An intelligent diagnosis method for mechanical and electrical key components based on time-frequency sub-band central statistical value enhancement

Abstract

The application discloses a kind of based on time-frequency banding center statistical value enhancement electromechanical key component intelligent diagnosis method, belong to electromechanical equipment health monitoring and fault diagnosis technical field, to solve the problem of insufficient feature extraction of fault signal under complex working condition in prior art, low identification accuracy and poor model robustness.Proceeds time-frequency transformation to original vibration signal and obtains time-frequency spectrum;Extract the feature row time-frequency coefficient corresponding to each frequency band center frequency;Envelope spectrum analysis is carried out to each feature row time-frequency coefficient;Based on the envelope spectrum of each sensitive frequency band, the center statistical value is calculated;Select neighborhood frequency row, calculate and form two-dimensional envelope spectrum feature map in neighborhood;Two-dimensional envelope spectrum feature map is input into convolutional neural network, and the fault category is output.The method realizes the effective enhancement and extraction of early weak fault by introducing fault physical mechanism information, improves the accuracy, stability and engineering application value of fault diagnosis under complex working condition.

Description

Technical Field

[0001] This invention belongs to the field of health monitoring and fault diagnosis technology of electromechanical equipment, and specifically relates to an intelligent diagnostic method for key electromechanical components based on time-frequency zone center statistical value enhancement. Background Technology

[0002] Aeronautical electromechanical systems are a crucial component of modern aviation equipment, undertaking key functions such as energy conversion, power transmission, attitude control, actuation, and auxiliary support. Typical aeronautical electromechanical systems include aircraft motors, fuel pumps, hydraulic pumps, electromechanical actuators, transmission mechanisms, power generation units, and various rotating bearing components. These components typically operate under high-speed rotation, alternating loads, wide temperature ranges, strong vibrations, and complex electromagnetic environments. Their operational status directly affects the safety, reliability, and mission execution capabilities of aviation equipment. Rolling bearings, as common critical support components, are widely present in aircraft motors, pumps, and transmission mechanisms. Early damage to rolling bearings often transmits the impact of the failure through the structure to the entire system, causing abnormal vibration responses, changes in energy distribution, and characteristic frequency modulation. If not identified and warned of in a timely manner, this can further lead to component performance degradation, system efficiency decline, and even mission interruption or safety accidents. Therefore, research on early fault feature extraction and intelligent diagnosis of key components in aeronautical electromechanical systems has significant engineering application value.

[0003] However, fault signals in aviation electromechanical systems are typically characterized by non-stationarity, weak impact, and strong background interference. Affected by complex operating conditions, structural transmission paths, noise interference, and multi-source excitation coupling, early fault characteristics often manifest as weak amplitude, dispersed frequency components, and complex modulation relationships. Existing technologies are clearly insufficient in extracting and identifying these early weak fault characteristics. Specifically, while traditional time-domain or frequency-domain feature methods are computationally simple, they lack sensitivity to early, subtle faults and are prone to fault feature overload under strong noise and varying operating conditions. Some time-frequency analysis methods, although capable of describing the frequency distribution characteristics of signals over time, still suffer from drawbacks in practical applications, such as strong dependence on parameter selection, energy leakage, inaccurate frequency band division, and difficulty in automatically filtering fault-related components. Existing deep learning methods directly use raw vibration signals, spectrograms, or time-frequency diagrams as input, lacking explicit feature constraints related to the fault mechanisms of aerospace electromechanical systems. When training samples are limited, fault types are similar, or early fault features are not obvious, the model is easily affected by noise, changes in operating conditions, and irrelevant frequency components, leading to decreased recognition accuracy and insufficient physical interpretability of the model's diagnostic results. Therefore, how to effectively extract and enhance fault-sensitive features while fully utilizing the advantages of time-frequency analysis in representing non-stationary signals, and achieve accurate identification and diagnosis of early, subtle faults, is a key technical challenge that urgently needs to be addressed in the field of aerospace electromechanical system fault diagnosis.

[0004] Based on this, the present invention proposes an intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement, in order to solve the problems existing in the prior art. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement, thereby solving the problems of insufficient fault signal feature extraction, low identification accuracy, and poor model robustness in existing technologies under complex operating conditions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for intelligent diagnosis of key electromechanical components based on time-frequency band center statistical value enhancement includes the following steps: Step 1: Receive the original vibration signal of the key electromechanical components of the aircraft using the signal decomposition module, perform time-frequency transformation on the original vibration signal to obtain the time spectrum, and then divide the frequency range of the time spectrum into bands evenly according to the preset number of bands, and extract the characteristic line time-frequency coefficients corresponding to the center frequency of each band. Step 2: Utilize the pattern filtering module to receive the time-frequency coefficients of the characteristic rows output by the signal decomposition module. Perform normalized envelope spectrum analysis on the time-frequency coefficients of each characteristic row. Determine the theoretical fault characteristic frequencies of potential fault modes based on the structural parameters and operating conditions of key electromechanical components. Identify the sensitive frequency bands corresponding to each potential fault mode using the frequency band with the largest response amplitude at the theoretical fault characteristic frequency. Calculate the center statistical value based on the envelope spectrum of each sensitive frequency band and filter the target fault mode according to the maximum center statistical value criterion. Select a neighboring frequency row centered on the sensitive frequency band characteristic row corresponding to the target fault mode, and calculate the normalized envelope spectrum for each frequency row within the neighboring area. Segment the spectrum along the frequency row direction to form a two-dimensional envelope spectrum feature map, which serves as the enhanced fault feature. Step 3: Input the enhanced two-dimensional envelope spectrum feature map output from Step 2 into the two-dimensional convolutional neural network model to achieve automatic classification and output of fault categories.

