A neural network model, training method and device for rotating object detection

Abstract

The present invention discloses a neural network model, training method and device for rotating object detection, which belong to the fields of artificial intelligence, computer vision and visual vibration displacement measurement. The present invention divides the target detection neural network model into three modules: a lightweight module for extracting shallow features, a residual feature pyramid module for extracting deep features, and simultaneously fusing shallow information and deep information for semantic feature; a head module for finally determining and outputting target position information. Based on the above, the model proposed by the present invention is used in the monitoring stage of visual vibration measurement of rotating objects, which effectively improves the speed of monitoring the safety performance of rotating objects while ensuring monitoring accuracy; by comparing the vibration signals generated by data of other target detection models, the practical engineering application value of the present invention is proved.

Description

Technical Field

[0001] The present invention relates to a neural network model, training method and device for detecting a rotating object, and belongs to the fields of artificial intelligence, computer vision and visual vibration displacement measurement. Background Art

[0002] Rotating machinery is characterized by high complexity, random operating conditions, and unpredictable conditions, making it highly susceptible to failures during operation. Forced shutdown due to severe vibration can lead to production downtime, economic losses, and even safety hazards such as harm to personnel. Therefore, condition monitoring and fault diagnosis of rotating machinery are crucial. The rotor is a crucial component of rotating machinery, yet current research in this area remains lacking.

[0003] Deep learning-based fault diagnosis algorithms for rotating machinery primarily use time-domain signals, frequency-domain signals, or time-frequency signals as model inputs. However, due to the presence of noise, the signals collected by contact sensors can be affected by noise, which in turn degrades the diagnostic performance of the model. The operating status of a rotating body can be assessed for safety performance by measuring its position offset. Vision-based displacement monitoring offers the advantages of non-contact, non-destructive, and multi-point simultaneous measurement. Deep learning-based target detection algorithms have attracted widespread attention from scholars due to their high real-time performance and excellent accuracy.

[0004] As the demand for detection accuracy increases, more and more complex and large networks are being designed. Although larger and deeper networks can extract deeper target features to improve detection accuracy, they also introduce a huge number of parameters and computational complexity. This makes algorithm training more difficult and increases energy consumption, which in turn leads to reduced performance, slow processing speed when detecting targets, and difficulty adapting to real-world industrial scenarios. In addition, these methods require a large amount of computing power to select qualified models, which is time-consuming. When analyzing large amounts of real-time video data online, there will be a large number of images that need to be quickly detected. Therefore, proposing a high-precision, lightweight target detection method and system for the vibration displacement measurement task of rotating objects is of great significance to vision-based engineering measurement projects. Summary of the Invention

[0005] The present invention provides a rotating object target detection neural network model, training method and device. Based on reasonable structure and coordination, a new rotating object target detection neural network model is constructed, and is further used for displacement measurement of vibrating rotors to realize safety status monitoring of the rotor, realizing target detection and vibration displacement measurement of vibrating rotating objects in a non-contact manner.

[0006] The technical solution of the present invention is:

[0007] According to one aspect of the present invention, a training method for a rotating object target detection neural network model is provided, which uses a lightweight module to extract shallow features, a residual feature pyramid module to extract deep features, and a head module to locate the target to jointly construct a target detection neural network model.

[0008] The loss function used by the target detection neural network model is the improved CIoU loss function. The improved CIoU loss function is based on the CIoU loss function and modifies the penalty term to:

[0009]

[0010] Where: R TDIoU represents the penalty term of the improved CIoU loss function; b is the prediction box PB, b gt is the standard box GT, ρ C (b,b gt ) is the Euclidean distance between the two box centers, ρ L (b,b gt ) is the Euclidean distance between the upper left corners of the two boxes, ρ R (b,b gt ) is the Euclidean distance between the lower right corners of the two boxes, and c is the diagonal length of the smallest rectangle covering the two boxes.

