based on the signal sampling rate and fault impact period, typically ranging from 512 to 2048 sampling points. Windows overlap by 50% to ensure the continuity of the calculation results. For each signal segment within a time window, its spectrum is calculated. Then, for each frequency point in the spectrum, the kurtosis value of that frequency point across all time windows is calculated. The kurtosis value is calculated as the ratio of the fourth-order central moment of the signal amplitude to the square of the second-order central moment. The kurtosis values ​​corresponding to each frequency point are arranged in frequency order to create a spectral kurtosis plot. In the spectral kurtosis plot, the frequency range where the kurtosis value is significantly higher than other regions is identified; this range is the resonant demodulation band. The upper and lower limits of the resonant demodulation band are determined: the lower limit is the frequency point where the kurtosis value begins to increase significantly, and the upper limit is the frequency point where the kurtosis value begins to decrease significantly, ensuring that the band completely encompasses the high-kurtosis value region.

[0101] Furthermore, based on the determined resonant demodulation frequency band, an infinite impulse response (IIR) bandpass filter is constructed. The filter type chosen is a Chebyshev Type I filter, characterized by equal-ripple amplitude-frequency characteristics in the passband and monotonic amplitude-frequency characteristics in the stopband, effectively suppressing interference signals outside the passband. The filter design parameters include the passband frequency range, passband ripple, and stopband attenuation. The passband frequency range is set to the upper and lower limits of the resonant demodulation frequency band. The passband ripple is set to 0.1 dB to ensure minimal amplitude distortion within the passband. The stopband attenuation is set to 40 dB to ensure effective suppression of interference signals within the stopband. The second vibration dataset is input into the constructed bandpass filter. The filter processes the input signal, allowing signals within the resonant demodulation frequency band to pass while suppressing low-frequency operating noise and high-frequency random interference outside the passband, outputting a narrowband filtered signal. This signal waveform typically exhibits a distinct periodic envelope.

[0102] Furthermore, based on the narrowband filtered signal, an overcomplete dictionary matrix is ​​trained using the K-SVD algorithm. The training process first initializes the overcomplete dictionary matrix by randomly selecting several signal segments from the narrowband filtered signal as initial atoms. The length of each atom is determined by the duration of the fault impact signal, typically 32-128 sampling points. The size of the dictionary matrix is ​​set to the atom length multiplied by the number of atoms, usually 2-4 times the atom length, ensuring the overcompleteness of the dictionary matrix. Then, the orthogonal matching pursuit algorithm is used to perform sparse decomposition on the narrowband filtered signal, obtaining a sparse coefficient vector. The sparse decomposition process iteratively selects the atom that best matches the signal residual, stores its corresponding coefficient in the sparse coefficient vector, and updates the signal residual simultaneously until the residual is less than a residual threshold or the maximum number of iterations is reached. Based on the sparse coefficient vector and the signal residual, the singular value decomposition algorithm is used to update the atoms in the overcomplete dictionary matrix, updating one atom at a time, optimizing the atom shape by minimizing the signal residual. Repeat the sparse decomposition and atom update process until the update amount of the dictionary matrix is ​​less than the set update amount requirement or the maximum number of training iterations is reached, resulting in a trained overcomplete dictionary matrix. Using the trained overcomplete dictionary matrix, the narrowband filtered signal is again sparsely decomposed using the orthogonal matching pursuit algorithm to obtain a sparse coefficient vector. Most elements in this vector are zero, with only a few non-zero elements. The non-zero elements correspond to the atoms in the dictionary that match the fault impact signal.

[0103] Finally, the sparsity of the obtained sparse coefficient vector is compared with a preset sparsity threshold. The sparsity is calculated by dividing the number of non-zero elements in the sparse coefficient vector by the total length of the vector. If the sparsity is less than or equal to the preset sparsity threshold, the sparse coefficient vector is retained; if the sparsity is greater than the preset sparsity threshold, the sparse coefficient vector is thresholded by setting non-zero elements with smaller amplitudes to zero until the sparsity meets the threshold requirement. The retained sparse coefficient vector is reconstructed by linear combination with an overcomplete dictionary matrix. The reconstruction method involves multiplying each non-zero element in the sparse coefficient vector by the corresponding atom in the overcomplete dictionary matrix, and then summing all the product results to obtain the reconstructed signal. A Hilbert transform is performed on the reconstructed signal, which involves calculating the convolution of the reconstructed signal with 1 / t to obtain the analytic signal. The envelope signal is extracted from the analytical signal. The amplitude change of the envelope signal can reflect the intensity change of the fault impact. The envelope signal is subjected to spectral analysis to obtain the envelope spectrum. Significant peaks will appear at the fault characteristic frequency and its harmonics in the envelope spectrum. The envelope signal and the envelope spectrum together constitute the enhanced signal feature set.

[0104] It is worth noting that the window length and overlap rate for spectral kurtosis calculation can be adjusted according to the characteristics of the signal. The higher the fault impulse frequency in the signal, the smaller the window length can be. The type of bandpass filter can also be a finite impulse response (FIR) filter, selected according to the actual filtering performance requirements. When the K-SVD algorithm is trained with a complete dictionary matrix, the length and number of atoms can be optimized according to the characteristics of the fault impulse signal. The longer the duration of the fault impulse signal, the larger the atom length should be. The more atoms there are, the stronger the representation capability of the dictionary matrix, but the computational complexity will also increase accordingly. The maximum number of iterations and the residual threshold of the orthogonal matching pursuit algorithm can be adjusted according to the sparsity of the signal to ensure accurate decomposition of the signal.

