A method for generating a grinding track model of a hobbing cutter grinding machine based on deep learning

Abstract

This invention provides a method for generating grinding trajectory models for hob grinders based on deep learning, belonging to the field of data processing technology. The method includes: establishing multi-dimensional feature channels based on a basic dataset; fusing features from the tooth surface region, grinding wheel boundary region, and historical trajectory region in the basic dataset; calculating the complexity of the tooth surface structure in local regions to obtain mega-entropy; calculating the extension degree of the hob between different segments to obtain chain amplitude; establishing a dual-adjustment structure to adjust and fuse the chain amplitude structure matrix to obtain an adjustment structure matrix; constructing a deep learning model based on a convolutional neural network; constructing a training sample set to train the deep learning model to obtain a grinding trajectory model; and inputting the basic dataset into the grinding trajectory model to generate grinding trajectory data. This invention enables trajectory nodes to automatically avoid regions where interference occurs or has interference characteristics during generation.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for generating grinding trajectory models for hobbing grinders based on deep learning. Background Technology

[0002] The hob grinder is a key piece of equipment for precision machining the back face of gear hobs. Its core function lies in generating a grinding trajectory that controls the relative motion between the grinding wheel and the workpiece. Existing technologies mainly rely on calculation methods based on geometric models. Based on the hob's design parameters, such as the rake angle, clearance angle, and chip groove profile, a precise geometric model and kinematic relationship between the grinding wheel and the workpiece are constructed through the principles of spatial meshing and envelope theory. Subsequently, by solving this mathematical model, the interpolation commands for each axis required by the CNC system are calculated to ensure that the grinding wheel can enclose the theoretically desired tool tooth profile.

[0003] However, in actual machining, since the grinding wheel itself has fixed physical dimensions, when grinding to the tooth root region according to the trajectory calculated by the geometric model, the edge or base of the grinding wheel may interfere with or collide with the adjacent already machined tooth surface. To avoid this problem, a larger safety distance is usually set in the trajectory planning or the tooth root connection curve is simplified. However, this may cause the material at the tooth root to not be completely ground according to the theoretical profile, destroying the integrity and continuity of the cutting edge shape and affecting the performance and life of the hob. Summary of the Invention

[0004] The purpose of this invention is to provide a method for generating grinding trajectory models for hobbing grinders based on deep learning, aiming to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A method for generating a grinding trajectory model for a hobbing grinder based on deep learning, the method comprising: Obtain the hob tooth surface structure, grinding wheel boundary dimensions, and historical machining samples to obtain the basic dataset; A multidimensional feature channel is established based on the basic dataset, and features of the tooth surface region, grinding wheel boundary region and historical trajectory region in the basic dataset are fused based on the multidimensional feature channel to obtain the structural feature matrix; Based on the structural feature matrix, the tooth is divided into segments according to the structural features of the tooth tip, tooth vent, and tooth root, and the complexity of the tooth surface structure in the local region is calculated to obtain the mega-entropy. Construct the chain amplitude structure matrix based on the mega-entropy and calculate the degree of extension of the hob between different segments to obtain the chain amplitude. A dual-adjustment structure is established based on the mega-entropy and chain amplitude, and the chain amplitude structure matrix is ​​adjusted and fused according to the dual-adjustment structure to obtain the adjustment structure matrix. A deep learning model is constructed based on a convolutional neural network. The deep learning model includes an input layer, a feature encoding layer, a structural association layer, a chained fusion layer, a control and adjustment layer, and an output layer. A training sample set is constructed based on historical processing samples and their interference markers. The deep learning model is then trained based on the training sample set and the adjustment structure matrix to obtain the grinding trajectory model. Input the basic dataset into the grinding trajectory model to generate grinding trajectory data.

[0006] Furthermore, a multi-dimensional feature channel is established based on the basic dataset, and features from the tooth surface region, grinding wheel boundary region, and historical trajectory region in the basic dataset are fused using the multi-dimensional feature channel to obtain a structural feature matrix, including: By mapping tooth surface structure data, grinding wheel boundary size data, and historical trajectory sample data to tooth surface region dataset, boundary region dataset, and trajectory region dataset, respectively, region classification data is obtained; Based on the regional classification data, the tooth surface region dataset, boundary region dataset, and trajectory region dataset are mapped to the first feature channel, the second feature channel, and the third feature channel to obtain channel mapping data. Based on the channel mapping data, the data in the three types of feature channels are aligned according to the structure index order to ensure that the data at the same index position form a corresponding relationship, thus obtaining the aligned channel data; Based on the alignment channel data, the data in the three feature channels at each index position are combined to obtain the combined feature data; Based on the combined feature data, the combined features at different index positions are arranged according to the row and column correspondence to obtain the structural feature matrix.

[0007] Furthermore, based on the structural feature matrix, segments are divided according to the structural features of the tooth tip, tooth vent, and tooth root, and the complexity of the tooth surface structure in local regions is calculated to obtain the mega-entropy, including: By extracting the data from the structural feature matrix according to the positions of the tooth tip, tooth vent, and tooth root, and dividing them into several local segments, segment division data is obtained. Based on the segment division data, the strength of the change in tooth surface structure in each local segment is calculated to obtain the local difference term; the overall complexity within different local segments is calculated to obtain the regional complexity term; the overall change intensity of complexity between each local segment is calculated to obtain the regional jump term; and the degree of correlation between different tooth surface structures in the same local segment is calculated to obtain the segment correlation term. By fusing the local difference term, the regional complexity term, the regional jump term, and the segment correlation term, we obtain the mega-entropy quantity.