[0011] The lightweight module includes: performing convolution operations and pooling operations on the input image data in sequence to obtain a first shallow output feature S1; performing a convolution block 1 operation on the first shallow output feature S1 to obtain a second shallow output feature S2; performing a convolution block 2 operation on the second shallow output feature S2 to obtain a third shallow output feature S3; performing one convolution block 1 operation and two convolution block 2 operations on the third shallow output feature S3 to obtain a fourth shallow output feature S4; performing one convolution block 1 operation and one convolution block 2 operation on the fourth shallow output feature S4 in sequence to obtain a fifth shallow output feature S5.

[0012] The convolution block 1 operates by sequentially performing a depth-separable convolution operation, a BatchNorm operation, a convolution operation, a BatchNorm operation, and an activation function operation on the input features to obtain the feature M. 11 At the same time, the input features are sequentially subjected to convolution operation, depthwise separable convolution operation, BatchNorm operation, convolution operation, BatchNorm operation, and activation function operation to obtain feature M 12 ; For feature M 11 , Feature M 12 The second shallow output feature S2 is obtained by performing addition and channel information interaction operations in sequence.

[0013] The convolution block 2 operates by dividing the input features into equal channels, and obtains two features M with the same number of channels. 13 , Feature M 14 ; For feature M 14 After the convolution operation, BatchNorm operation, activation function operation, depthwise separable convolution operation, BatchNorm operation, convolution operation, BatchNorm operation, activation function operation, the feature M is obtained. 15 ; The feature M 13 With feature M 15 The features are added together, and then the added features are subjected to channel information interaction operations to obtain the third shallow output feature S3.

[0014] The residual feature pyramid module includes: performing a dilated spatial pyramid pooling operation on the fifth shallow output feature S5 of the lightweight module to obtain a feature N 11 ; For feature N 11 After the convolution operation, feature N is obtained 12 ; Feature N 12 After upsampling, feature N is obtained 13 ; Combine the fourth shallow output feature S4 of the lightweight module with feature N 13 Splicing to get feature N 14 ; For feature N 14 After depth-wise separable convolution operation, feature N is obtained 21 ; Set feature N 21 After the convolution operation, feature N is obtained 22 ; Feature N 22 After upsampling, feature N is obtained 23 ; Set feature N 23 Concatenate with the third shallow output feature S3 of the lightweight module to obtain feature N 24 ; For feature N 24 After depth-wise separable convolution, the first residual feature pyramid output feature N1 is obtained; the first residual feature pyramid output feature N1 is convolved with the feature N 22 The second residual feature pyramid output feature N2 is obtained by splicing and undergoing depth-separable convolution; the second residual feature pyramid output feature N2 is convolved with the feature N 12 The third residual feature pyramid output feature N3 is obtained by splicing and undergoing depth-wise separable convolution; the first residual feature pyramid output feature N1, the second residual feature pyramid output feature N2, and the third residual feature pyramid output feature N3 are spliced ​​to obtain the fourth residual feature pyramid output feature N4.

[0015] The dilated spatial pyramid pooling operation includes: the input features are subjected to dilated convolution operations with convolution kernels of 3×3 and dilated rates of 6, 12, and 18 to obtain the feature M 21 , Feature M 22 , Feature M 23 ; At the same time, the input features are convolved to obtain feature M 24 At the same time, the input features are sequentially subjected to pooling, convolution, and upsampling operations to obtain feature M 25 ; For feature M 21 , Feature M 22 , Feature M 23 , Feature M 24 , Feature M 25 Perform the addition operation to obtain the output feature N 11 .

[0016] The head positioning module includes: performing convolution operations, BatchNorm operations, and activation function operations on the fourth residual feature pyramid output feature N4 of the residual feature pyramid module to obtain feature H1; and sequentially performing convolution operations, BatchNorm operations, activation function operations, convolution operations, BatchNorm operations, activation function operations, and convolution operations on feature H1 to obtain a classification branch, a regression branch, and a background branch.