[0105] For example, spectral kurtosis was calculated on the second vibration dataset. A time window of 1024 sampling points was set, with a window overlap of 50%. A spectral kurtosis plot was drawn, revealing a kurtosis value of 6.8 in the band near the center frequency of 2.5 kHz, while other regions had a kurtosis of only around 3. This band was identified as the resonant demodulation band, with a lower limit of 2 kHz and an upper limit of 3 kHz. A Chebyshev type I IIR bandpass filter was constructed, with a passband frequency range of 2-3 kHz, a passband ripple of 0.1 dB, and a stopband attenuation of 40 dB. The second vibration dataset was input into this filter, and a narrowband filtered signal was output. The filtered signal waveform exhibited a clear periodic envelope. Based on the narrowband filtered signal, an overcomplete dictionary matrix was trained using the K-SVD algorithm. The initial atom length was 64 sampling points, and the number of atoms was 256. After 100 iterations, the overcomplete dictionary matrix was obtained. Some atoms in this matrix exhibited a rapidly decaying damped oscillation pattern, which highly matched the instantaneous impact in the stacking equipment. The orthogonal matching pursuit algorithm was used to perform sparse decomposition on the narrowband filtered signal. With a maximum iteration count of 50 and a residual threshold of 0.001, a sparse coefficient vector with only 15 non-zero coefficients was obtained, resulting in a sparsity of 15 / 256 ≈ 0.058. The preset sparsity threshold was 0.06, and this sparse coefficient vector met the requirements. The sparse coefficient vector is linearly combined with the overcomplete dictionary matrix to reconstruct the signal. The reconstructed signal is then subjected to Hilbert transform to extract the envelope signal and perform spectral analysis. Significant peaks appear in the envelope spectrum at 125 Hz (inner ring fault frequency) and its harmonics, with the amplitude increased to more than 3 times that of the original signal. The envelope signal and envelope spectrum are used as the feature set of the enhanced signal.

[0106] In step S14, the enhanced signal feature set needs to be pattern matched with a preset reference frequency template library, and the matching results need to be quantitatively analyzed to determine the distribution range of potential fault signals, including:

[0107] Calculate the matching degree between the enhanced signal feature set and each template in the preset reference frequency template library, and generate a multidimensional similarity matrix;

[0108] Filter out local maxima points in the multidimensional similarity matrix whose amplitude exceeds a preset matching threshold, and identify the corresponding specific frequency patterns;

[0109] Calculate the energy entropy value of the signal segment corresponding to the specific frequency mode to obtain the quantization evaluation sequence;

[0110] A fault probability distribution histogram is constructed based on the quantitative evaluation sequence to determine the distribution range of potential fault signals.

[0111] It should be noted that the preset reference frequency template library is a database storing standard characteristic spectra under various known fault modes. Different types of faults will generate specific frequency characteristic spectra. By matching these spectra with the standard characteristic spectra, the fault type in the current signal can be identified. The template library adopts a hierarchical storage architecture, divided into a fault type layer, a characteristic spectrum layer, and a parameter layer. The fault type layer contains all possible fault types that the equipment may experience, such as inner ring defects, outer ring defects, and overlapping loosening. The characteristic spectrum layer stores the standard characteristic spectra corresponding to each fault type, including time-domain and frequency-domain characteristic spectra. The parameter layer stores the relevant parameters of each characteristic spectrum, such as characteristic frequency and amplitude range. The data sources for the template library include two parts: first, simulated data of different fault types obtained through simulation experiments, where vibration signals under fault conditions are collected by simulating equipment faults, and characteristic spectra are extracted as standard templates; second, real data obtained from actual fault cases in the field, collecting actual fault data that occurs during equipment operation, which is then preprocessed and feature extracted before being added to the template library. The template library's update mechanism combines periodic and real-time updates. The periodic update cycle is 6 months, during which statistical analysis is performed on the data in the template library to remove outdated or inaccurate templates and add new fault type templates. Real-time updates mean that when a new fault type is detected and verified, the feature spectrum of that fault type is automatically added to the template library to ensure the integrity and timeliness of the template library.

[0112] The preset matching threshold is a critical value used to filter valid matching results. Based on the statistical results of matching degree of a large amount of normal operation data and fault data, the matching degree is calculated by collecting multiple sets of normal operation data and fault data, and the maximum value of the matching degree of normal data is taken. It is generally set to 0.8. This threshold can be adjusted according to the update of the template library and the operating conditions of the equipment.

[0113] Energy entropy is an indicator of the uniformity of signal energy distribution. Fault signals often have energy concentrated in a specific frequency range, resulting in low energy entropy values. Energy entropy can be used to further verify the validity of matching results. A quantitative evaluation sequence is a sequence composed of the energy entropy values ​​of signal segments corresponding to specific frequency patterns. This sequence is established because it can quantitatively reflect the energy distribution characteristics of fault signals, providing data support for the construction of fault probability distribution histograms. A fault probability distribution histogram is a graphical representation of the distribution range of potential fault signals. Through the histogram, the distribution probability of fault signals in different frequency ranges can be clearly seen, identifying the main concentrated areas of potential faults.

[0114] In this step, the envelope spectrum of the enhanced signal feature set is first compared with each standard feature spectrum in the preset reference frequency template library to calculate the matching degree. The matching degree is calculated using the cosine similarity method, which can effectively measure the similarity between two vectors. Specifically, the envelope spectrum and standard feature spectrum of the enhanced signal are first normalized to the same frequency and amplitude range to ensure comparability. Then, the normalized envelope spectrum and standard feature spectrum are treated as two vectors, and the dot product of the two vectors is divided by the product of their magnitudes to obtain the matching degree value. The matching degree value ranges from 0 to 1, with a higher degree of matching closer to 1 and a lower degree of matching closer to 0. The matching degree values ​​of the enhanced signal and all standard feature spectra in the template library are arranged in template order to generate a multidimensional similarity matrix. The rows of the matrix correspond to the frequency points of the enhanced signal, the columns correspond to the template indices in the template library, and the element values ​​are the matching degree between that frequency point and the template.

