Machine vision detection method and system for helical gear wear
By combining a directional adjustable Gabor filter bank and Hough transform with non-subsampled Contourlet decomposition, the misjudgment problem in helical gear wear detection is solved, and the accurate positioning and prediction of the wear area is achieved, making it suitable for rapid detection in industrial settings.
Patent Information
- Application Number
- CN202511889368.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-15
- Publication Date
- 2026-03-17
AI Technical Summary
Existing technologies make it difficult to visualize the microscopic morphology of the wear area of helical gears. Traditional edge detection algorithms are prone to misjudging the helical boundary as the wear area, and it is difficult to distinguish the elastic deformation area, resulting in inaccurate wear detection.
An directionally adjustable Gabor filter bank and Hough transform are used to detect the main direction of the tooth surface in real time. Combined with non-subsampled Contourlet decomposition and a frequency band cross-attention module, the spiral texture and wear features are separated. Through physical field-constrained deep learning and dynamic deformation compensation model, the wear area can be accurately located and predicted.
It achieves precise separation and positioning of the wear area of helical gears, reduces the deformation misjudgment rate, meets the needs of rapid detection in industrial sites, and provides real-time and scalable wear detection.
Smart Images
Figure CN121686084A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of helical gear wear detection technology, and in particular to a machine vision detection method and system for helical gear wear. Background Technology
[0002] In mechanical transmission systems, helical gears, due to long-term exposure to cyclic contact stress, poor lubrication, or particle intrusion, gradually develop failure modes such as abrasive wear, fatigue pitting, or scuffing on their tooth surfaces. This morphological deterioration not only leads to decreased transmission accuracy and increased vibration and noise but can even cause catastrophic failures such as tooth breakage. Therefore, there is an urgent need to establish an efficient and accurate wear detection mechanism. Machine vision technology, through high-resolution imaging and algorithm analysis, can non-contactly acquire the three-dimensional morphology of the tooth surface. Combined with edge detection and texture analysis, it can accurately locate wear areas and classify the degree of wear and predict lifespan through deep learning models. Its advantages lie in real-time performance, objectivity, and scalability, especially in significantly reducing misjudgments under complex operating conditions. It provides data-driven support for gear health monitoring, aligning with the Industry 4.0 intelligent maintenance trend.
[0003] While existing methods based on vibration spectrum analysis can capture macroscopic fault characteristics, they struggle to visualize the microscopic morphology of wear areas. Contact measurement techniques (such as coordinate measuring machines) are limited by probe size and measurement speed, failing to meet the rapid inspection needs of industrial sites. Especially when using machine vision inspection, the unique helical texture of helical gears exhibits severe spectral aliasing with wear feature frequency bands in spatial images, causing traditional edge detection algorithms to misidentify helical boundaries as wear areas. Furthermore, during gear meshing, the elastic deformation of the tooth surface due to contact stress alters the instantaneous tooth profile shape. Without a dynamic deformation compensation model, existing algorithms may incorrectly identify elastic deformation areas as permanent wear or miss true wear boundaries. Therefore, a machine vision inspection method and system for helical gear wear is proposed. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a machine vision inspection method and system for helical gear wear.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A machine vision inspection method for helical gear wear includes the following steps: Step 1: Image Acquisition: Start the mechanical device, drive the helical gear to rotate to the initial position, acquire the gear image, and transmit the image data to the computer; Step 2: Dynamic Helical Feature Enhancement Processing: Instantaneous helical angle tracking, using Hough transform to detect the main direction of the tooth surface in real time, constructing a direction-adjustable Gabor filter bank (performing three-level Gabor filtering, fundamental frequency layer: extracting the fundamental frequency of the standard helix (Δf=λ / 3), harmonic layer: capturing 2nd / 3rd order nonlinear harmonics, transient layer: detecting impact pulse characteristics (Δf=λ / 6)), compensating for gear dynamic deformation, and highlighting wear-related features; Step 3: Multi-resolution feature decoupling: Separate wear features at different scales, perform non-downsampling Contourlet decomposition, preserve edge details, construct a channel attention module (CAM) to enhance spiral direction features, design a cross-band attention module (CBAM) to establish multi-scale associations, strengthen the association of key features, and directly input the output multi-scale features into the deep learning model in Step 4, while complementing the band separation in Step 2; Step 4: Physics-constrained deep learning: Integrating physical mechanisms and data-driven approaches, the enhanced image and strain field data are input into ResNet-101, which correlates the visual and strain fields to output a wear probability heatmap. Step 5: Dynamic Deformation Sensing Registration: Collect real-time strain data sequences, calculate the three-dimensional deformation, use the three-dimensional deformation as a rigid constraint term of the ICP algorithm to compensate for dynamic deformation, perform point cloud adaptive alignment and registration, and feed the alignment and registration results back to Step 7. Step Six: Adaptive Decision Making for Operating Conditions: Using UMAP+t-SNE, the operating condition parameters are mapped to an 8-dimensional prototype library. The membership degree of the current operating condition is calculated. The dynamic threshold is calculated based on the TS fuzzy system. The combined probability map threshold, physical field residual, and deformation anomaly value are used for comprehensive decision making. The decision results directly trigger the warning in Step Seven, which depends on the operating condition parameters and probability map in Step One and Four. Step 7: Health status assessment and early warning: Calculate the wear index and give the maintenance window period by combining the remaining life prediction model (LSTM).
