Gear geometric parameter rapid detection method and system based on wireless transmission

Abstract

The invention provides a gear geometric parameter rapid detection method and system based on wireless transmission, and relates to the technical field of gear detection, and the method comprises the steps: for a to-be-detected gear, extracting gear features based on a deep learning network, and determining a position combination of optimal measurement points in combination with a gear precision grade requirement; collecting geometric parameter data of the gear based on the position combination; and processing the geometric parameter data by adopting a cascade neural network to obtain actual geometric parameters, sending the actual geometric parameters to a remote monitoring terminal through a wireless transmission module with a self-adaptive channel selection function, and generating a detection result report.

Description

Technical Field

[0001] The present invention relates to gear detection technology, and in particular to a fast detection method and system for gear geometric parameters based on wireless transmission. Background Art

[0002] The accurate detection of gear geometric parameters is of great significance for ensuring the performance and reliability of gear transmission systems. Currently, the industrial community generally uses fixed measurement point positions for sampling detection. This method is difficult to meet the measurement requirements of gears with different specifications and accuracy grades, and there are problems such as unreasonable measurement point distribution and low sampling efficiency when detecting complex tooth profiles. At the same time, the traditional measurement point selection method lacks the adaptive ability to gear characteristics and accuracy requirements, and cannot guarantee the representativeness and integrity of measurement data.

[0003] The data processing methods of existing gear detection systems mainly rely on traditional signal processing algorithms, and have limited ability to handle measurement noise and system errors. Especially when dealing with complex features such as pitch cumulative error, tooth profile deviation, and tooth direction deformation, it is difficult to accurately identify and separate various deviation features. In addition, the existing methods lack the effective use of gear manufacturing process knowledge and cannot fully consider the influence of process factors in the processing process on gear geometric features.

[0004] Data transmission in industrial sites mainly uses wired methods, which have problems such as complex wiring, high maintenance costs, and poor system scalability. Although wireless transmission technology has the advantage of flexible layout, it is easily affected by various electromagnetic interferences in complex industrial environments, such as interference signals generated by machine tool operation, welding equipment, high-frequency switching power supplies, etc., resulting in unstable data transmission quality and affecting the reliability of the detection system. Summary of the Invention

[0005] Embodiments of the present invention provide a fast detection method and system for gear geometric parameters based on wireless transmission, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiment of the present invention, a method for quickly detecting gear geometric parameters based on wireless transmission is provided, including: for a gear to be measured, based on a deep learning network, extracting gear features through a circular convolutional layer, a radial attention module, and an axial convolutional layer, and determining a position combination of optimal measurement points in combination with the gear accuracy grade requirements; controlling a measurement mechanism to collect gear geometric parameter data based on the position combination; a data processing unit processes the geometric parameter data by using a cascaded neural network, wherein the first-level neural network uses a polar coordinate convolutional structure for gear feature extraction and noise reduction processing, and the second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains actual geometric parameters; sending the actual geometric parameters to a remote monitoring terminal through a wireless transmission module with an adaptive channel selection function, the wireless transmission module constructs an interference situation map based on the industrial field electromagnetic interference intensity, and performs an availability score according to the interference level of the grid where the channel is located, to achieve adaptive switching of the optimal transmission channel; the remote monitoring terminal generates a detection result report according to the actual geometric parameters.

[0007] Optionally, based on a deep learning network, extracting gear features through a circular convolutional layer, a radial attention module, and an axial convolutional layer, and determining a position combination of optimal measurement points in combination with the gear accuracy grade requirements includes: collecting an image of the gear to be measured and inputting it into a gear measurement point optimization network; the gear measurement point optimization network includes: a feature extraction module, which respectively extracts the circumferential feature, radial feature, and axial feature of the gear by using a circular convolutional layer, a radial attention module, and an axial convolutional layer; a measurement point evaluation module, which constructs the gear accuracy grade requirements into a hierarchical conditional vector, fuses it with the output features of the feature extraction module, and generates an evaluation feature map of candidate measurement points; a measurement point positioning module, which outputs a probability heat map of the measurement point position and an accessibility confidence map based on the evaluation feature map, the probability heat map characterizes the spatial distribution of the optimal measurement points, and the accessibility confidence map characterizes the signal quality reliability of the measurement points; a constraint optimization module, under the guidance of the probability heat map and the accessibility confidence map, selects a position combination of optimal measurement points that satisfies the spatial distribution constraint through a differentiable non-maximum suppression algorithm.

[0008] Optionally, the gear precision level requirements are constructed into a hierarchical condition vector and fused with the output features of the feature extraction module to generate an evaluation feature map of candidate measurement points, including: constructing the gear precision level requirements into a hierarchical condition vector containing global precision parameters and local precision parameters, and respectively mapping the global precision parameters and local precision parameters through non-linear transformation to obtain conditional features; aligning the circumferential features, radial features and axial features; constructing an annular-radial-axial interaction module, which fuses the circumferential features with the conditional features of the global precision parameters, interacts the radial features with the conditional features of the local precision parameters, and performs precision compensation on the axial features to obtain fused features; constructing a precision adaptive attention map based on the conditional features of the global precision parameters and local precision parameters, and recalibrating the fused features using the precision adaptive attention map; calculating a precision sensitivity score, a geometric feature fitness score and a measurement feasibility score based on the recalibrated fused features, and performing weighted fusion to obtain an evaluation feature map of candidate measurement points.

[0009] Optionally, under the guidance of the probability heat map and the reachability confidence map, the optimal measurement point position combination that meets the spatial distribution constraint is selected through a differentiable non-maximum suppression algorithm, including: performing weighted fusion on the probability heat map and the reachability confidence map to obtain a comprehensive score map of the measurement point candidate region; constructing a differentiable non-maximum suppression model based on the comprehensive score map, and the differentiable non-maximum suppression model uses a Gaussian kernel function to calculate the spatial suppression weight between candidate measurement points, and the spatial suppression weight decreases as the distance between candidate measurement points increases; using the score value in the comprehensive score map as the initial selection probability of the candidate measurement point, and iteratively updating the initial selection probability according to the spatial suppression weight to obtain the selection probability of the candidate measurement point considering spatial suppression; setting the minimum allowable distance constraint between measurement points based on the gear module, and calculating the spatial uniformity measure of the measurement point distribution; constructing an optimization objective function, which includes a measurement point comprehensive score term, a non-maximum suppression term and a spatial uniform distribution term based on the spatial uniformity measure; iteratively optimizing the optimization objective function through the gradient descent method to update the selection probability of the candidate measurement point; determining the candidate measurement points with the optimized selection probability greater than the preset probability threshold as the optimal measurement point position combination.

[0010] Optionally, the data processing unit processes the geometric parameter data using a cascaded neural network. The first-level neural network uses a polar coordinate convolution structure for gear feature extraction and noise reduction processing. The second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains the actual geometric parameters, including: The first-level neural network uses a polar coordinate convolution layer. The convolution kernel of the polar coordinate convolution layer is designed along the circumferential direction of the gear pitch circle, and the convolution step size matches the tooth pitch. The first-level neural network also includes a rotation-invariant feature extraction module. The rotation-invariant feature extraction module extracts tooth profile features through rotation-equivariant convolution operations and an involute adaptive sampling mechanism. The features output by the rotation-invariant feature extraction module are input into a dynamic noise decomposition module. The dynamic noise decomposition module uses a combination of wavelet transform and attention mechanism to perform frequency-domain decomposition on the measurement noise to obtain a systematic error pattern. The features corresponding to the systematic error pattern are input into the second-level neural network. The second-level neural network constructs a gear topological graph structure. The graph nodes in the gear topological graph structure represent geometric feature points, and the graph edges represent geometric constraint relationships. A graph convolutional network is applied on the topological graph structure to extract structured features. An attention mechanism is constructed based on gear manufacturing process knowledge to enhance the pitch cumulative error, tooth profile deviation, and tooth direction deformation features in the structured features, and a parameterized gear geometric model is constructed. The gear processing process rules are encoded as a parameter prior distribution, and the parameterized gear geometric model is constrained and optimized based on the parameter prior distribution to obtain the actual geometric parameters of the gear.

[0011] Optionally, an attention mechanism is constructed based on gear manufacturing process knowledge to enhance the pitch cumulative error, tooth profile deviation, and tooth direction deformation features in the structured features. Constructing a parameterized gear geometric model includes: constructing a multi-level process feature extraction network to extract tool wear features, heat treatment deformation features, and clamping deformation features respectively, and calculating the response distributions of the pitch cumulative error, tooth profile deviation, and tooth direction deformation features based on the extracted features; constructing a process correlation matrix based on the response distributions. The process correlation matrix characterizes the spatial distribution and temporal variation characteristics of the pitch cumulative error, tooth profile deviation, and tooth direction deformation features between adjacent tooth positions; constructing a multi-head attention mechanism according to the process correlation matrix. The multi-head attention mechanism calculates local attention scores at each scale level; adaptively fusing the process correlation matrix with the local attention scores to obtain process-guided attention weights, and fusing and enhancing the feature expressions of the pitch cumulative error, tooth profile deviation, and tooth direction deformation features at different scales through a residual connection structure; inputting the enhanced features into a multi-layer perceptron to construct an implicit geometric function. The implicit geometric function maps position encoding and network parameters into geometric expressions, and simultaneously applies a signed distance field constraint, a gradient consistency constraint, a normal constraint, and a process constraint including geometric tolerances of machining accuracy grades on the geometric expressions to obtain a parameterized gear geometric model.