[0115] It should be noted that the calculation of matching degree is not limited to cosine similarity. Those skilled in the art may also choose other similarity measurement methods such as Pearson correlation coefficient and reciprocal of Euclidean distance according to the actual situation.

[0116] Furthermore, a sliding window peak detection method is used to filter local maxima in the multidimensional similarity matrix. The sliding window size is set to 3×3 to ensure accurate identification of local peaks. For each element in the matrix, it is compared with the other 8 elements in the window. If the value of the element is greater than the values ​​of all other elements in the window and exceeds a preset matching threshold, the element is determined to be a local maximum. The template number and frequency range corresponding to each local maximum are recorded. Based on the template number, the corresponding standard feature spectrum is searched in a preset reference frequency template library. The frequency feature pattern represented by this standard feature spectrum is the specific frequency pattern. At the same time, the fault type corresponding to the specific frequency pattern is determined, such as inner ring defects or lamination loosening.

[0117] Furthermore, for the identified specific frequency patterns, signal segments within the corresponding frequency range and time interval are extracted from the enhanced signal feature set. The time interval of the signal segment is determined based on the periodic characteristics of the specific frequency pattern to ensure that it includes multiple fault impact cycles. The extracted signal segments are divided into multiple non-overlapping time windows. The window length is determined based on the length of the signal segment and the fault impact cycle, generally 10-50 fault impact cycles, with no overlap between windows. For the signal segments within each time window, its energy distribution is calculated. The energy distribution is calculated by dividing the spectrum of the signal segment into multiple frequency sub-bands and calculating the energy proportion of each sub-band. Based on the energy proportion distribution of each time window, the energy entropy value is calculated. The energy entropy is calculated by taking the logarithm of the energy proportion of each sub-band, multiplying it by the corresponding energy proportion, summing the results, and taking the negative value. The smaller the energy entropy value, the more concentrated the energy distribution and the more obvious the fault characteristics. The energy entropy values ​​of all time windows are arranged in chronological order to form a quantitative evaluation sequence.

[0118] Finally, a fault probability distribution histogram is constructed, and the distribution range of potential fault signals is determined: Statistical analysis is performed on the energy entropy values ​​in the quantitative evaluation sequence to determine the range of energy entropy values. This range is divided into multiple equally spaced intervals, the number of which is determined based on the distribution characteristics of the energy entropy values, typically 10-20 intervals. The frequency of energy entropy values ​​within each interval is counted, and the proportion of each interval's frequency to the total frequency is calculated; this proportion represents the fault probability. A fault probability distribution histogram is plotted with the energy entropy value intervals on the x-axis and the fault probability on the y-axis. In the histogram, the energy entropy value intervals with higher fault probabilities are identified; the signal segments corresponding to these intervals are the potential fault signals. Based on the frequency range corresponding to the potential fault signals, the distribution range of the potential fault signals is determined. The lower limit frequency of the distribution range is the lowest frequency point of a specific frequency pattern, and the upper limit frequency is the highest frequency point of a specific frequency pattern, ensuring that the range completely includes all frequency components of the potential fault signals.

[0119] It is worth noting that the matching degree can also be calculated using other similarity calculation methods such as the correlation coefficient method, and the choice should be made based on the actual matching effect; the window size of the sliding window peak detection can be adjusted according to the density of the multidimensional similarity matrix, and the denser the matrix elements, the larger the window size can be; the truncation length of the signal segment can be optimized according to the duration of the fault impact to ensure that the fault characteristics can be fully reflected; the number of frequency sub-bands for energy entropy calculation can be adjusted according to the frequency range of the signal, and the wider the frequency range, the more sub-bands can be.

[0120] For example, the envelope spectrum in the enhanced signal feature set is compared with the standard feature spectrum in a preset reference frequency template library to calculate the matching degree. The template library contains standard feature spectra for 10 fault types, including inner ring defects, outer ring defects, and overlapping loosening. Using the cosine similarity method, the similarity between the enhanced signal and the inner ring fault template at 125 Hz and its harmonics reaches 0.92, while the similarity with other templates is only about 0.45, generating a multidimensional similarity matrix. A 3×3 sliding window peak detection method is used to filter local maxima points. A preset matching threshold of 0.8 is set, identifying three significant local maxima points at 125 Hz, 250 Hz, and 375 Hz. The corresponding templates are the inner ring fault templates, and the specific frequency modes are the characteristic frequencies of the inner ring fault and their harmonics. A segment of the enhanced signal within the 100-400 Hz frequency range was extracted and divided into 20 time windows, each containing 20 fault impact cycles. The energy entropy value of each window was calculated to obtain a quantitative evaluation sequence. The lowest energy entropy value in the sequence was 2.1, while the energy entropy value of the random noise segment reached above 4.5. The energy entropy value range was divided into 10 intervals, and the fault probability of each interval was statistically analyzed. A fault probability distribution histogram was plotted. The histogram showed that the fault probability in the low entropy region (below 2.5) accounted for 65%, corresponding to the high-probability fault interval. The distribution range of the potential fault signal was determined to be 100-400 Hz, mainly concentrated in the inner ring impact mode.

[0121] In step S15, the statistical dispersion of the distribution range of the potential fault signal needs to be calculated. If it exceeds a preset dispersion threshold, the corresponding non-stationary signal segment is extracted and deep feature extraction is performed through multi-scale wavelet packet decomposition and singular value decomposition to obtain a fault feature dataset, including:

[0122] Calculate the statistical dispersion of the potential fault signal distribution range. If the statistical dispersion exceeds a preset dispersion threshold, then extract the corresponding non-stationary signal segment.