[0006] The above plan further includes: Furthermore, in step one, for small helical gears, a high-resolution area scan camera with a suitable frame rate is selected; for large helical gears, a line scan camera is used in conjunction with scanning motion to acquire a complete image, and at the same time, calibration is performed using a 16-channel fiber Bragg grating (FBG) array sensor.
[0007] Furthermore, in step two, the specific steps of the dynamic spiral feature enhancement processing are as follows: Local orientation field estimation of tooth surface: Anisotropic diffusion filtering is applied to the tooth surface image to preserve edges while suppressing noise, as shown below. ,in, To determine the edge-sensitive diffusion coefficient, the tooth surface is divided into 16×16 pixel local windows. Hough line detection is performed on each window, and the angle histogram is calculated. The principal direction angle is determined by parabolic fitting, and is expressed as... The theoretical helix angle offset is calculated based on the rotational speed n and torque F, and is expressed as follows: The final filter direction angle is obtained. ; Directional tunable Gabor filter bank design: Construct a directional tunable Gabor filter bank to perform three-stage Gabor filtering, including a fundamental frequency layer, a harmonic layer, and a transient layer. The fundamental frequency layer filter extracts the standard spiral fundamental frequency with a bandwidth of [missing information]. The harmonic layer filter captures second-order / third-order nonlinear harmonics with a bandwidth of [missing information]. The transient layer filter detects impulse pulse characteristics and has a bandwidth of ; Multi-scale feature fusion: for preprocessed images Perform a three-stage Gabor filter: ; According to the fiber Bragg grating (FBG) strain field Calculate the local deformation field, expressed as The filtering result is deformed and compensated, and is expressed as follows: ,in, For deformation mapping function, For the impulse filter, calculate the multi-scale energy map using the following formula: Wear features are extracted through adaptive threshold segmentation.
[0008] Furthermore, in step three, the specific steps for decoupling the multi-resolution features are as follows: Non-subsampled Contourlet Decomposition (NSCT): Decomposes the gear image into multi-scale, multi-directional sub-bands, preserving the details of the spiral edge. Scale separation is performed using the Non-subsampled Pyramid (NSP). (Original image) After a 6-level decomposition, low-frequency and high-frequency components are obtained. Each high-frequency component is further decomposed into 8 directional sub-bands (to adapt to the multi-directionality of the spiral). After non-subsampled pyramid (NSP) decomposition, directional filter bank (DFB) is used to capture the directional features of the spiral. The Channel Attention Module (CAM) enhances features related to the spiral direction: In the frequency band channel dimension, it adaptively enhances features related to the spiral direction, calculates global statistics for each directional sub-band, and generates channel weights through a fully connected layer (FC) and sigmoid activation. This results in weighted features. Band Cross-Attention Module (CBAM) establishes multi-scale associations: for adjacent scales and The spatial similarity of the high-frequency subbands is calculated, and cross-scale features are fused by weighting based on the similarity matrix; Multi-resolution feature reconstruction: Inverse non-subsampling Contourlet decomposition (NSCT) is performed on each weighted subband for reconstruction, and wear probability maps are generated through convolutional layers.