[0012] Optionally, the wireless transmission module constructs an interference situation map based on the electromagnetic interference intensity in the industrial field, and performs an availability score based on the interference level of the grid where the channel is located, and the adaptive switching of the optimal transmission channel is realized as follows: obtaining the electromagnetic interference intensity in the industrial environment, extracting the characteristics of multiple interference sources and random background noise, and obtaining an interference feature vector including periodic characteristics, frequency characteristics and duration characteristics; mapping the positions of the interference sources to a two-dimensional plane grid based on the interference feature vector and the industrial field layout information, calculating the superimposed interference intensity of each grid point within multiple sliding time windows of different scales, and generating an interference situation map reflecting the availability of each channel at different spatial positions; evaluating each available channel, determining the interference level of the grid where the channel is located based on the interference situation map, and calculating the channel availability score by combining the channel quality parameters and historical performance records, wherein the weight coefficients of each evaluation dimension are adaptively adjusted according to the channel performance change rate and the interference level of the grid where the channel is located; performing statistical analysis on the channel availability score within the sliding time window, and when the current channel availability score is lower than the preset threshold and there is a candidate channel whose availability score is greater than the product of the current channel availability score and the preset parameter, selecting the channel with the highest channel availability score and meeting the minimum switching interval requirement as the target channel.

[0013] In the second aspect of the embodiments of the present invention, a system is provided, including: a first unit for, for the gear to be measured, based on a deep learning network, extracting gear features through a circular convolutional layer, a radial attention module and an axial convolutional layer and determining the position combination of the optimal measurement points in combination with the gear accuracy level requirements; a second unit for controlling the measurement mechanism to collect the geometric parameter data of the gear based on the position combination; a third unit for the data processing unit to process the geometric parameter data by using a cascaded neural network, wherein the first-level neural network uses a polar coordinate convolutional structure for gear feature extraction and noise reduction processing, and the second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains the actual geometric parameters; a fourth unit for sending the actual geometric parameters to a remote monitoring terminal through a wireless transmission module with an adaptive channel selection function, the wireless transmission module constructs an interference situation map based on the electromagnetic interference intensity in the industrial field, and performs an availability score based on the interference level of the grid where the channel is located, and realizes the adaptive switching of the optimal transmission channel; a fifth unit for the remote monitoring terminal to generate a detection result report according to the actual geometric parameters.

[0014] In the third aspect of the embodiments of the present invention, an electronic device is provided, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0015] In a fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0016] Through deep learning networks and multi-module collaborative optimization, the present invention realizes the intelligent determination of the optimal measurement points of gears, and solves the problems that traditional measurement point selection methods rely on manual experience and lack adaptability.

[0017] The cascaded neural network data processing technical solution of the present invention innovatively combines polar coordinate convolution, rotation-invariant feature extraction, wavelet denoising and graph convolutional networks to realize the intelligent extraction and analysis of gear geometric parameters; by introducing an attention mechanism and parameter prior distribution constructed based on gear expertise, the influence of measurement noise is effectively suppressed, and the accuracy and reliability of geometric parameter extraction are improved.

[0018] By real-time obtaining the electromagnetic interference characteristics in the industrial environment, constructing a dynamically updated interference situation map, accurately associating electromagnetic interference with the physical space, and providing comprehensive interference distribution information for channel evaluation; adopting a multi-dimensional and multi-time-scale channel evaluation mechanism, comprehensively considering the interference level, channel quality and historical performance, realizing an accurate scoring of channel availability; the statistical analysis based on a sliding window effectively balances the timeliness and stability of channel switching, and avoids frequent ineffective switching. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic flow chart of a method for rapid detection of gear geometric parameters based on wireless transmission according to an embodiment of the present invention; Figure 2 It is a comparison chart of the cascaded neural network of the present invention and the traditional method in gear geometric parameter measurement. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0021] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0022] Figure 1The figure is a schematic flowchart of a method for quickly detecting gear geometric parameters based on wireless transmission according to an embodiment of the present invention. As Figure 1 shown, the method includes: for a gear to be measured, based on a deep learning network, extracting gear features through a circular convolutional layer, a radial attention module, and an axial convolutional layer, and determining a position combination of optimal measurement points in combination with the gear accuracy grade requirements; controlling a measurement mechanism to collect gear geometric parameter data based on the position combination; a data processing unit processing the geometric parameter data using a cascaded neural network, where the first-level neural network uses a polar coordinate convolutional structure for gear feature extraction and noise reduction processing, and the second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains actual geometric parameters; sending the actual geometric parameters to a remote monitoring terminal through a wireless transmission module with an adaptive channel selection function, where the wireless transmission module constructs an interference situation map based on the industrial field electromagnetic interference intensity, and performs an availability score according to the interference level of the grid where the channel is located, to achieve adaptive switching of the optimal transmission channel; the remote monitoring terminal generating a detection result report according to the actual geometric parameters.

[0023] Optionally, for a gear to be measured, based on a deep learning network, extracting gear features through a circular convolutional layer, a radial attention module, and an axial convolutional layer, and determining a position combination of optimal measurement points in combination with the gear accuracy grade requirements includes: collecting an image of the gear to be measured and inputting it into a gear measurement point optimization network; the gear measurement point optimization network includes: a feature extraction module that respectively extracts the circumferential feature, radial feature, and axial feature of the gear using a circular convolutional layer, a radial attention module, and an axial convolutional layer; a measurement point evaluation module that constructs the gear accuracy grade requirements as a hierarchical conditional vector, fuses it with the output features of the feature extraction module, and generates an evaluation feature map of candidate measurement points; a measurement point positioning module that outputs a probability heat map of the measurement point position and an accessibility confidence map based on the evaluation feature map, where the probability heat map characterizes the spatial distribution of the optimal measurement points, and the accessibility confidence map characterizes the signal quality reliability of the measurement points; a constraint optimization module that, under the guidance of the probability heat map and the accessibility confidence map, selects an optimal measurement point position combination that satisfies the spatial distribution constraint through a differentiable non-maximum suppression algorithm.

[0024] Exemplarily, an industrial camera can be used for image acquisition. After acquisition, through preprocessing, including operations such as grayscale conversion, histogram equalization, and normalization, a standard input format image is generated.

[0025] Feature extraction module. The circular convolution layer is designed with a circular convolution kernel. First, the gear image is transformed from the Cartesian coordinate system to the polar coordinate system, and the transformation resolution is 32 angular intervals × 64 radial intervals. A convolution kernel of size 5×5 is designed in the polar coordinate space, and convolution operations are performed along the circumferential direction. The depth of the convolution kernel is 32, and the convolution stride is 2, generating a feature map of 16×32×32. The circular convolution layer is cascaded in three layers, and each layer is connected by a ReLU activation function and batch normalization operations are added to improve the training stability.

[0026] The radial attention module uses the attention mechanism to strengthen the features in the radial direction. This module enhances the contrast of the image to highlight the gear radial contour features, and the enhancement coefficient is 1.2. Then, Gabor filters in 8 directions are used to extract the radial edge features. The parameters of the Gabor filter are set as follows: wavelength 8 pixels, standard deviation 4, phase offset 0, and aspect ratio 0.5. The extracted edge features are processed by an attention network, which consists of three fully connected layers with the number of neurons being 128, 64, and 32 respectively, and a GELU activation function is connected after each layer. The attention network outputs a 32-channel attention weight map, which is multiplied by the original features to obtain the enhanced radial features.

[0027] The axial convolution layer adopts a 1×7 convolution kernel structure and focuses on the features perpendicular to the gear axis direction. Similarly, a three-layer cascaded structure is set, and feature maps of 32, 64, and 32 channels are output respectively. This layer is specially designed with a depthwise separable convolution structure, first performing a 1×7 depth convolution and then a 1×1 point convolution, effectively reducing the computational amount.

[0028] The measurement point evaluation module constructs the gear precision level requirements into a hierarchical conditional vector and fuses it with the output features of the feature extraction module.

[0029] Measurement Point Location Module: The probability heatmap is generated by a convolutional - upsampling network. It adopts a three - layer convolutional structure with a convolution kernel size of 3×3. Then, through transposed convolution, the features are upsampled to the original image size, and a single - channel heatmap is output. The value range is 0 - 1, representing the suitability of each position as a measurement point. The reachability confidence chart characterizes the signal quality reliability of the measurement points. It is a spatial distribution map that reflects the stability of the measurement signals at each position on the gear surface through a quantization score (between 0 and 1). This reliability is usually quantified and calculated from multiple aspects: when the angle between the surface normal and the measurement direction is less than 30°, a high - reliability score of 0.9 - 1.0 is given; between 30° and 60°, it is a medium - reliability score of 0.5 - 0.8; when it is greater than 60°, it is less than 0.5. In curvature evaluation, the reliability of areas where the radius of curvature is more than 10 times the measurement accuracy requirement is higher than 0.8, and the reliability of areas where the curvature change rate in the edge transition zone exceeds 0.5 per millimeter is usually less than 0.4. At the same time, the best working distance deviation of the measurement device is considered (the reliability is greater than 0.8 within ±2mm and 0.5 - 0.7 within ±5mm) and the field - of - view angle limit.

[0030] Under the guidance of the probability heatmap and the reachability confidence map, the Constraint Optimization Module selects the optimal combination of measurement point positions that meet the spatial distribution constraints through a differentiable non - maximum suppression algorithm.

[0031] Control the multi - degree - of - freedom measurement mechanism to perform data acquisition according to the optimal measurement point position combination data. The measurement mechanism can be a contact - type measurement device, such as a five - axis linkage coordinate measuring machine, which uses a ruby probe to contact the tooth surface to collect coordinate data; or a non - contact measurement device, such as a structured light scanner, which obtains the 3D point cloud of the gear by projecting fringe light spots; or a laser triangulation system, which uses the principle of scattering and reflection of line laser on the tooth surface to collect the tooth profile. Automatically select the appropriate measurement speed and sampling frequency according to the measurement accuracy requirements to ensure that the data density at key positions meets the analysis requirements.