[0123] Multi-scale wavelet packet decomposition is performed on the non-stationary signal segment to reconstruct a high-resolution time-frequency energy spectrum;

[0124] Singular value decomposition is performed on the high-resolution time-frequency energy spectrum to extract the non-zero singular value sequence and construct the singular value feature vector.

[0125] The singular value eigenvectors are dimensionality reduced by kernel principal component analysis to generate independent eigenvectors, which are then structured and arranged to obtain a fault feature dataset.

[0126] It should be noted that statistical dispersion is an indicator that measures the degree of concentration or dispersion of potential fault signals. It reflects the stationarity of the signal; the greater the dispersion, the more unstable the signal and the more likely it is to contain fault characteristics. Statistical dispersion is calculated using standard deviation or variance. In this embodiment, standard deviation is used as the measure of statistical dispersion.

[0127] The preset dispersion threshold is a critical value used to determine whether potential fault signals need further in-depth feature extraction. Based on the statistical dispersion results of the distribution range of potential fault signals under a large number of normal operating conditions, the statistical dispersion of the distribution range of potential fault signals is calculated by collecting multiple sets of normal operating data of equipment. The threshold is taken as 1.5-2 times the statistical average value. In this embodiment, it is set to 10. This threshold can be adjusted according to the actual situation such as equipment type and operating conditions.

[0128] Singular value decomposition (SVD) is a matrix factorization method. High-resolution time-frequency energy spectra can be represented in matrix form. SVD can extract the main characteristic components of the matrix, and the sequence of non-zero singular values ​​reflects the core features of the matrix. Singular value eigenvectors (SVVs) are vectors composed of sequences of non-zero singular values ​​and are important representations of fault characteristics. Kernel principal component analysis (KPCA) is a nonlinear dimensionality reduction algorithm. Singular value eigenvectors may have high dimensionality and a lot of redundant information. KPCA can remove redundant information, retain key features, and generate independent eigenvectors with lower dimensionality.

[0129] In this step, the frequency range of the potential fault signal distribution is first determined, and all signal data points within this frequency range are extracted. The statistical dispersion (standard deviation) of these data points is calculated. The calculation process involves first calculating the mean of all data points, then calculating the difference between each data point and the mean, squared the differences, summing them, dividing by the number of data points, and finally taking the square root to obtain the statistical dispersion value. The calculated statistical dispersion value is compared with a preset dispersion threshold. If the statistical dispersion value exceeds the threshold, it indicates that the signal distribution is relatively dispersed and exhibits non-stationary characteristics, requiring the extraction of corresponding non-stationary signal segments. The method for extracting non-stationary signal segments is based on the time distribution of the potential fault signal, extracting signals within the time period where the statistical dispersion exceeds the threshold. The length of the segment is determined based on the duration of the non-stationary signal, generally 1-5 seconds, to ensure that the non-stationary characteristics are fully contained.

[0130] Furthermore, the extracted non-stationary signal segments are subjected to multi-scale wavelet packet decomposition using the db8 wavelet basis. The decomposition level is set to 5 levels, determined based on the frequency range of the non-stationary signal segments and the detailed features to be captured. Five levels divide the frequency band into 32 sub-bands, enabling more detailed capture of signal features at different scales. The decomposition process follows the Mallat algorithm for wavelet packet decomposition, iteratively decomposing the non-stationary signal segments to obtain wavelet packet coefficients for each scale and frequency band. The wavelet packet coefficients for each scale and frequency band are then reconstructed to obtain the sub-band signals at each scale. The energy of the sub-band signals at each scale is calculated as the sum of the squares of the sub-band signal amplitudes. Based on the energy values ​​at various scales and frequency bands, a high-resolution time-frequency energy spectrum is constructed. The horizontal axis of the time-frequency energy spectrum is time, and the vertical axis is frequency. The color depth represents the energy density. The higher the energy density, the darker the color. This spectrum can clearly show the energy distribution of the signal at different times and frequencies. Here, the energy refers to the energy density on the time-frequency plane, which is different from the energy entropy index used in step S14 to measure the uniformity of energy distribution.

[0131] Furthermore, the high-resolution time-frequency energy spectrum is transformed into a matrix form, where rows correspond to time points, columns to frequency points, and element values ​​represent the energy density at that time and frequency point. Singular value decomposition (SVD) is then performed on this matrix, which decomposes it into the product of three matrices: a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix. The diagonal elements of the singular value diagonal matrix are the singular values, arranged in descending order. Non-zero singular values ​​are extracted from the singular value diagonal matrix to form a sequence of non-zero singular values. The number of non-zero singular values ​​is typically much smaller than the number of rows and columns of the matrix, effectively condensing the core features of the matrix. The sequence of non-zero singular values ​​is then arranged sequentially to construct a singular value eigenvector. This vector's dimension is equal to the number of non-zero singular values, reflecting the main characteristics of the high-resolution time-frequency energy spectrum.