[0009] Furthermore, in step four, the specific steps of the physical field-constrained deep learning are as follows: Multimodal data alignment and preprocessing: synchronous triggering of the camera and fiber Bragg grating (FBG) array sensor via gear meshing cycle; A dual-branch feature extraction network is used, comprising visual features and strain field features. The visual features are obtained by modifying ResNet-101, removing the last fully connected layer, and inserting a spatial attention module (SAM) after the res4 layer to output a feature map. The strain field features are obtained by constructing a fully connected network (FCN), which consists of: an input layer of 16-channel strain data (corresponding to 16 FBG measurement points); hidden layers of a 3-layer MLP (256-128-64 neurons); and an output feature vector. Multimodal feature fusion: Calculate the strain field attention weights, perform weighted fusion of visual features (channel-by-channel weighting), construct the physical field constraint matrix, perform constrained feature mapping, and obtain the physical field constrained features; Design a joint loss function: including classification loss (cross-entropy), physical field constraint loss and strain field reconstruction loss, and perform weighted fusion to obtain the total loss function; Wear probability map generation: Construct a transposed convolutional network, input the physical field-constrained features, output the wear probability map, apply total variational regularization to post-process and optimize the wear probability map to obtain a wear probability heatmap, and obtain a binary wear mask through threshold segmentation.
[0010] Furthermore, in step five, the specific steps of the dynamic deformation sensing registration are as follows: FBG sensor data acquisition and preprocessing: The strain field distribution in the gear tooth root region is captured in real time by a 16-channel FBG array to reflect the elastic deformation during meshing. Moving average filtering is used to eliminate high-frequency vibration noise (such as mechanical shock and electromagnetic interference). 3D Deformation Field Reconstruction: By using pre-calibrated constitutive equations, discrete strain data is converted into a continuous 3D deformation field, a quantitative relationship between strain and displacement is established, and a high-resolution deformation field is generated by radial basis function (RBF) interpolation to compensate for the spatial discreteness of the sensor and support point cloud-level registration constraints. Strain-constrained ICP algorithm: Introducing deformation constraint terms into the traditional ICP objective function to compensate for elastic deformation during gear meshing; Adaptive registration execution: The convergence threshold is dynamically adjusted according to real-time operating conditions (speed, load) to balance accuracy and computational efficiency. Through KD-Tree acceleration and GPU parallel computing, millisecond-level registration is achieved to meet online detection requirements.
[0011] Furthermore, in step six, the specific steps of the adaptive decision-making process based on operating conditions are as follows: Operating Parameter Acquisition: Real-time acquisition of physical parameters (speed, load, temperature, etc.) reflecting the operating status of the equipment to construct a high-dimensional feature space; Hybrid Manifold Learning Dimensionality Reduction (UMAP+t-SNE): Maps 8-dimensional parameters to 2D / 3D space, preserves key topological structures, defines typical working condition patterns by clustering historical data, and achieves standardization of working condition classification; Membership calculation (Gaussian kernel function): quantifies the similarity between the current working condition and each pattern in the prototype library, and realizes soft classification (fuzzy classification) of the working condition. TS fuzzy dynamic threshold generation: dynamically adjusts the warning threshold based on the working condition prototype; Multi-indicator decision fusion: Combining probability maps (local damage), physical field residuals (global stress anomalies), and deformation anomalies (structural deformation), the fused indicators are compared with dynamic thresholds to provide early warning; Early warning triggering mechanism: The early warning level is divided according to the number of consecutive exceedances and the threshold range. The instantaneous fluctuations are smoothed out through the time window to avoid false triggering.
[0012] Furthermore, in step seven, the specific steps for the health status assessment and early warning are as follows: Multimodal feature fusion calculation of wear index: Transform multi-source heterogeneous data from machine vision, mechanical sensing and working condition parameters into a unified quantitative indicator of health status. The wear degree is comprehensively reflected by a weighted formula with clear physical meaning (stress variance + high frequency energy + working condition load), and adaptive threshold adjustment is performed based on historical statistical average. LSTM Remaining Life Prediction Model: Utilizes the memory cells of LSTM to handle the long-term dependency characteristics of wear evolution, establishes a nonlinear mapping relationship from the current state to the future failure time, and balances real-time performance and contextual information through a sliding window mechanism; Dynamic maintenance window decision: The three states of emergency, planned and monitoring are divided by dual thresholds. The threshold parameters are dynamically adjusted according to economic models such as spare parts cost and downtime loss. After the decision results are fed back to the production system, the actual maintenance records are used to back-optimize the LSTM remaining life prediction model. Online iterative optimization: The maintained health status data is incorporated into the training set to form a reinforcement learning loop of prediction-decision-validation-optimization.