[0032] This technical solution realizes the intelligent determination of the optimal measurement points of the gear through a deep - learning network and multi - module collaborative optimization, solves the problem that traditional measurement point selection methods rely on manual experience and lack adaptability; can adaptively adjust the measurement strategy according to the gear accuracy grade requirements, and at the same time consider the signal quality reliability of the measurement points to ensure the accuracy and reliability of the measurement results; improves the gear detection efficiency, reduces the number of measurement points while ensuring the measurement accuracy, significantly reduces the measurement time and resource consumption, and provides strong support for precision gear manufacturing.

[0033] Optionally, the gear precision level requirements are constructed as a hierarchical conditional vector and fused with the output features of the feature extraction module to generate an evaluation feature map of candidate measurement points, including: constructing the gear precision level requirements as a hierarchical conditional vector containing global precision parameters and local precision parameters, and respectively mapping the global precision parameters and local precision parameters through non-linear transformation to obtain conditional features; aligning the circumferential features, radial features, and axial features; constructing an annular-radial-axial interaction module, which fuses the circumferential features with the conditional features of the global precision parameters, interacts the radial features with the conditional features of the local precision parameters, and performs precision compensation on the axial features to obtain fused features; constructing a precision adaptive attention map based on the conditional features of the global precision parameters and local precision parameters, and recalibrating the fused features using the precision adaptive attention map; calculating a precision sensitivity score, a geometric feature fitness score, and a measurement feasibility score based on the recalibrated fused features, and performing weighted fusion to obtain an evaluation feature map of candidate measurement points.

[0034] Exemplarily, the gear precision level requirements are first constructed as a hierarchical conditional vector containing global precision parameters and local precision parameters. The global precision parameters include overall gear precision level, radial runout limit, total cumulative error of gear teeth, and other overall indicators; the local precision parameters include pitch error, tooth profile error, position of tooth contact pattern, and other regional indicators. The precision level (0-7 levels) is encoded and converted into a digital vector according to national standards. For example, for a gear with a precision level of 6, its global precision parameters can be encoded as an 8-dimensional vector [0, 0, 0, 0, 0, 1, 0, 0], representing its distribution in the precision level dimension; at the same time, numerical parameters such as a pitch error of 25μm and a tooth profile error of 20μm for the local precision parameters are normalized and organized into a 16-dimensional vector.

[0035] The constructed precision parameter vector is mapped to conditional features through non-linear transformation. A two-layer fully connected network structure is adopted. The global precision parameters are mapped through the first fully connected layer (input dimension 8, output dimension 32), and the ReLU activation function is used to introduce non-linearity; the local precision parameters are mapped through the second fully connected layer (input dimension 16, output dimension 64), and the ReLU activation function is also used. The parameter matrices of these two fully connected layers are initialized using the He initialization method, and the initial value of the bias term is set to 0. In this way, the global precision parameters are mapped to a 32-dimensional conditional feature vector, and the local precision parameters are mapped to a 64-dimensional conditional feature vector.

[0036] Perform feature alignment operations on the circumferential features, radial features, and axial features output by the feature extraction module. The initial sizes of the three types of features may be different. They are uniformly adjusted to the same spatial dimension of 32×32 through bilinear interpolation, and the number of channels is adjusted through a 1×1 convolutional layer, so that the circumferential features become 32×32×32, the radial features become 32×32×64, and the axial features become 32×32×32. In this way, the three types of features are aligned in the spatial dimension.

[0037] Construct a ring-radial-axial interaction module. The fusion of the circumferential features and the global accuracy parameter conditional features adopts a channel attention mechanism. The 32-dimensional global accuracy conditional features are mapped into 32-dimensional channel attention weights through a fully connected layer, and each weight value corresponds to a channel of the circumferential features. Multiply these weights with the corresponding channels of the circumferential features to achieve the selective enhancement of the circumferential features by the accuracy conditions. For example, when the gear accuracy level is high, the channel weights related to the fine tooth profile will increase; when the accuracy level is low, the channel weights related to the overall shape will increase.

[0038] The interaction between the radial features and the local accuracy parameter conditional features adopts a feature modulation method. Reshape the 64-dimensional local accuracy conditional features into a shape of 1×1×64, and then expand them to 32×32×64 through spatial broadcasting, and perform an element-wise addition operation with the radial features. This addition operation enables the information of the local accuracy parameters to affect each spatial position in the radial features, especially having a greater impact on the regions closely related to the local accuracy (such as the tooth tip, tooth root, or tooth side).

[0039] The accuracy compensation of the axial features is achieved through learnable scaling factors. According to the comprehensive situation of the global and local accuracy parameters, apply dynamic scaling factors to each channel of the axial features. After connecting the global and local accuracy conditional features, generate 32 scaling factors through a fully connected layer, with a value range of 0.5 to 2.0, which are respectively applied to the 32 channels of the axial features. For cases where high axial accuracy is required, the scaling factors of the relevant channels will increase; for cases where the axial accuracy requirement is not high, the scaling factors will decrease.

[0040] Through the above interaction operations, the modified circumferential features, radial features, and axial features are obtained, and then they are connected in the channel dimension to form a fused feature of 32×32×128. This fused feature contains the comprehensive information of the gear geometric features and accuracy requirements.

[0041] After connecting the global and local precision condition features, a single-channel spatial attention map of size 32×32 is generated through a two-layer convolutional network (with a convolutional kernel size of 3×3 and the number of channels being 64 and 1 respectively). The value range of the attention map is from 0 to 1, representing the degree of attention to each position in the fused features. In the case of high precision requirements, the regions related to the key precision indicators will obtain higher attention values. For example, for gears with high pitch precision requirements, the attention value of the region near the pitch circle of the gear will increase; for gears with high tooth profile precision requirements, the attention value of the gear tooth profile region will increase.

[0042] The attention map is expanded to a shape of 32×32×128 (duplicated 128 times in the channel dimension), and then an element-wise multiplication operation is performed with the fused features. The multiplication operation enhances the features in the regions with high attention values and suppresses the features in the regions with low attention values, achieving feature recalibration. The recalibrated features are more focused on the key regions related to the precision requirements.

[0043] Based on the recalibrated fused features, three scores are calculated: precision sensitivity score, geometric feature fitness score, and measurement feasibility score. The precision sensitivity score measures the contribution of each position to the precision evaluation and is obtained by processing the recalibrated features through a convolutional layer (with a convolutional kernel size of 1×1 and an output channel number of 1). The geometric feature fitness score reflects the significance of geometric features and is obtained by combining the processing of the recalibrated features through another convolutional layer (with a convolutional kernel size of 3×3 and an output channel number of 1) and a max pooling layer (with a pooling kernel size of 3×3 and a stride of 1). The measurement feasibility score considers the feasibility of actual measurement and is generated based on the measurement device parameters and gear geometric constraints. The sizes of the three scores are all single-channel feature maps of 32×32.

[0044] Weight coefficients of 0.4, 0.4, and 0.2 are set for the precision sensitivity score, geometric feature fitness score, and measurement feasibility score respectively, and a weighted summation operation is performed. The generated evaluation feature map has a size of 32×32 and a value range of 0 to 1, where the higher the value at a position, the more suitable it is as a measurement point.

[0045] This technical solution realizes the deep fusion of gear precision requirements and gear features by constructing a hierarchical conditional vector and a multi-module fusion mechanism, and solves the problem that it is difficult for the traditional method to adapt to different precision requirements in the selection of measurement points; by introducing a precision adaptive attention mechanism, the measurement strategy can be dynamically adjusted according to the precision requirements, thereby optimizing the measurement efficiency while ensuring the measurement precision; this precision adaptive measurement point evaluation method based on deep learning greatly improves the pertinence and effectiveness of gear geometric parameter measurement, can adapt to the measurement needs of gears with different precision levels, and provides a more efficient and reliable solution for gear precision detection.

[0046] Optionally, guided by the probability heatmap and the reachability confidence map, selecting the optimal combination of measurement point positions that satisfy the spatial distribution constraint through a differentiable non-maximum suppression algorithm includes: performing weighted fusion on the probability heatmap and the reachability confidence map to obtain a comprehensive score map of the candidate regions of the measurement points; constructing a differentiable non-maximum suppression model based on the comprehensive score map, where the differentiable non-maximum suppression model uses a Gaussian kernel function to calculate the spatial suppression weights between candidate measurement points, and the spatial suppression weights decrease as the distance between candidate measurement points increases; taking the score values in the comprehensive score map as the initial selection probabilities of the candidate measurement points, and iteratively updating the initial selection probabilities according to the spatial suppression weights to obtain the selection probabilities of the candidate measurement points considering spatial suppression; setting the minimum allowable distance constraint between measurement points based on the gear module, and calculating the spatial uniformity metric of the measurement point distribution; constructing an optimization objective function, where the optimization objective function includes a term for maximizing the comprehensive score of the measurement points, a non-maximum suppression term, and a spatial uniform distribution term based on the spatial uniformity metric; iteratively optimizing the optimization objective function by the gradient descent method to update the selection probabilities of the candidate measurement points; and determining the candidate measurement points with the optimized selection probabilities greater than the preset probability threshold as the optimal combination of measurement point positions.