[0132] Finally, kernel principal component analysis (KPC) is performed for dimensionality reduction and fault feature dataset construction. The Gaussian kernel function is chosen as the kernel function for KPC because it has good nonlinear mapping capabilities, enabling it to map high-dimensional singular value eigenvectors to a low-dimensional feature space. The parameters (kernel width) of the Gaussian kernel function are determined based on the distribution characteristics of the singular value eigenvectors. The optimal parameter values ​​are determined using cross-validation, typically ranging from 0.1 to 10. The singular value eigenvectors are then input into the KPC algorithm. The algorithm first calculates the kernel matrix, whose elements are the kernel function values ​​between samples in the singular value eigenvectors. The kernel matrix is ​​then centered to eliminate the influence of the mean. Next, eigenvalue decomposition is performed on the centered kernel matrix to obtain eigenvalues ​​and corresponding eigenvectors. The primary eigenvectors are selected based on the magnitude of the eigenvalues, generally chosen as those with a cumulative contribution rate of 85% or higher. Singular value eigenvectors are projected onto selected principal eigenvectors to obtain dimensionality-reduced independent eigenvectors. The dimensionality of these independent eigenvectors is significantly lower than that of the singular value eigenvectors, removing redundant information and retaining key features. The independent eigenvectors are then structured according to time order and frequency ranges. Each data unit contains a time label, a frequency label, and the corresponding eigenvector value, forming a fault feature dataset. This dataset is stored in CSV format for easy training and use by subsequent fault diagnosis models.

[0133] It is worth noting that other indicators such as variance and range can also be used to calculate statistical dispersion, and the choice should be made according to the actual signal characteristics. The wavelet basis and the number of decomposition levels in multi-scale wavelet packet decomposition can be optimized according to the characteristics of non-stationary signal segments. Different wavelet bases have different advantages in processing different types of signals. In the process of singular value decomposition, the number of non-zero singular values ​​can be determined according to the decay rate of singular values. Usually, the first 10-20 singular values ​​can reflect the main characteristics of the matrix. The kernel function of kernel principal component analysis can also be selected from other types such as polynomial kernel function and sigmoid kernel function, and the choice should be made according to the nonlinearity of the singular value eigenvectors.

[0134] For example, the potential fault signal distribution ranges from 100 to 400 Hz. Signal data points within this frequency range are extracted, and their statistical dispersion (standard deviation) is calculated. The dispersion value is 15.2, significantly exceeding the preset dispersion threshold of 10, indicating a relatively dispersed signal distribution and non-stationary characteristics. A corresponding non-stationary signal segment is extracted, with a duration of 2 seconds. This non-stationary signal segment is decomposed into 32 sub-bands using a 5-level multi-scale wavelet packet decomposition based on the db8 wavelet basis. The wavelet packet coefficients of each sub-band are reconstructed, and the energy of each sub-band is calculated to construct a high-resolution time-frequency energy spectrum. A significant energy clustering phenomenon appears near 150 Hz in the spectrum. The high-resolution time-frequency energy spectrum is transformed into a 1000×300 matrix (time point × frequency point). Singular value decomposition is performed on this matrix, yielding 60 non-zero singular values. The first 10 singular values ​​are significantly higher than the others. The sequence of the first 10 non-zero singular values ​​is extracted to construct a singular value feature vector. A Gaussian kernel function was selected for kernel principal component analysis (KPCA). The kernel width was determined to be 2 using cross-validation. Singular value eigenvectors (SVVs) were input into the KPCA algorithm to calculate the kernel matrix and perform centering. Eigenvalue decomposition was then performed on the centered kernel matrix to obtain eigenvalues ​​and eigenvectors. The cumulative contribution of the first five eigenvalues ​​reached 90%, and these five eigenvectors were selected as the principal eigenvectors. The SVVs were projected onto these five principal eigenvectors to obtain independent eigenvectors of dimension 5. These independent eigenvectors were then structured according to time order and frequency range to form a fault feature dataset. The dataset contained eigenvalues ​​for characteristic frequencies such as 125 Hz and 250 Hz.

[0135] In step S16, a hyperplane decision model needs to be constructed based on the fault feature dataset. By comparing and analyzing the local evolution trends of the features, a fault warning result is output, including:

[0136] Based on the independent feature vectors in the fault feature dataset, a hyperplane decision model is constructed.

[0137] The feature deviation of the independent feature vector in the hyperplane decision model is calculated in real time, and an evolution time series is constructed based on the feature deviation at consecutive time points;

[0138] The local evolution trend of the evolution time series is extracted, the rate of change and the anomaly confidence are calculated, and the early failure risk index is obtained by mapping.

[0139] If the early failure risk index exceeds the preset warning threshold, a failure warning result containing a risk level identifier will be output.

[0140] It should be noted that the hyperplane decision model is a model that constructs an optimal hyperplane in a high-dimensional feature space to separate samples of different categories. The independent feature vectors in the fault feature dataset can distinguish between the normal and fault states of equipment. By constructing a hyperplane decision model, the classification of equipment states can be achieved. The hyperplane decision model is constructed using support vector machines, which have the advantage of finding the optimal hyperplane in a high-dimensional feature space and exhibiting good generalization ability. Feature deviation refers to the distance between the real-time independent feature vectors and the normal state sample clustering region in the hyperplane decision model. Feature deviation quantifies the degree of difference between real-time features and normal features; the greater the difference, the greater the feature deviation, and the higher the probability of equipment failure.

[0141] The early failure risk index is a comprehensive index derived by mapping the rate of change and the confidence level of anomalies. It is established because this index can comprehensively reflect the degree of equipment failure risk, providing an intuitive basis for early warning decisions. The preset early warning threshold is a critical value used to determine whether an early warning result needs to be output. Its value is based on the statistical results of the early failure risk index from a large amount of normal operation data and failure data. By collecting multiple sets of normal operation data and failure data, their early failure risk indices are calculated, and the maximum value of the risk index of the normal data is taken as the threshold. In this embodiment, it is set to 0.6. This threshold can be adjusted according to the importance of the equipment and maintenance requirements.

