A video data processing method and a gearbox RUL prediction method

Abstract

The present invention discloses a video data processing method and a gearbox RUL prediction method, which relate to the fields of computer vision technology and mechanical fault diagnosis. The method comprises the following steps: constructing an initial background image through shallow clustering, and correcting the background image by increasing the number of cluster centers to obtain an approximate background image of a real scene; then defining the boundary of a moving object and reducing the similarity between a defocused blur area and the background; then determining a reconstruction time node and constructing an adjacent frame matrix, performing long-axis quantization frame difference on the adjacent frame matrix to obtain a valid information set of sedimentary materials; finally, calculating the mean of the corresponding valid information in the adjacent frame matrix and reconstructing it in the background image; dividing the reconstructed feature map into a training set and a validation set according to a certain ratio, and then sending the result into a convolutional neural network training prediction model to achieve a gearbox RUL prediction task.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision technology and mechanical fault diagnosis, and in particular to a video data processing method and a gearbox RUL prediction method. Background Art

[0002] With the rapid development of artificial intelligence, computer vision has been widely used in industrial scenarios such as defect detection, fault diagnosis, and health monitoring. Currently, it is possible to complete specific tasks by analyzing video or image data collected from industrial sites. However, the initial data obtained from industrial sites often contains a lot of noise, such as complex backgrounds and image distortion, which dilutes the effective information. Before meeting industrial needs based on computer vision, it is often desirable to process data with distinct features. Therefore, it is important to reconstruct the features of monitored objects in a clean background under complex backgrounds, thereby improving the upper limit of subsequent task completion.

[0003] Taking gearbox wear monitoring as an example, gearbox wear monitoring initially involved analyzing the mechanical structure of the gears, with mathematical modeling and analysis based on the wear mechanism. This approach suffers from poor portability and numerous unstable factors, making it difficult to achieve stability in practical applications. With the development of condition-based maintenance (CBM) technology, researchers began using vibration sensors to collect vibration signals generated by gearboxes during operation and to evaluate and predict the gearbox wear state through time-domain, frequency-domain, and wavelet analysis. However, the vibration signals are contaminated by a significant amount of environmental noise, and the data features are highly abstract, resulting in poor performance in RUL prediction tasks. With the advent of the On-Line Visual Ferrograph (OLVF) device, researchers introduced computer vision to gearbox wear monitoring. However, the video data captured by the OLVF device contains a large number of bubbles, as well as motion interference such as blurring caused by defocusing of wear debris and bubbles in the foreground. This dilutes the actual, effective wear debris information. Therefore, the first aspect of the present invention aims to address the problem of effective information in video data being diluted by the complex background. Furthermore, currently, videos captured through OLVF are primarily used to monitor the particle size of larger wear debris for fault classification, lacking a method directly related to gearbox RUL prediction. Compared to vibration signals, video data has a more distinct physical meaning and is more closely correlated with the wear process. To this end, the second aspect of this invention proposes a method for predicting gearbox RUL based on video data. Summary of the Invention

[0004] The present invention aims to provide a method for processing video data and a method for predicting the RUL of a gearbox, which at least solves the problem that effective information in video data is diluted by complex background. Furthermore, a method for predicting the RUL of a gearbox based on video data is provided.

[0005] To achieve the above objectives, the first aspect of the present application provides a video data processing method, which is a method for reconstructing the features of time-sequential deposited materials in a video under a complex background. Based on the effective information of consecutive adjacent frames of the video, the method distinguishes and clusters moving targets and non-moving targets through inter-tensor operations, obtains effective information through intra-class operations, and reconstructs it in a clean background. The method specifically includes the following steps:

[0006] Step 1: Obtain a primary background image by shallow clustering the initial video frame set of the video data, then increase the number of cluster centers to obtain a revised set of the primary background image, and use the data of the revised set to revise the primary background image and construct the video background image.

[0007] As a possible implementation method, the primary background image and the correction set in step 1 are tensors obtained by selecting 1 and 3 cluster centers using the K-mean algorithm, respectively.

[0008] Step 2: Calculate the grayscale difference between the video data and the reconstructed background image frame by frame, and replace the defocused blurred area and the original background area with the reconstructed new background.

