A deep learning-based industrial equipment fault diagnosis prediction method and system

Abstract

The application discloses a kind of industrial equipment fault diagnosis prediction method and system based on deep learning, comprising the following steps: S1, install high-definition shooting equipment beside industrial equipment, video recording is carried out to industrial equipment in running state, electronic image data is obtained, and electronic image is divided into training image set and test image set;S2, carry out noise reduction processing to electronic image, obtain the electronic image after noise reduction;S3, vibration signal feature extraction is carried out to the electronic image after noise reduction, and vibration data is obtained;S4, vibration parameter gradient division and intelligent compensation are carried out to the vibration data collected;S5, the vibration data collected is uploaded and summarized to cloud for data proofreading, and deviation data group is screened out, and production efficiency data features in the deviation data group are analyzed.The application introduces the relationship between vibration and production efficiency, divides according to vibration parameter gradient and compensates production efficiency, selects the industrial equipment with too large production efficiency deviation, determines that the equipment is abnormal equipment, labels the fault, and reminds the staff to check it.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a method and system for industrial equipment fault diagnosis and prediction based on deep learning. Background Technology

[0002] Traditional factory operations mainly rely on people to monitor the status and parameters of equipment. Operators monitor constantly changing data 24 hours a day, making it difficult to detect slow changes in parameters in the early stages of equipment failure. They can only deal with the problem after the equipment fails. Operators face high work pressure and high labor intensity.

[0003] Industrial equipment fault diagnosis requires extremely high professional skills. Only excellent operators can discern equipment abnormalities from thousands of data points and take correct action in the early stages of a fault. This is quite troublesome and difficult to meet the needs of staff. Summary of the Invention

[0004] To address the aforementioned technical problems, this paper provides a method and system for industrial equipment fault diagnosis and prediction based on deep learning. This technical solution solves the problem mentioned in the background that industrial equipment fault inspection requires extremely high professional skills from personnel. Only excellent operators can discern equipment anomalies from thousands of data points and take correct measures in the early stages of a fault, which is quite troublesome.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A deep learning-based method for fault diagnosis and prediction of industrial equipment includes the following steps:

[0007] S1. Install high-definition shooting equipment next to industrial equipment to record video of industrial equipment in operation, acquire electronic image data, and divide the electronic images into training image set and test image set.

[0008] S2, Denoise the electronic image to obtain the denoised electronic image;

[0009] S3, extract vibration signal features from the denoised electronic image to obtain vibration data;

[0010] S4 performs vibration parameter gradient division and intelligent compensation on the collected vibration data;

[0011] S5 uploads and aggregates the collected vibration data to the cloud for data verification, filters out data groups with verification deviations, and analyzes the production efficiency data characteristics of industrial equipment operation within these data groups; and

[0012] S6 identifies and labels anticipated faults in industrial equipment with abnormal data characteristics within the calibration deviation data group.

[0013] Preferably, step S1, which involves dividing the electronic images into a training image set and a test image set, specifically includes the following steps:

[0014] S11, acquire the red, green and blue color components of each pixel in the electronic image;

[0015] S13, obtain the weighted coefficients of human eye sensitivity in red, green and blue colors;

[0016] S15, calculate the gray value of each pixel in the electronic image according to the gray value formula, wherein the gray value range is [0, 255];

[0017] S17, using the grayscale values ​​of each pixel to replace the pixels in the electronic image; and

[0018] S19, divide the electronic images corresponding to all existing industrial equipment in the database into two equal parts, and set them as the training image set and the test image set.

[0019] Preferably, the grayscale value formula is:

[0020] Gray(i,j)=ε1R(i,j)+ε2G(i,j)+ε3B(i,j),

[0021] In the formula, Gray(i,j) is the gray level of the pixel in the i-th row and j-th column of the electronic image, ε1, ε2, and ε3 are the weighting coefficients of human eye sensitivity for red, green, and blue, respectively, and R(i,j), G(i,j), and B(i,j) are the red, green, and blue components of the pixel in the i-th row and j-th column of the electronic image, respectively, where i and j are both positive integers.