[0013] A machine vision inspection system used in a machine vision inspection method for helical gear wear includes: Multimodal data acquisition module: Real-time synchronous acquisition of multi-source heterogeneous data on gear operating status, outputting strain field time series data, high dynamic range image stream, and operating environment parameters; Dynamic feature enhancement module: compensates for motion blur and enhances wear-related features, outputting orientation-corrected images and frequency domain enhanced feature maps; Multi-resolution decoupling network: Separates wear patterns of different morphological features across scales and outputs multi-scale feature tensors (low-frequency base texture + high-frequency edge details). Physics-constrained inference engine: It integrates physical mechanisms and data-driven detection to output wear probability maps and contact stress field predictions; Dynamic Deformation Registration Module: Compensates for elastic deformation during gear meshing and outputs a 3D topography reconstruction after deformation compensation; Operating condition adaptive decision-making module: dynamically adjusts detection thresholds to reduce false alarm rate.
[0014] The present invention has the following beneficial effects: 1. In this invention, a direction-adjustable Gabor filter bank dynamically tracks changes in the helix angle, and Hough transform is used to detect the principal direction of the tooth surface in real time, achieving accurate separation of helical gear helical texture and wear characteristics. Specifically, the real-time detection of the principal direction of the tooth surface using Hough transform overcomes the sensitivity of traditional edge detection algorithms to helical texture, and the directional and scale selectivity of the Gabor filter can be adaptively adjusted, effectively suppressing helical texture interference and highlighting wear-related frequency band characteristics.
[0015] 2. This invention employs a non-contact machine vision + dynamic deformation compensation technology approach. Through real-time strain data sequence acquisition and three-dimensional deformation calculation, dynamic deformation is introduced as a rigid constraint into the ICP algorithm to achieve adaptive alignment and registration of point clouds, eliminating rigid body displacement errors and ensuring accurate positioning of the wear area boundary. This solution improves detection speed and eliminates the need for physical contact, making it suitable for harsh industrial environments.
[0016] 3. In this invention, a dynamic deformation compensation model is constructed. The Gabor filter bank tracks the change of helix angle and combines the rigid constraint term of the ICP algorithm to correct the feature offset caused by elastic deformation in real time, thereby reducing the deformation misjudgment rate. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the steps of a machine vision inspection method for helical gear wear proposed in this invention. Figure 2 This is a system block diagram of a machine vision inspection system for helical gear wear proposed in this invention. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] Please see Figure 1 As shown, this invention is a machine vision inspection method for helical gear wear, comprising the following steps: Step 1: Image Acquisition: Start the mechanical device, drive the helical gear to rotate to the initial position, acquire the gear image, and transmit the image data to the computer; Step 2: Dynamic Helical Feature Enhancement Processing: Instantaneous helical angle tracking, using Hough transform to detect the main direction of the tooth surface in real time, constructing a direction-adjustable Gabor filter bank (performing three-level Gabor filtering, fundamental frequency layer: extracting the fundamental frequency of the standard helix (Δf=λ / 3), harmonic layer: capturing 2nd / 3rd order nonlinear harmonics, transient layer: detecting impact pulse characteristics (Δf=λ / 6)), compensating for gear dynamic deformation, and highlighting wear-related features; Step 3: Multi-resolution feature decoupling: Separate wear features at different scales, perform non-downsampling Contourlet decomposition, preserve edge details, construct a channel attention module (CAM) to enhance spiral direction features, design a cross-band attention module (CBAM) to establish multi-scale associations, strengthen the association of key features, and directly input the output multi-scale features into the deep learning model in Step 4, while complementing the band separation in Step 2; Step 4: Physics-constrained deep learning: Integrating physical mechanisms and data-driven approaches, the enhanced image and strain field data are input into ResNet-101, which correlates the visual and strain fields to output a wear probability heatmap. Step 5: Dynamic Deformation Sensing Registration: Collect real-time strain data sequences, calculate the three-dimensional deformation, use the three-dimensional deformation as a rigid constraint term of the ICP algorithm to compensate for dynamic deformation, perform point cloud adaptive alignment and registration, and feed the alignment and registration results back to Step 7. Step Six: Adaptive Decision Making for Operating Conditions: Using UMAP+t-SNE, the operating condition parameters are mapped to an 8-dimensional prototype library. The membership degree of the current operating condition is calculated. The dynamic threshold is calculated based on the TS fuzzy system. The combined probability map threshold, physical field residual, and deformation anomaly value are used for comprehensive decision making. The decision results directly trigger the warning in Step Seven, which depends on the operating condition parameters and probability map in Step One and Four. Step 7: Health status assessment and early warning: Calculate the wear index and give the maintenance window period by combining the remaining life prediction model (LSTM).