[0142] In this step, independent feature vector samples of the equipment's normal operating state and known fault state are first collected to form a training dataset. The ratio of normal samples to fault samples in the training dataset is set to 1:1 to ensure good classification performance of the model. The training dataset is input into a support vector machine (SVM), and a radial basis function is selected as the kernel function. The parameters of the kernel function are determined by a grid search method, with a search range of 0.01-100. The optimal parameters are selected through 5-fold cross-validation. The SVM searches for the optimal hyperplane that maximizes the margin between normal and fault samples on both sides of the hyperplane. The equation of the hyperplane is determined by the support vectors, which are the sample points in the training dataset that are closest to the hyperplane. After the model is trained, the classification accuracy of the model is verified using a test dataset consisting of normal and fault samples that were not used in training. If the classification accuracy reaches 90% or higher, the model training is considered complete; otherwise, the parameters are adjusted and retrained until the accuracy requirement is met, resulting in the final hyperplane decision model.

[0143] Furthermore, vibration signals from the generator core stacking equipment are acquired in real time, and real-time independent feature vectors are extracted according to the processing flow of steps S11-S15. These real-time independent feature vectors are then input into the hyperplane decision model to calculate the feature deviation. The feature deviation is calculated as the distance from the real-time independent feature vector to the hyperplane, using Euclidean distance. A larger distance indicates a greater difference between the real-time feature and the normal feature. A calculation period for the feature deviation is set, determined based on the equipment's operating speed and fault development speed, typically 0.1-1 seconds, to ensure timely capture of feature changes. The feature deviations calculated at consecutive time points are arranged chronologically to construct an evolution time series. The series length is determined based on actual monitoring needs, typically 100-1000 data points, reflecting the trend of feature deviation changes over a period of time.

[0144] Furthermore, a sliding window method is employed to extract the local evolution trend of the evolution time series. The length of the sliding window is determined based on the length of the evolution time series and the fault development cycle, typically 20-50 data points. The window overlap rate is set to 50% to ensure the continuity of trend extraction. Linear fitting is performed on the feature deviation data within each sliding window to obtain the slope of the fitted line. This slope represents the rate of change of the local evolution trend: a positive slope indicates increasing feature deviation and deteriorating equipment condition; a negative slope indicates decreasing feature deviation and improving equipment condition; and a slope close to zero indicates stable feature deviation and normal equipment condition. The residuals between the feature deviation data within each sliding window and the fitted line are calculated. Anomaly confidence is calculated based on the distribution of the residuals; the smaller the residual, the higher the anomaly confidence, indicating stronger reliability of the local evolution trend. The anomaly confidence is determined by the ratio of the standard deviation of the residuals to a preset residual threshold; a smaller standard deviation results in a higher anomaly confidence, with a value ranging from 0 to 1. The rate of change and anomaly confidence are normalized to the range of 0-1, and then a weighted summation is used to map the early failure risk index. The weights are assigned based on the degree of influence of the rate of change and anomaly confidence on failure risk. The weight of the rate of change is set to 0.6, and the weight of the anomaly confidence is set to 0.4. The weights can be adjusted according to actual failure statistics. If the anomaly confidence has a greater impact on failure judgment, its weight can be appropriately increased. The weights are empirical values ​​derived from the statistical analysis of historical failure data, reflecting that trend changes are more indicative of early failure risk than single-point confidence. In actual deployment, the weights can be fine-tuned by using optimization methods such as grid search and cross-validation based on prior knowledge of different devices and different failure modes.

[0145] Finally, the fault warning result is output. The calculated early fault risk index is compared with the preset warning threshold. If the early fault risk index does not exceed the preset warning threshold, the equipment is considered to be operating normally, and no warning result is output. If the early fault risk index exceeds the preset warning threshold, the risk level is divided according to the magnitude of the early fault risk index: a risk index between 0.6 and 0.8 is considered medium risk, and a risk index above 0.8 is considered high risk. Combined with the specific frequency pattern identified in step S14, possible fault types are determined, such as inner ring defects and loosening of stacking. A fault warning result containing risk level identification, possible fault type, and characteristic deviation change trend is generated. The output methods of the warning result include local display and remote push. Local display shows the warning information in real time on the device's screen, while remote push sends the warning information to the mobile terminal of maintenance personnel via the network, so that maintenance personnel can receive the warning prompts in a timely manner and take corresponding maintenance measures.

[0146] It is worth noting that the construction of the hyperplane decision model can also employ other classification algorithms such as logistic regression and decision trees, selected based on the characteristics of the fault feature dataset; the calculation of feature deviation can also use other distance metrics such as Mahalanobis distance, which can consider the correlation between features and has better performance in some cases; the length and overlap rate of the sliding window can be adjusted according to the characteristics of the evolution time series, with the window length appropriately reduced as the series changes more drastically; weight allocation can be determined using methods such as the analytic hierarchy process to ensure the rationality of the weights; the number and threshold of risk level classifications can be adjusted according to the actual operation and maintenance needs of the equipment, such as adding low-risk levels to provide more detailed references for operation and maintenance personnel.