[0009] Step 3: Extract and convert label information from the video number to obtain relevant information such as time, speed, and load conditions.

[0010] Step 4: Select the reconstruction time node for reconstructing the data with constant time and constant flow as the conditions based on the time, speed, load conditions and other information of the video matching, and locate the reconstruction node to the specific frame according to the refresh rate of the video data.

[0011] Step 5: With the reconstruction node as the center, select multiple consecutive adjacent frames to construct an adjacent frame matrix, and select a quantization factor to quantize the adjacent frame matrix.

[0012] Step 6: Calculate the mode value of the adjacent frame matrix in the sample dimension after quantization, and record the position of the corresponding sample as the valid information set. Select the data at the position in the valid information set corresponding to the adjacent frame matrix before quantization, calculate the mean value in the sample dimension to represent the wear debris information of this dimension, and reconstruct it in the background image.

[0013] Based on the above method, the second aspect of the present application provides a method for predicting the RUL of a gearbox, which further includes the following steps:

[0014] Step 7: Reconstruct the feature map of all video data using the methods in steps 2 to 6.

[0015] Step 8: Obtain traditional wear debris characteristic indicators through mathematical modeling.

[0016] Step 9: Divide the feature map reconstructed in step 7 into a training set and a validation set, and train the deep learning network to achieve RUL prediction of the gearbox.

[0017] As a possible implementation method, the training set and validation set are divided into 7:3 in step 9, the AlexNet network is used, the MSE loss function is selected, and SGD is used to update the network parameters.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] (1) The secondary background reconstruction method proposed in this invention constructs a correction set based on the prior knowledge of background interference to correct the “primary background image” in a similar way to feedforward compensation. The obtained background image is closer to the true value than the traditional method.

[0020] (2) The wear particle feature reconstruction method proposed in the present invention has strong robustness and can be used to reconstruct the wear particle features of videos collected at different wear stages. It also shows very strong suppression ability in terms of occlusion and motion interference.

[0021] (3) The video-based gearbox RUL prediction method proposed in this paper addresses the lack of video data in gearbox RUL prediction tasks and provides a new approach to gearbox RUL prediction. In addition, the gearbox RUL prediction model proposed in this paper does not require the use of wear mechanisms or empirical knowledge when constructing it, making it easier to promote in actual industrial scenarios.

[0022] Experiments on the collected video data show that the reconstruction algorithm of the present invention can effectively suppress the influence of moving objects such as bubbles and defocus blur in the foreground on the characteristic information of deposited wear debris. The reconstructed feature map retains the basic mathematical characteristics of the wear debris and has achieved good results in the RUL estimation of monitored objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a structural schematic diagram of the OLVF device of the present invention.

[0024] Figure 2 This is a flow chart of wear debris feature reconstruction according to the present invention.

[0025] Figure 3 This is a flowchart of the secondary background reconstruction of the present invention.

[0026] Figure 4 This is a convolutional neural network training flow chart of the present invention.

[0027] Figure 5 It is the intermediate result diagram of the WDFR process of the present invention; among them, (a) original video frame, (b) secondary background reconstruction result diagram, (c) original video frame after background reconstruction, and (d) feature diagram of wear debris feature reconstruction result.

[0028] Figure 6 This is a flow chart of the video-based gearbox RUL prediction method of the present invention.

[0029] Figure 7 This is a comparison chart of the gearbox RUL prediction results based on the video of the present invention.

[0030] Figure 1 Among them, gear box lubricating oil bypass 1, monitoring branch 2, valve 3, monitoring port 4, light source 5, electromagnet 6, video collector 7. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0032] It should be noted that the words first, second, third and similar terms in the description and claims of this application are only used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that the specific order or sequence can be interchanged where permitted so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein.

[0033] In the following description, the numbers representing the steps, such as S301, S302, etc., do not necessarily mean that the steps must be executed in this manner. If permitted, the order of the steps can be interchanged, or the steps can be executed simultaneously.

[0034] The term "comprising" as used in the specification and claims should not be construed as limiting what is listed thereafter; it does not exclude other components or steps. Thus, it should be interpreted as specifying the presence of the features, integers, steps, or components mentioned, but not excluding the presence or addition of one or more other features, integers, steps, or components, or groups thereof. Thus, the expression "a device comprising unit A and unit B" should not be limited to a device consisting solely of unit A and unit B.