[0008] Furthermore, a chain amplitude structure matrix is ​​constructed based on the mega-entropy, and the extent of the cutter's extension between different segments is calculated to obtain the chain amplitude, including: Based on the segment division data and mega-entropy, the characteristic changes between different local segments are combined in pairs, and the expansion relationship between each segment pair is sequentially filled into the data array to obtain the chain amplitude structure matrix. Based on the chain length structure matrix, the degree of cumulative difference between adjacent segments is calculated, the expansion trend caused by local structural differences between different segments of the hob is identified, and the inter-segment expansion term is obtained; the relative position of each segment in the direction from tooth tip to tooth root is calculated, the degree of influence of different segment positions on the overall expansion is identified, and the position modulation term is obtained. The overall complexity is calculated to amplify or compress the expansion results of each segment, resulting in a complexity coupling term. The degree of mutual influence between expansion results of different segments is calculated, resulting in a chain amplitude coupling term. The inter-segment expansion term, position modulation term, complexity coupling term, and chain amplitude coupling term are fused together to obtain the chain amplitude quantity.

[0009] Furthermore, a dual-adjustment structure is established based on the mega-entropy and chain amplitude, and the chain amplitude structure matrix is ​​adjusted and fused according to the dual-adjustment structure to obtain the adjustment structure matrix, including: By converting mega-entropy into complex amplification factors and chain amplitude into extended balance factors, the adjustment data is obtained; Based on the adjustment data, a complex amplification coefficient and an extended balance coefficient are simultaneously superimposed numerically to ensure that the two form a corresponding relationship within the same adjustment framework, thus obtaining dual-adjustment structure data. Based on the dual-adjustment structure data, the adjustment effect of complex amplification coefficient and extended balance coefficient is applied simultaneously to each matrix element of the chain axis structure matrix to obtain the correction matrix data; Based on the correction matrix data, the data related to the tooth tip, tooth vent, and tooth root are summarized and arranged in the processing order to obtain the adjustment structure matrix.

[0010] Furthermore, deep learning models are built based on convolutional neural networks, including: Construct an input layer to receive the basic dataset, divide the basic dataset into tooth surface input data, boundary input data, and trajectory input data, and align them to obtain the input data; A feature coding layer is constructed to extract features of tooth surface shape change, grinding wheel shape constraint, and historical motion trajectory based on the input data, and combine them into fusion features to obtain feature coding data; A structural association layer is constructed to divide the fused features into tooth crest partition features, tooth ventral partition features, and tooth root partition features based on the feature encoding data, and arrange them into structural association data.

[0011] Furthermore, building deep learning models based on convolutional neural networks also includes: A chain-like fusion layer is constructed to combine the structural features corresponding to the tooth tip, tooth vent, and tooth root into multiple chain-like combinations according to the segment order based on the structural association data, thereby obtaining chain-like fused data; A control and adjustment layer is constructed to adjust the chained fusion data based on mega-entropy and chain amplitude. Mega-entropy is used to assign weights to each segment, and chain amplitude is used to limit the expansion range of each segment, thus obtaining control and adjustment data. An output layer is constructed to extract nodes of the motion trajectory based on the control and adjustment data, and sort the nodes according to the processing order to obtain the trajectory node output data.

[0012] Furthermore, a training sample set is constructed based on historical processing samples and their interference markers. The deep learning model is then trained using the training sample set and the adjustment structure matrix to obtain a grinding trajectory model, including: By matching each historical processed sample with its corresponding interference marker, labeled sample data is obtained; Based on the labeled sample data, the continuous processing records are divided into training input sequences, and the corresponding interference marks are organized into training target sequences to obtain the training sample set; Predicted trajectory data is obtained by feeding the training input sequence into a deep learning model; Error data is obtained by comparing the predicted trajectory data with the training target sequence and calculating the degree of deviation between the two. Based on the error data and the adjustment structure matrix, the parameters of the deep learning model are updated until the error data meets the preset convergence condition, thus obtaining the grinding trajectory model.

[0013] Furthermore, by comparing the predicted trajectory data with the training target sequence and calculating the degree of deviation between the two, error data is obtained, including: By matching the predicted trajectory data with the training target sequence on the trajectory node index, aligned prediction data and aligned target data are obtained; The node deviation data is obtained by calculating the numerical difference between each pair of trajectory nodes in the alignment prediction data and the alignment target data. By grouping and summarizing the node deviation data according to the processing order, grouped error data is obtained; By normalizing and aggregating the grouped error data, the errors of each group are converted into the overall deviation, thus obtaining the error data.

[0014] Furthermore, based on the error data and the adjusted structure matrix, the parameters of the deep learning model are updated until the error data meets the preset convergence condition, resulting in a grinding trajectory model, including: Based on the error data, determine the adjustment parameters in the deep learning model and obtain the parameter data to be updated; The error data is weighted according to the adjustment structure matrix, and different adjustment weights are assigned to the errors of different segments and different regions to obtain weighted error data. The updated parameter data is obtained by performing update calculations based on the weighted error data; The updated parameter data is applied to the deep learning model, and the error calculation and parameter update steps are repeated. When the error data meets the preset convergence condition, the parameter update step is stopped, and the grinding trajectory model is obtained.