[0017] According to another aspect of the present invention, a training method for a rotating body target detection neural network model is provided, comprising: collecting first non-contact monitoring data and second non-contact monitoring data under rotating body vibration; inputting the first non-contact monitoring training data and target information into the target detection neural network model for pre-training to obtain a pre-trained target detection neural network model; inputting the first non-contact monitoring test data and target information into the pre-trained target detection neural network model for testing, and screening the optimal weight based on the second non-contact monitoring data, and generating the target detection neural network model based on the optimal weight.

[0018] According to another aspect of the present invention, a training device for a rotating body target detection neural network model is provided, comprising: an acquisition module for acquiring first non-contact monitoring data and second non-contact monitoring data under rotating body vibration; a training module for inputting the first non-contact monitoring training data and target information into the target detection neural network model for pre-training, thereby obtaining a pre-trained target detection neural network model; a generation module for inputting the first non-contact monitoring test data and target information into the pre-trained target detection neural network model for testing, and screening the optimal weight based on the second non-contact monitoring data, thereby generating a target detection neural network model based on the optimal weight.

[0019] The beneficial effects of the present invention are: in order to solve the problem that the existing methods require a lot of computing power support, large amount of calculation, high training difficulty, huge energy consumption, which makes the deep learning model insufficient in real-time fault diagnosis in industrial scenarios and leads to large fluctuations in accuracy, the present invention proposes a rotating body target detection neural network model, which has high accuracy and is lightweight, and better realizes real-time monitoring in industrial scenarios under the premise of high-precision positioning of vibration targets. Specifically: the present invention divides the target detection neural network model into three modules, the lightweight module is used to extract shallow features, the residual feature pyramid module is used to extract deep features, and at the same time, the shallow information and the deep information are semantically fused; the head module is used to finally determine and output the target position information. Based on the above, the model proposed by the present invention is used in the monitoring stage of visual vibration measurement of rotating bodies, which effectively improves the speed of monitoring the safety performance of rotating bodies while ensuring monitoring accuracy; by comparing the vibration signals generated by data of other target detection models, the practical engineering application value of the present invention is proved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a model structure diagram of the present invention;

[0021] Figure 2 This is the effect diagram of the marking of the rotating body using the marking software;

[0022] Figure 3 Lightweight module structure diagram;

[0023] Figure 4 This is the principle diagram of convolution block 1;

[0024] Figure 5 This is the principle diagram of convolution block 2;

[0025] Figure 6 This is the specific step diagram of the residual feature pyramid module;

[0026] Figure 7 This is the principle diagram of void space pyramid pooling;

[0027] Figure 8 Schematic diagram of the head positioning module;

[0028] Figure 9 This is a specific training flow chart;

[0029] Figure 10 It is the time domain signal diagram of eddy current sensor and image displacement;

[0030] Figure 11 This is the frequency domain signal diagram of the eddy current sensor and image displacement. DETAILED DESCRIPTION

[0031] The invention will be further described below with reference to the accompanying drawings and embodiments, but the content of the present invention is not limited to the scope of the drawings.

[0032] Example 1: Figure 1-11 As shown, according to one aspect of an embodiment of the present invention, a neural network model for detecting rotating objects is provided, which uses a lightweight module to extract shallow features, a residual feature pyramid module to extract deep features, and a head module to locate the target to jointly construct a target detection neural network model, as shown in FIG. Figure 1 shown.

[0033] Furthermore, the loss function used by the target detection neural network model is an improved CIoU loss function. The improved CIoU loss function is based on the CIoU loss function, and the penalty term is modified as follows:

[0034]

[0035] Where: R TDIoU represents the penalty term of the improved CIoU loss function; b is the prediction box PB, b gt is the standard box GT, ρ C (b,b gt ) is the Euclidean distance between the center points of the two boxes (the two boxes are the predicted box and the standard box), ρ L (b,b gt ) is the Euclidean distance between the upper left corners of the two boxes, ρ R (b,b gt ) is the Euclidean distance between the lower right corners of the two boxes, and c is the diagonal length of the minimum rectangle covering the two boxes. The improved CIoU loss function TD IoU takes into account the overlapping area, center point distance, and distance between the upper left and lower right corners of PB and GT, making the target positioning more accurate.