[0047] Exemplarily, the probability heatmap reflects the suitability of each position on the gear surface as a measurement point, with a size of 32×32 and a pixel value range from 0 to 1. The reachability confidence map characterizes the quality reliability of the measurement signals at each position, also with a size of 32×32 and a pixel value range from 0 to 1. When performing weighted fusion, a weight of 0.7 is assigned to the probability heatmap and a weight of 0.3 is assigned to the reachability confidence map, and the comprehensive score map is generated by pixel-by-pixel weighted summation. For gears with higher precision requirements, the weight of the probability heatmap can be appropriately increased; for gears with complex surfaces, the weight of the reachability confidence map can be increased. For example, for a gear with a precision level of 6, if the probability heatmap value at a specific point is 0.85 and the reachability confidence map value is 0.6, then the comprehensive score of this point is 0.85×0.7 + 0.6×0.3 = 0.775.

[0048] Taking the 32×32 comprehensive score map as the input layer of the non-maximum suppression model, where each pixel position (i, j) and its corresponding score value S(i, j) represent a potential candidate position of the measurement point and its initial importance score. The construction process adopts a two-stage method: in the first stage, an initial candidate set is generated based on the comprehensive score map, and the positions with score values exceeding the adaptive threshold (set as the score mean plus 0.8 times the standard deviation) are retained as valid candidate points to form a set N of candidate measurement points P = {p1, p2,..., p n} and its corresponding score set S = {S1, S2,..., S n}; In the second stage, an N×N spatial suppression matrix W is constructed to represent the mutual suppression relationship between candidate points. For any two candidate points pi and pj, the Gaussian kernel function is used to calculate their spatial suppression weights: W(i, j) = S(i) × exp[-d(i, j)² / (2σ²)], where d(i, j) is the Euclidean distance between the two points in physical space, and σ is the Gaussian kernel standard deviation parameter, which is set to 0.8 times the gear module. When the distance between the two points is 0, the suppression weight is equal to the score value of the source point; as the distance increases, the suppression weight decays exponentially. When the distance reaches 3σ, the suppression weight drops to about 1% of the initial value, achieving a local suppression effect. Since the Gaussian function is differentiable everywhere, the entire suppression process is differentiable and supports gradient optimization algorithms, enabling the non-maximum suppression process to be incorporated into an end-to-end optimization framework. For example, for a gear with a module of 2mm, σ is set to 1.6mm, corresponding to about 4.8 pixels in the image space, and the effective suppression range reaches 14.4 pixels, effectively avoiding the problem of excessive clustering of measurement points.

[0049] Take the score value in the comprehensive score map as the initial selection probability of the candidate measurement points, and iteratively update the initial selection probability according to the spatial suppression weights. Use 10 rounds of iterative processing. In each round of iteration, for each point i in the comprehensive score map, accumulate the product of its initial probability and the suppression weight with all other points j to obtain the total suppression amount received by point i. Then subtract the total suppression amount from the initial probability and normalize it to the range of 0 - 1 through a logical function as the updated selection probability. The maximum selection probability of each point is retained as the final value during the iterative process.

[0050] Set the minimum allowable distance constraint between measurement points based on the gear module. The minimum allowable distance is usually set to 2 times the gear module. For example, for a gear with a module of 2mm, the minimum allowable distance is 4mm. The spatial uniformity metric is calculated in two steps: first, construct a 32×32 uniformly distributed grid, and weightedly assign the candidate measurement points to the nearest grid points according to their selection probabilities; then calculate the standard deviation of the occupancy rate of the grid points. The smaller the standard deviation, the more uniform the distribution. Divide the comprehensive score map into 4×4 regions, calculate the sum of the selection probabilities within each region, and then calculate the standard deviation of the sum of the probabilities of these 16 regions. The smaller the standard deviation, the more uniform the distribution.

[0051] Construct an optimization objective function. The comprehensive score term is the weighted sum of the selection probabilities and comprehensive scores of all candidate points; the non-maximum suppression term is the weighted sum of the products of the selection probabilities and suppression weights between all pairs of candidate points, and the weight coefficient is negative to achieve the suppression effect; the spatial uniform distribution term is a penalty term based on the spatial uniformity metric, and the weight is proportional to the uniformity measure. The weight of the comprehensive score term is set to 1.0, the weight of the non-maximum suppression term is set to -0.5, and the weight of the spatial uniform distribution term is set to 0.3.

[0052] The optimization objective function is iteratively optimized by the gradient descent method to update the selection probability of candidate measurement points. The Adam optimizer is used in the optimization process, and the learning rate is set to 0.01. A total of 50 rounds of iteration are carried out. In each round of iteration, first calculate the gradient of the objective function with respect to the selection probability of each candidate point, then update the selection probability according to the gradient, and ensure that the selection probability is within the range of 0-1 through the logistic function. To avoid local optimal solutions, a simulated annealing mechanism is introduced. A certain degree of random perturbation is allowed at the beginning of the optimization, and the perturbation intensity decreases with the increase of the number of iteration rounds. The initial perturbation intensity is 0.1 and finally drops to 0.01.

[0053] The candidate measurement points with the optimized selection probability greater than the preset probability threshold are determined as the optimal measurement point position combination. The probability threshold is usually set to 0.7, but it can be dynamically adjusted according to the required number of measurement points. For example, if 20 measurement points are to be selected, and the probability threshold of 0.7 results in 25 points being selected, the threshold can be adjusted to 0.75; if only 15 points are selected, the threshold can be reduced to 0.65. If the requirements are still not met after adjustment, then sort according to the selection probability from high to low and select the top 20 points. The finally selected measurement points are output in the form of their three-dimensional coordinates in the gear coordinate system, accurate to 0.01mm.

[0054] In special cases, such as when there are unmeasurable areas on the gear surface, post-processing of the selection results is required. The values in the reachability confidence map of unmeasurable areas are usually very low (less than 0.2). If no measurement points can be selected in some necessary areas (such as key tooth surfaces), adjustments can be made by reducing the probability threshold in the local area or manually specifying the priority area. For standard gears, it is recommended to select at least 3 measurement points on each tooth surface; for special gears (such as modified gears, bevel gears), the distribution requirements need to be adjusted according to their geometric characteristics.

[0055] The measurement point optimization method based on differentiable non-maximum suppression and spatial distribution constraints in this technical solution cleverly solves the problem of insufficient flexibility caused by the fixed mode in traditional measurement point selection; by comprehensively considering the suitability of measurement points, signal reliability and spatial distribution characteristics, it realizes the adaptive optimization layout of measurement points, enabling the selected measurement points to accurately reflect the gear geometric characteristics, ensuring the quality of measurement signals, and meeting the requirement of uniform spatial distribution at the same time; improving the reliability and comprehensiveness of gear geometric parameter measurement, applicable to the precision measurement applications of various gears, and providing more intelligent technical support for quality control in industrial manufacturing.

[0056] Optionally, the data processing unit processes the geometric parameter data using a cascaded neural network. The first-level neural network uses a polar coordinate convolution structure for gear feature extraction and noise reduction processing. The second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains the actual geometric parameters, including: The first-level neural network uses a polar coordinate convolution layer. The convolution kernel of the polar coordinate convolution layer is designed along the circumferential direction of the gear pitch circle, and the convolution step size matches the tooth pitch. The first-level neural network further includes a rotation-invariant feature extraction module. The rotation-invariant feature extraction module extracts tooth profile features through rotation-equivariant convolution operations and an involute adaptive sampling mechanism. The features output by the rotation-invariant feature extraction module are input into a dynamic noise decomposition module. The dynamic noise decomposition module uses a combination of wavelet transform and attention mechanism to perform frequency-domain decomposition on the measurement noise to obtain a systematic error pattern. The features corresponding to the systematic error pattern are input into the second-level neural network. The second-level neural network constructs a gear topological graph structure. The graph nodes in the gear topological graph structure represent geometric feature points, and the graph edges represent geometric constraint relationships. A graph convolutional network is applied on the topological graph structure to extract structured features. An attention mechanism is constructed based on gear manufacturing process knowledge to enhance the pitch cumulative error, tooth profile deviation, and tooth direction deformation features in the structured features, and a parameterized gear geometric model is constructed. The gear processing process rules are encoded as a parameter prior distribution, and the parameterized gear geometric model is constrained and optimized based on the parameter prior distribution to obtain the actual geometric parameters of the gear.

[0057] Exemplarily, the first-level neural network of the data processing unit uses a specially designed polar coordinate convolution layer for gear feature extraction. First, the measured gear surface point cloud data is transformed into the polar coordinate system, where the radial coordinate range is 0.8 to 1.2 times the pitch circle radius, and the angular coordinate range is 0 to 360 degrees. The convolution kernel of the polar coordinate convolution layer is fan-shaped and is designed along the circumferential direction of the pitch circle. The radial size is 0.5 times the module, and the angular size matches the tooth pitch. For example, for a standard gear with 36 teeth, the angle corresponding to each tooth is 10 degrees, so the angular convolution kernel size is set to 10 degrees. The convolution step size is set to 1 / 4 of the tooth pitch, i.e., 2.5 degrees, ensuring that the convolution operation can smoothly transition between adjacent tooth regions. Each polar coordinate convolution layer contains 32 convolution kernels, which can extract different features of the gear. Three polar coordinate convolution layers are stacked, with a ReLU activation function and a batch normalization layer added in the middle to form a basic feature extraction network structure.

[0058] The first-level neural network also includes a rotation-invariant feature extraction module, which extracts tooth profile features through rotation-equivariant convolution operations and involute adaptive sampling mechanisms. The rotation-equivariant convolution operations use a filter bank with 8 fixed rotation angles, namely 0, 45, 90, 135, 180, 225, 270, and 315 degrees, making the feature extraction insensitive to the rotation posture of the gear. The involute adaptive sampling mechanism calculates the positions of reference sampling points according to the standard involute equation and sets the sampling window size to 1 / 20 of the involute arc length. For a gear with a module of 2 mm, the involute working section arc length is approximately 6 mm, and the corresponding sampling window is approximately 0.3 mm. In high-curvature regions such as the tooth tip and tooth root transition zones, the sampling density is dynamically increased to 2 times the standard density; in the involute straight-line segment region, the standard sampling density is maintained. This adaptive sampling mechanism improves the ability to extract key gear features. The rotation-invariant feature extraction module outputs a feature map with 64 channels, representing the rotation-invariant features of the gear.