[0147] For example, 1000 independent feature vector samples each of the normal operating state and fault states such as inner ring defects and loosening of superimposed layers are collected to form a training dataset, with a 1:1 ratio of normal samples to fault samples. The training dataset is input into a support vector machine, and a radial basis function is selected as the kernel function. The optimal kernel function parameters are determined through grid search and 5-fold cross-validation. After the model is trained, the classification accuracy of the test dataset reaches 95%, resulting in a hyperplane decision model. The equipment vibration signal is collected in real time, and real-time independent feature vectors are extracted according to the process. The feature deviation calculation period is set to 0.5 seconds, and 100 feature deviation data points are continuously calculated to construct an evolution time series. The sequence data are 1.2, 1.5, 2.1, 2.8, 3.5... showing a gradually increasing trend. A sliding window with a length of 30 data points and an overlap rate of 50% is used to extract the local evolution trend from the evolution time series. The rate of change for each window is obtained through linear fitting. The rate of change for some windows reaches 0.7, while the historical normal rate is within 0.2. The anomaly confidence level for each window was calculated, with some windows reaching 95%. After normalizing the rate of change and the anomaly confidence level, the results were weighted and summed with weights of 0.6 and 0.4, respectively, yielding an early failure risk index of 0.82. The preset warning threshold was 0.6, and this index exceeded the threshold. Based on the risk index, the risk level was classified as medium risk. Combined with a specific frequency pattern, the possible failure type was determined to be slight wear on the bearing inner ring. A fault warning result was generated, including a medium risk indicator, a slight wear fault indication on the bearing inner ring, and an increasing trend in characteristic deviation. This warning was displayed locally on the device screen and remotely pushed to the mobile terminals of maintenance personnel.

[0148] Furthermore, the various preset thresholds involved in this invention (such as noise tolerance threshold, dispersion threshold, matching threshold, warning threshold, etc.) can be configured to have adaptive update or periodic calibration functions. For example, during periods confirmed to be operating normally, data can be automatically collected and relevant statistics (such as permutation entropy mean, dispersion mean, risk index baseline, etc.) can be recalculated to dynamically update the thresholds, or the operation and maintenance personnel can manually reset them based on the new baseline after the equipment overhaul. The template library also supports expansion and optimization based on newly confirmed fault cases.

[0149] In summary, this invention discloses a fault detection method for generator core stacking equipment, comprising: acquiring raw vibration signals and obtaining a first vibration dataset through synchronous calibration and feature extraction; obtaining a second vibration dataset through time-frequency analysis decomposition and denoising; constructing a model to enhance periodic weak impact signals and obtaining an enhanced signal feature set; matching and quantizing with a preset reference frequency template library to determine the distribution range of potential fault signals; obtaining a fault feature dataset through deep feature extraction when a threshold is exceeded, and constructing a hyperplane decision model based on this to analyze the local evolution trend of features and output early warning results. This method achieves accurate identification and graded early warning of early equipment faults, effectively reducing the missed detection rate and false detection rate, and improving equipment operational reliability and production continuity.

[0150] Reference Figure 2 The second embodiment of the present invention provides a fault detection system for generator core stacking equipment, comprising:

[0151] The data acquisition module is used to acquire the original vibration signal during the operation of the generator core stacking equipment, and to perform synchronous calibration and feature extraction processing on the original vibration signal to obtain the first vibration dataset.

[0152] The decomposition and denoising module is used to decompose and denoise the first vibration dataset by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components and the low-frequency effective components, and reconstructing the second vibration dataset.

[0153] The feature enhancement module is used to filter the second vibration dataset by constructing a bandpass filter through the spectral kurtosis map, and to reconstruct and enhance it by training a complete dictionary matrix, so as to obtain an enhanced signal feature set.

[0154] The fault matching module is used to perform pattern matching between the enhanced signal feature set and the preset reference frequency template library, perform quantitative analysis on the matching results, and determine the distribution range of potential fault signals.

[0155] The feature extraction module is used to calculate the statistical dispersion of the distribution range of the potential fault signal. If it exceeds the preset dispersion threshold, the corresponding non-stationary signal segment is extracted and deep feature extraction is performed through multi-scale wavelet packet decomposition and singular value decomposition to obtain the fault feature dataset.

[0156] The judgment and early warning module is used to construct a hyperplane decision model based on the fault feature dataset, and output fault early warning results by comparing and analyzing the local evolution trend of the features.

[0157] It should be noted that the generator core stacking equipment fault detection system provided in this embodiment of the invention is used to execute all the process steps of the generator core stacking equipment fault detection method in the above embodiment. The working principle and beneficial effects of the two are one-to-one, so they will not be described again.

[0158] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0159] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for fault detection in generator core stacking equipment, characterized in that, include: The original vibration signal of the generator core stacking equipment during operation is acquired, and the original vibration signal is synchronously calibrated and feature extracted to obtain the first vibration dataset. The first vibration dataset is decomposed and denoised by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components and the low-frequency effective components, and reconstructing the second vibration dataset. For the second vibration dataset, a bandpass filter is constructed using the spectral kurtosis map for filtering, and the enhanced signal is reconstructed and enhanced using a trained complete dictionary matrix to obtain an enhanced signal feature set; The enhanced signal feature set is matched with a preset reference frequency template library for pattern matching, and the matching results are quantitatively analyzed to determine the distribution range of potential fault signals. Calculate the statistical dispersion of the distribution range of the potential fault signal. If it exceeds the preset dispersion threshold, extract the corresponding non-stationary signal segment and perform deep feature extraction through multi-scale wavelet packet decomposition and singular value decomposition to obtain the fault feature dataset. A hyperplane decision model is constructed based on the fault feature dataset. By comparing and analyzing the local evolution trends of the features, fault warning results are output. The first vibration dataset is decomposed and denoised by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components from the low-frequency effective components, and reconstructing the second vibration dataset, including: The time-frequency distribution matrix of the first vibration dataset is calculated using the Wignerville distribution algorithm; Based on the energy ridge line of the time-frequency distribution matrix, the instantaneous frequency and instantaneous amplitude are extracted, and the first vibration dataset is decomposed into a sequence of intrinsic mode components; For the intrinsic mode component sequence, calculate the permutation entropy value. If the permutation entropy value exceeds the preset noise tolerance threshold, then the corresponding intrinsic mode component sequence is determined to be a high-frequency aliasing component; otherwise, it is determined to be a low-frequency effective component. After performing soft threshold shrinkage on the high-frequency aliasing components, they are linearly superimposed and reconstructed with the low-frequency effective components to obtain the second vibration dataset.