[0036] Furthermore, if Figure 3 As shown, the lightweight module includes: taking the original rotor image input to the network as the input of the lightweight module. The input image data is sequentially subjected to a convolution operation with a convolution kernel size of 3×3 and a pooling operation to obtain a first shallow output feature S1; the first shallow output feature S1 is subjected to a convolution block 1 operation to obtain a second shallow output feature S2; the second shallow output feature S2 is subjected to a convolution block 2 operation to obtain a third shallow output feature S3; the third shallow output feature S3 is subjected to one convolution block 1 operation and two convolution block 2 operations to obtain a fourth shallow output feature S4; the fourth shallow output feature S4 is sequentially subjected to one convolution block 1 operation and one convolution block 2 operation to obtain a fifth shallow output feature S5. The fifth shallow output feature S5 is used as the output feature of the shallow feature extraction by the lightweight module.

[0037] Furthermore, if Figure 4 As shown, the convolution block 1 operates on the input features in sequence through a convolution operation with a step size of 2 and a convolution kernel size of 3×3 depth-separable convolution operation, a BatchNorm operation, a convolution operation with a convolution kernel size of 1×1, a BatchNorm operation, and a ReLU activation function operation to obtain the feature M 11 At the same time, the input features are sequentially subjected to a convolution operation with a convolution kernel size of 1×1, a depthwise separable convolution operation with a convolution kernel size of 3×3 and a step size of 2, a BatchNorm operation, a convolution operation with a convolution kernel size of 1×1, a BatchNorm operation, and a ReLu activation function operation to obtain the feature M 12 ; For feature M 11 , Feature M 12 The second shallow output feature S2 is obtained by performing addition and channel information interaction operations in sequence.

[0038] Furthermore, if Figure 5 As shown, the convolution block 2 operates to divide the input features into channels, and obtains two features M with the same number of channels. 13 , Feature M 14 ; For feature M 14 After the convolution operation with a convolution kernel size of 1×1, the BatchNorm operation, the ReLU activation function operation, the depth-separable convolution operation with a convolution kernel size of 3×3 and a step size of 1, the BatchNorm operation, the convolution operation with a convolution kernel size of 1×1, the BatchNorm operation, and the ReLU activation function operation, the feature M is obtained. 15 ; The feature M 13 With feature M 15 The features are added together, and then the added features are subjected to channel information interaction operations to obtain the third shallow output feature S3.

[0039] Furthermore, if Figure 6 As shown, the residual feature pyramid module includes: the fifth shallow output feature S5 of the lightweight module is used as the input feature of the residual feature pyramid module for deep feature extraction; the fifth shallow output feature S5 of the lightweight module is subjected to the hollow space pyramid pooling operation to obtain the feature N 11 ; For feature N 11 After the convolution operation with a convolution kernel size of 3×3, feature N is obtained 12 ; Feature N 12 After 2 times upsampling operation, feature N is obtained 13 ; Combine the fourth shallow output feature S4 of the lightweight module with feature N 13 Splicing to get feature N 14 ; For feature N 14The feature N is obtained by performing a depth-separable convolution operation with a kernel size of 3×3 and a stride of 1. 21 ; Set feature N 21 After the convolution operation with a convolution kernel size of 3×3, feature N is obtained 22 ; Feature N 22 After 2 times upsampling operation, feature N is obtained 23 ; Set feature N 23 Concatenate with the third shallow output feature S3 of the lightweight module to obtain feature N 24 ; For feature N 24 The first residual feature pyramid output feature N1 is obtained by depth-separable convolution with a convolution kernel size of 3×3 and a step size of 1; the first residual feature pyramid output feature N1 is convolved with a convolution kernel size of 3×3 and the feature N 22 The concatenation is performed and the depth-separable convolution with a convolution kernel size of 3×3 and a step size of 1 is performed to obtain the second residual feature pyramid output feature N2; the second residual feature pyramid output feature N2 is convolved with a convolution kernel size of 3×3 and the feature N 12 The third residual feature pyramid output feature N3 is obtained by splicing and performing a depth-separable convolution with a convolution kernel size of 3×3 and a step size of 1; the first residual feature pyramid output feature N1, the second residual feature pyramid output feature N2, and the third residual feature pyramid output feature N3 are spliced ​​to obtain the fourth residual feature pyramid output feature N4.

