A gear damage detection method and apparatus incorporating damage propagation speculation

Abstract

The application discloses a gear damage detection method and device introducing damage propagation prediction, the gear damage detection method synchronously collects multi-angle images of the front and side of the gear through a high-resolution camera, and inputs the images into a gear damage detection model to identify the damage type and damage degree of the gear; by constructing a comprehensive rating model including a damage type layer and a comprehensive evaluation layer, precise quantitative evaluation of the gear damage grade and the remaining service life is realized, the integration of automatic, high-precision diagnosis and life prediction of gear damage is realized, and the maintenance efficiency of the gear is significantly improved. Meanwhile, the gear damage detection model of the application sequentially extracts features, fuses features and makes decisions through a backbone network, a neck network and a detection head, and introduces a damage propagation prediction attention module in the backbone network for feature extraction, captures long-range dependency between different positions through a multi-head self-attention mechanism, and effectively highlights the key area in gear damage detection.

Description

Technical Field

[0001] This invention belongs to the field of damage detection technology, specifically relating to a gear damage detection method and apparatus that incorporates damage propagation prediction. Background Technology

[0002] Aero engines are high-performance, complex electromechanical systems, and their gear systems play a crucial role in power transmission, torque regulation, and speed matching. Because aero engine gears operate under high temperature, high pressure, and high speed conditions, and undergo prolonged service, their components are highly susceptible to fatigue damage, plastic deformation, and material wear. Gear failure can trigger a major safety accident affecting the entire engine system.

[0003] Currently, the condition assessment of aircraft gears mainly relies on periodic manual inspections. This inspection method is inefficient, dependent on human experience, and prone to missed detections. It cannot achieve quantitative judgment and trend prediction of gear damage. In addition, some equipment uses non-destructive testing methods such as X-rays and ultrasound, but these devices are expensive, the testing process is complex, and it is difficult to achieve efficient and automatic identification of multiple damage types.

[0004] With the rapid development of artificial intelligence, deep learning, and computer vision, more and more research and applications are beginning to explore the use of image recognition and machine learning technologies in manufacturing and inspection scenarios, especially demonstrating excellent performance in defect identification and lifespan estimation of complex structures. Therefore, developing a gear inspection device based on a large model, which can combine high-quality image acquisition with the powerful analytical capabilities of deep neural networks to achieve multi-damage type identification and lifespan assessment of aircraft gears, is of great significance for ensuring the safe operation of aviation equipment. Summary of the Invention

[0005] The purpose of this invention is to provide a gear damage detection method and apparatus that incorporates damage propagation prediction.

[0006] In a first aspect, the present invention provides a gear damage detection method incorporating damage propagation prediction, the method comprising:

[0007] A gear damage detection model is constructed, comprising a backbone network, a neck network, and a detection head. The backbone network extracts features from the input image sequentially through a series of convolutional layers and multi-layer feature extraction modules. The neck network fuses the extraction results of the feature extraction modules through a series of multi-layer feature fusion modules, and inputs the output features of the last feature fusion module into the detection head for processing to obtain the output result of the gear damage detection model.

[0008] In the backbone network, a damage propagation inference attention module is introduced into some feature extraction modules; the damage propagation inference attention module combines the damage propagation inference module and the attention enhancement module for feature extraction; the damage propagation inference module includes a physical feature extraction layer, a damage inference layer and an output decision layer connected in sequence.

[0009] The image of the gear under test is acquired and input into the gear damage detection model for detection to obtain the damage type and corresponding damage degree of the gear under test;

[0010] The damage severity of different damage types is weighted and fused, and the remaining life of the gear under test is obtained based on the fusion result, thus completing the damage detection of the gear under test.

[0011] Preferably, a damage inference network is introduced into the damage inference layer for feature processing; the damage inference network is used to quantify the damage length and damage width, and an attention kernel is generated through a mapping function to process the input features of the damage inference network.