[0059] The features output by the rotation-invariant feature extraction module are input into the dynamic noise decomposition module, which uses a combination of wavelet transform and attention mechanism to perform frequency-domain decomposition on the measurement noise. The Daubechies wavelet (DB4) is used to decompose the feature signal into 5 levels, decomposing the signal into different frequency-band components. For example, for measurement data with a sampling rate of 10 points / mm, the decomposed frequency bands are: 0 - 0.156 Hz, 0.156 - 0.313 Hz, 0.313 - 0.625 Hz, 0.625 - 1.25 Hz, and 1.25 - 2.5 Hz. Then, a frequency-domain attention mechanism is introduced to weight different frequency bands through learnable weight coefficients. For gear error features, usually the middle frequency band (0.313 - 0.625 Hz) has a higher weight (0.4), the low frequency band (0 - 0.313 Hz) is the second (0.3), and the high frequency band (0.625 - 2.5 Hz) is the lowest (0.1), effectively suppressing high-frequency measurement noise. The weighted frequency-domain components are recombined to obtain a systematic error pattern, which retains the main error feature information while suppressing random noise.

[0060] Input the features corresponding to the systematic error pattern into the second-level neural network, which first constructs the gear topology graph structure. The nodes in the graph structure represent the gear geometric feature points, including the points on the addendum circle, the points on the dedendum circle, the points on the left and right involutes, and the points on the pitch circle, etc. There are a total of 36 teeth multiplied by 8 feature points per tooth, for a total of 288 nodes. The edges of the graph represent geometric constraint relationships, including three categories: adjacent constraints between feature points on the same tooth (such as the symmetry relationship between the left and right involutes), periodic constraints between adjacent teeth (such as the adjacent tooth pitch relationship), and circular constraints between similar feature points (such as the circular arc constraint formed by all pitch circle points). The weights of the edges in the topology graph are set according to the importance of the geometric relationships. For example, the constraint weight between feature points on the same tooth is 1.0, the constraint weight between adjacent teeth is 0.8, and the circular constraint weight is 0.6.

[0061] Apply a graph convolutional network to the gear topology graph structure. The graph convolutional network consists of three layers, and each layer uses different convolutional kernel sizes to capture structural information in different ranges. The first layer uses 1-hop convolution to aggregate only the information of directly connected nodes; the second layer uses 2-hop convolution to capture second-order neighbor information; the third layer uses 3-hop convolution to extract a wider range of structural relationships. The number of output channels for each layer of graph convolution is 32, 64, and 128 respectively. The specific implementation of the graph convolution operation is that each node synthesizes its own features and the weighted features of adjacent nodes, and the weights are determined by the weights of the edges. The features output by the graph convolutional network retain the key topological information of the gear geometric structure, providing a basis for subsequent parametric modeling.

[0062] Construct an attention mechanism based on gear manufacturing process knowledge to enhance the pitch cumulative error, tooth profile deviation, and tooth direction deformation features in the structured features, and construct a parametric gear geometric model.

[0063] For the constraint optimization process of the gear parametric geometric model, first, a prior distribution library of process-related parameters is constructed, including four main processes: hobbing, milling, grinding, and shaving. For the hobbing process, the eccentricity parameter follows a normal distribution with a mean of 0 and a standard deviation of 0.01 mm. The cumulative pitch error adopts a prior mode that combines a first-order periodic function and a second-order periodic function, with a composite ratio of 2:1 and an amplitude upper limit set to 0.03 mm. For the milling process, the pitch error follows a distribution model that linearly accumulates with the tooth position, with an accumulation rate of approximately 0.002 mm / tooth. For the grinding process, the tooth profile error follows a normal distribution with a standard deviation of 0.005 mm in the middle region of the involute, and the error increases to 0.008 mm in the tooth tip region. For the shaving process, the helix deviation is set to follow a parabolic distribution, with the smallest deviation in the middle region and a linear increase in the deviation at both ends, with a growth rate of 0.001 mm / mm. Secondly, a process feature weight factor is introduced to automatically configure the constraint strength of various parameters according to the gear processing method. For example, the constraint weight of the tooth profile error under the grinding process is 0.7, while this weight is only 0.3 under the hobbing process. The parametric gear geometric model includes global geometric parameters (10 parameters such as pitch diameter, pressure angle, base circle diameter, etc.) and local geometric parameters (4 parameters per tooth such as tooth thickness and addendum height for 36 teeth, a total of 144). During optimization, a prior distribution is introduced as a soft constraint condition to construct a composite objective function: F = 0.8 × fitting error + 0.2 × prior deviation. The fitting error calculates the average distance between the gear profile generated by the calculation model and the measurement points; the prior deviation calculates the degree of deviation between the current parameters and the process prior distribution. The specific constraint steps are as follows: In each iteration, the theoretical gear profile of the parametric geometric model is generated using the current parameter values; the fitting degree with the measurement points is calculated; at the same time, it is checked whether the parameters conform to the prior distribution characteristics. For example, when the eccentricity parameter tries to be adjusted to 0.03 mm, a large penalty value will be generated based on the prior distribution (standard deviation 0.01 mm), and the parameters are constrained to adjust to a reasonable range. For the area where measurement data is missing, the prior distribution plays a dominant role; in the area where data is sufficient, the fitting term is the main one. The optimization uses the gradient descent method with a learning rate of 0.01, and the parameters are adjusted according to the comprehensive guidance of the prior distribution and the measurement data. For a 36-tooth gear with a module of 2 mm, this method ensures that the finally obtained parametric geometric model not only maintains a high measurement fitting degree but also has its parameter distribution characteristics consistent with the process rules, avoiding overfitting and ensuring the physical rationality of the geometric model.

[0064] Figure 2This is a comparison chart of the cascaded neural network of the present invention and traditional methods in gear geometric parameter measurement. The figure compares the performance differences between the present invention and traditional convolutional networks, the least squares fitting method, and the standard coordinate measuring method. It can be clearly seen from the figure that the measured values of the present invention for key geometric parameters such as tooth profile deviation, pitch cumulative error, helix deformation, and eccentricity are significantly lower than those of the other three methods, indicating higher measurement accuracy; at the same time, it shows the best performance in terms of noise suppression rate and also has an obvious advantage in processing speed. This proves the superior performance of the cascaded neural network in the field of gear measurement. In particular, the combination of the polar coordinate convolution structure and the topological graph structure provides an effective solution for high-precision gear geometric parameter measurement. The cascaded neural network data processing technical solution of this technical solution innovatively combines polar coordinate convolution, rotation-invariant feature extraction, wavelet noise reduction, and graph convolutional networks to achieve intelligent extraction and analysis of gear geometric parameters; by introducing an attention mechanism and parameter prior distribution constructed based on gear professional knowledge, the influence of measurement noise is effectively suppressed, and the accuracy and reliability of geometric parameter extraction are improved.

[0065] Optionally, an attention mechanism is constructed based on gear manufacturing process knowledge to enhance the pitch cumulative error, tooth profile deviation, and helix deformation features in the structured features. Constructing a parametric gear geometric model includes: constructing a multi-level process feature extraction network to extract tool wear features, heat treatment deformation features, and clamping deformation features respectively, and calculating the response distribution of the pitch cumulative error, tooth profile deviation, and helix deformation features based on the extracted features; constructing a process correlation matrix based on the response distribution, and the process correlation matrix characterizes the spatial distribution and temporal change characteristics of the pitch cumulative error, tooth profile deviation, and helix deformation features between adjacent tooth positions; constructing a multi-head attention mechanism according to the process correlation matrix, and the multi-head attention mechanism calculates local attention scores at each scale level; adaptively fusing the process correlation matrix with the local attention scores to obtain process-guided attention weights, and enhancing the feature expressions of the pitch cumulative error, tooth profile deviation, and helix deformation features at different scales through a residual connection structure; inputting the enhanced features into a multi-layer perceptron to construct an implicit geometric function, and the implicit geometric function maps position encoding and network parameters into geometric expressions, and simultaneously imposing a signed distance field constraint, a gradient consistency constraint, a normal constraint, and a process constraint including geometric tolerances of machining accuracy grades on the geometric expressions to obtain a parametric gear geometric model.

[0066] Exemplarily, a multi-level process feature extraction network is constructed, which includes three parallel branches for extracting tool wear features, heat treatment deformation features, and clamping deformation features respectively. The tool wear feature extraction branch adopts a three-layer convolutional neural network structure. The first layer uses 32 convolutional kernels with a size of 3×3 and a stride of 1; the second layer uses 64 convolutional kernels with a size of 3×3 and a stride of 2; the third layer uses 128 convolutional kernels with a size of 3×3 and a stride of 1. After each layer of convolution, a batch normalization layer and a ReLU activation function are connected. This branch mainly focuses on the minute changes in the tooth profile, especially the systematic changes that occur with the increase of the cutting distance. For example, when cutting 36 teeth, a gradual wear feature in the tooth height direction can be observed, and the wear in the tooth top area is usually 0.005 - 0.01 mm more serious than that in the tooth root area. The heat treatment deformation feature extraction branch adopts an attention residual network structure, which includes 4 residual blocks. Each residual block consists of two convolutional layers and a skip connection, and the number of convolutional kernels is 64, 64, 128, and 128 respectively. This branch focuses on the tooth direction deformation features, such as the typical drum-shaped deformation after quenching, where the middle area protrudes 0.01 - 0.02 mm more than the two ends. The clamping deformation feature extraction branch adopts a graph convolutional network, which includes three layers of graph convolution, and the feature dimensions are 32, 64, and 128 respectively. This branch captures the tooth ring deformation features caused by the clamping force, which are mainly manifested as radial periodic changes. For the three-jaw chuck clamping, it is manifested as a radial runout with a 120-degree period and an amplitude of about 0.01 - 0.03 mm.