[0040] Furthermore, if Figure 7 As shown, the dilated spatial pyramid pooling operation includes: the input features are subjected to dilated convolution operations with convolution kernels of 3×3 and dilated rates of 6, 12, and 18 to obtain the feature M 21 , Feature M 22 , Feature M 23 At the same time, the input feature is convolved with a convolution kernel size of 1×1 to obtain the feature M 24 At the same time, the input features are sequentially subjected to pooling operations, convolution operations with a convolution kernel size of 1×1, and 2 times upsampling operations to obtain feature M 25 ; For feature M 21 , Feature M 22 , Feature M 23 , Feature M 24 , Feature M 25 Perform the addition operation to obtain the output feature N 11 .

[0041] Furthermore, if Figure 8As shown, the head localization module includes: the fourth residual feature pyramid output feature N4 of the residual feature pyramid module is used as the input feature of the head localization module. The fourth residual feature pyramid output feature N4 of the residual feature pyramid module is subjected to a convolution operation with a convolution kernel size of 1×1, a BatchNorm operation, and a SiLU activation function operation to obtain feature H1; feature H1 is sequentially subjected to a convolution operation with a convolution kernel size of 3×3, a BatchNorm operation, a SiLU activation function operation, a convolution operation with a convolution kernel size of 3×3, a BatchNorm operation, a SiLU activation function operation, and a convolution operation with a convolution kernel size of 1×1 to obtain a classification branch, a regression branch, and a background branch. The classification branch is a binary classification of background and target, with 2n outputs (in this embodiment of the present invention, the target is a rotor). The regression branch is used to determine the predicted box position information of the target area and has 4n outputs. The background branch is used to determine whether the target category contained in the predicted box is a rotor and has 2n outputs. The classification branch includes two parameters [0, 1], representing background and rotor, respectively. The regression branch includes four parameters [x1, y1, x2, y2], which represent the horizontal and vertical coordinates (x1, y1) of the upper left corner of the prediction box and the horizontal and vertical coordinates (x2, y2) of the lower right corner of the prediction box, respectively. The background branch includes two parameters [0, 1], which represent whether the prediction box contains a rotor or not.

[0042] According to another aspect of an embodiment of the present invention, a training method for a rotating body target detection neural network model is provided, comprising: collecting first non-contact monitoring data and second non-contact monitoring data under rotating body vibration; inputting the first non-contact monitoring training data and target information into the target detection neural network model for pre-training to obtain a pre-trained target detection neural network model; wherein the target information includes a labeling box and a category; inputting the first non-contact monitoring test data and target information into the pre-trained target detection neural network model for testing, and screening the optimal weight based on the second non-contact monitoring data, and generating the target detection neural network model based on the optimal weight.

[0043] According to another aspect of an embodiment of the present invention, a training device for a rotating body target detection neural network model is provided, including: an acquisition module for acquiring first non-contact monitoring data and second non-contact monitoring data under rotating body vibration; a training module for inputting the first non-contact monitoring training data and target information into the target detection neural network model for pre-training to obtain a pre-trained target detection neural network model; a generation module for inputting the first non-contact monitoring test data and target information into the pre-trained target detection neural network model for testing, and screening the optimal weight based on the second non-contact monitoring data, and generating the target detection neural network model based on the optimal weight.