[0007] In a preferred embodiment, in step 1, the time-frequency transformation is a short-time Fourier transform, and the time spectrum of the original signal is obtained through the short-time Fourier transform. for: ; in, Indicates experimental signal, t Represents a time variable. τ Indicates the time shift parameter. ω ( t - τ ) represents a window function. f Represents frequency variables. j Represents the imaginary unit. e Represents the natural constant.

[0008] In a preferred embodiment, in step 1, when the frequency range of the time spectrum is uniformly divided into bands according to a preset number of bands, the bandwidth is... The calculation formula is: ; in, f max Indicates the maximum sampling frequency. f min Indicates the minimum sampling frequency. n This indicates the preset number of strips.

[0009] In a preferred embodiment, in step 1, when extracting the characteristic row time-frequency coefficients corresponding to the center frequency of each frequency band, the first... k The center frequency of each frequency band for: .

[0010] In a preferred embodiment, in step 2, when performing normalized envelope spectrum analysis on the feature row data, the Hilbert transform method is used to calculate the envelope signal to obtain the normalized envelope spectrum. for: ; in, Indicates Fourier transform, Indicates the first k The center frequency coefficient of each frequency band Indicates the envelope signal. τ Indicates the time shift parameter. Indicates the first k The center frequency of each frequency band.

[0011] In a preferred embodiment, in step 2, when determining the sensitive frequency band of the envelope spectrum, the amplitude corresponding to each frequency is extracted. The calculation formula is: ; in, Indicates the first The frequency band in the first The normalized envelope spectrum amplitude at the fault characteristic frequency of the potential fault mode theory. Indicates the first The envelope spectrum is obtained by enveloping and demodulating the time-frequency coefficients of the characteristic rows corresponding to each frequency band. Indicates the first The theoretical fault characteristic frequencies corresponding to the potential fault modes Represents the set of frequencies of potential fault characteristics. Indicates frequency band index. , Indicates the potential failure mode number. , , , These represent the theoretical characteristic frequencies of cage failure, rolling element failure, outer ring failure, and inner ring failure, respectively.

[0012] In a preferred embodiment, in step 2, when calculating the center statistics of the sensitive frequency band, the center statistics of the normalized envelope spectrum are... The calculation formula is: ; in, Indicates the first Central statistics corresponding to the potential failure modes Indicates the first Sensitive frequency band indexes corresponding to the potential failure modes Indicates the first The normalized envelope spectrum of the sensitive frequency band corresponding to the potential fault modes at the theoretical fault characteristic frequency The amplitude at that point, This represents the average background amplitude of the normalized envelope spectrum of the sensitive frequency band within the selected analysis frequency range. This represents the frequency variable in the envelope spectrum.

[0013] In a preferred embodiment, in step 3, the two-dimensional convolutional neural network model includes the following structural layers: The input layer is used to receive a two-dimensional envelope spectral feature map of size H×W; Where H represents the number of sampling points on the frequency axis, and W represents the number of frequency rows; Convolutional layers, immediately following the input layer, are used to extract local fault texture features; Pooling layers, which follow convolutional layers, are used to reduce feature dimensionality, reduce computational cost, and enhance the translation invariance of features. Fully connected layers, located in the middle and later parts of the network, are used to integrate and map the local features extracted by the convolutional and pooling layers to generate a global feature representation. The output layer, which follows the fully connected layer, uses the Softmax function to classify and output the fault categories.

[0014] In a preferred embodiment, in step 3, a Dropout layer is inserted before or in the middle of the fully connected layer.

[0015] Compared with existing technologies, this invention provides an intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement, which has the following beneficial effects: This invention effectively demodulates fault modulation impulse components by combining time-frequency banding and normalized envelope spectrum analysis. Combined with the central statistical value to quantify the deviation of the spectral peak amplitude at the fault characteristic frequency, it achieves effective enhancement and extraction of early weak fault characteristics, solving the problem of insufficient sensitivity of traditional methods to early faults.

[0016] This invention divides the time spectrum into continuous and non-overlapping frequency bands along the frequency direction through uniform banding processing. Combined with the automatic screening mechanism of sensitive frequency bands and the maximum center statistical value criterion, it screens target fault modes, effectively suppressing the interference of background noise and fault-free frequency components.

[0017] This invention explicitly introduces theoretical fault characteristic frequency information related to the physical mechanism of the fault during signal processing, and establishes a quantitative mapping relationship from the physical mechanism of the fault to the saliency of the features through central statistics. This overcomes the deficiency of traditional deep learning methods as black box models that lack physical interpretability, and makes the diagnostic results have clear engineering physical significance.

[0018] This solves the problems of insufficient feature extraction of fault signals under complex working conditions, low recognition accuracy, and poor model robustness in existing technologies. Attached Figure Description

[0019] Figure 1 This is a flowchart of the intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement according to the present invention.

[0020] Figure 2 This is a waveform diagram of the vibration signal throughout the entire life cycle according to an embodiment of the present invention.

[0021] Figure 3 This is a trend chart of the root mean square (RMS) index in an embodiment of the present invention.

[0022] Figure 4 This is a time spectrum and frequency band mode division diagram of an embodiment of the present invention.

[0023] Figure 5 This is a comparison diagram of the feature envelope spectra of various modes in an embodiment of the present invention.

[0024] Figure 6 This is a comparison curve of the potential fault mode amplitudes in an embodiment of the present invention.

[0025] Figure 7 This is a line graph comparing the central statistical values ​​of an embodiment of the present invention.

[0026] Figure 8 This is an enhanced envelope spectrum feature map of an embodiment of the present invention.

[0027] Figure 9 This is a confusion matrix diagram for the verification set fault diagnosis in an embodiment of the present invention.

