Main shaft fault diagnosis method based on deep learning

Abstract

The invention belongs to the technical field of equipment fault early warning and diagnosis, and relates to a main shaft fault diagnosis method based on deep learning. The main shaft fault diagnosis method comprises the following steps: acquiring lathe operation sound wave data during operation of a main shaft through a plurality of sound acquisition devices, and filtering the sound wave data acquired by the sound acquisition devices near the main shaft to obtain main shaft operation sound wave data; preprocessing the main shaft operation sound wave data to obtain high-resolution sound wave data; analyzing the high-resolution sound wave data according to the residual neural network, and performing sliding analysis by adopting a window in the analysis process; and judging whether the spindle breaks down based on the analysis result. The deep learning model of the residual neural network is adopted to carry out feature extraction and mode recognition on the processed sound wave data, the difference between normal operation and abnormal state of the main shaft can be automatically learned and recognized, and the accuracy and reliability of diagnosis are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of equipment fault early warning diagnosis, and relates to a spindle fault diagnosis method based on deep learning. Background Art

[0002] A machine tool spindle is the shaft that rotates the workpiece or tool on a machine tool. It typically consists of a spindle, bearings, and transmission components. It primarily supports transmission components, such as gears and pulleys; transmits motion and torque, as in the case of a machine tool spindle; and clamps workpieces, such as a mandrel. With the exception of machine tools with linear motion, such as planers and broaching machines, most other machine tools have spindles. The integrity of the spindle assembly is a crucial factor in determining machining quality and cutting efficiency.

[0003] Subtle changes in the sound produced during spindle operation may indicate potential failures, but these details are often submerged in conventional sound wave data. There are many sources of noise generated during machine tool operation, making it difficult to distinguish the sound of the spindle operation.

[0004] Traditional methods rely on manual experience to judge faults, which are subjective and have high misjudgment rates, and cannot meet the real-time and high efficiency requirements in industrial production. Summary of the Invention

[0005] In order to solve the problems existing in the background technology, the present invention proposes a spindle fault diagnosis method based on deep learning.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows: A spindle fault diagnosis method based on deep learning, comprising: Collecting lathe operation sound wave data when the spindle is running by multiple sound collection devices, wherein the multiple sound collection devices are installed at different positions, and each sound generating device when the lathe is running has only one sound collection device within a preset range; Filtering the sound wave data collected by the sound collection device near the main shaft to obtain the main shaft operation sound wave data; Pre-process the spindle operation acoustic wave data to obtain high-resolution acoustic wave data; High-resolution acoustic wave data is analyzed based on residual neural networks, and sliding analysis is performed using windows during the analysis process. Based on the analysis results, determine whether the spindle has a fault.

[0007] Furthermore, the specific method of filtering the sound wave data collected by the sound collection device near the main shaft to obtain the main shaft operation sound wave data is: Performing Fourier transform on the sound wave data collected by the sound collection device to convert the time domain signal into a frequency domain signal; Separating the frequency domain signals corresponding to the sound wave data collected by the sound collection devices other than the sound collection device corresponding to the main axis from the frequency domain signals corresponding to the sound wave data collected by the sound collection device corresponding to the main axis; Determine the frequency range of the spindle signal and other component signals in the lathe, and separate the spindle operation sound wave data.

[0008] Furthermore, the specific method of pre-processing the spindle operation acoustic wave data to obtain high-resolution acoustic wave data is: The noise in the spindle running acoustic wave data is removed by wavelet transform; The spindle running acoustic wave data after noise removal is fitted, the fitted acoustic wave data is selected, and the difference points are amplified to obtain high-resolution acoustic wave data.

[0009] Furthermore, the specific method for analyzing high-resolution acoustic wave data based on the residual neural network is: Define a window. The window height is user-defined under the condition that it is smaller than the average amplitude of high-resolution acoustic wave data. The window width is also user-defined. The high-resolution sound wave data is segmented into three categories based on the window size, and the segmented images are passed into the residual neural network; The residual neural network performs convolution, pooling, and normalization on the high-resolution sound wave data in each window, and retains the features of the high-resolution sound wave data in each window through the residual block in the residual neural network; The similarity of the high-resolution acoustic wave data in each window is determined by the mean square error of the features in all windows. If the difference exceeds the fault threshold, the spindle is judged to be faulty.

[0010] Furthermore, the specific method of dividing the high-resolution sound wave data into three categories according to the window size is: Establish a Cartesian coordinate system with the x-axis located at the horizontal center axis of the acoustic wave data; The first type of window is divided by arranging the window range horizontally at positions symmetrical about the x-axis; The second type of window is the first type of window, which is shifted upwards by the window height and arranged horizontally to split the window. The third type of window is the first type of window, which is shifted downward by the window height and arranged horizontally to split it.

[0011] Compared with the prior art, the present invention has the following beneficial effects: The present invention performs frequency domain analysis on data collected by sound collection devices installed at different locations and uses Fourier transform to isolate the frequency range corresponding to the main axis, effectively filtering out other interfering signals. Wavelet transform is also used to remove noise and amplify differences, thereby obtaining high-resolution sound wave data, enabling even subtle changes to be clearly captured.