[0012] Preferably, the physical feature extraction layer includes a series of convolutional layers and a C2f module; the damage inference layer includes a first branch and a second branch in parallel; the first branch includes a C3 module, an activation function, and a convolutional layer connected in sequence; the second branch includes a series of damage inference networks and an activation function; the output decision layer is used to process the fusion result of the output features of the first branch and the second branch in the damage inference layer; the output decision layer includes a C2fPSA module, a partial self-attention module, a convolutional layer, and an activation function connected in sequence.

[0013] Preferably, the partial self-attention module calculates the global contextual relationship of the input features through a multi-head self-attention mechanism, and concatenates the calculation result with the input features to obtain the output features of the partial self-attention module.

[0014] Preferably, the damage propagation inference attention module includes a first convolutional block, a second convolutional block, a damage propagation inference module, an attention enhancement module, a feature weighting module, a third convolutional block, and a convolutional layer connected in sequence. The attention enhancement module uses a self-attention mechanism to perform global context optimization on the output of the damage propagation inference module. The feature weighting module is used to normalize the output of the attention enhancement module and multiply the normalized result element-wise with the output of the attention enhancement module to enhance damage-related features and suppress non-damage features.

[0015] Preferably, the feature extraction module in the backbone network has five layers; the feature fusion module in the neck network has four layers. In the neck network, the input of the first-layer feature fusion module is the concatenation result of the output features of the third-layer and last-layer feature extraction modules after convolution processing; the input of the second-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the second-layer feature extraction module; the input of the third-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the first-layer feature fusion module; the input of the last-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the last-layer feature extraction module.

[0016] Preferably, in the backbone network, except for the last layer feature extraction module, the lower layer feature extraction modules have the same structure, all including cascaded convolutional layers and damage propagation inference attention modules; the higher layer feature extraction modules have the same structure, all including cascaded convolutional layers and feature extraction sub-blocks; the last layer feature extraction module includes cascaded fast spatial pyramid pooling modules and damage propagation inference attention modules.

[0017] Preferably, both the feature extraction sub-block and the feature fusion module use the C3k2 module.

[0018] Secondly, the present invention provides a gear damage detection device that incorporates damage propagation prediction, which is used to perform the gear damage detection method described above; the gear damage detection device includes an image acquisition mechanism and a gear damage detection module; the image acquisition mechanism is used to acquire an image of the gear under test and input it to the gear damage detection module to obtain the remaining life of the gear under test.

[0019] Preferably, the image acquisition mechanism includes a body and a fixed module, an image acquisition module, and a motion module mounted on the body; the fixed module is used to mount the gear under test; the image acquisition module is mounted on the motion module and is used to acquire images of the gear under test; the motion module includes a horizontal linear module and a vertical linear module; the horizontal linear module is used to drive the vertical linear module, thereby causing the image acquisition module on the vertical linear module to move in the horizontal and vertical directions.

[0020] The beneficial effects of this invention are:

[0021] 1. This invention introduces a damage propagation inference attention module into the feature extraction module for feature extraction. It can capture long-range dependencies between different locations through a multi-head self-attention mechanism, achieve fine-grained adjustment of attention weights, and strengthen damage-related features while suppressing non-damage features, thereby effectively highlighting key areas in gear damage detection. At the same time, this invention uses a damage inference network in the damage propagation inference attention module to quantify damage length and damage width, and qualitatively calculates the possible extension range of the damage through a mapping function, thereby spreading the model's attention to the entire range.

[0022] 2. This invention upgrades damage assessment from a single threshold method to a multi-dimensional quantitative model through a multi-parameter weighted fusion method, overcoming the limitations of traditional methods in inaccurate assessment under complex working conditions. This supports interpretable damage evolution analysis, provides a reliable data foundation for life prediction, and significantly reduces false alarm and false negative rates.