[0022] Preferably, step 2, which involves denoising the electronic image to obtain the denoised electronic image, specifically includes the following steps:

[0023] S21, obtain the number of pixels M*N of the electronic images in the training image set;

[0024] S22, use element 0 to fill the blank part of a 2M*2N rectangle;

[0025] S23, use the centralization transformation formula to calculate the centralization function corresponding to the 2M*2N rectangle;

[0026] S24. Perform a discrete Fourier transform on the centered function corresponding to the 2M*2N rectangle to obtain the intermediate function;

[0027] S25, obtain the ideal low-pass filter function;

[0028] S26, using the inverse discrete Fourier transform formula, calculate the output function; and

[0029] S27 extracts the top-left M*N pixels of the image corresponding to the output function and outputs them as the denoised electronic image.

[0030] Preferably, the centralization transformation formula is:

[0031] h(x, y) = (-1) x+y f(x, y),

[0032] In the formula, h(x, y) is the centering function corresponding to the 2M*2N rectangle, f(x, y) is the input function corresponding to the 2M*2N rectangle, and (x, y) are the coordinates of the pixel in the 2M*2N rectangle;

[0033] The intermediate function is:

[0034]

[0035] In the formula, F(u, v) is an intermediate function, j is the imaginary unit, and (u, v) are frequency domain coordinates.

[0036] Preferably, the ideal low-pass filter function is:

[0037]

[0038] In the formula, H(u,v) is the ideal low-pass filter function, d(u,v) is the distance from point (u,v) to the center point of the frequency domain, and r is the preset frequency domain radius;

[0039] The output function is:

[0040]

[0041] In the formula, f′(x, y) is the output function.

[0042] Preferably, step S3, which involves extracting vibration signal features from the denoised electronic image to obtain vibration data, specifically includes the following steps:

[0043] S31, decompose the electronic image frame by frame and convert it to the frequency domain;

[0044] S33, perform power spectrum analysis on the frequency domain image to obtain preliminary location of the vibration area;

[0045] S35: By performing multi-directional frequency filtering on the time-domain signals superimposed on the frequency-domain image in different frames and comparing the vibration intensity of the signals, the vibration direction is determined.

[0046] S37, after filtering the initial vibration region of the frequency domain image according to the vibration direction, the precise vibration region is obtained by comparing the amplitude intensity in the frequency and spatial domains after filtering; and

[0047] S39. By performing power spectrum analysis in the time domain on the phase values ​​of points within the precise vibration region on the frequency domain image of the precise vibration region, the final vibration data is obtained.

[0048] Preferably, the step of performing vibration parameter gradient division and intelligent compensation on the collected vibration data specifically includes the following steps:

[0049] Obtain historical production efficiency data of industrial equipment corresponding to each set of gradient vibration parameter combinations within a set time window, exhaustively enumerate compensation strategies for the historical production efficiency data of all industrial equipment, adopt the same compensation strategy for historical production efficiency data with the same gradient vibration parameter combinations, perform full permutation of strategies for historical production efficiency data with different gradient vibration parameter combinations, and perform comprehensive compensation calculation for each strategy permutation combination.

[0050] The vibration parameters are then incorporated into the compensation strategy algorithm for comprehensive compensation calculation. The specific process is as follows:

[0051] The selected k-th strategy is labeled as s. k Given c strategies and m sets of gradient vibration parameter combinations, there are a total of c m There are several strategy permutations and combinations, where the strategy chosen in the v-th permutation and combination is initially set to zero. m In matrix *m, v is represented by the base c value. Based on this representation, the specific compensation expression for the v-th strategy permutation and combination is constructed:

[0052]

[0053] In the formula, S v This is the total value of current production efficiency data compensation. This is the compensation strategy algorithm function corresponding to the i-th group of gradient vibration parameter combinations, a i It is the combined value of the i-th group of gradient vibration parameters, w i It represents the number of industrial equipment sets that are compensated for the i-th group of gradient vibration parameter combinations;

[0054] Calculate the absolute value of the deviation between the compensated historical production efficiency data and the total production efficiency, select the strategy combination with the smallest absolute value of deviation and mark it as the optimal strategy combination, and establish the correlation between the gradient vibration parameter combination and the optimal compensation strategy based on the gradient vibration parameter combination and the optimal strategy combination.

[0055] Based on the correlation between the gradient vibration parameter combination and the optimal compensation strategy, the corresponding compensation strategy algorithm is selected to compensate the real-time data and then uploaded to the cloud.