[0028] Figure 10 This is a comparison diagram of the CNN classification confusion matrix for three input methods according to an embodiment of the present invention.

[0029] Among them, Figure 5 middle, Figure 5 (a) is a comparison diagram of the envelope spectrum of mode 1. Figure 5 (b) is a comparison diagram of the envelope spectrum of mode 2. Figure 5 (c) is a comparison diagram of the envelope spectrum of mode 9. Figure 5 (d) is a comparison diagram of the envelope spectrum of mode 10.

[0030] exist Figure 10 middle, Figure 10 (a) is the classification confusion matrix diagram of the original signal directly input into the CNN. Figure 10 (b) is the classification confusion matrix diagram of the original atlas directly input into the CNN. Figure 10 (c) is the classification confusion matrix diagram of the enhanced graph input CNN of the present invention. Detailed Implementation

[0031] 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.

[0032] As per the instruction manual Figure 1 -Appendix Figure 10 As shown in the figure, this invention proposes an intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement. The overall architecture of this method is as follows: Figure 1 As shown, it consists of three cascaded modules, with data flowing unidirectionally from left to right. The first module is a signal decomposition module based on time-frequency analysis and fault physical information. It receives the raw vibration signals of key aircraft electromechanical components, performs short-time Fourier transform or wavelet transform to obtain the time-frequency spectrum, and then performs uniform banding according to a preset number of bands N, extracting the feature row time-frequency coefficients corresponding to the center frequencies of each band. The second module is a pattern filtering module based on center statistical values. It receives the feature row data output from the signal decomposition module and, after a series of processing steps including extracting the normalized envelope spectrum of the bands, selecting the most characteristic frequency bands of each fault, calculating the center statistical values ​​of each fault, and selecting the neighborhood frequency band of the maximum center value, outputs a two-dimensional envelope spectrum feature map as the enhanced fault features. The third module is a fault diagnosis module based on a two-dimensional convolutional neural network. It receives the two-dimensional envelope spectrum feature map as input, and after feature extraction and fusion through convolutional layers and fully connected layers, outputs the fault category diagnosis results for the key aircraft electromechanical components.

[0033] In this embodiment, the XJTU-SYBearingDatasets bearing life-cycle dataset, publicly released by Xi'an Jiaotong University and Suntech Power, is selected as the experimental data. This dataset is widely used for performance verification of bearing fault diagnosis algorithms and has high authority and representativeness. The dataset contains bearing life-cycle vibration signals under various operating conditions, with each CSV file recording a continuous vibration signal segment during the bearing life-cycle test. Specifically, in this embodiment, the vibration data of the Bearing1_3 test bearing under the operating condition of 35Hz speed and 12kN radial load is selected as the illustrative object. This bearing has an inner ring pitch circle diameter of 33mm, a rolling element diameter of 7.92mm, 8 rolling elements, and a contact angle of 0 degrees. Under this operating condition, the sampling frequency is set to 25600Hz. Bearing1_3 records the complete process of the bearing gradually deteriorating from normal operation until outer ring failure, covering the entire life-cycle data of the bearing's health gradually deteriorating until failure.

[0034] Figure 2 This example displays the time-domain waveform of the full-life-cycle vibration signal of the Bearing1_3 bearing selected in this embodiment, where the horizontal axis represents the acquisition file number and the vertical axis represents the vibration signal amplitude. Figure 2 It can be observed that the vibration signal exhibits a clear three-stage evolution characteristic: in the range of document number 0-80, the amplitude is close to zero, and the signal shows a stable noise state, indicating that the bearing is in the normal operation stage; in the range of document number 80-130, the amplitude gradually increases and periodic fluctuations occur, suggesting that the bearing has begun to experience performance degradation; in the range of document number 140-155, the amplitude oscillates violently, and the peak value exceeds ±20, showing typical impact fault characteristics, indicating that the bearing has experienced a significant fault.

[0035] To accurately pinpoint the timing of early bearing failures, this embodiment of the invention uses the root mean square value as a time-domain statistical indicator to monitor the bearing degradation process.

[0036] Figure 3 The graph shows the trend of the root mean square (RMS) index as a function of file number, with the horizontal axis representing the file number and the vertical axis representing the RMS value. The main graph shows that the RMS value remains stable at approximately 0.5 before file number 50; it rises slowly between file numbers 50 and 100; and it spikes sharply after file number 140. To accurately capture the early fault initiation point, a magnified sub-graph is embedded in the upper left corner of the main graph. The horizontal axis displays files 57-61, and the vertical axis displays a fine range of 0.5-0.58. The red circle in the graph precisely marks the location of file number 59, and the red arrow points to the inflection point of the curve, showing the RMS value rising from approximately 0.52 to approximately 0.56. Based on the above analysis, this embodiment of the invention selects file number 59 as a representative sample of the early fault characteristic appearance stage for subsequent time-frequency analysis and feature enhancement processing.

[0037] Based on the above experimental data, the intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement, as described in this invention, includes the following steps: Step 1: Receive the raw vibration signals of key aircraft electromechanical components using a signal decomposition module. Perform time-frequency transformation on the selected raw vibration signal from document No. 59 to obtain the time spectrum. Then, evenly divide the frequency range of the time spectrum according to a preset number of bands and extract the characteristic row time-frequency coefficients corresponding to the center frequency of each band. Specifically, this includes: Step 1.1: Collect the raw vibration signals of key aircraft electromechanical components, perform time-frequency transformation on the raw vibration signals, and obtain the time spectrum. Specifically: The time-frequency transformation employs a short-time Fourier transform (SFT). This SFT uses a sliding window function to obtain the energy distribution of the original vibration signal at different time and frequency positions, characterizing the non-stationary characteristics of fault signals in critical aircraft electromechanical components. The SFT yields the time spectrum of the original signal. Its formula is expressed as: (1); in, Represents the original experimental vibration signal. t Represents a time variable. τ Indicates the time shift parameter. ω ( t - τ ) represents a window function. f Represents frequency variables. j Represents the imaginary unit. e Represents the natural constant.