[0023] 3. This invention achieves high-definition image acquisition of gear surfaces without blind spots through a high-precision motion module and a finely adjustable industrial camera system. Combined with built-in large-model intelligent analysis, it can accurately identify and assess the severity of various types of damage, such as pitting, wear, scuffing, deformation, and breakage, significantly improving the comprehensiveness and accuracy of fault diagnosis. Its detection capabilities adapt to various operating conditions, solving the problem of insufficient diagnostic accuracy under varying conditions using traditional methods. Simultaneously, this invention integrates intelligent damage detection and remaining life prediction functions. Relying on large-model technology, it achieves coordinated linkage between multiple tasks such as condition monitoring, fault diagnosis, and life prediction, enabling diagnostic and prediction tasks to mutually promote each other, improving overall performance, and overcoming the limitations of traditional dedicated models that are single-function and lack linkage, providing equipment with a "one-stop" health management service.

[0024] 4. This invention optimizes the predictive model using actual operational data and supports rapid adaptation to new operating conditions or new fault types through fine-tuning in scenarios such as small-sample learning and continuous learning. This demonstrates excellent scalability and continuous evolution capabilities, extending the technical lifecycle of the device. Simultaneously, through early and accurate fault warnings and remaining life predictions, this invention effectively avoids unplanned downtime caused by sudden faults, providing precise data support for enterprises to formulate predictive maintenance strategies. This extends gear lifespan, improves equipment operating efficiency, and reduces maintenance costs and safety risks.

[0025] 5. This invention features a high degree of automation throughout the entire process, from automatic gear clamping and positioning, image acquisition to analysis and prediction. This reduces reliance on human observation and expert experience. The human-computer interaction module is user-friendly, the results are displayed intuitively, parameter settings are convenient, and early warning information is provided, thus lowering the barrier to entry and reducing technical dependence. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the overall structure of Embodiment 1 of the present invention.

[0027] Figure 2 This is a schematic diagram of the fixed module in Embodiment 1 of the present invention.

[0028] Figure 3 This is a schematic diagram of the horizontal straight line module in Embodiment 1 of the present invention.

[0029] Figure 4 This is a schematic diagram of the longitudinal linear module in Embodiment 1 of the present invention.

[0030] Figure 5 This is a schematic diagram of the second camera installation method in Embodiment 1 of the present invention.

[0031] Figure 6 This is an overall flowchart of Embodiment 2 of the present invention.

[0032] Figure 7 This is a structural diagram of the gear damage detection model in Embodiment 2 of the present invention.

[0033] Figure 8 This is a structural diagram of the damage propagation inference attention module in Embodiment 2 of the present invention.

[0034] Figure 9 This is a structural diagram of the damage propagation reasoning module in Embodiment 2 of the present invention.

[0035] Reference numerals: 1. Fixed module; 2. Image acquisition module; 3. Motion module; 4. Human-computer interaction module; 5. Fixed axis; 6. Load-bearing device; 7. Servo motor; 8. Vibration damping flange; 9. Rotary axis; 10. Gear under test; 11. Stepper motor; 12. Lead screw slide; 13. Slide rail; 14. Lead screw; 15. First camera; 16. Second camera; 17. Servo steering mechanism. Detailed Implementation

[0036] The present invention will be further described below with reference to the accompanying drawings.

[0037] Example 1

[0038] like Figure 1 As shown, a gear damage detection device incorporating damage propagation prediction includes an image acquisition mechanism and a gear damage detection module. The image acquisition mechanism is used to acquire images of the gear under test and input them into the gear damage detection module to obtain the damage score and remaining life of the gear under test.

[0039] The image acquisition mechanism includes a main body and, mounted on the main body, a fixed module 1, an image acquisition module 2, a motion module 3, and a human-machine interaction module 4. The fixed module 1 is used to mount the gear under test and ensure its stable rotation during the testing process. The image acquisition module 2, mounted on the motion module 3, is used to acquire data from the gear under test. The motion module 3 is used to move the image acquisition module 2 on a two-dimensional plane. The human-machine interaction module 4 assists the user in operating the intelligent gear damage detection device and displays the test results of the gear under test in real time.