[0035] References in this specification to "some embodiments" or "an embodiment" mean that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least some embodiments of the present invention. Therefore, the phrases "in some embodiments" or "in an embodiment" appearing throughout this specification do not necessarily refer to the same embodiment, but may do so. Furthermore, in one or more embodiments, the particular features, structures, or characteristics may be combined in any suitable manner, as would be apparent to one of ordinary skill in the art from this disclosure.

[0036] Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which this application belongs. In the event of any inconsistency, the meaning described in this specification or the meaning derived from the contents recorded in this specification shall prevail. In addition, the terms used in the present invention are only for the purpose of describing the embodiments of the present application and are not intended to limit this application.

[0037] First, the method for obtaining video data of this application is as follows: In order to realize the RUL prediction of gearbox wear, the On-Line Visual Ferrograph (abbreviated as OLVF) device is used to obtain the required video data. The structural diagram of the OLVF device is as follows: Figure 1 As shown, a monitoring branch 2 is connected to the gearbox lubricating oil bypass 1. This branch is equipped with a valve 3 and a transparent monitoring port 4, through which video of the lubricating oil flowing through the monitoring branch can be captured. Below the monitoring port are a light source 5 and an electromagnet 6, and above it is a video capture device 7 (CMOS). The OLVF device works by using the video capture device to record the deposition process of metal debris captured by the applied magnetic field in the monitoring branch under a constant external magnetic field.

[0038] The working process of the OLVF device includes the following steps:

[0039] 1. Turn off the power supply to the electromagnet, open the valve, and flush the abrasive residue on the wall of the monitoring branch pipe;

[0040] 2. Turn on the light source and electromagnet power supply;

[0041] 3. Turn on the video collector to record data. After 29 seconds, turn off the video collector and save the data. Turn off the electromagnet and light source.

[0042] A total of 5461 video data were collected through the OLVF device, and the obtained video refresh rate was 24 frames per second, with a single video length of 29 seconds. The data acquisition time used in the present invention started at 13:38:37 on December 5, 2020, and ended at 16:27:20 on December 25, 2020, which lasted 20 days. During the initial 0 seconds to 263951 seconds, the gear operated at 1 load condition to debug the OLVF device, and the wear process was relatively slow. From 265700 seconds to the end of the experiment, the wear experiment was carried out by slowly increasing the load from 1 to 6. During the experiment, the gear speed was increased according to experience under a constant load of 6 to accelerate wear, and the gear speed was lowered in the later stage of the experiment to ensure experimental safety.

[0043] The video data used in this invention is collected during the whole life cycle of a new gearbox from operation to serious wear. The video collected by the OLVF device is grayscale video. The specific video frames are as follows: Figure 3 As shown in (a), the wear debris in the video frame is valid information, but the video data is mixed with a large number of bubbles. Furthermore, the wear debris and bubbles in the foreground are blurred due to defocusing, which dilutes the actual information of the deposited wear debris. Therefore, the first aspect of the present invention aims to solve the problem of the effective information in the video data being diluted by the complex background.

[0044] In order to solve the above technical problems, the present invention reconstructs the feature map of video data through the wear chip feature reconstruction method (abbreviated as WDFR), and then obtains the wear chip concentration index (IDC) and wear chip area index (IPCA) characteristics of the wear chips in the feature map through mathematical analysis methods. After that, the feature map is divided into training set and validation set in a certain proportion to train the convolutional neural network, thereby realizing the RUL prediction task of the gearbox.

[0045] For details, please refer to Figure 2 It is understood that the wear debris feature reconstruction method in the present invention is as follows: first, the video background image is extracted by the secondary background reconstruction method, and then the defocused blur area is defined by the pixel difference between the background image and the original video frame. After that, the reconstruction time node and the corresponding video frame are obtained by converting the label information, and finally, the adjacent frame matrix is constructed to obtain the wear debris information and reconstruct the feature map by the long-axis quantization frame difference method.