[0056] Preferably, step S6, which involves identifying and labeling industrial equipment with abnormal data characteristics within the calibration deviation data group as having anticipated faults, specifically includes the following steps:

[0057] S61, taking a single industrial device as the calculation unit, performs Fourier transform on the data collected by each industrial device in the calibration interval within the calibration deviation data group, calculates the spectrum of the uploaded data of each industrial device, and averages the spectrum values ​​of the uploaded data of all industrial devices at the same frequency to establish a comparison spectrum.

[0058] S63, obtain the reference spectrum of the calibration deviation data group, convert the uploaded data spectrum and the reference spectrum of all industrial equipment in the calibration deviation data group into vectors to calculate the Euclidean distance, perform a Sigmoid function and reciprocal transformation on the calculation result, and obtain the expected failure coefficient of the industrial equipment; and

[0059] S65 sets the expected failure coefficient threshold, marks the abnormal failure risk of industrial equipment with an expected failure coefficient greater than or equal to the expected failure coefficient threshold, and issues an early warning.

[0060] A deep learning-based industrial equipment fault diagnosis and prediction system includes:

[0061] The acquisition module is used to install high-definition shooting equipment next to industrial equipment to record video of industrial equipment in operation, acquire electronic image data, and divide the electronic images into training image set and test image set.

[0062] The data processing module is used to perform noise reduction processing on the electronic image to obtain the noise-reduced electronic image;

[0063] The feature extraction module is used to extract vibration signal features from the denoised electronic image to obtain vibration data;

[0064] The filtering module is used to perform vibration parameter gradient division and intelligent compensation on the collected vibration data, upload and aggregate the collected vibration data to the cloud for data verification, filter out data groups with verification deviations, and analyze the production efficiency data characteristics of industrial equipment operation within the data groups with verification deviations; and

[0065] The fault identification module is used to identify and label expected faults in industrial equipment with abnormal data characteristics within the calibration deviation data group.

[0066] Compared with existing technologies, this invention provides a method and system for fault diagnosis and prediction of industrial equipment based on deep learning, which has the following beneficial effects:

[0067] This invention uses video monitoring of industrial equipment in operation. Deep learning is employed to divide electronic images into training and testing image sets. Noise reduction is performed on the electronic images, and vibration signal features are extracted from the images to obtain vibration data. Since industrial equipment vibration is closely related to production efficiency, this invention introduces the relationship between vibration and production efficiency. Vibration parameters are divided according to gradients, and production efficiency is compensated for. Industrial equipment with excessively high production efficiency is screened out, identified as abnormal equipment, and marked with fault labels to remind staff to inspect it, thus meeting the needs of the staff. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the fault diagnosis and prediction method of the present invention;

[0069] Figure 2 This is a schematic diagram illustrating the method of dividing electronic images into a training image set and a test image set in this invention;

[0070] Figure 3 This is a schematic diagram of the method for obtaining a denoised electronic image in this invention;

[0071] Figure 4 This is a schematic diagram of the method for obtaining vibration data in this invention;

[0072] Figure 5 This is a schematic diagram of the method for identifying and labeling expected faults in industrial equipment with abnormal data characteristics within a calibration deviation data group, as described in this invention.

[0073] Figure 6 This is a block diagram of the deep learning-based industrial equipment fault diagnosis and prediction system of the present invention. Detailed Implementation

[0074] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0075] Please refer to Figures 1-5 As shown, a deep learning-based method for fault diagnosis and prediction of industrial equipment includes the following steps:

[0076] S1. Install high-definition shooting equipment next to industrial equipment to record video of industrial equipment in operation, acquire electronic image data, and divide the electronic images into training image set and test image set.

[0077] S2, Denoise the electronic image to obtain the denoised electronic image;

[0078] S3, extract vibration signal features from the denoised electronic image to obtain vibration data;

[0079] S4 performs vibration parameter gradient division and intelligent compensation on the collected vibration data;

[0080] S5 uploads and aggregates the collected vibration data to the cloud for data verification, filters out data groups with verification deviations, and analyzes the production efficiency data characteristics of industrial equipment operation within these data groups; and

[0081] S6 identifies and labels anticipated faults in industrial equipment with abnormal data characteristics within the calibration deviation data group.