[0067] Integrate the outputs of the three feature extraction branches through a feature fusion module. The feature fusion module uses fully connected layers to map the feature vectors of different branches to the same dimension (set to 128 dimensions), and then calculates the fusion weights through an attention gating mechanism. For each feature vector, a weight coefficient is calculated, and the value range is 0-1. For example, for a gear with obvious heat treatment deformation, the weight coefficient of the heat treatment feature branch is set to 0.5-0.7, while the weights of other branches are lower. After multiplying the weight coefficient by the corresponding feature vector and summing, the fused feature vector is obtained. Then, the fused feature vector is passed through a feature decoder network, which consists of three fully connected layers with the number of neurons being 256, 512, and 1024 respectively, and the activation function is ReLU. The output dimension of the last layer is 36×3, corresponding to the pitch cumulative error, tooth profile deviation, and helix deformation features of 36 tooth positions respectively. For the pitch cumulative error, the errors of adjacent teeth are accumulated to form a cumulative error curve. For example, for a 36-tooth gear processed by hobbing, the error accumulation can be manifested as a sine wave with a period equal to the number of teeth, and the amplitude is about 0.02-0.04mm. For the tooth profile deviation, 15 control points are used to represent the deviation distribution of the involute profile, and the control points are evenly distributed along the tooth height direction. For example, for a gear processed by grinding, the tooth profile deviation is usually larger in the tooth tip area (about 0.01-0.015mm) and smaller near the pitch circle (about 0.003-0.005mm). For the helix deformation feature, 11 control points are used to represent the deformation distribution of the helix profile. For example, for a quenched gear, the typical barrel-shaped deformation is manifested as a protrusion in the middle (about 0.01-0.02mm), and the two ends are straight or slightly concave (about 0.003-0.005mm). These features are output in the form of a response distribution map with a resolution of 36×15×11, covering 36 tooth positions, 15 tooth profile control points, and 11 helix control points.

[0068] Construct a process correlation matrix based on the response distribution. The size of this matrix is N×N, where N is the total number of gear features. For a 36-tooth gear, considering the three types of features of the pitch cumulative error, tooth profile deviation, and helix deformation of each tooth, the value of N is 36×3 = 108. Each element in the matrix represents the correlation strength between the corresponding features, and the value range is [-1, 1]. For example, for a gear processed by hobbing, the correlation of the tooth profile deviation between adjacent teeth is usually greater than 0.8, while the correlation between teeth half a turn apart drops below 0.3; for heat treatment deformation, the correlation of the helix deformation between adjacent teeth is as high as 0.95, reflecting the continuity of heat treatment deformation. The process correlation matrix also contains time series change characteristics. For continuously processed gears, the matrix contains sub-blocks representing time dependence, such as the cumulative effect caused by tool wear, which is manifested as a gradually changing correlation strength along the diagonal direction.

[0069] Construct a multi - head attention mechanism based on the process correlation matrix, set 8 attention heads, and each head is responsible for extracting different types of process - related features. Each attention head contains three fully - connected layers: query transformation, key transformation, and value transformation, with the hidden dimension set to 64. For the 108 feature points of the 36 - tooth gear, each attention head generates a 108×108 attention score matrix. The multi - head attention calculates local attention scores at three different scale levels: the global scale focuses on the feature distribution of the entire gear; the regional scale focuses on the feature correlation within a 90 - degree sector; the local scale focuses on the feature correlation of adjacent 3 - 5 teeth. The process - guided attention weights are obtained by adaptively fusing the process correlation matrix with the local attention scores. The fusion formula uses the weighted average method, and the initial weight ratio is process correlation matrix: attention score = 0.7:0.3, which can be automatically adjusted as the training progresses. The residual connection structure is used to fuse and enhance the feature expressions of pitch cumulative error, tooth profile deviation, and helix deviation features at different scales. The residual connection adopts the form of identity mapping plus feature transformation to ensure that the original feature information will not be lost in the deep network.

[0070] The enhanced feature vector (with a dimension of 1024) through the attention mechanism is input into a multi-layer perceptron network, which consists of 5 fully connected layers. The number of neurons in each layer is 256, 512, 512, 256, and 128 respectively. The Swish activation function is used after each layer and Dropout (rate 0.1) is applied to prevent overfitting. At the input end, position encoding technology is adopted to map the three-dimensional space coordinates (x, y, z) into a high-dimensional feature vector. Frequency encoding with 10 frequency components is used to map each coordinate component into a 20-dimensional vector, and the three coordinates total 60 dimensions. The core of the implicit geometric function lies in the organic combination of these 60-dimensional position encoding and the 1024-dimensional enhanced feature vector to construct a conditional geometric representation. The implementation method is to perform feature fusion of the position encoding vector and the enhanced feature vector before the first fully connected layer of the network. The fusion process adopts the feature modulation method, that is, for any query point in space, according to its tooth position, the corresponding local feature is extracted from the 1024-dimensional global feature (achieved through the attention mechanism), and then combined with the position encoding of this point by concatenation or weighting as the network input. This design enables the network to precisely adjust the implicit expression for specific process characteristics in different regions (such as the tooth profile deviation characteristics of teeth No. 12 - 15 and the heat treatment deformation characteristics of teeth No. 25 - 28), so as to accurately reflect the minute shape changes in each region when predicting the distance from any point in space to the gear surface. For a 36-tooth gear, 36×72×10 = 25920 points are evenly sampled in three-dimensional space to construct a training data set. Each sampled point is associated with the process feature information corresponding to this position in addition to the associated three-dimensional coordinates, ensuring that the generated geometric model not only maintains the overall shape regularity of the gear but also accurately represents local process deformations. Four types of constraints are imposed during the training of the implicit geometric function: the signed distance field constraint ensures that the function output correctly represents the distance from a point to the surface, with an error not exceeding 0.01 mm; the gradient consistency constraint guarantees that the gradient direction is consistent with the surface normal, with an angular error not exceeding 2 degrees; the normal constraint ensures that the normal change in adjacent regions is smooth, with a normal change rate not exceeding 0.05; the process constraint introduces form and position tolerance limits. According to the GB / T 10095-2008 standard, for a gear of grade 7 accuracy, the pitch cumulative error is limited within 0.05 mm, the tooth profile error is within 0.025 mm, and the helix error is within 0.02 mm. The training uses the Adam optimizer, with a learning rate of 0.0001, a batch size of 128, and 5000 iterations.

[0071] The point cloud representation of the gear surface is obtained by using the zero isosurface extraction algorithm represented by an implicit function (such as the gradient descent method). One million initial points are uniformly sampled in space, and these points are projected onto the nearest zero isosurface by the gradient descent method, and points with a distance less than 0.001 mm are retained as surface points. Then, the obtained point cloud is meshed to generate a triangular patch model. The meshing uses the Poisson surface reconstruction algorithm, and the octree depth is set to 10, resulting in approximately 100,000 triangular patches. Next, the parametric representation of the gear is extracted from the triangular mesh model. Through the gear cross-section analysis, the basic parameters are extracted: the pitch circle diameter is approximately 72.05 mm (standard value 72 mm), the pressure angle is approximately 19.8 degrees (standard value 20 degrees), and the modulus is approximately 1.99 mm (standard value 2 mm). Based on the distribution of surface points, the actual geometric parameters of each tooth are calculated: the pitch cumulative error curve shows periodic changes, with a maximum value of about 0.038mm; the tooth profile deviation is the smallest near the pitch circle (about 0.004mm) and the largest in the tooth top area (about 0.015mm); the tooth deformation is a typical drum shape, with a bulge of about 0.017mm in the middle. Finally, these parameters are organized into a structured parameter description file, which contains global parameters (pitch circle diameter, pressure angle, module, etc.) and local parameters (pitch error of each tooth, tooth profile deviation distribution, tooth deformation distribution, etc.), forming a complete parametric gear geometry model.

[0072] This technical solution achieves a deep understanding and enhanced expression of gear geometric features through a multi-level process feature extraction network and an attention mechanism guided by a process association matrix, breaking through the limitation that traditional geometric modeling methods are difficult to express small process deviations. The multiple constraints built based on process knowledge ensure that the parameterized geometric model is consistent with both the measured data and the process laws, and improves the robustness of the model in the presence of noise and missing data. The introduction of implicit geometric functions enables the model to have continuous expression capabilities and can accurately reconstruct gear geometry at any resolution, providing a high-precision geometric foundation for subsequent finite element analysis, contact analysis, and performance evaluation, and has important engineering application value for improving the quality of gear design and manufacturing.

[0073] Optionally, the wireless transmission module constructs an interference situation map based on the electromagnetic interference intensity in the industrial field, and performs an availability score according to the interference level of the grid where the channel is located, and realizes the adaptive switching of the optimal transmission channel, including: obtaining the electromagnetic interference intensity in the industrial environment, extracting features of multiple interference sources and random background noise, and obtaining an interference feature vector including periodic features, frequency features and duration features; mapping the positions of the interference sources to a two-dimensional plane grid based on the interference feature vector and the industrial field layout information, calculating the superimposed interference intensity of each grid point within multiple sliding time windows of different scales, and generating an interference situation map reflecting the availability of each channel at different spatial positions; evaluating each available channel, determining the interference level of the grid where the channel is located based on the interference situation map, and calculating the channel availability score in combination with the channel quality parameters and historical performance records, wherein the weight coefficients of each evaluation dimension are adaptively adjusted according to the channel performance change rate and the interference level of the grid where the channel is located; performing statistical analysis on the channel availability score within the sliding time window, and when the current channel availability score is lower than the preset threshold and there is a candidate channel whose availability score is greater than the product of the current channel availability score and the preset parameter, selecting the channel with the highest channel availability score and meeting the minimum switching interval requirement as the target channel.