[0038] In this embodiment, the window size of the short-time Fourier transform is set to 40, and the sampling frequency is... f s Set to 25600. The choice of window size requires a trade-off between time resolution and frequency resolution. A shorter window achieves higher time resolution but lower frequency resolution, while a longer window achieves higher frequency resolution but lower time resolution. For bearing fault diagnosis applications, a window setting of 40 sampling points effectively captures the temporal location of fault impacts while maintaining sufficient frequency resolution to distinguish different fault characteristic frequencies.

[0039] Figure 4 The time-frequency spectrum of the vibration signal from document No. 59 after short-time Fourier transform is displayed. The horizontal axis represents the sampling point number, the vertical axis represents the frequency, and the color intensity indicates energy intensity. The non-stationary concentration characteristics of the signal energy in the time-frequency plane can be observed from the time-frequency spectrum; different frequency components exhibit different energy distributions at different times. The frequency axis is divided into multiple continuous and non-overlapping frequency bands by yellow horizontal dashed lines, labeled as mode 1 to mode 10. This banding method achieves a structured frequency domain division of the time-frequency spectrum.

[0040] Step 1.2: Divide the frequency range of the time spectrum evenly according to the preset number of bands, dividing the time spectrum into continuous and non-overlapping frequency bands along the frequency direction, and extract the characteristic row time-frequency coefficients corresponding to the center frequency of each band. Bandwidth The calculation formula is: (2); in, f max Indicates the maximum sampling frequency. fmin Indicates the minimum sampling frequency. n This indicates the preset number of strips.

[0041] In this embodiment, the preset number of segments is... n Set to 10, therefore the bandwidth Equals 1280Hz. Number of bands. n The selection needs to be determined based on the frequency distribution range of bearing failure characteristics and the required analysis accuracy. Larger... n A higher value can achieve a finer frequency band division, but it increases computational complexity; a smaller value... n While this value can reduce the computational burden, it may not accurately capture the sensitive frequency band where fault characteristics are located.

[0042] Step 1.3: After completing the frequency band division, it is necessary to extract the characteristic row time-frequency coefficients of each frequency band as representative signal components of that frequency band. The basic principle of characteristic row selection is to calculate the center frequency of each frequency band and select the frequency row closest to each center frequency in the time-frequency spectrum as the characteristic row time-frequency coefficients of the corresponding frequency band. k The center frequency of each frequency band The formula is: (3); In this embodiment, the time-frequency spectrum range is 0 to 12800 Hz, and the preset number of bands is [not specified]. n The value is 10, and the bandwidth Δ f The frequency is 1280Hz. Therefore, using the above formula, the center frequencies of the 10 frequency bands are calculated to be 640Hz, 1920Hz, 3200Hz, 4480Hz, 5760Hz, 7040Hz, 8320Hz, 9600Hz, 10880Hz, and 12160Hz. Since the frequency axis obtained by the short-time Fourier transform is a discrete frequency point, for each center frequency, the frequency point with the smallest distance from the center frequency is found in the time-spectrum frequency axis, and the time-spectrum row corresponding to this frequency point is used as the characteristic row time-frequency coefficient of the corresponding frequency band. In this way, the original vibration signal can be transformed into a set of characteristic row time-frequency coefficients discretely distributed along the frequency direction, providing input for subsequent normalized envelope spectrum analysis and center statistical value screening.

[0043] Step 1.4: Repeat step 1.3 until all frequency bands have been decomposed. (The result is the first...) k The time-frequency coefficients of each frequency band are .

[0044] Through the specific steps of the above signal decomposition module, the original vibration signal is converted into a set of characteristic row time-frequency coefficients discretely distributed along the frequency direction, realizing the frequency domain structured characterization of non-stationary signals, and providing a standardized data foundation for subsequent envelope spectrum analysis and fault mode screening.

[0045] Step 2: Utilize the pattern filtering module to receive the time-frequency coefficients of the characteristic rows output by the signal decomposition module. Perform normalized envelope spectrum analysis on the time-frequency coefficients of each characteristic row. Determine the theoretical fault characteristic frequencies of potential fault modes based on the structural parameters and operating conditions of key electromechanical components. Identify the sensitive frequency bands corresponding to each potential fault mode using the frequency band with the largest response amplitude at the theoretical fault characteristic frequency. Calculate the center statistical value based on the envelope spectrum of each sensitive frequency band, and filter the target fault mode according to the maximum center statistical value criterion. Select a neighboring frequency row centered on the characteristic row of the sensitive frequency band corresponding to the target fault mode, and calculate the normalized envelope spectrum for each frequency row within the neighboring area. Concatenate these normalized envelope spectra along the frequency row direction to form a two-dimensional envelope spectrum feature map, which serves as the enhanced fault feature. Specifically, this includes: Step 2.1: Normalized Envelope Spectrum Analysis: Perform envelope spectrum or modulation spectrum analysis on the time-frequency coefficients of each characteristic row, and then normalize the spectrum to obtain the normalized envelope spectrum. Specifically: The envelope signal is calculated using the Hilbert transform method. The formula for the Hilbert transform is: (4); in, .

[0046] Calculating the envelope signal using the Hilbert transform method The formula is: (5); in, This represents the Hilbert transform.

[0047] The normalized envelope spectrum was finally obtained. The formula is: (6); in, Indicates Fourier transform, , Represents the Hilbert transform. τ Indicates the time shift parameter. Indicates the first k The center frequency of each frequency band Indicates the first k The time-frequency coefficients of each frequency band Indicates the first k The center frequency of each frequency band.