[0082] This invention utilizes video monitoring of industrial equipment in operation. Deep learning is employed to divide electronic images into training and testing image sets. Noise reduction processing is applied to the electronic images, and vibration signal features are extracted from the images to obtain vibration data. Since industrial equipment vibration is closely related to production efficiency, excessive vibration can lead to accelerated equipment wear, increased failure frequency, and reduced uptime. Furthermore, vibration can affect product quality, resulting in higher defect rates, increased scrap and rework costs. Therefore, maintaining equipment within an appropriate vibration range helps improve production efficiency, extend equipment life, and reduce maintenance costs. This invention introduces the relationship between vibration and production efficiency, classifying equipment according to vibration parameter gradients and compensating for production efficiency. Industrial equipment with excessively high production efficiency deviations is identified as abnormal equipment, marked with fault codes, and alerted to staff for inspection, thus meeting the needs of workers.

[0083] Step S1, which involves dividing the electronic images into a training image set and a test image set, specifically includes the following steps:

[0084] S11, acquire the red, green and blue color components of each pixel in the electronic image;

[0085] S13, obtain the weighted coefficients of human eye sensitivity in red, green and blue colors;

[0086] S15, calculate the gray value of each pixel in the electronic image according to the gray value formula, and the gray value range is [0, 255];

[0087] S17, using the grayscale values ​​of each pixel to replace the pixels in the electronic image; and

[0088] S19, divide the electronic images corresponding to all existing industrial equipment in the database into two equal parts, and set them as the training image set and the test image set.

[0089] The pixels in an electronic image are composed of red, green, and blue. By replacing each pixel in the electronic image with its grayscale value, they are brought to the same standard. The image is then divided into a training image set and a test image set to facilitate image analysis and research.

[0090] The formula for grayscale value is:

[0091] Gray(i,j)=ε1R(i,j)+ε2G(i,j)+ε3B(i,j),

[0092] In the formula, Gray(i,j) is the gray level of the pixel in the i-th row and j-th column of the electronic image, ε1, ε2, and ε3 are the weighting coefficients of human eye sensitivity for red, green, and blue, respectively, and R(i,j), G(i,j), and B(i,j) are the red, green, and blue components of the pixel in the i-th row and j-th column of the electronic image, respectively, where i and j are both positive integers.

[0093] The above formula can determine the grayscale of each pixel in an electronic image based on the red, green, and blue color components of each pixel and the weighted coefficients of human eye sensitivity to these three colors. Step S2, denoising the electronic image to obtain the denoised electronic image, specifically includes the following steps:

[0094] S21, obtain the number of pixels M*N of the electronic images in the training image set;

[0095] S22, use element 0 to fill the blank part of a 2M*2N rectangle;

[0096] S23, use the centralization transformation formula to calculate the centralization function corresponding to the 2M*2N rectangle;

[0097] S24. Perform a discrete Fourier transform on the centered function corresponding to the 2M*2N rectangle to obtain the intermediate function;

[0098] S25, obtain the ideal low-pass filter function;

[0099] S26, using the inverse discrete Fourier transform formula, calculate the output function; and

[0100] S27 extracts the top-left M*N pixels of the image corresponding to the output function and outputs them as the denoised electronic image.

[0101] The centralization transformation formula is:

[0102] h(x, y) = (-1) x+y f(x, y),

[0103] In the formula, h(x, y) is the centering function corresponding to the 2M*2N rectangle, f(x, y) is the input function corresponding to the 2M*2N rectangle, and (x, y) are the coordinates of the pixel in the 2M*2N rectangle;

[0104] The intermediate function is:

[0105]

[0106] In the formula, F(u, v) is an intermediate function, j is the imaginary unit, and (u, v) are frequency domain coordinates.

[0107] The ideal low-pass filter function is:

[0108]

[0109] In the formula, H(u,v) is the ideal low-pass filter function, d(u,v) is the distance from point (u,v) to the center point of the frequency domain, and r is the preset frequency domain radius;

[0110] The output function is:

[0111]

[0112] In the formula, f′(x, y) is the output function.

[0113] Specifically, the image processing task is accomplished using the NumPy and SciPy libraries. The code includes steps such as obtaining the pixel count from the image, creating and centering a 2M*2N matrix, calculating the discrete Fourier transform, applying an ideal low-pass filter, performing the inverse transform, and cropping the final image. The Python code framework is shown below:

[0114] import numpy as np

[0115] from scipy.fftpack import fft2,ifft2

[0116] import matplotlib.pyplot as plt

[0117] from scipy import ndimage

[0118] #Assuming you already have a training image, you can use matplotlib's imread function to read it.