[0015] The above-described solution of the present invention has at least the following beneficial effects: This invention establishes multi-dimensional feature channels, mapping tooth surface region data, boundary region data, and trajectory region data to different channels and aligning them under the same indexing system. Through channelized representation, different data sources form tensors with completely independent dimensions but synchronized indices at the neural network input, enabling the convolutional kernel to extract local features from each channel during computation and achieve cross-channel feature combination through subsequent weight aggregation mechanisms. This allows the neural network to read structural, constraint, and historical information within the same convolutional window, forming a tensor structure that can be further processed by the feature encoding layer. The result is that the data is divided into a structured, convolutional, multi-dimensional array in the input space, providing a stable foundation for subsequent convolution, pooling, and feature encoding, and ensuring consistent index and area features during forward propagation of the neural network.

[0016] This invention aligns and combines three types of channel mapping data according to structural indices into a structural feature matrix. This matrix gives the positions of the tooth tip, tooth ventral region, and tooth root a fixed row and column range in the matrix space. The structural feature matrix allows the convolutional kernel to perform local scanning within a fixed two-dimensional index domain and accurately identify the location of structural change regions. Since the tooth surface is a three-dimensional curved surface structure, directly inputting it into the network often makes it difficult to guarantee neighborhood correspondence. However, after matrixing, each row and column index represents the structural position of the tooth shape in a certain local area, allowing the convolution operation to maintain topological correspondence. This provides the local computation of the neural network with a clear spatial reference basis. The convolutional kernel can move continuously in the matrix and calculate local features, ensuring that stage features are not distorted by differences in geometric coordinates, forming a fixed-domain representation that the neural network can recognize and propagate.

[0017] This invention transforms complex geometric relationships into continuous variables by calculating mega-entropy, enabling convolutional layers, feature encoding layers, and control adjustment layers to directly process these numerical values ​​representing complex forms through weights and activation functions. Since convolutional neural networks inherently perform weighted summations and nonlinear transformations, their inputs must be continuously processable matrices or vectors. Mega-entropy encodes geometric or topological characteristics that are difficult to process directly through convolution into one-dimensional or multi-dimensional numerical values, allowing them to participate in forward propagation along with the original input matrix. By mapping structural changes to network-computable features, abstract features acquire numerical forms suitable for entering the neural network, allowing convolutional kernels to directly access structural change information through kernel weights during computation.

[0018] This invention constructs a chain amplitude structure matrix by segmenting and using mega-entropy, and calculates the chain amplitude from it. Traditional convolution can only process local neighborhoods, but the chain amplitude structure matrix, through segment pair combination and matrix filling, enables regions with no direct adjacency to establish numerical connections in the matrix space. The chain amplitude further extracts the overall expansion trend from the matrix, allowing the network to access cross-segment structural relationships through this variable during forward propagation. This establishes a numerical framework that can propagate non-local structural information within the convolutional network, so that long-distance structural associations no longer rely on manual features but are read by the network in the form of matrix values. This allows network components such as convolutional layers and chain fusion layers to automatically combine information from different regions during computation.

[0019] This invention maps mega-entropy to a complex amplification factor and chain amplitude to an extended balance factor, and simultaneously superimposes these two variables onto each element of the chain amplitude structure matrix to obtain an adjusted structure matrix. This ensures that local complexity information and cross-regional relationship information are no longer scattered across different data structures, but rather work together on the corresponding elements of a single matrix, forming a unified feature container that can be directly read by the neural network. This allows the convolutional kernel to simultaneously acquire the degree of local structural change and global structural constraints when processing any local region, enabling the data to form a unified tensor structure with complete correlation at the input layer, facilitating continuous computation and feature propagation in deep networks. Attached Figure Description

[0020] Figure 1 This is a flowchart of a method for generating a grinding trajectory model for a hobbing grinder based on deep learning, provided by an embodiment of the present invention. Detailed Implementation

[0021] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0022] like Figure 1 As shown, an embodiment of the present invention proposes a method for generating a grinding trajectory model for a hobbing grinder based on deep learning, the method comprising: Obtain the hob tooth surface structure, grinding wheel boundary dimensions, and historical machining samples to obtain the basic dataset; A multidimensional feature channel is established based on the basic dataset, and features of the tooth surface region, grinding wheel boundary region and historical trajectory region in the basic dataset are fused based on the multidimensional feature channel to obtain the structural feature matrix; Based on the structural feature matrix, the tooth is divided into segments according to the structural features of the tooth tip, tooth vent, and tooth root, and the complexity of the tooth surface structure in the local region is calculated to obtain the mega-entropy. Construct the chain amplitude structure matrix based on the mega-entropy and calculate the degree of extension of the hob between different segments to obtain the chain amplitude. A dual-adjustment structure is established based on the mega-entropy and chain amplitude, and the chain amplitude structure matrix is ​​adjusted and fused according to the dual-adjustment structure to obtain the adjustment structure matrix. A deep learning model is constructed based on a convolutional neural network. The deep learning model includes an input layer, a feature encoding layer, a structural association layer, a chained fusion layer, a control and adjustment layer, and an output layer. A training sample set is constructed based on historical processing samples and their interference markers. The deep learning model is then trained based on the training sample set and the adjustment structure matrix to obtain the grinding trajectory model. Input the basic dataset into the grinding trajectory model to generate grinding trajectory data.