[0020] In one embodiment, in step one, for small helical gears, a high-resolution area scan camera with a suitable frame rate is selected; for large helical gears, a line scan camera is used in conjunction with scanning motion to acquire a complete image, and at the same time, calibration is performed using a 16-channel fiber Bragg grating (FBG) array sensor.
[0021] In one embodiment, the specific steps of the dynamic spiral feature enhancement process in step two are as follows: Local orientation field estimation of tooth surface: Anisotropic diffusion filtering is applied to the tooth surface image to preserve edges while suppressing noise, as shown below. ,in, To determine the edge-sensitive diffusion coefficient, the tooth surface is divided into 16×16 pixel local windows. Hough line detection is performed on each window, and the angle histogram is calculated. The principal direction angle is determined by parabolic fitting, and is expressed as... The theoretical helix angle offset is calculated based on the rotational speed n and torque F, and is expressed as follows: The final filter direction angle is obtained. ; Directional tunable Gabor filter bank design: Construct a directional tunable Gabor filter bank to perform three-stage Gabor filtering, including a fundamental frequency layer, a harmonic layer, and a transient layer. The fundamental frequency layer filter extracts the standard spiral fundamental frequency with a bandwidth of [missing information]. The harmonic layer filter captures second-order / third-order nonlinear harmonics with a bandwidth of [missing information]. The transient layer filter detects impulse pulse characteristics and has a bandwidth of ; Multi-scale feature fusion: for preprocessed images Perform a three-stage Gabor filter: ; According to the fiber Bragg grating (FBG) strain field Calculate the local deformation field, expressed as The filtering result is deformed and compensated, and is expressed as follows: ,in, For deformation mapping function, For the impulse filter, calculate the multi-scale energy map using the following formula: Wear features are extracted through adaptive threshold segmentation.
[0022] In one embodiment, the specific steps of multi-resolution feature decoupling in step three are as follows: Non-subsampled Contourlet Decomposition (NSCT): Decomposes the gear image into multi-scale, multi-directional sub-bands, preserving the details of the spiral edge. Scale separation is performed using the Non-subsampled Pyramid (NSP). (Original image) After a 6-level decomposition, low-frequency and high-frequency components are obtained. Each high-frequency component is further decomposed into 8 directional sub-bands (to adapt to the multi-directionality of the spiral). After non-subsampled pyramid (NSP) decomposition, directional filter bank (DFB) is used to capture the directional features of the spiral. The Channel Attention Module (CAM) enhances features related to the spiral direction: In the frequency band channel dimension, it adaptively enhances features related to the spiral direction, calculates global statistics for each directional sub-band, and generates channel weights through a fully connected layer (FC) and sigmoid activation. This results in weighted features. Band Cross-Attention Module (CBAM) establishes multi-scale associations: for adjacent scales and The spatial similarity of the high-frequency subbands is calculated, and cross-scale features are fused by weighting based on the similarity matrix; Multi-resolution feature reconstruction: Inverse non-subsampling Contourlet decomposition (NSCT) is performed on each weighted subband for reconstruction, and wear probability maps are generated through convolutional layers.
[0023] In one embodiment, the specific steps of the physical field-constrained deep learning in step four are as follows: Multimodal data alignment and preprocessing: synchronous triggering of the camera and fiber Bragg grating (FBG) array sensor via gear meshing cycle; A dual-branch feature extraction network is used, comprising visual features and strain field features. The visual features are obtained by modifying ResNet-101, removing the last fully connected layer, and inserting a spatial attention module (SAM) after the res4 layer to output a feature map. The strain field features are obtained by constructing a fully connected network (FCN), which consists of: an input layer of 16-channel strain data (corresponding to 16 FBG measurement points); hidden layers of a 3-layer MLP (256-128-64 neurons); and an output feature vector. Multimodal feature fusion: Calculate the strain field attention weights, perform weighted fusion of visual features (channel-by-channel weighting), construct the physical field constraint matrix, perform constrained feature mapping, and obtain the physical field constrained features; Design a joint loss function: including classification loss (cross-entropy), physical field constraint loss and strain field reconstruction loss, and perform weighted fusion to obtain the total loss function; Wear probability map generation: Construct a transposed convolutional network, input the physical field-constrained features, output the wear probability map, apply total variational regularization to post-process and optimize the wear probability map to obtain a wear probability heatmap, and obtain a binary wear mask through threshold segmentation.