[0119] #image=plt.imread('path_to_your_image.jpg')

[0120] #For demonstration purposes, we'll create a simple random image.

[0121] image = np.random.randint(0, 256, (100, 100), dtype = np.uint8) # Example image 100x100;

[0122] # Obtain the number of pixels M*N in the electronic images of the training image set.

[0123] M*N = image.shape

[0124] # Use element 0 to fill the blank areas of a 2M*2N rectangle.

[0125] padded_image=np.zeros((2*M,2*N),dtype=np.uint8)

[0126] padded_image[M / / 2:M / / 2+M,N / / 2:N / / 2+N]=image

[0127] # Calculate the centralization function of a 2M*2N rectangle using the centralization transformation formula.

[0128] centered_image=np.fft.fftshift(padded_image.astype(np.float32),axes=(0,1))

[0129] # Perform a Discrete Fourier Transform on the centered function corresponding to the 2M*2N rectangle to obtain the intermediate function.

[0130] f_image=fft2(centered_image)

[0131] # Obtain the ideal low-pass filter function, assuming the cutoff frequency is D0.

[0132] D0 = 30# cutoff frequency, which can be adjusted.

[0133] [u,v]=np.meshgrid(np.arange(-N,N),np.arange(-M,M))

[0134] D = np.sqrt((uN)**2 + (vM)**2)

[0135] H = D <= D0

[0136] H = H.astype(np.float32)

[0137] #The output function is calculated using the inverse discrete Fourier transform formula.

[0138] F_filtered = f_image * H

[0139] output_image=np.real(ifft2(F_filtered))

[0140] #Centralization is back

[0141] output_image=np.fft.ifftshift(output_image,axes=(0,1))

[0142] # Extract the top-left M*N pixels of the image corresponding to the output function and output it as the denoised electronic image.

[0143] final_image=np.clip(output_image[M / / 2:M / / 2+M,N / / 2:N / / 2+N].astype(np.uint8),0,255)

[0144] # Display the original image and the image after noise reduction

[0145] plt.figure(figsize=(12,6))

[0146] plt.subplot(1,2,1)

[0147] plt.title('Original Image')

[0148] plt.imshow(image,cmap='gray')

[0149] plt.axis('off')

[0150] plt.subplot(1,2,2)

[0151] plt.title('Denoised Image')

[0152] plt.imshow(final_image,cmap='gray')

[0153] plt.axis('off')

[0154] plt.show().

[0155] Step S3 involves extracting vibration signal features from the denoised electronic image to obtain vibration data. This includes the following steps:

[0156] S31, decompose the electronic image frame by frame and convert it to the frequency domain;

[0157] S33, perform power spectrum analysis on the frequency domain image to obtain preliminary location of the vibration area;

[0158] S35: By performing multi-directional frequency filtering on the time-domain signals superimposed on the frequency-domain image in different frames and comparing the vibration intensity of the signals, the vibration direction is determined.

[0159] S37, after filtering the initial vibration region of the frequency domain image according to the vibration direction, the precise vibration region is obtained by comparing the amplitude intensity in the frequency and spatial domains after filtering; and

[0160] S39. By performing power spectrum analysis in the time domain on the phase values ​​of points within the precise vibration region on the frequency domain image of the precise vibration region, the final vibration data is obtained.

[0161] Specifically, to achieve the above steps, we need to perform a series of processing steps on the electronic image, including frame decomposition, frequency domain transformation, power spectrum analysis, time domain signal superposition, multi-directional frequency filtering, vibration direction determination, precise vibration region localization, and final vibration data analysis. Below is a simplified Python code framework:

[0162]

[0163]

[0164]

[0165] Step S4 involves performing vibration parameter gradient partitioning and intelligent compensation on the collected vibration data, specifically including the following steps:

[0166] Obtain historical production efficiency data of industrial equipment corresponding to each set of gradient vibration parameter combinations within a set time window, exhaustively enumerate compensation strategies for the historical production efficiency data of all industrial equipment, adopt the same compensation strategy for historical production efficiency data with the same gradient vibration parameter combinations, perform full permutation of strategies for historical production efficiency data with different gradient vibration parameter combinations, and perform comprehensive compensation calculation for each strategy permutation combination.