[0023] In this embodiment of the invention, the hob tooth surface structure, grinding wheel boundary dimensions, and historical machining samples are acquired to obtain a basic dataset. This enables isomorphic storage of tooth surface geometry, grinding wheel entity constraint information, and machining process motion information, providing a unique identifier for subsequent processing. A multi-dimensional feature channel is established based on the basic dataset, and features from the tooth surface region, grinding wheel boundary region, and historical trajectory region in the basic dataset are fused using this multi-dimensional feature channel to obtain a structural feature matrix. This allows the neural network to simultaneously read these three types of information within a single convolution window during local convolution, avoiding information loss caused by a single geometric or boundary input. Based on the structural feature matrix... The array is segmented according to the structural features of the tooth tip, tooth belly, and tooth root, and the complexity of the tooth surface structure in local regions is calculated to obtain the mega-entropy. This allows the local complexity and interrelationships of the tooth tip, tooth belly, and tooth root to be quantifiable at the data level, mapping local micro-features and cross-segment macro-changes into a unified numerical quantity. Based on the mega-entropy, a chain amplitude structure matrix is ​​constructed, and the degree of extension of the hob between different segments is calculated to obtain the chain amplitude. This gives the cross-segment structural influence and the overall structural extension a matrix and scalar expression form in the data space that can directly participate in deep learning calculations, providing a data foundation for the subsequent introduction of structural extension constraints.

[0024] A dual-adjustment structure is established based on mega-entropy and chain amplitude, and the chain amplitude structure matrix is ​​adjusted and fused according to the dual-adjustment structure to obtain an adjusted structure matrix. This allows the neural network to consider local structural features and overall expansion constraints within the same data framework when performing convolution, feature encoding, or chain fusion, establishing a complete structural adjustment mechanism along the data processing path and providing a foundation for subsequent structural control. A deep learning model is constructed based on the convolutional neural network. The deep learning model includes an input layer, a feature encoding layer, a structural association layer, a chain fusion layer, a control adjustment layer, and an output layer, enabling the feature representation inside the convolutional neural network to directly interface with the actual structural group of the hob tooth profile and grinding wheel constraints. The process of weaving lays the data representation foundation for the chain generation and control of subsequent trajectory nodes. A training sample set is constructed based on historical processing samples and their interference markers. The deep learning model is then trained using the training sample set and the adjustment structure matrix to obtain the grinding trajectory model. During the training process, the model learns the correspondence between trajectory nodes and interference occurrences. During parameter updates, it forms an implicit division between collision and non-collision regions, avoiding feature patterns that have experienced interference. The basic dataset is input into the grinding trajectory model to generate grinding trajectory data. The model automatically outputs trajectory node data that meets structural constraints and historical experience constraints, realizing end-to-end data mapping from multi-source structural data to trajectory data.

[0025] This involves obtaining the hob tooth surface structure, grinding wheel boundary dimensions, and historical machining samples to create the basic dataset, which specifically includes: First, for the target hob, three-dimensional structural data of the hob tooth surface is acquired with the help of a coordinate measuring machine or a tooth profile measuring instrument. The hob is mounted on a special fixture of the measuring device. By rotating the spindle and feeding axially, the sensor probe sequentially performs contact or non-contact scanning along the tooth tip, tooth flank, and tooth root of each tooth to obtain a discrete measurement point coordinate sequence in the machine tool coordinate system. Based on the original point cloud data obtained from the measurement, interpolation fitting and curve reconstruction are used to transform each tooth profile into a sequence of profile points uniformly sampled according to standardized tooth height parameters. At the same time, the geometric attributes such as the normal vector, radius of curvature, and local slope of each profile point are calculated. The above data are associated with the tooth number, tooth direction position, and tooth height parameter index to form tooth surface structure data.

[0026] For grinding wheel data, the grinding wheel to be used is installed on a grinding wheel inspection fixture. A contour projector, contour meter, or line-scan laser sensor is used to perform a circumferential scan of the grinding wheel working layer and its adjacent substrate area, collecting contour point sets of the outer circle, end face, fillet transition section, and edge of the grinding wheel substrate. Based on the nominal outer diameter, thickness, and boundary position of the abrasive layer and the substrate, the scanned point set is coordinate registered and circularly fitted to extract the outer diameter parameters, end face shape parameters, fillet radius parameters, and outer edge parameters of the grinding wheel substrate. The grinding wheel boundary contour is recorded in a unified grinding wheel coordinate system. The outer edge contour of the grinding wheel is discretized into a set of boundary point sequences with angle indices to obtain the grinding wheel boundary dimension data.

[0027] For historical machining samples, the trajectory record file of the existing machining task is retrieved from the CNC system or host computer monitoring system of the hobbing grinder. The machine tool multi-axis position commands, execution feedback positions and corresponding timestamps of each interpolation cycle during the entire grinding process are read. The above data is converted into trajectory node coordinates represented in the hobbing workpiece coordinate system or the machine tool unified coordinate system. Furthermore, the interference mark information corresponding to each historical trajectory is read from the interference detection module, the in-machine monitoring module or the manual annotation results. The trajectory segments that have rubbed, overcut or too close are attached to the corresponding trajectory nodes in the form of Boolean marks or multi-level risk marks to form historical machining sample data with interference marks.