[0012] The present invention uses a deep learning model of a residual neural network to perform feature extraction and pattern recognition on the processed acoustic wave data, which can automatically learn and identify the difference between normal operation and abnormal state of the spindle, thereby improving the accuracy and reliability of diagnosis.

[0013] The present invention also adopts a sliding window mechanism, which allows real-time analysis of continuously inflowing data streams, and divides the window into three categories, which can meet the analysis requirements in amplitude, phase and frequency. At the same time, it judges whether the mean square error value of the features in the window exceeds the set threshold, thereby realizing the instant fault diagnosis function. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the method of the present invention.

[0015] Figure 2 It is a schematic diagram of window segmentation of the present invention.

[0016] The marks in the figure are: 1 is the first type of window, 2 is the second type of window, and 3 is the third type of window. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] like Figure 1-Figure 2 As shown, the technical solution adopted by the present invention is as follows: a spindle fault diagnosis method based on deep learning, comprising: The lathe operation sound wave data when the spindle is running is collected by multiple sound collecting devices, the multiple sound collecting devices are installed at different positions, and each sound generating device when the lathe is running has only one sound collecting device within a preset range.

[0019] The sound wave data collected by the sound collection device near the main shaft is filtered to obtain the main shaft operation sound wave data.

[0020] The spindle operation acoustic wave data is preprocessed to obtain high-resolution acoustic wave data.

[0021] High-resolution acoustic wave data are analyzed based on residual neural network, and window sliding analysis is used during the analysis process.

[0022] Based on the analysis results, determine whether the spindle has a fault.

[0023] The lathe operation sound wave data is collected by multiple sound collection devices. The multiple sound collection devices are installed in different locations, and the components of the lathe that collect the sound sources are also different. When the lathe is running, each sound-generating location, that is, the component on the lathe that can generate sound, has only one sound collection device within a preset range. An ellipse is drawn with the diagonal as the focus for each component, and the sound collection device should be deployed within the ellipse. No sound collection device is deployed in the overlapping area of ​​the ellipses of different components. Deploying the sound collection device in this way can distinguish different sound-generating locations within the sound collection device, ensuring that when analyzing the sound wave data of the spindle during operation, it is not interfered with by the sounds emitted by other sound-generating locations.

[0024] After the lathe operation sound wave data are collected by multiple sound collection devices, the sound wave data collected by the sound collection device near the main spindle are filtered to retain only the main spindle operation sound wave data.

[0025] The sound wave data collected by the sound acquisition device is Fourier transformed to convert the time domain signal into a frequency domain signal. The frequency domain signal can highlight the characteristics of the sound in terms of frequency. Based on the different sound characteristics, the sound wave data when the spindle is running and the sound wave data when the spindle is not running can be distinguished.

[0026] The frequency domain signals corresponding to the sound wave data collected by the sound collection devices other than the sound collection device corresponding to the main axis are separated from the frequency domain signals corresponding to the sound wave data collected by the sound collection device corresponding to the main axis.

[0027] Set the spindle corresponding sound wave data to , the sound wave data when the spindle is not running is set to ,pass The spindle operation sound wave data can be obtained, where m is the distance attenuation factor.

[0028] Determine the frequency range of the spindle signal and other lathe component signals, isolating the spindle operating acoustic wave data. This reduces interference from other equipment when performing spindle diagnostics during subsequent analysis, effectively filtering out other interfering signals and ensuring accurate data during spindle diagnostics.

[0029] After the spindle operation acoustic wave data is obtained, the spindle operation acoustic wave data is preprocessed to obtain high-resolution acoustic wave data.

[0030] Wavelet transforms are used to further remove noise from the spindle's operating acoustic wave data. Spindle faults can manifest early in the acoustic wave signal as subtle changes. The presence of noise can cause these early fault characteristics to be misjudged or go undetected. De-noising using wavelet transforms can highlight abnormal components in the signal. For example, early pitting corrosion in the spindle bearings generates pulses of a specific frequency in the acoustic wave signal. Removing background noise using wavelet transforms makes these pulses easier to capture and analyze, enabling earlier detection of bearing faults, preventing them from further deteriorating and reducing repair costs and downtime.

[0031] Select an appropriate fitting algorithm, such as polynomial regression or linear regression, to fit the spindle's operating acoustic wave data points, constructing a mathematical model that closely reproduces the spindle's actual operating acoustic wave characteristics and generating fitted acoustic wave data. Based on this, an intelligent algorithm accurately locates differences between the fitted data and the original denoised data. These subtle differences may indicate potential spindle anomalies.

[0032] By using the difference features between the main axis running sound wave data and the fitted sound wave data, the expression of the difference features is expanded, and these difference points are amplified at a specific ratio, so that the originally imperceptible weak fluctuations are clearly displayed, and finally high-resolution sound wave data is obtained.

[0033] After obtaining high-resolution sound wave data, the high-resolution sound wave data is analyzed based on the residual neural network.

[0034] The following operations are performed on the waveform image formed by the high-resolution acoustic wave data.