[0040] like Figure 2 As shown, the fixing module 1 includes a fixing shaft 5, a load-bearing device 6, a servo motor 7, a vibration-damping flange 8, and a rotating shaft 9. The fixing shaft 5 is used to mount the gear 10 under test; the fixing shaft is designed with a small-angle grinding feature to achieve self-locking fixation of the gear under test and prevent axial movement. The fixing shaft 5 is connected to the rotating shaft 9 driven by the servo motor 7 through the vibration-damping flange 8 fixed on the load-bearing device 6, and is used to drive the rotating shaft 9 through the servo motor 7 to drive the gear under test to rotate stably during the testing process.

[0041] like Figure 3 and Figure 4 As shown, motion module 3 includes two horizontal linear modules and two vertical linear modules. The horizontal linear modules correspond one-to-one with the vertical linear modules, and the vertical linear modules are mounted on their corresponding horizontal linear modules, thereby driving the vertical linear modules to move along the direction of the horizontal linear modules. The horizontal and vertical linear modules have the same structure, each including a stepper motor 11, a slide table 12, a slide rail 13, and a lead screw 14. The slide table 12 is slidably connected to the lead screw; one end of the lead screw 14 is connected to the stepper motor 11, and under the drive of the stepper motor 11, it drives the slide table 12 to slide along the slide rail 13.

[0042] like Figure 5 As shown, the image acquisition module includes a first camera 15, a second camera 16, and a servo steering mechanism 17. The first camera 15 and the second camera 16 are respectively mounted on two longitudinal linear modules; the first camera 15 is fixedly mounted on the slide table 12 of the longitudinal linear module, and the lens of the first camera 15 is vertically downward; the second camera 16 is mounted on the slide table 12 of the longitudinal linear module through the servo steering mechanism 17, and the lens of the second camera 16 is in the installation direction of the gear being measured; the angle of the second camera 16 can be finely adjusted through the servo steering mechanism 17 to ensure the elimination of blind spots and the acquisition of high-definition images of the gear surface.

[0043] Example 2

[0044] like Figure 6 As shown, a gear damage detection method incorporating damage propagation prediction is presented, employing the gear damage detection device described in Example 1; this gear damage detection method includes the following steps:

[0045] Step 1: Camera Positioning

[0046] A coordinate system is established with the vertex of the fixed axis as the origin and the direction of movement of the longitudinal linear module as the horizontal axis. The optimal horizontal coordinate x of the second camera 16 is obtained, and its expression is:

[0047]

[0048] Where f is the focal length of the second camera 16; d is the pitch circle diameter of the gear.

[0049] The slide table is driven by a lead screw to move in a corresponding direction. The displacement formula can be calculated using the following formula:

[0050]

[0051] in, denoted as denoted as , where n is the displacement of the slide table; n is the number of rotation steps of the stepper motor; and p is the pitch of the lead screw.

[0052] By calculating the motor's rotation speed and the lead screw pitch, the position of the slide can be precisely controlled. Combined with the camera's focal length and gear parameters, the appropriate coordinates for the industrial camera can be calculated, allowing for the acquisition of clear gear images.

[0053] Step 2: Constructing the dataset

[0054] To accurately and efficiently construct the gear dataset, datasets such as the BSData gear pitting image dataset and the HUST-gearbox-dataset were used. The labels in the datasets include the damage type, damage extent, damage area, and severity of the gear. The damage extent indicates the location, range, and geometry of the damage in the image.

[0055] Step 3: Construct a gear damage detection model

[0056] like Figure 7 As shown, the gear damage detection model includes a backbone network, a neck network, and a detection head. The backbone network consists of convolutional layers and a five-layer feature extraction module; the convolutional layers are used to downsample the input image, and the five-layer feature extraction module sequentially extracts features from the processing results. The neck network consists of four sequentially connected feature fusion modules; the feature fusion module fuses the extraction results from the feature extraction modules and inputs the output features of the last feature fusion module into the detection head for processing, thus obtaining the output result of the gear damage detection model.