[0028] The basic dataset is input into the grinding trajectory model to generate grinding trajectory data, specifically including: After training the convolutional neural network structure and obtaining a converged grinding trajectory model, based on the actual situation of the hob to be processed, the basic dataset of the hob is retrieved from the database or file system. The tooth surface structure data, grinding wheel boundary dimension data, and historical machining sample data are loaded into the tooth surface input tensor, boundary input tensor, and trajectory input tensor respectively according to a preset channel mapping rule. During the loading process, firstly, based on the standardized tooth height parameters of the hob tooth profile, the discrete tooth profile points are sorted by tooth number and tooth height index, and filled into the spatial dimension of the input tensor, so that one spatial dimension of the tensor corresponds to the sampling point sequence in the tooth height direction, and the other... One spatial dimension corresponds to the tooth direction or sample sequence index; then, the grinding wheel boundary dimension data is mapped to the same spatial grid according to the same parameterization rules as the tooth surface structure, and the boundary information composed of the grinding wheel outer diameter point series, fillet contour point series and base boundary point series is written into the second input channel, so that each spatial position has the corresponding grinding wheel constraint information at the same time; for historical processing sample data, according to the trajectory sequence index and time step index, the recorded trajectory node coordinates and interference mark information are mapped to the parameter space that is consistent with the tooth surface structure and grinding wheel boundary, and the trajectory path value and interference label value are filled in at the corresponding spatial position to form the third input channel.

[0029] The three channels are stacked into a unified multi-channel input tensor along the channel dimension. Necessary normalization and dimensionality pruning are performed according to the size requirements of the network input layer. This multi-channel input tensor and the corresponding adjustment structure matrix are then synchronously fed into the grinding trajectory model. During forward inference, the grinding trajectory model first receives the multi-channel tensor and adjustment structure matrix through the input layer and passes them to the feature encoding layer. Multi-layer convolution operations are used to extract tooth surface shape patterns, grinding wheel boundary constraint patterns, and historical trajectory distribution patterns within the local receptive field. In the subsequent structural association layer, the encoded features are rearranged according to the segment structure of the tooth tip, tooth flank, and tooth root, forming a feature mapping with spatial partitioning significance. Then, under the action of chain fusion and control adjustment structure, the feature responses corresponding to different segments and regions are numerically adjusted according to the input adjustment structure matrix. Feature combinations similar to historical interference patterns are suppressed, while feature combinations consistent with safe trajectory patterns are strengthened. Finally, the grinding trajectory data corresponding to each tooth and each machining stage is output in the form of a sequence of trajectory nodes at the output layer.

[0030] In a preferred embodiment of the present invention, a multi-dimensional feature channel is established based on the basic dataset, and features from the tooth surface region, grinding wheel boundary region, and historical trajectory region in the basic dataset are fused based on the multi-dimensional feature channel to obtain a structural feature matrix, including: By mapping tooth surface structure data, grinding wheel boundary size data, and historical trajectory sample data to tooth surface region dataset, boundary region dataset, and trajectory region dataset, respectively, region classification data is obtained; Based on the regional classification data, the tooth surface region dataset, boundary region dataset, and trajectory region dataset are mapped to the first feature channel, the second feature channel, and the third feature channel to obtain channel mapping data. Based on the channel mapping data, the data in the three types of feature channels are aligned according to the structure index order to ensure that the data at the same index position form a corresponding relationship, thus obtaining the aligned channel data; Based on the alignment channel data, the data in the three feature channels at each index position are combined to obtain the combined feature data; Based on the combined feature data, the combined features at different index positions are arranged according to the row and column correspondence to obtain the structural feature matrix.

[0031] In this embodiment of the invention, by mapping tooth surface structure data, grinding wheel boundary size data, and historical trajectory sample data to tooth surface region datasets, boundary region datasets, and trajectory region datasets respectively, region classification data is obtained. This avoids data mixing or semantic ambiguity within channels, ensuring that the meaning represented by each channel at the neural network input is fixed and the boundaries are clear. Based on the region classification data, the tooth surface region dataset, boundary region dataset, and trajectory region dataset are mapped to the first feature channel, the second feature channel, and the third feature channel, respectively, to obtain channel mapping data. This achieves cross-channel feature fusion, providing an ordered and convolutional multidimensional data foundation for the subsequent construction of the structural feature matrix. Based on the channel mapping data... The data in the three types of feature channels are aligned according to the structural index order to ensure that the data at the same index position form a corresponding relationship, thus obtaining aligned channel data. This ensures that the multi-source data at each position has a strict alignment relationship and can be aggregated. Based on the aligned channel data, the data in the three types of feature channels at each index position are combined to obtain combined feature data, realizing the local integrated expression of multi-source data. This makes it easier for the network to perceive complete structural, boundary, and historical trajectory information within the same receptive field. Based on the combined feature data, the combined features at different index positions are arranged according to the row and column correspondence to obtain the structural feature matrix, which provides a unified data carrier and a clear spatial division basis for subsequent processing.