[0044] Furthermore, the optional specific embodiments of the present invention are described as follows:

[0045] S1. On a high-speed rotor vibration test bench, synchronously collect 2 seconds of image data (image resolution of 512×512) and eddy current data of the mechanical component using a high-speed camera (sampling frame rate set to 500 frames / second and 1000 frames / second) and an eddy current sensor, and use the eddy current data as reference data for subsequent experimental comparisons; wherein the image data is the first non-contact monitoring data, and the eddy current data is the second non-contact monitoring data;

[0046] S2. The images collected in the image dataset with a frame rate of 500 frames per second for 1 second are divided into a training set and a test set according to a ratio of 8:2. The training dataset and the test dataset are annotated with a uniformly sized annotation box using the LabelImg annotation tool to obtain a labeled training set and a labeled verification set, as shown in the following example: Figure 2 The image shown is a labeled image. The image collected for 1 second with a frame rate of 1000 frames per second in the collected image data set is used as the data to be detected;

[0047] S3, use the lightweight module to extract shallow features, use the residual feature pyramid module to extract deep features, use the head module to locate the rotor to build the target detection neural network model;

[0048] S4. Before formal pre-training, set the parameters of the configuration file to obtain pre-training parameters. In this embodiment, the configuration file sets the pre-training parameters as Batch Size = 8 and the training eopchs as 200.

[0049] S5, training flow chart as follows Figure 8 As shown, the training dataset and configuration file are used to pre-train the target detection neural network model. The essence of training involves obtaining target weight information and calculating loss during the network's forward propagation. A loss function is then added during the backward propagation process to adjust the performance. When the minimum loss value is reported during this iteration, the weight file is saved for subsequent target detection. At the end of training, the optimal candidate weights are selected. Images are loaded for training in batches based on the BatchSize value. Weights are output for each iteration according to the set parameters, and the weight with the lowest loss value is selected. These weights are then loaded into the model. The mechanical component's axis trajectory data is then monitored (all data is normalized to obtain a time-domain signal curve of the vibrating rotor; the time-domain signal is fast Fourier transformed to obtain a frequency-domain signal curve of the vibrating rotor). The data is then compared with the axis trajectory data collected in real time by the online monitoring component. The weight corresponding to the model with the closest matching between the two data sets is selected as the optimal weight and loaded into the target detection neural network model.

[0050] S6. Based on the data to be detected, a comparative experiment was conducted on multiple target detection neural network models based on different principles, and the performance indicators of mAP@0.5:0.95, FPS, and fitting error with the reference signal were used to evaluate the performance of the detection results. The performance of the detection results was compared with target detection neural network models based on multiple different positioning principles. The rotor time domain signal curve obtained according to the present invention was fitted and compared with the eddy current data. The rotor frequency domain signal curve obtained according to the present invention was fitted and compared with the eddy current data. The target detection neural network models based on different positioning principles mainly include Faster R-CNN based on dual-stage anchor frames, YOLOv3, SSD, RetinaNet based on single-stage anchor frames, and CenterNet based on single-stage anchor frames without anchor frames.