[0048] Figure 5 This section displays comparative data envelopment spectra (DES) of characteristic rows for some frequency band modes, including four sub-plots: Mode 1, Mode 2, Mode 9, and Mode 10. The horizontal axis of each sub-plot represents frequency, and the vertical axis represents amplitude. Each sub-plot contains four red vertical dashed lines, labeled at the top. fd1 , f d2 , f d3 and f d4 The numbers and represent the four potential fault characteristic frequencies of rolling bearings: cage fault characteristic frequency, rolling element fault characteristic frequency, outer ring fault characteristic frequency, and inner ring fault characteristic frequency. As can be observed from the figure, the amplitude responses of different modes at these theoretical fault characteristic frequencies differ significantly. Some modes exhibit obvious peaks at specific fault characteristic frequencies, while other modes do not. This phenomenon provides a basis for subsequent sensitive frequency band screening and target fault mode identification.

[0049] The potential failure characteristic frequency of a rolling bearing needs to be calculated based on the bearing's structural parameters, motion parameters, and failure mechanism.

[0050] In this embodiment, the relevant parameters of the bearing under test are: number of rolling elements Z=8, rolling element diameter... d =7.92mm, pitch circle diameter D =33mm, contact angle α =0°, frequency f rot =35Hz. Based on these parameters, the characteristic frequencies of the four potential fault modes can be calculated.

[0051] Cage Failure Characteristic Frequency f d1 The calculation formula is: ; Rolling element failure characteristic frequency f d2 The calculation formula is: ; Outer ring fault characteristic frequency f d3 The calculation formula is: ; Inner ring fault characteristic frequency f d4 The calculation formula is: ; Substituting the bearing parameters into the above formula, the theoretical characteristic frequencies of each potential failure mode in this embodiment are calculated: f d1 Approximately 13.3 Hz f d2 Approximately 68.7 Hz f d3 Approximately 107.9 Hzf d4 Approximately 173.6 Hz. Among them... f d3 The outer ring fault characteristic frequency of 107.9 Hz corresponds to the actual bearing fault mode, which will be verified in the subsequent mode screening process.

[0052] Step 2.2: Sensitive frequency band determination: Determine the theoretical fault characteristic frequency set of various potential fault modes based on the structural parameters, motion parameters and fault mechanisms of key aircraft electromechanical components.

[0053] The set of potential fault characteristic frequencies is calculated based on the structural parameters, motion parameters, and fault mechanisms of key aero-mechanical components. When the key aero-mechanical component is a rolling bearing, the potential fault characteristic frequencies include at least one of the following: inner ring fault characteristic frequency, outer ring fault characteristic frequency, rolling element fault characteristic frequency, and cage fault characteristic frequency, calculated from the number of rolling elements, rolling element diameter, pitch circle diameter, contact angle, and rotational frequency. The amplitude corresponding to each theoretical fault characteristic frequency is extracted from the normalized envelope spectrum, and the sensitive frequency band corresponding to the potential fault is determined by the frequency band containing the maximum amplitude of each type of potential fault characteristic frequency. The amplitude corresponding to each frequency is extracted. The calculation formula is: (7); in, Indicates the first The frequency band in the first The normalized envelope spectrum amplitude at the fault characteristic frequency of a potential fault mode theory is used to characterize the response intensity of the frequency band to this type of fault characteristic. Indicates the first The envelope spectrum is obtained by enveloping demodulation and normalization of the time-frequency coefficients corresponding to each frequency band. Indicates the first The theoretical fault characteristic frequencies corresponding to the potential fault modes; Represents the set of frequencies of potential fault characteristics; Indicates frequency band index. ; Indicates the potential failure mode number; , , , These represent the theoretical characteristic frequencies of cage failure, rolling element failure, outer ring failure, and inner ring failure, respectively.

[0054] The potential failure frequency is calculated based on the bearing structure and parameters, and specific operating conditions.

[0055] Find the frequency band index that carries the most information about each potential fault. The formula is: (8); in, To extract the maximum index value.

[0056] The frequency band feature line is represented as follows .

[0057] In this embodiment, the amplitude at the fundamental frequency and second harmonic is mainly considered, i.e., m takes values ​​of 0 and 1. The frequency band index carrying the most potential fault information is identified. k m The formula is This formula identifies the frequency band most sensitive to this type of fault by comparing the amplitude response of different frequency bands at the theoretical fault characteristic frequency.

[0058] Figure 6 This diagram illustrates a comparison of fault information carried by various potential faults in the envelope spectrum across different frequency bands. The horizontal axis represents the mode number, the vertical axis represents the characteristic amplitude, and the legend includes four curves corresponding to... f d1 , f d2 , f d3 and f d4 As can be seen from the figure, f d3 The curve reaches a local peak near mode 6, indicating that the frequency band corresponding to mode 6 contains the richest outer ring fault information; f d2 The curves are higher at modes 1 and 6, indicating that the inner ring fault characteristics are more frequent than the outer ring fault characteristic frequencies. f d2 Responses were observed in both high-frequency and low-frequency modes; the amplitudes of each fault frequency exhibited a non-monotonic distribution across different frequency bands, indicating that different fault modes had varying sensitivities to frequency band division. These results provide a basis for subsequently determining the sensitive frequency bands of each potential fault mode, but the final target fault mode still requires further assessment based on central statistical values.

[0059] Step 2.3: Calculation of Central Statistics: The central statistics measure the deviation of the spectral peak amplitude at the potential fault characteristic frequency from the average background amplitude of the envelope spectrum. A larger central statistics value indicates a more prominent fault characteristic peak and weaker background interference, making the fault mode more likely to be the actual fault mode of a critical aircraft electromechanical component. Based on the sensitive frequency band envelope spectrum corresponding to various potential fault modes, calculate their average background amplitude and then calculate the central statistics value.