[0032] Specifically, by mapping tooth surface structure data, grinding wheel boundary dimension data, and historical trajectory sample data to tooth surface region datasets, boundary region datasets, and trajectory region datasets, respectively, region classification data is obtained, including: First, the system reads raw data related to tooth surface structure, grinding wheel boundary dimensions, and historical machining trajectories from the basic dataset and organizes the data from different sources according to a preset format. For tooth surface structure data, the system reads the 3D measurement file of the hob tooth surface or the discrete curve generated by the design parameters, and collects the spatial coordinates, tooth surface normal vector, local curvature, tangential direction, and the structural segment to which the point belongs (tooth tip, tooth vent, tooth root) for each discrete point on the tooth profile. A unified structural index is added to this data, so that each discrete point on the tooth surface has a unique number on the entire tooth profile, thus accurately representing the relative position of the point on the tooth surface through the number. When processing the grinding wheel boundary dimension data, the system extracts the outer circle contour of the grinding wheel, the end face shape of the grinding wheel, the outer shape of the grinding wheel base, and the coordinates of its calibration points in the machine tool coordinate system according to the geometric design parameters or 3D scanning data of the grinding wheel. The outer boundary of the grinding wheel is parameterized so that the boundary points of the grinding wheel are also labeled with the same structural index, so that the grinding wheel boundary can be matched with the tooth surface structure in subsequent steps through the index. For historical trajectory sample data, the system reads the motion trajectory nodes of each machining process from the machining log recorded by the CNC system, including the spatial coordinates of the trajectory points, the normal direction, the trajectory time sequence number, and the machining area to which the trajectory segment belongs. Then, through the tooth surface matching algorithm, each trajectory node is mapped to the structural position corresponding to the tooth surface, and a structural index consistent with the tooth surface is assigned to each trajectory node.

[0033] Then, based on the data source category, data belonging to discrete points on the tooth surface, data belonging to the grinding wheel boundary, and data belonging to the historical trajectory are respectively classified into the tooth surface region dataset, the boundary region dataset, and the trajectory region dataset, making each of the three types of data an independent set with a unified structure index and a unified vector format. For missing points in the tooth surface region dataset, interpolation methods based on the tooth surface parametric equations are used to fill in the missing points; for missing points in the grinding wheel boundary region dataset, the contours are filled in using extrapolation methods based on the grinding wheel geometric parameters; for structural locations in the historical trajectory region dataset without trajectory mapping, they are represented by specific trajectory-free labeled vectors, ultimately obtaining the region classification data.

[0034] Specifically, based on the region classification data, the tooth surface region dataset, boundary region dataset, and trajectory region dataset are mapped to the first feature channel, the second feature channel, and the third feature channel to obtain channel mapping data, which includes: First, three parallel channel structures are created, with the index range of each channel strictly consistent with the structure index set, forming three parallel data sequences in terms of shape. For the first feature channel, the system sequentially fills the channel with the vector corresponding to each structure index in the tooth surface region dataset. This vector typically contains the three-dimensional coordinates, normal vector, curvature, rate of change of curvature, segment number, and normalized position parameters of that location in the tooth profile parameter space at a certain position on the tooth surface. Each index position corresponds to a complete tooth surface feature vector, enabling the channel to present the complete local morphology of the tooth surface structure during convolution. When constructing the second feature channel, the system writes the grinding wheel boundary features at each index position in the boundary region dataset into the channel according to the structure index order. This feature typically includes the minimum distance from the outer edge of the grinding wheel to the structure position, the relative azimuth angle of the grinding wheel boundary to that position, the contour parameters of the grinding wheel shape at that position, the approximate constraint vector of the grinding wheel outer diameter, and the boundary influence value of the matrix shape, etc. If a certain index position does not contain grinding wheel boundary data (e.g., the tooth surface has no direct correspondence with the grinding wheel boundary), the system will fill in a zero vector or boundary placeholder at that position to ensure the continuity of the entire channel along the structure index axis. The third feature channel is used to fill the trajectory region dataset. In this channel, the vector corresponding to each structure index position consists of the trajectory node parameters that are closest to that region in the historical trajectory or determined by the mapping algorithm. These parameters include the trajectory node coordinates, trajectory direction, trajectory node sequence number, trajectory velocity information, and the distance measure between the trajectory and the tooth surface structure. When trajectory mapping data is missing at a certain structure index position, the system fills it in using a trajectory missing marker, making the entire channel a trajectory feature sequence with complete index coverage. The system performs consistency checks on the index positions of the three channels to ensure that each structure index position has a data vector in all three channels, constituting the channel mapping data.

[0035] In a preferred embodiment of the present invention, the tooth surface structure is segmented according to the structural features of the tooth tip, tooth vent, and tooth root based on the structural feature matrix, and the complexity of the tooth surface structure in a local region is calculated to obtain a mega-entropy quantity, including: By extracting the data from the structural feature matrix according to the positions of the tooth tip, tooth vent, and tooth root, and dividing them into several local segments, segment division data is obtained. Based on the segment division data, the strength of the change in tooth surface structure in each local segment is calculated to obtain the local difference term; the overall complexity within different local segments is calculated to obtain the regional complexity term; the overall change intensity of complexity between each local segment is calculated to obtain the regional jump term; and the degree of correlation between different tooth surface structures in the same local segment is calculated to obtain the segment correlation term. By fusing the local difference term, the regional complexity term, the regional jump term, and the segment correlation term, we obtain the mega-entropy quantity.