[0074] Exemplarily, when obtaining the electromagnetic interference intensity in the industrial environment, multiple electromagnetic interference detectors are deployed, and the detector spacing is set to 5-10 meters to form a detection network. Each detector is equipped with a spectrum analyzer, the scanning range covers the 2.4 GHz and 5 GHz industrial, scientific and medical frequency bands, the sampling rate is set to 5 MHz, and the sampling interval is 100 milliseconds. Perform preprocessing on the collected original electromagnetic signals, including band-pass filtering, signal smoothing and normalization. The filtering bandwidth is 1.2 times the channel bandwidth, the smoothing window length is 20 sampling points, and the signal intensity range after normalization is 0-100 dB. When extracting features of multiple interference sources and random background noise, use short-time Fourier transform to extract frequency domain features, the window length is 256 points, the overlap rate is 50%, and Hanning window weighting is used; use autocorrelation function analysis to extract periodic features, and the maximum delay is 1000 sampling points; use envelope detection to extract time domain duration features, and the threshold is set to the background noise mean plus 3 times the standard deviation. The finally obtained interference feature vector includes three parts: the periodic feature includes three indicators of main period, period intensity and period stability; the frequency feature includes five indicators of center frequency, bandwidth, power spectral density; the duration feature includes three indicators of average duration, duty cycle and burst frequency.

[0075] When mapping the interference source location to a two-dimensional plane grid based on the above-extracted interference feature vectors and industrial site layout information, a coordinate system for the industrial site is established. The X-axis and Y-axis correspond to the length and width of the factory building respectively, and the origin is set at the southwest corner of the factory building. The industrial site is divided into equally spaced grids with a grid size of 2 meters × 2 meters. To determine the interference source location using triangulation, at least the signal strength data of three detectors are required. The relative distance is calculated through the signal strength inverse law, and the positioning accuracy is approximately one-fourth of the grid size. For mobile interference sources, Kalman filtering is used for trajectory smoothing, with the process noise covariance set to 0.5 and the observation noise covariance set to 2.0. After the interference source location is completed, the interference superposition intensity of each grid point is calculated within multiple sliding time windows of different scales. Three time windows are set: the short window is 10 seconds, the medium window is 60 seconds, and the long window is 10 minutes. For each grid point, considering the contributions of all identified interference sources, and combining the frequency feature and duration feature in the interference feature vector extracted in the first step, the impact degree on each wireless channel is evaluated. The interference intensity decays with distance following an exponential decay model, with the decay exponent being 2.5 - 3.5 (adjusted according to the environmental complexity). The grid point interference intensity is the weighted sum of the intensities of each interference source, and the weights are determined based on the periodic feature and duration feature extracted in the previous step. Interference sources with stronger periodicity and longer duration have higher weights. The finally generated interference situation map is represented in the form of a heat map, with the color ranging from blue to red corresponding to the interference intensity from low to high, and the refresh rate is 1 second / time, and the resolution is consistent with the grid. This interference situation map intuitively reflects the availability status of each channel at different spatial positions.

[0076] Based on the generated interference situation map, each available channel is evaluated. The channels supported by the wireless transmission module include 11 channels in the 2.4GHz band (channels 1 - 11) and 9 channels in the 5GHz band (channels 36, 40, 44, 48, 149, 153, 157, 161, 165). Based on the interference situation map generated in the previous step, determine the interference level of the grid where the channel is located. The interference level is divided into five levels: level 1 indicates negligible interference (interference intensity < -90dBm), level 2 indicates slight interference (-90dBm to -75dBm), level 3 indicates moderate interference (-75dBm to -65dBm), level 4 indicates severe interference (-65dBm to -55dBm), and level 5 indicates extreme interference (> -55dBm). The interference level assessment takes into account the periodic and frequency characteristics in the interference feature vector. For interference with a high degree of frequency overlap and strong periodicity with the channel, its impact level on the channel will be correspondingly increased. Calculate the channel availability score by combining the channel quality parameters and historical performance records. The channel quality parameters include signal-to-noise ratio, bit error rate, and link margin. The signal-to-noise ratio is divided into five grades: excellent (> 25dB, score 100), good (20 - 25dB, score 80), medium (15 - 20dB, score 60), poor (10 - 15dB, score 40), and inferior (< 10dB, score 20); the bit error rate is also divided into five grades: excellent (< 0.01%, score 100), good (0.01% - 0.1%, score 80), medium (0.1% - 1%, score 60), poor (1% - 5%, score 40), and inferior (> 5%, score 20); the link margin is divided into five grades: excellent (> 20dB, score 100), good (15 - 20dB, score 80), medium (10 - 15dB, score 60), poor (5 - 10dB, score 40), and inferior (< 5dB, score 20). The historical performance records include the average availability, stability, and throughput performance in the past 24 hours. The calculation of the channel availability score uses the weighted average method. The initial weights of each evaluation dimension are: interference level 30%, signal-to-noise ratio 25%, bit error rate 20%, link margin 15%, and historical performance 10%. The weight coefficients are adaptively adjusted according to the channel performance change rate and the interference level of the grid where the channel is located. When the channel performance change rate (the percentage difference between the current value and the 10-minute average value) exceeds 20%, increase the weight of the corresponding dimension; when the grid interference level is level 4 or 5, increase the weight of the interference level by 10 percentage points; the total weight normalization process ensures that the sum is 100%. This multi-dimensional evaluation comprehensively considers the spatial interference distribution provided by the interference situation map and the real-time measured channel performance indicators to form a comprehensive channel availability score.

[0077] Based on the channel availability score calculated in the previous step, when performing statistical analysis on it within a sliding time window, three sliding windows with different lengths are set: a short window (30 seconds, weight 40%), a medium window (5 minutes, weight 35%), and a long window (30 minutes, weight 25%). Calculate the mean, variance, and trend of the channel scores within each window. This multi-time-scale analysis can simultaneously capture the effects of instantaneous interference and continuous interference, corresponding to different time characteristics extracted in the interference feature vector. When the current channel availability score is lower than the preset threshold and there is a candidate channel whose availability score is greater than the product of the current channel availability score and the preset parameter, the channel switching mechanism is triggered. The preset threshold is set to 70 points (out of 100), and the preset parameter is set to 1.2, that is, the candidate channel score needs to be more than 20% higher than the current channel to consider switching, avoiding frequent ineffective switching. Select the channel with the highest channel availability score and meeting the minimum switching interval requirement as the target channel. The minimum switching interval is defaulted to 60 seconds and can be adjusted to 30 - 300 seconds according to the stability requirements of the application. The channel selection process will refer to the spatial information in the interference situation map and preferentially select channels with lower interference levels at the current location and on the predicted future short-term movement path. After determining the target channel, execute the seamless switching process: establish a parallel connection on the target channel in advance. After verifying that the connection quality meets the standard (test 10 data packets, success rate > 95%), complete the service migration from the current channel to the target channel within 20 milliseconds.

[0078] The remote monitoring terminal automatically generates a detection report based on a preset template, including the basic parameters of the gear, the measured parameters, the error analysis (tooth profile deviation, helix deviation, pitch cumulative error), and the qualified judgment result.

[0079] This technical solution obtains the electromagnetic interference characteristics in the industrial environment in real time, constructs a dynamically updated interference situation map, accurately associates the electromagnetic interference with the physical space, and provides comprehensive interference distribution information for channel evaluation; adopts a multi-dimensional and multi-time-scale channel evaluation mechanism, comprehensively considers the interference level, channel quality, and historical performance, and realizes an accurate score for channel availability; the statistical analysis based on the sliding window effectively balances the timeliness and stability of channel switching and avoids frequent ineffective switching.

[0080] In the second aspect of the embodiments of the present invention, a rapid detection system for gear geometric parameters based on wireless transmission is provided, including: a first unit for, for a gear to be measured, based on a deep learning network, extracting gear features through a circular convolutional layer, a radial attention module, and an axial convolutional layer and determining a position combination of optimal measurement points in combination with gear accuracy level requirements; a second unit for controlling a measurement mechanism to collect gear geometric parameter data based on the position combination; a third unit for a data processing unit to process the geometric parameter data using a cascaded neural network, where the first-level neural network uses a polar coordinate convolutional structure for gear feature extraction and noise reduction processing, and the second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains actual geometric parameters; a fourth unit for sending the actual geometric parameters to a remote monitoring terminal through a wireless transmission module with an adaptive channel selection function, where the wireless transmission module constructs an interference situation map based on the electromagnetic interference intensity in the industrial field and performs an availability score according to the interference level of the grid where the channel is located, realizing adaptive switching of the optimal transmission channel; a fifth unit for the remote monitoring terminal to generate a detection result report based on the actual geometric parameters.

[0081] In the third aspect of the embodiments of the present invention, an electronic device is provided, including: a processor; a memory for storing processor-executable instructions; wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0082] In the fourth aspect of the embodiments of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0083] The present invention can be a method, device, system, and / or computer program product. The computer program product may include a computer-readable storage medium on which computer-readable program instructions for executing various aspects of the present invention are loaded.