[0024] In one embodiment, the specific steps of the dynamic deformation sensing registration in step five are as follows: FBG sensor data acquisition and preprocessing: The strain field distribution in the gear tooth root region is captured in real time by a 16-channel FBG array to reflect the elastic deformation during meshing. Moving average filtering is used to eliminate high-frequency vibration noise (such as mechanical shock and electromagnetic interference). 3D Deformation Field Reconstruction: By using pre-calibrated constitutive equations, discrete strain data is converted into a continuous 3D deformation field, a quantitative relationship between strain and displacement is established, and a high-resolution deformation field is generated by radial basis function (RBF) interpolation to compensate for the spatial discreteness of the sensor and support point cloud-level registration constraints. Strain-constrained ICP algorithm: Introducing deformation constraint terms into the traditional ICP objective function to compensate for elastic deformation during gear meshing; Adaptive registration execution: The convergence threshold is dynamically adjusted according to real-time operating conditions (speed, load) to balance accuracy and computational efficiency. Through KD-Tree acceleration and GPU parallel computing, millisecond-level registration is achieved to meet online detection requirements.
[0025] In one embodiment, the specific steps of the adaptive decision-making process in step six are as follows: Operating Parameter Acquisition: Real-time acquisition of physical parameters (speed, load, temperature, etc.) reflecting the operating status of the equipment to construct a high-dimensional feature space; Hybrid Manifold Learning Dimensionality Reduction (UMAP+t-SNE): Maps 8-dimensional parameters to 2D / 3D space, preserves key topological structures, defines typical working condition patterns by clustering historical data, and achieves standardization of working condition classification; Membership calculation (Gaussian kernel function): quantifies the similarity between the current working condition and each pattern in the prototype library, and realizes soft classification (fuzzy classification) of the working condition. TS fuzzy dynamic threshold generation: dynamically adjusts the warning threshold based on the working condition prototype; Multi-indicator decision fusion: Combining probability maps (local damage), physical field residuals (global stress anomalies), and deformation anomalies (structural deformation), the fused indicators are compared with dynamic thresholds to provide early warning; Early warning triggering mechanism: The early warning level is divided according to the number of consecutive exceedances and the threshold range. The instantaneous fluctuations are smoothed out through the time window to avoid false triggering.
[0026] In one embodiment, the specific steps of the health status assessment and early warning in step seven are as follows: Multimodal feature fusion calculation of wear index: Transform multi-source heterogeneous data from machine vision, mechanical sensing and working condition parameters into a unified quantitative indicator of health status. The wear degree is comprehensively reflected by a weighted formula with clear physical meaning (stress variance + high frequency energy + working condition load), and adaptive threshold adjustment is performed based on historical statistical average. LSTM Remaining Life Prediction Model: Utilizes the memory cells of LSTM to handle the long-term dependency characteristics of wear evolution, establishes a nonlinear mapping relationship from the current state to the future failure time, and balances real-time performance and contextual information through a sliding window mechanism; Dynamic maintenance window decision: The three states of emergency, planned and monitoring are divided by dual thresholds. The threshold parameters are dynamically adjusted according to economic models such as spare parts cost and downtime loss. After the decision results are fed back to the production system, the actual maintenance records are used to back-optimize the LSTM remaining life prediction model. Online iterative optimization: The maintained health status data is incorporated into the training set to form a reinforcement learning loop of prediction-decision-validation-optimization.
[0027] A machine vision inspection system used in a machine vision inspection method for helical gear wear includes: Multimodal data acquisition module: Real-time synchronous acquisition of multi-source heterogeneous data on gear operating status, outputting strain field time series data, high dynamic range image stream, and operating environment parameters; Dynamic feature enhancement module: compensates for motion blur and enhances wear-related features, outputting orientation-corrected images and frequency domain enhanced feature maps; Multi-resolution decoupling network: Separates wear patterns of different morphological features across scales and outputs multi-scale feature tensors (low-frequency base texture + high-frequency edge details). Physics-constrained inference engine: It integrates physical mechanisms and data-driven detection to output wear probability maps and contact stress field predictions; Dynamic Deformation Registration Module: Compensates for elastic deformation during gear meshing and outputs a 3D topography reconstruction after deformation compensation; Operating condition adaptive decision-making module: dynamically adjusts detection thresholds to reduce false alarm rate.