[0046] For further information, please refer to Figure 3It is understood that the secondary background reconstruction method of the present application is as follows: in the present invention, due to the acquisition characteristics of the OLVF device, there is less interference at the beginning of each video; therefore, in the present invention, n videos are selected from the acquired 5461 video data, specifically 1000 videos are selected; the first m frames of each selected video are intercepted, specifically the first 10 frames can be intercepted; the video frames intercepted from all the selected videos are stacked together to obtain a tensor of h×w×(n·m), where h is the number of pixels in the vertical direction of the video frame, and w is the number of pixels in the horizontal direction of the video frame; the stacked tensor is subjected to the K-mean algorithm in the last dimension.

[0047] Due to the presence of wear debris and bubbles in the video frame, the background image obtained by direct clustering is often darker than the real background, and the pixel grayscale values in the video frame can be divided into the following categories: bubble reflective highlight area, bubble edge and wear debris low grayscale area, and background area. Therefore, this application uses 1 and 3 cluster centers to calculate the clustering results respectively. The clustering result obtained when the number of cluster centers is selected as 1 is defined as the primary background image, and its grayscale value is darker than the real background image; the clustering result obtained when the number of cluster centers is selected as 3 is defined as the correction set; the pixel difference between the correction set and the primary background image is compared, and the center value of the three corresponding centers of the correction set that is smaller than and closest to the pixel value of the corresponding position of the primary background image is selected to construct the background image.

[0048] The above correction method can be expressed by the following formula 1:

[0049]

[0050] In formula 1, f bg That is the background image obtained by secondary reconstruction, Ω1 and Ω3 are the results of selecting 1 and 3 cluster centers respectively, th is the revision threshold, specifically selected as 3, and correction will be used if it is smaller than the threshold.

[0051] The background image obtained by the above method does not contain any information. Next, define the out-of-focus blur area. The specific method is as follows:

[0052] Since bubbles and uncaptured wear debris will cause defocus blur of the camera when passing through the foreground area, these blurs are often similar to the background grayscale. Due to the universality and severity of defocus interference, the present invention proposes a method for defining the defocus blur area.

[0053] First, calculate the grayscale difference between the original video frame and the constructed background image, which can be expressed by the following formula 2:

[0054] de(x i ,y j )=|f(x i ,y j )-f bg (x i ,yj )|

[0055] (Formula 2)

[0056] In Formula 2, de is the frame difference result, and f is the original video frame.

[0057] Then, based on the grayscale difference calculation results, the pixels at each position in the original video frame are classified. The areas with grayscale differences less than the threshold T are identified as background areas, and the areas with grayscale differences greater than the threshold T are identified as bubbles or wear debris areas. T is the threshold, which can be selected as 15. This can be expressed as the following formula 3:

[0058]

[0059] Afterwards, the pixel values identified as background areas in the classification results are replaced with the pixel values of the corresponding coordinates in the reconstructed background image. The above operation can be expressed as follows:

[0060]

[0061] In Formula 4, Y is the background reconstruction result image after suppressing defocus blur of the original video frame.

[0062] Next, we construct the adjacent frame matrix: To take advantage of the obvious physical significance of video data, we construct the adjacent frame matrix based on the two points of fixed deposition time and constant flow. The reconstruction time node selection rule is shown in Formula 5:

[0063]

[0064] In formula 5, t0 is the time node selected based on constant deposition time, t1 is the reconstruction node selected based on constant flow rate, and t max The maximum length of the video. is the current gear speed, is the maximum gear speed.

[0065] Taking constant traffic as an example, the corresponding video frame k1 of feature reconstruction satisfies Formula 6.

[0066] k1=t1×fps

[0067] (Equation 6)

[0068] In Formula 6, fps is the refresh rate of the video, which is 24 frames per second in the present invention.

[0069] With the k1 frame as the center, the construction of the adjacent frame matrix ∑ satisfies Formula 7.

[0070] ∑(x i ,y j , z k )=∪Υn , n∈z∩[k1-num÷2,k1+num÷2]

[0071] (Equation 7)

[0072] In the present invention, num in Formula 7 is 50.