[0035] First, a window is defined. The window height is user-defined under the condition that it is smaller than the average amplitude of the high-resolution acoustic wave data, and the window width is also user-defined.

[0036] like Figure 2 As shown, the high-resolution acoustic wave data is divided into three categories based on the window size. A Cartesian coordinate system is established, with the x-axis located at the horizontal center axis of the acoustic wave data. The x-axis runs through the entire high-resolution acoustic wave data and divides the high-resolution acoustic wave data into two approximately equal parts.

[0037] Figure 2 As shown in , 1 is the first type of window, 2 is the second type of window, and 3 is the third type of window.

[0038] The first type of window is a window range that is arranged horizontally at positions symmetrical about the x-axis and is mainly used to extract frequency and phase data of high-resolution acoustic wave data.

[0039] The second type of window is the first type of window, shifted upward and arranged horizontally at the same height, and the third type of window is the first type of window, shifted downward and arranged horizontally at the same height. This mainly extracts amplitude information from high-resolution acoustic wave data.

[0040] The segmented image is passed into the residual neural network for subsequent analysis.

[0041] The residual neural network performs convolution, pooling and normalization on the high-resolution sound wave data in each window, and retains the features of the high-resolution sound wave data in each window through the residual block in the residual neural network.

[0042] Residual neural networks use residual blocks to process the complex relationships in high-resolution acoustic wave data. High-resolution acoustic wave data is inherently highly nonlinear. Traditional neural networks, when processing deep networks, experience vanishing or exploding gradients as the number of layers increases, making the network difficult to train. ResNet's residual connections alleviate this problem, allowing the features of the original high-resolution acoustic wave data to be preserved in deep networks.

[0043] The system determines the similarity of high-resolution acoustic wave data within each window. By analyzing the phase, frequency, and amplitude information of the high-resolution acoustic wave data and the differences during spindle operation, it identifies subtle differences in the sound during operation and uses this information to detect potential spindle failures. The system uses the mean square error (MSE) of the features within all windows to determine if the difference exceeds the failure threshold, indicating a spindle failure.

[0044] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A spindle fault diagnosis method based on deep learning, characterized in that: Includes: Collecting lathe operation sound wave data when the spindle is running by multiple sound collection devices, wherein the multiple sound collection devices are installed at different positions, and each sound generating device when the lathe is running has only one sound collection device within a preset range; Filtering the sound wave data collected by the sound collection device near the main shaft to obtain the main shaft operation sound wave data; Pre-process the spindle operation acoustic wave data to obtain high-resolution acoustic wave data; High-resolution acoustic wave data is analyzed based on residual neural networks, and sliding analysis is performed using windows during the analysis process. Based on the analysis results, determine whether the spindle has a fault.

2. The spindle fault diagnosis method based on deep learning according to claim 1, characterized in that: The specific method of filtering the sound wave data collected by the sound collection device near the main shaft to obtain the main shaft operation sound wave data is: Performing Fourier transform on the sound wave data collected by the sound collection device to convert the time domain signal into a frequency domain signal; Separating the frequency domain signals corresponding to the sound wave data collected by the sound collection devices other than the sound collection device corresponding to the main axis from the frequency domain signals corresponding to the sound wave data collected by the sound collection device corresponding to the main axis; Determine the frequency range of the spindle signal and other component signals in the lathe, and separate the spindle operation sound wave data.

3. The spindle fault diagnosis method based on deep learning according to claim 1, characterized in that: The specific method of pre-processing the spindle operation acoustic wave data to obtain high-resolution acoustic wave data is as follows: The noise in the spindle running acoustic wave data is removed by wavelet transform; The spindle running acoustic wave data after noise removal is fitted, the fitted acoustic wave data is selected, and the difference points are amplified to obtain high-resolution acoustic wave data.

4. The spindle fault diagnosis method based on deep learning according to claim 1, characterized in that: The specific method for analyzing high-resolution acoustic wave data based on the residual neural network is: Define a window. The window height is user-defined under the condition that it is smaller than the average amplitude of high-resolution acoustic wave data. The window width is also user-defined. The high-resolution sound wave data is segmented into three categories based on the window size, and the segmented images are passed into the residual neural network; The residual neural network performs convolution, pooling, and normalization on the high-resolution sound wave data in each window, and retains the features of the high-resolution sound wave data in each window through the residual block in the residual neural network; The similarity of the high-resolution acoustic wave data in each window is determined by the mean square error of the features in all windows. If the difference exceeds the fault threshold, the spindle is judged to be faulty.

5. The spindle fault diagnosis method based on deep learning according to claim 4 is characterized in that: The specific method for dividing high-resolution sound wave data into three categories based on window size is: Establish a Cartesian coordinate system with the x-axis located at the horizontal center axis of the acoustic wave data; The first type of window is divided by arranging the window range horizontally at positions symmetrical about the x-axis; The second type of window is the first type of window, which is shifted upwards by the window height and arranged horizontally to split the window. The third type of window is the first type of window, which is shifted downward by the window height and arranged horizontally to split it.

Images