[0060] Central statistic of normalized envelope spectrum The calculation formula is: (9); in, Indicates the first The central statistical value corresponding to a potential fault mode is used to measure the prominence of the spectral peak at the theoretical fault characteristic frequency of the fault mode relative to the average background amplitude of the envelope spectrum. Indicates the first Sensitive frequency band indexes corresponding to various potential failure modes; Indicates the first The normalized envelope spectrum of the sensitive frequency band corresponding to the potential fault modes at the theoretical fault characteristic frequency The amplitude at that point; This represents the average background amplitude of the normalized envelope spectrum of the sensitive frequency band within the selected analysis frequency range; This represents the frequency variable in the envelope spectrum. If this parameter is maximized, the corresponding row of time-frequency coefficients contains the most fault information.

[0061] By comparing the maximum central statistics of various potential faults, and based on the maximum central statistics criterion, the target fault mode is selected from multiple potential fault modes. The index of the potential fault mode with the maximum central statistics is extracted. The calculation formula is: (10); in, This indicates the target failure mode number obtained from the final screening. , , , These represent the central statistics corresponding to the four potential failure modes. They are arranged according to the failure characteristic frequency order defined in this paper. This represents the central statistic corresponding to the cage failure mode. This represents the central statistic corresponding to the rolling element failure mode. This represents the central statistics corresponding to the outer ring failure modes. This represents the central statistical value corresponding to the inner circle fault mode. The larger the central statistical value, the more obvious the spectral peak at the theoretical fault characteristic frequency corresponding to the fault mode, and the more likely the fault mode is the actual fault mode of the current electromechanical key component.

[0062] The potential failure characteristic frequency of a rolling bearing needs to be calculated based on the bearing's structural parameters, motion parameters, and failure mechanism.

[0063] Figure 7The graph compares the center statistics of the normalized envelope spectra corresponding to various potential failure modes. The horizontal axis represents the potential failure mode, and the vertical axis represents the center statistics. A blue line connects the four data points, and red dots mark the peak values. The graph shows that the center statistics of modes 1, 2, and 4 are concentrated around 0.175, with relatively similar values. However, the center statistics of mode 3 jumps significantly to approximately 0.40, forming a sharp triangular peak, much higher than the other three failure modes. Based on the maximum center statistics criterion, mode 3 can be determined as the most likely actual failure mode. In this embodiment, mode 3 corresponds to the outer ring failure characteristic frequency. f d3 Its theoretical value is 107.9 Hz, a result that matches the actual failure type of the bearing.

[0064] Step 2.4: Neighborhood Frequency Row Selection: Centering on the sensitive frequency band characteristic line corresponding to the target fault mode, select neighborhood frequency rows to form a neighborhood time-frequency coefficient set. The neighborhood is determined based on the diffusion range of fault impact energy near the target frequency band. Since early-stage rolling bearing faults typically manifest as modulated impact components with a certain bandwidth, they are not ideal single-spectrum lines in the envelope spectrum, but rather form local energy accumulation regions near the fault characteristic frequency. By statistically analyzing the energy distribution within the neighborhood of the fault characteristic frequency, the persistence characteristics of the fault impact in the time-frequency domain can be effectively characterized, and the ability to express weak fault information can be enhanced. The neighborhood... These are empirical parameters that can be selected based on sampling frequency, speed range, and the degree of fault impact diffusion. They are used to highlight the input intensity of fault information and improve the stability of pattern screening.

[0065] In this embodiment, a set of neighborhood time-frequency coefficients is formed by selecting a preset number of neighboring frequency rows centered on the sensitive frequency band feature behavior corresponding to the target fault mode. Specifically, a neighborhood set containing 20 frequency rows can be selected, where the number of frequency rows can be adjusted according to the frequency resolution of the short-time Fourier transform, the sampling frequency, the rotational speed range, and the degree of fault impact energy diffusion. Compared with directly using a single center frequency row, selecting neighborhood frequency rows can preserve the local energy diffusion characteristics of the fault impact in the time-frequency plane, improving the stability of the early weak fault feature expression.

[0066] Step 2.5: Construction of Two-Dimensional Envelope Spectrum Feature Map: Calculate the normalized envelope spectrum for each frequency row in the neighborhood time-frequency coefficient set, and then stitch them together along the frequency row direction to form a two-dimensional envelope spectrum feature map, which serves as the enhancement feature. The enhanced envelope spectrum feature map is formed by selecting the neighborhood of the sensitive frequency band corresponding to the target fault mode, calculating the normalized envelope spectrum for each frequency row, and then stitching them together along the frequency direction. This feature map expands the one-dimensional envelope spectrum into a two-dimensional image along the frequency row direction, preserving the neighborhood energy diffusion characteristics of the fault impact, highlighting the target fault features and their harmonic components, and suppressing irrelevant frequency components, thus providing high signal-to-noise ratio and strong physical interpretability input features for subsequent CNNs.

[0067] Figure 8 This image displays the enhanced envelope spectral feature map extracted after pattern filtering. The horizontal axis represents frequency, and the vertical axis represents the envelope or frequency band number. The background is dark blue, and multiple vertical bright yellow high-energy bands appear. The yellow arrow points to the first bright band, marked 109.2 Hz, which is approximately equal to 107.9 Hz. f d3 The second bright band, marked 2X, indicates the second harmonic frequency position. It can be clearly observed from the image that the energy... f d3 That is, the frequency bands are highly concentrated at 107.9 Hz and its harmonics, forming distinct characteristic spectral lines, while background noise and irrelevant frequency components are effectively suppressed. This result verifies the effectiveness of the method of the present invention in enhancing fault characteristics.