[0028] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A machine vision method for detecting wear in helical gears, characterized by, The method comprises the following steps: Step one: image acquisition: start the mechanical device, drive the helical gear to rotate to the initial position, collect the gear image, and transmit the image data to the computer; Step two: dynamic spiral feature enhancement processing: instantaneous spiral angle tracking, real-time detection of the main direction of the gear surface using Hough transform, construction of a direction-adjustable Gabor filter bank, compensation for dynamic deformation of the gear, and highlighting of wear-related features; Step three: multi-resolution feature decoupling: separate different scale wear features, non-subsampled Contourlet decomposition, preserve edge details, construct a channel attention module to strengthen the spiral line direction feature, design a frequency band cross-attention module to establish multi-scale correlation, and strengthen the key feature correlation; Step four: physical field constraint deep learning: fusion of physical mechanism and data-driven, input of enhanced images and strain field data into ResNet-101, correlation of vision and strain field, and output of wear probability heat map; Step five: dynamic deformation perception registration: collect real-time strain data sequence, calculate three-dimensional deformation, use three-dimensional deformation as a rigid constraint term for ICP algorithm to compensate for dynamic deformation, and perform point cloud adaptive alignment registration; Step six: working condition adaptive decision: map the working condition parameters to an 8-dimensional prototype library, calculate the current working condition membership degree, calculate the dynamic threshold according to the TS fuzzy system, and make a comprehensive decision based on the probability map threshold, physical field residual error, and deformation abnormal value; Step seven: health state evaluation and early warning: calculate the wear index, and give the maintenance window period in combination with the residual life prediction model.
2. A machine vision method for detecting wear in helical gears as claimed in claim 1, wherein, In step one, for small helical gears, a surface array camera with high resolution and appropriate frame rate is selected; for large helical gears, a line array camera is used to obtain complete images by scanning motion, and at the same time, a 16-channel fiber Bragg grating array sensor is used for calibration.
3. A machine vision method for detection of wear in helical gears as claimed in claim 1, wherein, In step two, the specific steps of the dynamic spiral feature enhancement processing are as follows: Local direction field estimation: anisotropic diffusion filtering is applied to the tooth surface image to preserve edges while suppressing noise, denoted as where, is the edge-sensitive diffusion coefficient, the tooth surface is divided into local windows of 16x16 pixels, Hough line detection is performed for each window, and the angle histogram is counted The main direction angle is determined by parabolic fitting, denoted as The theoretical helix angle offset is calculated according to the rotational speed n and the torque F, denoted as The final filter direction angle is obtained ; Direction-adjustable Gabor filter bank design: a direction-adjustable Gabor filter bank is constructed, and three-level Gabor filtering is performed, including a fundamental frequency layer, a harmonic layer, and a transient layer, the fundamental frequency layer filter extracts a standard spiral line fundamental frequency, the bandwidth , the harmonic layer filter captures 2nd / 3rd order nonlinear harmonics, the bandwidth , the transient layer filter detects impact pulse characteristics, the bandwidth ; Multi-scale feature fusion: on the pre-processed image Performing a three-level Gabor filtering: ； According to the strain field of fiber bragg grating The local deformation field is calculated and represented as The deformation compensation is performed on the filtering result, and is represented as Wherein, The deformation mapping function is The impact filter is used to calculate the multi-scale energy graph, and the calculation formula is The wear feature is extracted through adaptive threshold segmentation.
4. A machine vision method for detection of wear in helical gears as claimed in claim 1, wherein, In step three, the specific steps of the multi-resolution feature decoupling are as follows: Nonsubsampled contourlet transform: the gear image is decomposed into multi-scale and multi-direction subbands, the spiral edge details are reserved, the nonsubsampled pyramid is used for scale separation, the original image After 6-level decomposition, low-frequency components and high-frequency components are obtained, each high-frequency component is further decomposed into 8 direction subbands, and the spiral direction features are captured by using a direction filter bank after nonsubsampled pyramid decomposition; Channel attention module strengthens the spiral line direction: in the frequency band channel dimension, adaptively enhance the features related to the spiral line direction, calculate the global statistics for each direction subband, generate channel weights through a fully connected layer and a Sigmoid activation, and then obtain the weighted features; The band cross attention module establishes multiscale correlation: the spatial similarity of high-frequency subbands of adjacent scales and is calculated, and the cross-scale features are fused according to a similarity matrix. Multi-resolution feature reconstruction: perform inverse non-subsampled Contourlet decomposition and reconstruction on the weighted subbands, and generate a wear probability map through a convolution layer.