[0057] In the backbone network, the first two feature extraction modules each include cascaded convolutional layers and damage propagation inference attention modules; the third and fourth feature extraction modules each include cascaded convolutional layers and feature extraction sub-blocks; the last feature extraction module includes cascaded fast spatial pyramid pooling (SPPF) modules and damage propagation inference attention modules.

[0058] In the neck network, the input to the first-layer feature fusion module is the concatenation result of the output features of the third-layer and last-layer feature extraction modules after convolution processing; the input to the second-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the second-layer feature extraction module; the input to the third-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the first-layer feature fusion module; the input to the fourth-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the last-layer feature extraction module.

[0059] In this embodiment, both the feature extraction sub-block and the feature fusion module adopt the C3k2 module in YOLOv11.

[0060] like Figure 8 As shown, the damage propagation inference attention module includes a first convolutional block, a second convolutional block, a damage propagation inference module, an attention enhancement module, a feature weighting module, a third convolutional block, and convolutional layers connected in sequence. The first and third convolutional blocks have the same structure, both including cascaded convolutional layers and ReLU activation functions; the second convolutional block includes cascaded convolutional layers and a Sigmoid activation function. The attention enhancement module uses a self-attention mechanism to perform global context optimization on the output of the damage propagation inference module. The damage propagation inference attention module captures long-range dependencies between different locations through a multi-head self-attention mechanism, achieving fine-tuning of the initial attention weights. The feature weighting module normalizes the output of the attention enhancement module to the [0,1] interval using the Sigmoid function, and multiplies the normalized result element-wise with the output of the attention enhancement module, thereby enhancing damage-related features and suppressing non-damage features. This attention-guided feature recalibration mechanism can effectively highlight key areas in gear damage detection.

[0061] like Figure 9As shown, the damage propagation inference module comprises a physical feature extraction layer, a damage inference layer, and an output decision layer connected in sequence. The physical feature extraction layer includes a series of convolutional layers and a C2f module. The damage inference layer comprises a first branch and a second branch in parallel. The first branch includes a C3 module, a Tanh activation function, and a convolutional layer connected in sequence; the second branch includes a series of damage inference networks and a softmax activation function. The output decision layer processes the fusion result of the output features from the first and second branches in the damage inference layer; the output decision layer includes a C2fPSA module, a partial self-attention module (PSA), a convolutional layer, and a softmax activation function connected in sequence.

[0062] In this embodiment, the partial self-attention module calculates the global contextual relationship of the input features through a multi-head self-attention mechanism, and concatenates the calculation result with the input features to obtain the output features of the partial self-attention module.

[0063] The damage inference network is used to quantify damage length and width, and qualitatively calculate the potential extent of damage through a mapping function, thereby expanding the model's attention to the entire extent. In the damage inference network, an ellipse is fitted using the least squares method based on the damage contour, extracting parameters such as damage length (major axis of the ellipse), damage width (minor axis of the ellipse), and orientation angle. Based on these parameters, an attention kernel is generated, providing an accurate quantitative basis for damage propagation prediction. The specific process is as follows:

[0064] (1) Nonlinear mapping from damage width to extension distance

[0065] Transform the damage width into the extension distance of the attention kernel. Wider damage corresponds to a larger damage extension range, which is expressed as:

[0066]

[0067] in, To expand the proportional parameters, select according to the type of material; Extend the distance based on the base; These are the input features for the damage inference network.

[0068] (2) Generate direction-aware attention kernel

[0069] An elliptical orientation-aware attention kernel is generated for each damaged pixel, with the major axis along the length direction and the minor axis along the width direction, to achieve directional attention. The Gaussian weights at each position in the orientation-aware attention kernel are then obtained. for:

[0070]

[0071]

[0072]

[0073] in, The major axis direction angle is the deflection angle obtained from ellipse fitting; The standard deviation of the major axis; The standard deviation of the minor axis; The major axis extension factor, The minor axis extension factor is selected based on the type of material.