[0167] The vibration parameters are then incorporated into the compensation strategy algorithm for comprehensive compensation calculation. The specific process is as follows:

[0168] The selected k-th strategy is labeled as s. k Given c strategies and m sets of gradient vibration parameter combinations, there are a total of c m There are several strategy permutations and combinations, where the strategy chosen in the v-th permutation and combination is initially set to zero. m In matrix *m, v is represented by the base c value. Based on this representation, the specific compensation expression for the v-th strategy permutation and combination is constructed:

[0169]

[0170] In the formula, S v This is the total value of current production efficiency data compensation. This is the compensation strategy algorithm function corresponding to the i-th group of gradient vibration parameter combinations, a i It is the combined value of the i-th group of gradient vibration parameters, w i It represents the number of industrial equipment sets that are compensated for the i-th group of gradient vibration parameter combinations;

[0171] Calculate the absolute value of the deviation between the compensated historical production efficiency data and the total production efficiency, select the strategy combination with the smallest absolute value of deviation and mark it as the optimal strategy combination, and establish the correlation between the gradient vibration parameter combination and the optimal compensation strategy based on the gradient vibration parameter combination and the optimal strategy combination.

[0172] Based on the correlation between the gradient vibration parameter combination and the optimal compensation strategy, the corresponding compensation strategy algorithm is selected to compensate the real-time data and then uploaded to the cloud.

[0173] By classifying equipment according to vibration parameter gradients and compensating for production efficiency, industrial equipment with excessively high production efficiency was identified as abnormal equipment. This equipment was marked as faulty and staff were reminded to check it, thus meeting the staff's needs.

[0174] The step S5, which involves uploading and summarizing the collected vibration data to the cloud for data verification and filtering out the data groups with verification deviations, specifically includes the following steps:

[0175] Calculate the deviation ratio between the cumulative value of the production efficiency data set by the industrial equipment and the real-time uploaded data, set a deviation ratio threshold, filter out data with a deviation ratio greater than or equal to the set deviation ratio threshold, and mark them as data groups for calibration deviation.

[0176] Step S6 involves identifying and labeling industrial equipment with abnormal data characteristics within the calibration deviation data group for anticipated faults. This step specifically includes the following steps:

[0177] S61, taking a single industrial device as the calculation unit, performs Fourier transform on the data collected by each industrial device in the calibration interval within the calibration deviation data group, calculates the spectrum of the uploaded data of each industrial device, and averages the spectrum values ​​of the uploaded data of all industrial devices at the same frequency to establish a comparison spectrum.

[0178] S63, obtain the reference spectrum of the calibration deviation data group, convert the uploaded data spectrum and the reference spectrum of all industrial equipment in the calibration deviation data group into vectors to calculate the Euclidean distance, perform a Sigmoid function and reciprocal transformation on the calculation result, and obtain the expected failure coefficient of the industrial equipment; and

[0179] S65 sets the expected failure coefficient threshold, marks the abnormal failure risk of industrial equipment with an expected failure coefficient greater than or equal to the expected failure coefficient threshold, and issues an early warning.

[0180] Specifically, here is a simplified Python code framework:

[0181] import numpy as np

[0182] from scipy.fftpack import fft

[0183] from sklearn.preprocessing import MinMaxScaler

[0184] #Assume we have the following data:

[0185] #device_efficiency_data: A dictionary where the key is the device ID and the value is a list of set productivity data.

[0186] #real_time_data: A dictionary where the key is the device ID and the value is a list of data uploaded in real time (corresponding to the set data time points).

[0187]

[0188]

[0189]

[0190] This application also discloses a deep learning-based industrial equipment fault diagnosis and prediction system, comprising:

[0191] The acquisition module 10 is used to install a high-definition shooting device next to the industrial equipment, record video of the industrial equipment in operation, acquire electronic image data, and divide the electronic images into a training image set and a test image set.

[0192] Data processing module 30, the data processing module 30 is used to perform noise reduction processing on electronic images to obtain noise-reduced electronic images;

[0193] Feature extraction module 50 is used to extract vibration signal features from the denoised electronic image to obtain vibration data;

[0194] The filtering module 70 is used to perform vibration parameter gradient division and intelligent compensation on the collected vibration data, upload and aggregate the collected vibration data to the cloud for data verification, filter out the verification deviation data group, and analyze the production efficiency data characteristics of industrial equipment operation within the verification deviation data group; and

[0195] The fault identification module 90 is used to identify and label expected faults in industrial equipment with abnormal data characteristics within the calibration deviation data group.