5. A machine vision method for detection of wear in helical gears as claimed in claim 1, wherein, In step four, the specific steps of the physical field constraint deep learning are as follows: Multi-modal data alignment and preprocessing: trigger the camera and fiber Bragg grating array sensor synchronously through the gear meshing period; Dual-branch feature extraction network: includes visual features and strain field features, the visual features are obtained by modifying ResNet-101, removing the last fully connected layer, and inserting a spatial attention module after the res4 layer, and then outputting the feature map, and the strain field features are obtained by constructing a fully connected network; Multi-modal feature fusion: calculate the strain field attention weight, weight the visual features, construct a physical field constraint matrix, perform constraint feature mapping, and obtain the physical field constrained features; Design joint loss function: including classification loss, physical field constraint loss and strain field reconstruction loss, weighted fusion to get total loss function; Wear probability map generation: build transpose convolution network, input physical field constrained features, output wear probability map, apply total variation regularization to optimize wear probability map, get wear probability heat map, get binary wear mask through threshold segmentation.
6. A machine vision method for detection of wear in helical gears as claimed in claim 1, wherein, In step five, the specific steps of the dynamic deformation-aware registration are: FBG sensor data acquisition and preprocessing: real-time capture of strain field distribution in gear root area through 16-channel FBG array, reflecting elastic deformation in meshing process, using moving average filter to eliminate high-frequency vibration noise; Three-dimensional deformation field reconstruction: through the pre-calibrated constitutive equation, the discrete strain data is converted into continuous three-dimensional deformation field, the quantitative relationship between strain and displacement is established, and the radial basis function interpolation is used to generate high-resolution deformation field, which compensates for the spatial dispersion of sensors and supports point cloud level registration constraints; Strain-constrained ICP algorithm: introduce deformation constraint term in traditional ICP objective function, compensate for elastic deformation during gear meshing; Adaptive registration execution: dynamically adjust the convergence threshold according to real-time working conditions, balance accuracy and computational efficiency.
7. A machine vision method for detection of wear in helical gears as claimed in claim 1, wherein In step six, the specific steps of the working condition adaptive decision are: Working condition parameter acquisition: real-time acquisition of physical parameters reflecting the running state of the equipment, and construction of high-dimensional feature space; Mixed manifold learning dimension reduction: map 8-dimensional parameters to 2D / 3D space, preserve key topological structure, define typical working condition mode by clustering historical data, realize standardization of working condition classification; Membership calculation: quantize the similarity between current working condition and each mode in the prototype library, realize soft classification of working condition; TS fuzzy dynamic threshold generation: dynamically adjust the warning threshold according to the working condition prototype; Multi-index decision fusion: integrate probability map, physical field residual error and deformation anomaly, compare the fusion index with dynamic threshold, and perform early warning; Early warning trigger mechanism: according to the number of continuous exceedances and threshold interval, divide the warning level, smooth the instantaneous fluctuations through time window, and avoid false triggering.
8. A machine vision method for detection of wear in helical gears as claimed in claim 1, wherein, In step seven, the specific steps of the health state evaluation and early warning are: Multi-modal feature fusion to calculate wear index: convert multi-source heterogeneous data of machine vision, mechanical sensing and working condition parameters into unified health state quantitative index, comprehensively reflect wear degree through physical significance clear weighted formula, and adjust adaptive threshold based on historical statistical mean value; LSTM residual life prediction model: use the memory unit of LSTM to process the long-term dependence characteristics of wear evolution, establish the nonlinear mapping relationship from the current state to the future failure time, and balance real-time and context information through sliding window mechanism; Dynamic maintenance window period decision: divide three states of emergency / planning / monitoring through double threshold, and dynamically adjust the threshold parameter according to the economic model. After the decision result is fed back to the production system, the actual maintenance record is used to optimize the LSTM residual life prediction model; Online iterative optimization: include the health state data after maintenance in the training set, form the prediction-decision-verification-optimization reinforcement learning cycle.
9. A machine vision inspection system used in a machine vision inspection method for helical gear wear according to claim 1, characterized in that, Multi-modal data acquisition module: Real-time synchronous acquisition of multi-source heterogeneous data of gear running state; Dynamic feature enhancement module: Compensation of motion blur and strengthening of wear-related features; Multi-resolution decoupling network: Cross-scale separation of different morphological features of wear patterns; Physical field constraint reasoning engine: Fusion of physical mechanism and data-driven detection, output of wear probability map and contact stress field prediction; Dynamic deformation registration module: Compensation of elastic deformation during gear meshing; Working condition adaptive decision module: Dynamic adjustment of detection threshold to reduce false alarm rate.