[0051] In this embodiment, we use the data collected by the eddy current sensor as a reference, and the rotor's position information in the first frame of the image as the reference frame for calculating the displacement offset. By calculating the displacement offsets of the remaining frames relative to the first frame, we obtain the rotor's vibration displacement, and then obtain the rotor's time domain signal and frequency domain signal. As shown in Table 1, the target detection neural network model of the present invention achieves the highest detection accuracy and the best real-time performance. For example, while achieving the highest target prediction box detection accuracy, the detection speed of the present invention is 3.1 times that of CenterNet, 12.9 times that of Faster R-CNN, 4.6 times that of SSD, 4.2 times that of RetinaNet, and 1.7 times that of YOLOv3. This is because the network structure of the present invention utilizes a lightweight module and a residual feature pyramid structure. The lightweight module uses channel information interaction to achieve information fusion between different channels and also uses channel averaging to achieve higher efficiency without dense convolution. The residual feature pyramid structure fully utilizes global and local information about the rotor's position, further enhancing the model's ability to recognize the rotor at different times and positions. On the other hand, the depth-wise separable convolution adopted by the residual feature pyramid can ensure the detailed feature information of the rotor target while saving computational complexity. In terms of signal fitting error, the time domain and frequency domain fitting accuracy of the present invention is the highest. For example, in terms of time domain signal fitting error (NRMSE1), it is 6.47% higher than Faster R-CNN, 13.1% higher than SSD, 16.9% higher than RetinaNet, and 19.9% ​​higher than YOLOv3. CenterNet did not detect the position information of all rotors, resulting in no valid NRMSE value. The reason is that the TDIoU loss proposed in the present invention takes into account the overlapping area of ​​the predicted box and the standard box, the center point distance, and the distance between the upper left corner and the lower right corner. This method of directly considering the error between the corresponding corner points of the two boxes makes the positioning accuracy of the rotor higher. As Figure 10 、 Figure 11As shown in FIG, the rotor vibration data obtained by the target detection neural network model of the present invention is closest to the reference displacement signal. Therefore, the target detection neural network proposed in the present invention is particularly suitable for visual vibration displacement measurement of rotors.

[0052] Table 1. Performance comparison of detection results of different object detection neural networks

[0053] method CenterNet Faster R-CNN SSD RetinaNet YOLOv3 The present invention mAP@0.5:0.95 0.874 0.643 0.897 0.896 0.773 0.900 FPS 46 11 31 34 83 142 <![CDATA[NRMSE1]]> / 0.4233 0.4559 0.4766 0.4943 0.3959 <![CDATA[NRMSE2]]> / 0.0768 0.0846 0.0874 0.0922 0.0747

[0054] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A neural network model for rotating object detection, characterized in that: The lightweight module is used to extract shallow features, the residual feature pyramid module is used to extract deep features, and the head module is used to locate the target to jointly build a target detection neural network model; The lightweight module includes: Perform convolution and pooling operations on the input image data in sequence to obtain the first shallow output feature S1; Perform convolution block 1 operation on the first shallow layer output feature S1 to obtain the second shallow layer output feature S2; Perform convolution block 2 on the second shallow layer output feature S2 to obtain the third shallow layer output feature S3; The third shallow output feature S3 is subjected to one convolution block 1 operation and two convolution block 2 operations to obtain a fourth shallow output feature S4; the fourth shallow output feature S4 is subjected to one convolution block 1 operation and one convolution block 2 operation in sequence to obtain a fifth shallow output feature S5; The residual feature pyramid module includes: The fifth shallow output feature S5 of the lightweight module is subjected to the void space pyramid pooling operation to obtain the feature N 11 ; For feature N 11 After the convolution operation, feature N is obtained 12 ; Feature N 12 After upsampling, feature N is obtained 13 ; Combine the fourth shallow output feature S4 of the lightweight module with feature N 13 Splicing to get feature N 14 ; For feature N 14 After the depth-wise separable convolution operation, feature N is obtained 21 ; The feature N 21 After the convolution operation, feature N is obtained 22 ; Feature N 22 After upsampling, feature N is obtained 23 ; Set feature N 23 Concatenate with the third shallow output feature S3 of the lightweight module to obtain feature N 24 ; For feature N 24 After depth-wise separable convolution, the first residual feature pyramid output feature N1 is obtained; The first residual feature pyramid output feature N1 is convolved with feature N 22 After concatenation and depth-wise separable convolution, the second residual feature pyramid output feature N2 is obtained; The second residual feature pyramid output feature N2 is convolved with feature N 12 After concatenation and depth-wise separable convolution, the third residual feature pyramid output feature N3 is obtained; The first residual feature pyramid output feature N1, the second residual feature pyramid output feature N2, and the third residual feature pyramid output feature N3 are concatenated to obtain a fourth residual feature pyramid output feature N4; The head positioning module includes: performing convolution operations, BatchNorm operations, and activation function operations on the fourth residual feature pyramid output feature N4 of the residual feature pyramid module to obtain feature H1; and sequentially performing convolution operations, BatchNorm operations, activation function operations, convolution operations, BatchNorm operations, activation function operations, and convolution operations on feature H1 to obtain a classification branch, a regression branch, and a background branch.