[0196] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for fault diagnosis and prediction of industrial equipment based on deep learning, characterized in that, Including the following steps: S1. Install high-definition shooting equipment next to industrial equipment to record video of industrial equipment in operation, acquire electronic image data, and divide the electronic images into training image set and test image set. S2, Denoise the electronic image to obtain the denoised electronic image; S3, extract vibration signal features from the denoised electronic image to obtain vibration data; S4 performs vibration parameter gradient division and intelligent compensation on the collected vibration data; S5 uploads and aggregates the collected vibration data to the cloud for data verification, filters out the verification deviation data group, and analyzes the production efficiency data characteristics of industrial equipment operation within the verification deviation data group. as well as S6, identify and label expected faults in industrial equipment with abnormal data characteristics within the calibration deviation data group; Step S4 involves performing vibration parameter gradient division and intelligent compensation on the collected vibration data, specifically including the following steps: Obtain historical production efficiency data of industrial equipment corresponding to each set of gradient vibration parameter combinations within a set time window, exhaustively enumerate compensation strategies for the historical production efficiency data of all industrial equipment, adopt the same compensation strategy for historical production efficiency data with the same gradient vibration parameter combinations, perform full permutation of strategies for historical production efficiency data with different gradient vibration parameter combinations, and perform comprehensive compensation calculation for each strategy permutation combination. The vibration parameters are then incorporated into the compensation strategy algorithm for comprehensive compensation calculation. The specific process is as follows: The selected k-th strategy is labeled as s. k Given c strategies and m sets of gradient vibration parameter combinations, there are a total of c m There are several strategy permutations and combinations, where the strategy chosen in the v-th permutation and combination is initially set to zero. m In matrix *m, v is represented by the base c value. Based on this representation, the specific compensation expression for the v-th strategy permutation and combination is constructed: In the formula, S v This is the total value of current production efficiency data compensation. This is the compensation strategy algorithm function corresponding to the i-th group of gradient vibration parameter combinations, a i It is the combined value of the i-th group of gradient vibration parameters, w i It represents the number of industrial equipment sets that are compensated for the i-th group of gradient vibration parameter combinations; Calculate the absolute value of the deviation between the compensated historical production efficiency data and the total production efficiency, select the strategy combination with the smallest absolute value of deviation and mark it as the optimal strategy combination, and establish the correlation between the gradient vibration parameter combination and the optimal compensation strategy based on the gradient vibration parameter combination and the optimal strategy combination. Based on the correlation between the gradient vibration parameter combination and the optimal compensation strategy, the corresponding compensation strategy algorithm is selected to compensate the real-time data and then uploaded to the cloud.

2. The industrial equipment fault diagnosis and prediction method based on deep learning according to claim 1, characterized in that, The step S1, which involves dividing the electronic images into a training image set and a test image set, specifically includes the following steps: S11, acquire the red, green and blue color components of each pixel in the electronic image; S13, obtain the weighted coefficients of human eye sensitivity in red, green and blue colors; S15, calculate the gray value of each pixel in the electronic image according to the gray value formula, wherein the gray value range is [0, 255]; S17, using the grayscale value of each pixel to replace each pixel in the electronic image; as well as S19, divide the electronic images corresponding to all existing industrial equipment in the database into two equal parts, and set them as the training image set and the test image set.

3. The industrial equipment fault diagnosis and prediction method based on deep learning according to claim 2, characterized in that, The formula for grayscale value is: Gray(i,j)=ε1R(i,j)+ε2G(i,j)+ε3B(i,j), In the formula, Gray(i,j) is the gray level of the pixel in the i-th row and j-th column of the electronic image, ε1, ε2, and ε3 are the weighting coefficients of human eye sensitivity for red, green, and blue, respectively, and R(i,j), G(i,j), and B(i,j) are the red, green, and blue components of the pixel in the i-th row and j-th column of the electronic image, respectively, where i and j are both positive integers.