[0073] Next, we implement the long-axis quantization frame difference method. Because the motion of the object causes distortions such as refraction and scattering of the light source, the grayscale of the deposited wear particles fluctuates within a small range. Directly performing frame difference processing will reduce the weight of the effective information of interest. Therefore, we first need to quantize the adjacent frame matrices using the rules of Formula 8.

[0074]

[0075] ∑ in formula 8 q is the result after quantization, round() is the rounding function, σ is the quantization factor, and σ=3 can be selected.

[0076] Then, the valid information position set is determined by voting in the adjacent frame matrix. The voting rule is shown in Formula 9.

[0077] Ω=∪(x i ,y j , z k ),∑ q (x i ,y j , z k )=mode(∑ q (x i ,y j , z k ))

[0078] (Equation 9)

[0079] In formula 9, Ω is the effective wear debris information corresponding to ∑ q The set of positions in the middle, mode() is the function to find the mode.

[0080] Finally, the effective information at the corresponding set Ω position in the adjacent frame matrix before quantization is used for clustering to reduce the data loss caused by quantization and construct the final feature map. The construction principle is shown in Formula 10.

[0081]

[0082] In formula 10 is the result of feature reconstruction of the original video at time t1, and mean() is the mean function.

[0083] Next, the characteristic graph properties are analyzed as follows:

[0084] The traditional iron chip deposition chart has two mathematical indicators: IPCA and IDC, which are defined as follows:

[0085]

[0086]

[0087] IDC=max(C(t))

[0088] (Equation 13)

[0089] Among them A i represents the pixels identified as wear particles; w and h are the sizes of the acquired video frame; ρ is the density of the wear particles (g / cm - 3); R is the radius of the simplified spherical wear debris (μm); ξ is the optical magnification factor of the OLVF device; η is the debris capture efficiency of the OLVF device; Q is the flow rate of the deposition process (L / min -1 ); C(t) is the wear debris concentration at time t (mg / L -1 ).

[0090] This paper will derive a method for extracting IPCA and IDC from feature graphs, and propose a method for extracting abstract features from feature graphs using convolutional neural networks to achieve gearbox RUL prediction. Figure 4 、 Figure 6 understand.

[0091] The IPCA and IDC extraction methods in the feature graph are as follows:

[0092] Regarding the calculation of the IPCA index, wear debris can be directly identified by the feature map and the reconstructed background map using Formula 14, and then calculated using Formula 11.

[0093]

[0094] Regarding the calculation of the IDC index, since the information of multiple consecutive frames is integrated during the reconstruction of the feature map, it is difficult to calculate the instantaneous wear particle concentration through adjacent frames. Therefore, the present invention uses the average concentration to approximate the IDC, which is defined as Formula 15:

[0095]

[0096] Substituting the expression of C(t) in Equation 12 and proposing the time-independent integral term, we can express it as Equation 16:

[0097]

[0098] In formula 16, Q is the flow rate during the deposition process and satisfies Where k is a constant coefficient, is the gear speed, r is the pitch circle radius, and it is assumed that there is no time-varying characteristic within the same precipitation cycle, then formula 17 can be obtained:

[0099]

[0100] Based on the characteristics of video data acquisition, it is determined that no deposition occurs at time 0 of the video, and the average wear debris concentration index is finally expressed by formula 18:

[0101]

[0102] Next, the present invention proposes a gearbox RUL prediction algorithm based on WDFR as follows:

[0103] The feature maps reconstructed from video frames in this paper can be directly applied to the gearbox RUL prediction task after feature extraction through a convolutional neural network. AlexNet is selected as the convolutional neural network. Due to its relatively simple structure, the prediction effect is more dependent on the quality of the feature maps. The parameters of the convolution kernels at each layer are as follows:

[0104]

[0105] During model training, the true remaining life of the gears is calculated using a linear degradation metric based on the time labels of the feature maps. The optimizer's initial learning rate is set to 0.001, decreasing to 0.5 every five epochs. The momentum parameter is set to 0.9. The mean square error (MSE) is used as the loss function to represent model performance. The network parameters are updated using the stochastic gradient descent (SGD) algorithm. During training, the training set is reshuffled before each epoch, and the batch size is set to 1. Each batch reads an additional 25 feature maps sequentially following the time label of the current feature map. The median of the output is used for gradient updates to avoid gradient explosions caused by undesirable data fluctuations in OLVF acquisition.