[0036] In this embodiment of the invention, data from the structural feature matrix is ​​extracted according to the positions of the tooth tip, tooth vent, and tooth root, and then divided into several local segments to obtain segment division data. This avoids the mixing and statistical analysis on the global matrix, which would cause structural information aliasing. Local segments become the basic calculation units, and subsequent local differences, regional complexity, regional transitions, and segment relationships can all be performed around these units. Based on the segment division data, the strength of change in the tooth surface structure at each local segment is calculated to obtain a local difference term. This allows for the differentiation between segments with smooth internal structures and segments with significant internal structural changes, identifying which segments may correspond to areas with large changes in the machining path. The overall complexity within different local segments is calculated to obtain a regional complexity term, characterizing the diversity of the segment structure. The degree of coupling between sex and multidimensionality allows the mega-entropy to distinguish between segments with only local fluctuations but simple overall morphology and segments with dispersed overall distribution and significant multimodal characteristics. It calculates the overall intensity of complexity change between local segments, obtaining a regional jump term, and provides a quantitative description of the structural continuity or abrupt change at the boundaries of each segment. It calculates the degree of correlation between different tooth surface structures at the same local segment, obtaining a segment correlation term, which can identify whether the same segment maintains similar structural patterns across different tooth surface types. By fusing the local difference term, regional complexity term, regional jump term, and segment correlation term, the mega-entropy is obtained, reducing the redundancy of the neural network input dimension while retaining sensitivity to various structural features, providing a quantitative basis for the construction of chain amplitude structure matrices and dual-regulation structures.

[0037] In a preferred embodiment of the present invention, a chain width structure matrix is ​​constructed based on the mega-entropy, and the extent of the hob's extension between different segments is calculated to obtain the chain width, including: Based on the segment division data and mega-entropy, the characteristic changes between different local segments are combined in pairs, and the expansion relationship between each segment pair is sequentially filled into the data array to obtain the chain amplitude structure matrix. Based on the chain length structure matrix, the degree of cumulative difference between adjacent segments is calculated, the expansion trend caused by local structural differences between different segments of the hob is identified, and the inter-segment expansion term is obtained; the relative position of each segment in the direction from tooth tip to tooth root is calculated, the degree of influence of different segment positions on the overall expansion is identified, and the position modulation term is obtained. The overall complexity is calculated to amplify or compress the expansion results of each segment, resulting in a complexity coupling term. The degree of mutual influence between expansion results of different segments is calculated, resulting in a chain amplitude coupling term. The inter-segment expansion term, position modulation term, complexity coupling term, and chain amplitude coupling term are fused together to obtain the chain amplitude quantity.

[0038] In this embodiment of the invention, based on segment division data and mega-entropy, the feature changes between different local segments are combined in pairs, and the expansion relationships between each segment pair are sequentially filled into the data array to obtain the chain amplitude structure matrix. This realizes the transformation from a single segment feature sequence to a segment pair relationship matrix, providing a structured data foundation for subsequent calculations. Based on the chain amplitude structure matrix, the degree of cumulative difference between adjacent segments is calculated, and the expansion trend caused by local structural differences between different segments of the hob is identified to obtain the inter-segment expansion term, reflecting the cumulative effect of local regional structural changes in the segment column direction. The relative position of each segment in the tooth tip to tooth root direction is calculated, and the influence of different segment positions on the overall expansion is identified to obtain the position modulation term, so that the same expansion trend has different values ​​at different physical positions. The numerical representation of the tooth tip, tooth vent, and tooth root can form a clear hierarchical division in the data; the amplification or compression effect of the overall complexity on the expansion results of each segment is calculated to obtain the complexity coupling term, which realizes the embedding of structural complexity that is difficult to be directly used as input into the feature space of the network through a single numerical feature; the degree of mutual influence between the expansion results of different segments is calculated to obtain the chain amplitude coupling term, which reflects the correlation between different segments and surrounding segments in the expansion behavior; the inter-segment expansion term, position modulation term, complexity coupling term, and chain amplitude coupling term are fused to obtain the chain amplitude quantity, so that the local expansion information, position encoding information, overall complexity background information, and inter-segment interaction information that were originally scattered in multiple data structures are compressed into a single or low-dimensional feature vector at each segment.

[0039] In a preferred embodiment of the present invention, a dual-adjustment structure is established based on the mega-entropy and chain amplitude, and the chain amplitude structure matrix is ​​adjusted and fused according to the dual-adjustment structure to obtain an adjustment structure matrix, including: By converting mega-entropy into complex amplification factors and chain amplitude into extended balance factors, the adjustment data is obtained; Based on the adjustment data, a complex amplification coefficient and an extended balance coefficient are simultaneously superimposed numerically to ensure that the two form a corresponding relationship within the same adjustment framework, thus obtaining dual-adjustment structure data. Based on the dual-adjustment structure data, the adjustment effect of complex amplification coefficient and extended balance coefficient is applied simultaneously to each matrix element of the chain axis structure matrix to obtain the correction matrix data; Based on the correction matrix data, the data related to the tooth tip, tooth vent, and tooth root are summarized and arranged in the processing order to obtain the adjustment structure matrix.