4. The industrial equipment fault diagnosis and prediction method based on deep learning according to claim 3, characterized in that, Step S2 involves denoising the electronic image to obtain the denoised electronic image, specifically including the following steps: S21, obtain the number of pixels M*N of the electronic images in the training image set; S22, use element 0 to fill the blank part of a 2M*2N rectangle; S23, use the centralization transformation formula to calculate the centralization function corresponding to the 2M*2N rectangle; S24. Perform a discrete Fourier transform on the centered function corresponding to the 2M*2N rectangle to obtain the intermediate function; S25, obtain the ideal low-pass filter function; S26. The output function is calculated using the inverse discrete Fourier transform formula. as well as S27 extracts the top-left M*N pixels of the image corresponding to the output function and outputs them as the denoised electronic image.

5. The industrial equipment fault diagnosis and prediction method based on deep learning according to claim 4, characterized in that, The centralization transformation formula is: h(x,y)=(-1) x+y f(x,y), In the formula, h(x, y) is the centering function corresponding to the 2M*2N rectangle, f(x, y) is the input function corresponding to the 2M*2N rectangle, and (x, y) are the coordinates of the pixel in the 2M*2N rectangle; The intermediate function is: In the formula, F(u, v) is an intermediate function, j is the imaginary unit, and (u, v) are frequency domain coordinates.

6. The industrial equipment fault diagnosis and prediction method based on deep learning according to claim 4, characterized in that, The ideal low-pass filter function is: In the formula, H(u,v) is the ideal low-pass filter function, d(u,v) is the distance from point (u,v) to the center point of the frequency domain, and r is the preset frequency domain radius; The output function is: In the formula, f′(x, y) is the output function.

7. The industrial equipment fault diagnosis and prediction method based on deep learning according to claim 6, characterized in that, Step S3 involves extracting vibration signal features from the denoised electronic image to obtain vibration data, specifically including the following steps: S31, decompose the electronic image frame by frame and convert it to the frequency domain; S33, perform power spectrum analysis on the frequency domain image to obtain preliminary location of the vibration area; S35: By performing multi-directional frequency filtering on the time-domain signals superimposed on the frequency-domain image in different frames and comparing the vibration intensity of the signals, the vibration direction is determined. S37, after filtering the initial vibration region of the frequency domain image according to the vibration direction, the precise vibration region is obtained by comparing the amplitude intensity in the frequency and spatial domains after filtering; and S39. By performing power spectrum analysis in the time domain on the phase values ​​of points within the precise vibration region on the frequency domain image of the precise vibration region, the final vibration data is obtained.

8. The industrial equipment fault diagnosis and prediction method based on deep learning according to claim 1, characterized in that, Step S6 involves identifying and labeling industrial equipment with abnormal data characteristics within the calibration deviation data group for anticipated faults. This step specifically includes the following steps: S61, taking a single industrial device as the calculation unit, performs Fourier transform on the data collected by each industrial device in the calibration interval within the calibration deviation data group, calculates the spectrum of the uploaded data of each industrial device, and averages the spectrum values ​​of the uploaded data of all industrial devices at the same frequency to establish a comparison spectrum. S63, obtain the reference spectrum of the calibration deviation data group, convert the uploaded data spectrum and the reference spectrum of all industrial equipment in the calibration deviation data group into vectors to calculate the Euclidean distance, perform Sigmoid function and reciprocal conversion on the calculation results, and obtain the expected failure coefficient of the industrial equipment. as well as S65 sets the expected failure coefficient threshold, marks the abnormal failure risk of industrial equipment with an expected failure coefficient greater than or equal to the expected failure coefficient threshold, and issues an early warning.

9. A deep learning-based industrial equipment fault diagnosis and prediction system, used to implement the deep learning-based industrial equipment fault diagnosis and prediction method as described in any one of claims 1-8, characterized in that, include: The acquisition module (10) is used to install a high-definition shooting device next to the industrial equipment, record video of the industrial equipment in operation, acquire electronic image data, and divide the electronic images into a training image set and a test image set. The data processing module (30) is used to perform noise reduction processing on the electronic image to obtain the noise-reduced electronic image; The feature extraction module (50) is used to extract vibration signal features from the denoised electronic image to obtain vibration data; The filtering module (70) is used to perform vibration parameter gradient division and intelligent compensation on the collected vibration data, upload and summarize the collected vibration data to the cloud for data verification, filter out the verification deviation data group, and analyze the production efficiency data characteristics of industrial equipment operation within the verification deviation data group. as well as The fault identification module (90) is used to identify and label the expected faults of industrial equipment with abnormal data characteristics in the calibration deviation data group.

Images