2. The rotating object detection neural network model according to claim 1, characterized in that: The loss function used by the target detection neural network model is the improved CIoU loss function. The improved CIoU loss function is based on the CIoU loss function and modifies the penalty term to: ； Where: represents the penalty term of the improved CIoU loss function; b is the prediction box PB, b gt It is a standard frame GT. is the Euclidean distance between the two box centers, is the Euclidean distance between the upper left corners of the two boxes, is the Euclidean distance between the lower right corners of the two boxes, c is the length of the diagonal of the smallest rectangle that covers both boxes.

3. The rotating object detection neural network model according to claim 1, characterized in that: The convolution block 1 operates by sequentially performing a depth-separable convolution operation, a BatchNorm operation, a convolution operation, a BatchNorm operation, and an activation function operation on the input features to obtain the feature M. 11 At the same time, the input features are sequentially subjected to convolution operation, depthwise separable convolution operation, BatchNorm operation, convolution operation, BatchNorm operation, and activation function operation to obtain feature M 12 ; For feature M 11 , Feature M 12 The second shallow output feature S2 is obtained by performing addition and channel information interaction operations in sequence.

4. The neural network model for rotating object detection according to claim 1, characterized in that: The convolution block 2 operates by dividing the input features into equal channels, and obtains two features M with the same number of channels. 13 , Feature M 14 ; For feature M 14 After the convolution operation, BatchNorm operation, activation function operation, depthwise separable convolution operation, BatchNorm operation, convolution operation, BatchNorm operation, activation function operation, the feature M is obtained. 15 ; The feature M 13 With feature M 15 The features are added together, and then the added features are subjected to channel information interaction operations to obtain the third shallow output feature S3.

5. The neural network model for rotating object detection according to claim 1, characterized in that: The dilated spatial pyramid pooling operation includes: the input features are subjected to dilated convolution operations with convolution kernels of 3×3 and dilated rates of 6, 12, and 18 to obtain the feature M 21 , Feature M 22 , Feature M 23 ; At the same time, the input features are convolved to obtain feature M 24 At the same time, the input features are sequentially subjected to pooling, convolution, and upsampling operations to obtain feature M 25 ; For feature M 21 , Feature M 22 , Feature M 23 , Feature M 24 , Feature M 25 Perform the addition operation to obtain the output feature N 11 .

6. A training method for a neural network model for rotating object detection according to claim 1, characterized in that: include: Collecting first non-contact monitoring data and second non-contact monitoring data under vibration of the rotating body; Inputting the first non-contact monitoring training data and target information into the target detection neural network model for pre-training to obtain a pre-trained target detection neural network model; The first non-contact monitoring test data and target information are input into the pre-trained target detection neural network model for testing, and the optimal weight is selected according to the second non-contact monitoring data, and the target detection neural network model is generated according to the optimal weight.

7. A training device for a neural network model for rotating object detection according to claim 1, characterized in that: include: An acquisition module, used for acquiring first non-contact monitoring data and second non-contact monitoring data under vibration of the rotating body; A training module, configured to input the first non-contact monitoring training data and target information into a target detection neural network model for pre-training, thereby obtaining a pre-trained target detection neural network model; The generation module is used to input the first non-contact monitoring test data and target information into the pre-trained target detection neural network model for testing, and screen the optimal weight according to the second non-contact monitoring data, and generate the target detection neural network model according to the optimal weight.

Images