[0074] (3) Two-dimensional convolution to achieve attention diffusion

[0075] The damage mask is diffused along the direction, and the input features of the damage inference network are compared using a direction-aware attention kernel. Perform 2D convolution to generate a preliminary attention map. for:

[0076]

[0077] Step 4: Train the gear damage detection model using the dataset. Use a camera to capture images of the gears under test and input them into the gear damage detection model to obtain the type and extent of gear damage.

[0078] Step 5: Damage Assessment

[0079] Gear damage assessment is divided into two levels: the damage type level and the comprehensive assessment level. The specific methods are as follows:

[0080] (1) Damage type layer: Taking into account the severity of all damage types, a weighted average is calculated according to the damage type weight, and different weight coefficients are assigned to different damage types; for example, fracture and deformation have the highest weight (0.3), followed by adhesive (0.2), while pitting and wear have relatively low weights (0.1).

[0081] (2) Comprehensive Assessment Layer: For each type of damage, the parameter level is determined based on the ratio of measured values ​​to critical values. A five-level grading standard is used to obtain the gear damage score (comprehensive score): Level 1 (Minor Damage): Parameter value is 15% below the critical value; comprehensive score 0-20 points. The gear is in excellent condition with no significant damage characteristics. No special treatment is required; routine inspections can be performed as planned. Level 2 (Slight Damage): Parameter value reaches 15%-30% of the critical value; comprehensive score 20-40 points. The gear is in good condition with slight damage characteristics. It is recommended to increase the monitoring frequency and pay attention to the damage development trend. Level 3 (Moderate Damage): Parameter value reaches 30%-50% of the critical value; comprehensive score 40-60 points. The gear has obvious damage characteristics. The maintenance cycle needs to be shortened, a preventive maintenance plan should be developed, and a detailed inspection is recommended during the next scheduled inspection. Level 4 (Severe Damage): Parameter value reaches 50%-75% of the critical value; comprehensive score 60-80 points. The gear is severely damaged, affecting operational safety. Repairs need to be arranged as soon as possible, and replacement of damaged parts should be considered. The engine operating status should be restricted. Level 5 (Severe Damage): Parameter values ​​exceed the critical value by 75%; overall score 80-100 points. Gear damage is extremely severe; immediate grounding and repair are required. The damaged gear must be replaced, and a comprehensive fault analysis must be conducted. (The above overall score...) The specific calculation is derived from the following formula:

[0082]

[0083]

[0084] in, The comprehensive damage index; Weights for damage types; The degree of damage; This represents the number of damage types.

[0085] The remaining life of a single tooth is calculated using the following formula:

[0086]

[0087] in, For maximum lifespan; This represents the remaining lifespan factor.

[0088] The minimum remaining life of a single tooth is taken as the final remaining life of the gear under test.

Claims

1. A gear damage detection method incorporating damage propagation prediction, characterized in that: The method includes: A gear damage detection model is constructed, comprising a backbone network, a neck network, and a detection head. The backbone network extracts features from the input image sequentially through a series of convolutional layers and multi-layer feature extraction modules. The neck network fuses the extraction results of the feature extraction modules through a series of multi-layer feature fusion modules, and inputs the output features of the last feature fusion module into the detection head for processing to obtain the output result of the gear damage detection model. In the backbone network, a damage propagation inference attention module is introduced into some feature extraction modules; the damage propagation inference attention module combines the damage propagation inference module and the attention enhancement module for feature extraction; the damage propagation inference module includes a physical feature extraction layer, a damage inference layer and an output decision layer connected in sequence. The image of the gear under test is acquired and input into the gear damage detection model for detection to obtain the damage type and corresponding damage degree of the gear under test; The damage severity of different damage types is weighted and fused, and the remaining life of the gear under test is obtained based on the fusion result, thus completing the damage detection of the gear under test.