[0068] Step 3: Input the enhanced two-dimensional envelope spectrum feature map output from Step 2 into the two-dimensional convolutional neural network model to achieve automatic classification and output of fault categories.

[0069] The described two-dimensional convolutional neural network model is a deep learning model that can automatically learn hierarchical feature representations from image data, possessing powerful feature extraction and pattern recognition capabilities. Specifically, the two-dimensional convolutional neural network model includes the following structural layers: The input layer receives a two-dimensional envelope spectral feature map with a size of H×W, where H is the number of sampling points on the frequency axis and W is the number of frequency rows, i.e., the number of frequency rows contained in the neighborhood.

[0070] The convolutional layer follows the input layer and is used to extract local fault texture features. Multiple convolutional kernels are used to perform sliding convolution operations on the input feature map, and each convolutional kernel can detect a specific local feature pattern.

[0071] Pooling layers follow convolutional layers and are used to reduce feature dimensionality, reduce computational cost, and enhance the translation invariance of features.

[0072] Fully connected layers are located in the middle and later parts of the network and are used to integrate and map the local features extracted by the convolutional and pooling layers to generate a global feature representation.

[0073] Inserting a Dropout layer before or in the middle of a fully connected layer allows for the random discarding of some neurons' outputs with a certain probability, preventing network overfitting and improving generalization ability.

[0074] The output layer, which follows the fully connected layer, uses the Softmax function to predict the probability of multiple fault categories, and outputs the probability distribution of healthy status, inner circle faults, outer circle faults or other preset fault categories.

[0075] In this embodiment, the specific network parameters are configured as follows: the input feature map size is 128×20, i.e., 128 sampling points in the frequency direction and 20 rows in the frequency row direction. The convolutional layer contains 32 convolutional kernels of size 3×3, using the ReLU activation function. The pooling layer uses max pooling, with a pooling window size of 2×2. The second convolutional layer contains 64 convolutional kernels of size 3×3. The fully connected layer contains 128 neurons, with a dropout rate set to 0.5. The output layer corresponds to three fault categories: healthy state, inner circle fault, and outer circle fault. The network is trained using the cross-entropy loss function, and the Adam optimizer is used to update the network parameters. The learning rate is set to 0.001, and the number of training iterations is set to 100 epochs.

[0076] Figure 9 The confusion matrix for fault diagnosis using the method of this invention on the validation set is shown. The horizontal axis represents the predicted class, and the vertical axis represents the true class. Row labels include: 0 for healthy status, 1 for inner circle fault, and 2 for outer circle fault. The confusion matrix shows that diagonal elements are bright yellow with high values, while off-diagonal elements are dark purple with extremely low values, exhibiting a highly diagonally concentrated distribution. This indicates that the classifier has a high accuracy rate in identifying all types of faults, with the proportion of correctly classified healthy status, inner circle fault, and outer circle fault all approaching 100%.

[0077] To verify the effect of the enhanced features of this invention on the diagnostic performance of a two-dimensional convolutional neural network, three sets of comparative experiments were set up on the same dataset. The three sets of experiments used the same training and validation set partitioning method and the same class labeling system; the only difference was the type of features input to the CNN network. In the first set of experiments, the original vibration signal was directly input into the CNN network for classification. In the second set, the time-frequency spectrum obtained by time-frequency transformation of the original vibration signal was directly input into the CNN network for classification. In the third set, the enhanced envelope spectrum feature map obtained by time-frequency banding, center statistical value filtering, and neighborhood envelope spectrum concatenation of this invention was input into the CNN network for classification.

[0078] Figure 10The confusion matrix comparison results of three sets of comparative experiments are shown. The horizontal axis represents the predicted class, the vertical axis represents the true class, and the classification accuracy of each set of experiments is marked below. The comparison results clearly demonstrate the significant performance improvement brought about by the enhanced features: The first set of experiments used the original vibration signal directly input into the CNN network, and the accuracy rate was only 31.85%, which is close to the level of random guessing. This shows that the fault information contained in the original time domain signal is difficult to be directly and effectively utilized by the network. The second set of experiments used time-frequency spectra directly input into the CNN network, which improved the accuracy to 53.33%. Although it was an improvement over the original signal, it still did not achieve a satisfactory diagnostic effect, indicating that time-frequency spectra features lacking physical mechanism constraints are difficult to support high-precision fault classification. The third set of experiments used the enhanced envelope spectrum feature map of this invention as input to the CNN network, and the accuracy was significantly improved to 98.52%, approaching the level of completely accurate classification. This represents an improvement of 66.67% compared to the original signal input and 45.19% compared to the time-frequency spectrum input. This comparative result fully verifies that the feature enhancement achieved by the method of this invention through time-frequency banding, center statistical value screening, and neighborhood envelope spectrum splicing can significantly improve the recognition accuracy and stability of the fault diagnosis system.

[0079] From the complete description of the above embodiments, it can be seen that: The method of this invention combines time-frequency banding and normalized envelope spectrum analysis to effectively demodulate the fault modulation impulse component. Combined with the central statistical value to quantitatively characterize the deviation of the spectral peak amplitude at the fault characteristic frequency, it achieves effective enhancement and extraction of early weak fault characteristics.

[0080] The method of this invention divides the time spectrum into continuous and non-overlapping frequency bands along the frequency direction through uniform banding processing. Combined with the automatic screening mechanism of sensitive frequency bands and the maximum center statistical value criterion, it screens target fault modes, effectively suppressing the interference of background noise and fault-free frequency components.

[0081] The method of this invention explicitly introduces theoretical fault feature frequency information related to the physical mechanism of the fault during signal processing. It establishes a quantitative mapping relationship from the physical mechanism of the fault to the saliency of the features through central statistics, which overcomes the deficiency of traditional deep learning methods as black box models that lack physical interpretability and makes the diagnostic results have clear engineering physical significance.