[0084] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rapid detection method for gear geometric parameters based on wireless transmission, characterized in that, Including: For the gear to be measured, based on a deep learning network, gear features are extracted through a circular convolutional layer, a radial attention module, and an axial convolutional layer, and the position combination of the optimal measurement points is determined in combination with the gear accuracy grade requirements; the measurement mechanism is controlled to collect the geometric parameter data of the gear based on the position combination; the data processing unit processes the geometric parameter data using a cascaded neural network, where the first-level neural network uses a polar coordinate convolutional structure for gear feature extraction and noise reduction processing, and the second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains the actual geometric parameters; the actual geometric parameters are sent to a remote monitoring terminal through a wireless transmission module with an adaptive channel selection function, the wireless transmission module constructs an interference situation map based on the industrial site electromagnetic interference intensity, and performs an availability score according to the interference level of the grid where the channel is located to achieve adaptive switching of the optimal transmission channel; the remote monitoring terminal generates a detection result report according to the actual geometric parameters.

2. The method according to claim 1, characterized in that, For the gear to be measured, based on a deep learning network, the determination of the position combination of the optimal measurement points through a circular convolutional layer, a radial attention module, and an axial convolutional layer and in combination with the gear accuracy grade requirements includes: collecting an image of the gear to be measured and inputting it into a gear measurement point optimization network; the gear measurement point optimization network includes: a feature extraction module that extracts the circumferential feature, radial feature, and axial feature of the gear respectively using a circular convolutional layer, a radial attention module, and an axial convolutional layer; a measurement point evaluation module that constructs the gear accuracy grade requirements into a hierarchical conditional vector and fuses it with the output features of the feature extraction module to generate an evaluation feature map of the candidate measurement points; a measurement point positioning module that outputs a probability heat map of the measurement point position and an accessibility confidence map based on the evaluation feature map, the probability heat map characterizing the spatial distribution of the optimal measurement points, and the accessibility confidence map characterizing the signal quality reliability of the measurement points; a constraint optimization module that, under the guidance of the probability heat map and the accessibility confidence map, selects the optimal measurement point position combination that satisfies the spatial distribution constraint through a differentiable non-maximum suppression algorithm.

3. The method according to claim 2, wherein Construct the gear precision level requirements into a hierarchical conditional vector, and fuse it with the output features of the feature extraction module to generate an evaluation feature map of candidate measurement points, including: constructing the gear precision level requirements into a hierarchical conditional vector containing global precision parameters and local precision parameters, and respectively mapping the global precision parameters and local precision parameters through non-linear transformation to obtain conditional features; aligning the circumferential features, radial features and axial features; constructing an annular-radial-axial interaction module, which fuses the circumferential features with the conditional features of the global precision parameters, interacts the radial features with the conditional features of the local precision parameters, and compensates the precision of the axial features to obtain fused features; constructing a precision adaptive attention map based on the conditional features of the global precision parameters and local precision parameters, and recalibrating the fused features by using the precision adaptive attention map; calculating the precision sensitivity score, geometric feature fitness score and measurement feasibility score according to the recalibrated fused features, and performing weighted fusion to obtain the evaluation feature map of candidate measurement points.

4. The method according to claim 2, wherein Under the guidance of the probability heat map and the reachability confidence map, select the optimal measurement point position combination that meets the spatial distribution constraint through a differentiable non-maximum suppression algorithm, including: performing weighted fusion on the probability heat map and the reachability confidence map to obtain a comprehensive score map of the measurement point candidate area; constructing a differentiable non-maximum suppression model based on the comprehensive score map, and the differentiable non-maximum suppression model uses a Gaussian kernel function to calculate the spatial suppression weight between candidate measurement points, and the spatial suppression weight decreases as the distance between candidate measurement points increases; using the score value in the comprehensive score map as the initial selection probability of the candidate measurement point, and iteratively updating the initial selection probability according to the spatial suppression weight to obtain the candidate measurement point selection probability considering spatial suppression; setting the minimum allowable distance constraint between measurement points based on the gear module, and calculating the spatial uniformity measure of the measurement point distribution; constructing an optimization objective function, which includes a measurement point comprehensive score item, a non-maximum suppression item and a spatial uniform distribution item based on the spatial uniformity measure; iteratively optimizing the optimization objective function by the gradient descent method to update the candidate measurement point selection probability; determining the candidate measurement points with the optimized selection probability greater than the preset probability threshold as the optimal measurement point position combination.

5. The method according to claim 1, characterized in that The data processing unit processes the geometric parameter data using a cascaded neural network. Among them, the first-level neural network uses a polar coordinate convolution structure for gear feature extraction and noise reduction processing. The second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains the actual geometric parameters, including: The first-level neural network uses a polar coordinate convolution layer. The convolution kernel of the polar coordinate convolution layer is designed along the circumferential direction of the gear pitch circle, and the convolution step size matches the tooth pitch. The first-level neural network also includes a rotation-invariant feature extraction module. The rotation-invariant feature extraction module extracts tooth profile features through rotation-equivariant convolution operations and an involute adaptive sampling mechanism; Input the features output by the rotation-invariant feature extraction module into a dynamic noise decomposition module. The dynamic noise decomposition module uses a combination of wavelet transform and attention mechanism to perform frequency-domain decomposition on the measurement noise to obtain a systematic error pattern; Input the features corresponding to the systematic error pattern into the second-level neural network. The second-level neural network constructs a gear topological graph structure. The graph nodes in the gear topological graph structure represent geometric feature points, and the graph edges represent geometric constraint relationships. And apply a graph convolutional network on the topological graph structure to extract structured features; Construct an attention mechanism based on gear manufacturing process knowledge to enhance the pitch cumulative error, tooth profile deviation, and tooth direction deformation features in the structured features, and construct a parameterized gear geometric model; Encode the gear processing process rules as a parameter prior distribution, and optimize the constraints on the parameterized gear geometric model based on the parameter prior distribution to obtain the actual geometric parameters of the gear.

6. The method according to claim 5, characterized in that, Constructing an attention mechanism based on gear manufacturing process knowledge to enhance the pitch cumulative error, tooth profile deviation, and tooth direction deformation features in the structured features and constructing a parameterized gear geometric model includes: Constructing a multi-level process feature extraction network to extract tool wear features, heat treatment deformation features, and clamping deformation features respectively, and calculating the response distributions of the pitch cumulative error, tooth profile deviation, and tooth direction deformation features based on the extracted features; Constructing a process correlation matrix based on the response distribution. The process correlation matrix characterizes the spatial distribution and temporal variation characteristics of the pitch cumulative error, tooth profile deviation, and tooth direction deformation features between adjacent tooth positions; Constructing a multi-head attention mechanism according to the process correlation matrix. The multi-head attention mechanism calculates local attention scores at each scale level; Adaptive fusion of the process correlation matrix and the local attention scores to obtain process-guided attention weights, and fusing and enhancing the feature expressions of the pitch cumulative error, tooth profile deviation, and tooth direction deformation features at different scales through a residual connection structure; Inputting the enhanced features into a multi-layer perceptron to construct an implicit geometric function. The implicit geometric function maps position encoding and network parameters into geometric expressions, and simultaneously applies a signed distance field constraint, a gradient consistency constraint, a normal constraint, and a process constraint including geometric tolerances of machining accuracy grades on the geometric expressions to obtain a parameterized gear geometric model.

7. The method according to claim 1, characterized in that The wireless transmission module constructs an interference situation map based on the electromagnetic interference intensity in the industrial field, and performs an availability score according to the interference level of the grid where the channel is located, and realizes the adaptive switching of the optimal transmission channel, including: obtaining the electromagnetic interference intensity in the industrial environment, extracting features of multiple interference sources and random background noise, and obtaining an interference feature vector including periodic features, frequency features, and duration features; mapping the positions of the interference sources to a two-dimensional plane grid based on the interference feature vector and the industrial field layout information, calculating the superimposed interference intensity of each grid point within multiple sliding time windows at different scales, and generating an interference situation map reflecting the availability of each channel at different spatial positions; evaluating each available channel, determining the interference level of the grid where the channel is located based on the interference situation map, and calculating the channel availability score in combination with the channel quality parameters and historical performance records, where the weight coefficients of each evaluation dimension are adaptively adjusted according to the channel performance change rate and the interference level of the grid where the channel is located; performing statistical analysis on the channel availability score within the sliding time window, and when the current channel availability score is lower than the preset threshold and there is a candidate channel whose availability score is greater than the product of the current channel availability score and the preset parameter, selecting the channel with the highest channel availability score and meeting the minimum switching interval requirement as the target channel.

8. A rapid gear geometric parameter detection system based on wireless transmission, which is used to implement the method described in any one of the foregoing claims 1-7, characterized in that, Including: The first unit is used to, for the gear to be measured, based on a deep learning network, extract gear features through a circular convolutional layer, a radial attention module, and an axial convolutional layer, and determine the position combination of the optimal measurement points in combination with the gear accuracy level requirements; the second unit is used to control the measurement mechanism to collect the geometric parameter data of the gear based on the position combination. The third unit is used for the data processing unit to process the geometric parameter data by using a cascaded neural network, where the first-level neural network uses a polar coordinate convolutional structure for gear feature extraction and noise reduction processing, and the second-level neural network adaptively constructs a gear geometric model based on a topological graph structure and obtains the actual geometric parameters. The fourth unit is used to send the actual geometric parameters to a remote monitoring terminal through a wireless transmission module with an adaptive channel selection function. The wireless transmission module constructs an interference situation map based on the electromagnetic interference intensity in the industrial field, and performs an availability score according to the interference level of the grid where the channel is located, and realizes the adaptive switching of the optimal transmission channel. The fifth unit is used for the remote monitoring terminal to generate a detection result report according to the actual geometric parameters.

9. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images