2. The gear damage detection method incorporating damage propagation prediction according to claim 1, characterized in that: The damage inference layer introduces a damage inference network for feature processing; the damage inference network is used to quantify damage length and damage width, and an attention kernel is generated through a mapping function to process the input features of the damage inference network.

3. The gear damage detection method incorporating damage propagation prediction according to claim 1, characterized in that: The physical feature extraction layer includes a series of convolutional layers and a C2f module; the damage inference layer includes a first branch and a second branch in parallel; the first branch includes a C3 module, an activation function, and a convolutional layer connected in sequence; the second branch includes a series of damage inference networks and an activation function; the output decision layer is used to process the fusion result of the output features of the first branch and the second branch in the damage inference layer; the output decision layer includes a C2fPSA module, a partial self-attention module, a convolutional layer, and an activation function connected in sequence.

4. The gear damage detection method incorporating damage propagation prediction according to claim 3, characterized in that: The partial self-attention module calculates the global contextual relationship of the input features through a multi-head self-attention mechanism, and concatenates the calculation result with the input features to obtain the output features of the partial self-attention module.

5. The gear damage detection method incorporating damage propagation prediction according to claim 1, characterized in that: The damage propagation inference attention module includes a first convolutional block, a second convolutional block, a damage propagation inference module, an attention enhancement module, a feature weighting module, a third convolutional block, and a convolutional layer connected in sequence; the attention enhancement module uses a self-attention mechanism to perform global context optimization on the output of the damage propagation inference module. The feature weighting module is used to normalize the output of the attention enhancement module and then multiply the normalized result element-wise with the output of the attention enhancement module.

6. The gear damage detection method incorporating damage propagation prediction according to claim 1, characterized in that: The backbone network has five layers for the feature extraction module; the neck network has four layers for the feature fusion module. In the neck network, the input of the first-layer feature fusion module is the concatenation result of the output features of the third-layer and last-layer feature extraction modules after convolution processing; the input of the second-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the second-layer feature extraction module; the input of the third-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the first-layer feature fusion module; the input of the last-layer feature fusion module is the concatenation result of the output features of the previous-layer feature fusion module after convolution processing and the output features of the last-layer feature extraction module.

7. The gear damage detection method incorporating damage propagation prediction according to claim 1, characterized in that: In the backbone network, except for the last layer feature extraction module, the low-level feature extraction modules have the same structure, all of which include cascaded convolutional layers and damage propagation inference attention modules. The high-level feature extraction modules have the same structure, each including a series of convolutional layers and feature extraction sub-blocks; the last feature extraction module includes a series of fast spatial pyramid pooling modules and damage propagation inference attention modules.

8. The gear damage detection method incorporating damage propagation prediction according to claim 7, characterized in that: Both the feature extraction sub-block and the feature fusion module adopt the C3k2 module.

9. A gear damage detection device incorporating damage propagation prediction, characterized in that: The gear damage detection device is used to perform the gear damage detection method with damage propagation inference as described in claim 1. The gear damage detection device includes an image acquisition mechanism and a gear damage detection module. The image acquisition mechanism is used to acquire images of the gear under test and input them to the gear damage detection module to obtain the remaining life of the gear under test.

10. A gear damage detection device incorporating damage propagation prediction according to claim 9, characterized in that: The image acquisition mechanism includes a body and a fixed module (1), an image acquisition module (2), and a motion module (3) mounted on the body; the fixed module (1) is used to mount the gear under test; the image acquisition module (2) is mounted on the motion module (3) and is used to acquire images of the gear under test; the motion module (3) includes a horizontal linear module and a vertical linear module; the horizontal linear module is used to drive the vertical linear module, thereby driving the image acquisition module on the vertical linear module to move in the horizontal and vertical directions.

Images