[0082] The method of this invention combines a physics-driven feature enhancement method with a data-driven deep learning model. Each module has a clear function and adjustable parameters, and has high engineering application flexibility and generalization ability. It is applicable to fault diagnosis of various key aerospace electromechanical components such as aero motors, fuel pumps, hydraulic pumps, electromechanical actuators, transmission mechanisms and their supporting bearings under complex working conditions.

[0083] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent diagnosis of key electromechanical components based on time-frequency band center statistical value enhancement, characterized in that, Includes the following steps: Step 1: Receive the original vibration signal of the key electromechanical components of the aircraft using the signal decomposition module, perform time-frequency transformation on the original vibration signal to obtain the time spectrum, and then divide the frequency range of the time spectrum into bands evenly according to the preset number of bands, and extract the characteristic line time-frequency coefficients corresponding to the center frequency of each band. Step 2: Utilize the pattern filtering module to receive the characteristic row time-frequency coefficients output by the signal decomposition module, perform normalized envelope spectrum analysis on each characteristic row time-frequency coefficient, determine the theoretical fault characteristic frequency of the potential fault mode based on the structural parameters and operating conditions of the key electromechanical components, and determine the sensitive frequency band corresponding to each potential fault mode based on the frequency band with the largest response amplitude at the theoretical fault characteristic frequency. Based on the center statistical values ​​of the envelope spectrum of each sensitive frequency band, the target fault mode is screened according to the maximum center statistical value criterion. The neighboring frequency row is selected with the sensitive frequency band feature line corresponding to the target fault mode as the center, and the normalized envelope spectrum is calculated for each frequency row in the neighboring area. The two-dimensional envelope spectrum feature map is spliced ​​along the direction of the frequency row to form the enhanced fault feature. Step 3: Input the enhanced two-dimensional envelope spectrum feature map output from Step 2 into the two-dimensional convolutional neural network model to achieve automatic classification and output of fault categories.

2. The intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement as described in claim 1, characterized in that, In step 1, the time-frequency transformation is a short-time Fourier transform, and the time spectrum of the original signal is obtained through the short-time Fourier transform. for: ； in, Indicates experimental signal, t Represents a time variable. τ Indicates the time shift parameter. ω ( t - τ ) represents a window function. f Represents frequency variables. j Represents the imaginary unit. e Represents the natural constant.

3. The intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement as described in claim 1, characterized in that, In step 1, when the frequency range of the time spectrum is uniformly divided into bands according to a preset number of bands, the bandwidth... The calculation formula is: ； in, f max Indicates the maximum sampling frequency. f min Indicates the minimum sampling frequency. n This indicates the preset number of strips.

4. The intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement as described in claim 3, characterized in that, In step 1, when extracting the characteristic row time-frequency coefficients corresponding to the center frequency of each frequency band, the first... k The center frequency of each frequency band for: 。 5. The intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement as described in claim 1, characterized in that, In step 2, when performing normalized envelope spectrum analysis on the feature row data, the Hilbert transform method is used to calculate the envelope signal to obtain the normalized envelope spectrum. for: ； in, Indicates Fourier transform, Indicates the first k The center frequency coefficient of each frequency band Indicates the envelope signal. τ Indicates the time shift parameter. Indicates the first k The center frequency of each frequency band.

6. The intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement as described in claim 5, characterized in that, In step 2, when determining the sensitive frequency band of the envelope spectrum, the amplitude corresponding to each frequency is extracted. The calculation formula is: ； in, Indicates the first The frequency band in the first The normalized envelope spectrum amplitude at the fault characteristic frequency of the potential fault mode theory. Indicates the first The envelope spectrum is obtained by enveloping and demodulating the time-frequency coefficients of the characteristic rows corresponding to each frequency band. Indicates the first The theoretical fault characteristic frequencies corresponding to the potential fault modes Represents the set of frequencies of potential fault characteristics. Indicates frequency band index. , Indicates the potential failure mode number. , , , These represent the theoretical characteristic frequencies of cage failure, rolling element failure, outer ring failure, and inner ring failure, respectively.

7. The intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement as described in claim 6, characterized in that, In step 2, when calculating the center statistics of the sensitive frequency band, the center statistics of the normalized envelope spectrum are... The calculation formula is: ； in, Indicates the first Central statistics corresponding to the potential failure modes Indicates the first Sensitive frequency band indexes corresponding to the potential failure modes Indicates the first The normalized envelope spectrum of the sensitive frequency band corresponding to the potential fault modes at the theoretical fault characteristic frequency The amplitude at that point, This represents the average background amplitude of the normalized envelope spectrum of the sensitive frequency band within the selected analysis frequency range. This represents the frequency variable in the envelope spectrum.

8. The intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement as described in claim 1, characterized in that, In step 3, the two-dimensional convolutional neural network model includes the following structural layers: The input layer is used to receive a two-dimensional envelope spectral feature map of size H×W; Where H represents the number of sampling points on the frequency axis, and W represents the number of frequency rows; Convolutional layers, immediately following the input layer, are used to extract local fault texture features; Pooling layers, which follow convolutional layers, are used to reduce feature dimensionality, reduce computational cost, and enhance the translation invariance of features. Fully connected layers, located in the middle and later parts of the network, are used to integrate and map the local features extracted by the convolutional and pooling layers to generate a global feature representation. The output layer, which follows the fully connected layer, uses the Softmax function to classify and output the fault categories.

9. The intelligent diagnostic method for key electromechanical components based on time-frequency band center statistical value enhancement as described in claim 8, characterized in that, In step 3, a Dropout layer is inserted before or in the middle of the fully connected layer.

Images