2. The fault detection method for generator core stacking equipment according to claim 1, characterized in that, The process involves acquiring the original vibration signal during the operation of the generator core stacking equipment, performing synchronous calibration and feature extraction on the original vibration signal to obtain a first vibration dataset, including: The vibration signal of the generator core stacking equipment during operation is collected synchronously at multiple points by a sensor array to obtain the original vibration signal containing multiple frequency components and instantaneous change characteristics. The original vibration signal is processed by analog-to-digital conversion and phase calibration to eliminate channel transmission delay, thereby obtaining time-synchronized vibration data; The time-synchronized vibration data is decomposed using wavelet packet decomposition to separate steady-state background noise and transient impact components, and multi-frequency component feature vectors are extracted. The spatial coordinates of the sensor array are obtained and associated with the feature vector of the multi-frequency components to obtain the first vibration dataset.

3. The fault detection method for generator core stacking equipment according to claim 1, characterized in that, For the second vibration dataset, a bandpass filter is constructed using the spectral kurtosis map for filtering, and the signal is reconstructed and enhanced using a trained complete dictionary matrix to obtain an enhanced signal feature set, including: The spectral kurtosis plot of the second vibration dataset is obtained to determine the resonance demodulation frequency band. A bandpass filter is constructed to filter the second vibration dataset to obtain a narrowband filtered signal. Based on the overcomplete dictionary matrix trained by the narrowband filtered signal, the sparse coefficient vector of the narrowband filtered signal under the overcomplete dictionary matrix is ​​calculated. If the sparse coefficient vector satisfies a preset sparsity threshold, it is retained and linearly combined with the overcomplete dictionary matrix for reconstruction. A Hilbert transform is then performed on the reconstructed signal to obtain an enhanced signal feature set.

4. The fault detection method for generator core stacking equipment according to claim 1, characterized in that, The step of performing pattern matching between the enhanced signal feature set and a preset reference frequency template library, and then performing quantitative analysis on the matching results to determine the distribution range of potential fault signals includes: Calculate the matching degree between the enhanced signal feature set and each template in the preset reference frequency template library, and generate a multidimensional similarity matrix; Filter out local maxima points in the multidimensional similarity matrix whose amplitude exceeds a preset matching threshold, and identify the corresponding specific frequency patterns; Calculate the energy entropy value of the signal segment corresponding to the specific frequency mode to obtain the quantization evaluation sequence; A fault probability distribution histogram is constructed based on the quantitative evaluation sequence to determine the distribution range of potential fault signals.

5. The fault detection method for generator core stacking equipment according to claim 1, characterized in that, The statistical dispersion of the distribution range of the potential fault signal is calculated. If it exceeds a preset dispersion threshold, the corresponding non-stationary signal segment is extracted and deep feature extraction is performed through multi-scale wavelet packet decomposition and singular value decomposition to obtain a fault feature dataset, including: Calculate the statistical dispersion of the potential fault signal distribution range. If the statistical dispersion exceeds a preset dispersion threshold, then extract the corresponding non-stationary signal segment. Multi-scale wavelet packet decomposition is performed on the non-stationary signal segment to reconstruct a high-resolution time-frequency energy spectrum; Singular value decomposition is performed on the high-resolution time-frequency energy spectrum to extract the non-zero singular value sequence and construct the singular value feature vector. The singular value eigenvectors are dimensionality reduced by kernel principal component analysis to generate independent eigenvectors, which are then structured and arranged to obtain a fault feature dataset.

6. The fault detection method for generator core stacking equipment according to claim 5, characterized in that, The process of constructing a hyperplane decision model based on the fault feature dataset, and outputting fault warning results by comparing and analyzing the local evolution trends of the features, includes: Based on the independent feature vectors in the fault feature dataset, a hyperplane decision model is constructed. The feature deviation of the independent feature vector in the hyperplane decision model is calculated in real time, and an evolution time series is constructed based on the feature deviation at consecutive time points; The local evolution trend of the evolution time series is extracted, the rate of change and the anomaly confidence are calculated, and the early failure risk index is obtained by mapping. If the early failure risk index exceeds the preset warning threshold, a failure warning result containing a risk level identifier will be output.

7. A fault detection system for generator core stacking equipment, characterized in that, For implementing the method as described in any one of claims 1-6, comprising: The data acquisition module is used to acquire the original vibration signal during the operation of the generator core stacking equipment, and to perform synchronous calibration and feature extraction processing on the original vibration signal to obtain the first vibration dataset. The decomposition and denoising module is used to decompose and denoise the first vibration dataset by calculating the time-frequency distribution matrix, separating the high-frequency aliasing components and the low-frequency effective components, and reconstructing the second vibration dataset. The feature enhancement module is used to filter the second vibration dataset by constructing a bandpass filter through the spectral kurtosis map, and to reconstruct and enhance it by training a complete dictionary matrix, so as to obtain an enhanced signal feature set. The fault matching module is used to perform pattern matching between the enhanced signal feature set and the preset reference frequency template library, perform quantitative analysis on the matching results, and determine the distribution range of potential fault signals. The feature extraction module is used to calculate the statistical dispersion of the distribution range of the potential fault signal. If it exceeds the preset dispersion threshold, the corresponding non-stationary signal segment is extracted and deep feature extraction is performed through multi-scale wavelet packet decomposition and singular value decomposition to obtain the fault feature dataset. The judgment and early warning module is used to construct a hyperplane decision model based on the fault feature dataset, and output fault early warning results by comparing and analyzing the local evolution trend of the features.

Images