[0106] Wear debris deposits in videos captured by the OLVF device are independent of each other, and wear particle deposition in adjacent time periods is approximately identically distributed. The feature atlas obtained by feature reconstruction is divided into training and validation sets in a 7:3 ratio and fed into the network to train regression weights and predict the wear progression of gears. After 10 epochs of training, the model reached convergence.

[0107] Experimental results and analysis

[0108] The original video undergoes WDFR feature reconstruction process as follows Figure 5As shown in the figure. (a) is the frame extraction result of the original video data, which contains a lot of interference factors on the wear debris information; (b) is the background image obtained by secondary background reconstruction; (c) is the result after background reconstruction of the original video frame; (d) is the final generated wear debris feature reconstruction result. The final result of the gearbox RUL prediction is shown in the figure. Figure 7 As shown. The two methods of traditional IDC and directly feeding the original video frames into the convolutional neural network without using WDFR are compared. It can be seen from the experimental results that the present invention combines the effective information in multiple consecutive frames of the video and reconstructs the feature information with the background image, effectively suppressing the interference of bubbles, defocus blur and other factors on the wear debris information caused by moving objects in the original video. In the reconstructed feature map, not only the information of traditional indicators such as IPCA and IDC is retained, but also the abstract features in the feature map can be extracted through the convolutional neural network to complete the life monitoring of gearbox wear. The gearbox RUL prediction method implemented by the present invention through video data compares the method of directly using video frames without WDFR (the present invention) to train the convolutional neural network and the method of using traditional IDC indicators to characterize the wear process. In order to quantify the experimental results, MSE, MAE and RMSE are introduced as evaluation indicators, which are defined in Formula 19-Formula 21 respectively:

[0109]

[0110]

[0111]

[0112] The quantitative experimental results are shown in Table 1 below:

[0113] Table 1

[0114]

[0115] The comparative experimental results show that the prediction results of the present invention are better than those of IDC and those without WDFR. Obviously, the above embodiments are merely examples to clearly illustrate the technical solution of the present invention, and are not limitations on the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made on the basis of the above description. It is not necessary and impossible to list all the implementation methods here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A video data processing method, characterized in that: include: Step 1: shallowly cluster the initial video frame set of the video data to obtain a primary background image, then increase the number of cluster centers to obtain a revised set of the primary background image, and use the data of the revised set to revise the primary background image and construct the video background image; Step 2: Calculate the grayscale difference between the video data and the reconstructed background image frame by frame, and replace the defocused blurred areas with similar background grayscale and the original background areas with the reconstructed new background; Step 3: Extract and convert the label information from the video number to obtain information related to time, speed, and load conditions; Step 4: Select a reconstruction time node based on the time, speed, and load information of the video matching, and locate the reconstruction node to a specific frame based on the refresh rate of the video data; Step 5: With the reconstruction node as the center, select multiple consecutive adjacent frames to construct an adjacent frame matrix, and select a quantization factor to quantize the adjacent frame matrix; Step 6: Calculate the mode value of the adjacent frame matrix in the sample dimension after quantization, and record the position of the corresponding sample as the valid information set. Select the data at the position in the valid information set corresponding to the adjacent frame matrix before quantization, calculate the mean value in the sample dimension to represent the wear debris information of this dimension, and reconstruct it in the background image.

2. The video data processing method according to claim 1, wherein: The primary background image and correction set in step 1 are tensors obtained by using the K-mean algorithm to select cluster centers of 1 and 3, respectively.

3. A method for predicting the RUL of a gearbox, characterized in that: include: Step 7: Reconstruct a feature map of all video data using the method in steps 2 to 6 of claim 1; Step 8: Obtain traditional wear debris characteristic indicators through mathematical modeling; Step 9: Divide the feature map reconstructed in step 7 into a training set and a validation set, and train the deep learning network to achieve RUL prediction of the gearbox.

4. The method for predicting the RUL of a gearbox according to claim 3, characterized in that: In step 9, the ratio of training set to validation set is 7:3, AlexNet network is used, MSE loss function is used, and SGD is used to update network parameters.

Images