[0040] In this embodiment of the invention, by converting mega-entropy into complex amplification coefficients and chain amplitude into extended balance coefficients, adjustment data is obtained. This provides a calculable parameter form for subsequently superimposing dual adjustment effects on the chain amplitude structure matrix, enabling the entire system to have a unified scale and interface at the data representation level. Based on the adjustment data, complex amplification coefficients and extended balance coefficients are simultaneously superimposed numerically, ensuring that they form a corresponding relationship within the same adjustment framework, resulting in dual-adjustment structure data. This achieves synchronous adjustment of each matrix element, avoiding repeated indexing and searching in the calculation path. Based on the dual-adjustment structure data, the adjustment effects of complex amplification coefficients and extended balance coefficients are simultaneously applied to each matrix element of the chain amplitude structure matrix, resulting in correction matrix data. Each element in the matrix no longer represents only a single extension trend, but rather a comprehensive quantity under the combined effect of local structural information and cross-segment extended balance. Based on the correction matrix data, data related to tooth tip, tooth vent, and tooth root are summarized and arranged according to the processing order to obtain the adjustment structure matrix. This not only retains the comprehensive correction results of local complexity and inter-segment extension relationships, but also encodes these correction results according to the order of the actual processing path.

[0041] Specifically, by converting mega-entropy into complex amplification factors and chain amplitude into extended balance factors, adjustment data is obtained, including: First, a complex amplification coefficient mapping relationship is established for mega-entropy. A global statistical operation is performed on all mega-entropy samples to obtain the minimum, maximum, and mean values ​​of mega-entropy. Based on this, a monotonically increasing mapping function is set to map each mega-entropy value to a complex amplification coefficient within a preset interval. The input of the mapping function is the mega-entropy corresponding to a certain segment or segment pair, and the output is the complex amplification coefficient corresponding to that segment or segment pair. The mapped complex amplification coefficients are stored in array or matrix form, where the array or matrix index is consistent with the index of the segment or segment pair. For each chain amplitude, a similar statistical and mapping process is performed as for mega-entropy. An extended balance coefficient mapping function is set in the control module to normalize the amplitude and map the range of the chain amplitude, converting each chain amplitude into a corresponding extended balance coefficient. All extended balance coefficients are stored in an extended balance coefficient array or matrix in segment pair index order, so that each chain amplitude corresponds to a unique extended balance coefficient. The system then aligns the complex amplification coefficient array and the extended balance coefficient array one by one according to the pre-determined segment or segment pair index, and records the complex amplification coefficient and extended balance coefficient at the same index position to construct the adjustment data.

[0042] Specifically, based on the dual-adjustment structure data, a complex amplification factor and an extended balance factor are simultaneously applied to each element of the chain axis structure matrix to obtain the corrected matrix data, which includes: First, for any matrix element in the chain-width structure matrix, the corresponding record cell is retrieved in the double-adjustment structure data based on its row and column indices. The complex amplification factor and extended balance factor corresponding to that matrix element are read from this record cell. Then, the original matrix element is numerically corrected. The original matrix element, complex amplification factor, and extended balance factor at each position are used as a one-time input to obtain a uniquely determined correction value, which is then written to the correction matrix storage area corresponding to the index. During the traversal, all matrix elements of the chain-width structure matrix undergo reading, adjustment, and writing operations sequentially until correction values ​​at all positions are generated, ultimately forming a correction matrix data that completely corresponds to the original chain-width structure matrix in size and index.

[0043] In a preferred embodiment of the present invention, constructing a deep learning model based on a convolutional neural network includes: Construct an input layer to receive the basic dataset, divide the basic dataset into tooth surface input data, boundary input data, and trajectory input data, and align them to obtain the input data; A feature coding layer is constructed to extract features of tooth surface shape change, grinding wheel shape constraint, and historical motion trajectory based on the input data, and combine them into fusion features to obtain feature coding data; A structural association layer is constructed to divide the fused features into tooth crest partition features, tooth ventral partition features, and tooth root partition features based on the feature encoding data, and arrange them into structural association data.

[0044] In this embodiment of the invention, an input layer is constructed to receive a basic dataset, divide the basic dataset into tooth surface input data, boundary input data, and trajectory input data, and align them to obtain input data, avoiding input tensor dimension mismatch or spatial correspondence confusion caused by inconsistent data sources; a feature encoding layer is constructed to extract tooth surface shape change features, grinding wheel shape constraint features, and historical motion trajectory features based on the input data, and combine them into fusion features to obtain feature encoding data, so that the feature flow inside the deep learning model completes the comprehensive expression of different information sources in the encoding stage, and enters the subsequent structural association modeling process in a unified feature tensor form; a structural association layer is constructed to divide the fusion features into tooth tip partition features, tooth vent partition features, and tooth root partition features based on the feature encoding data, and arrange them into structural association data, so that the tooth tip, tooth vent, and tooth root have a structured feature sequence with clear partition boundaries and transition order in the feature space, providing an input foundation with regional semantics, positional semantics, and transition semantics for subsequent trajectory generation.

[0045] Specifically, a feature encoding layer is constructed to extract features such as tooth surface shape change, grinding wheel shape constraint, and historical motion trajectory from the input data, and combine them into fused features to obtain feature-encoded data, including: First, the three types of input data are numerically normalized and dimensionally organized within their respective channels. The tooth surface input data is arranged into a two-dimensional array according to the tooth direction and tooth height indices. The grinding wheel boundary input data is organized into a boundary constraint array according to the same indexing system as the tooth surface. The trajectory input data is organized into a trajectory array according to the trajectory node index and processing order. Missing positions are filled or outliers are removed in each array to ensure a continuous distribution of the input arrays in both spatial and index dimensions. Then, a tooth surface feature encoding sub-network is constructed using the tooth surface input data as input. Convolutional kernels are placed along the tooth direction and tooth height directions on the tooth surface array to locally scan the height, normal, and curvature changes of the tooth surface structure, extracting the first type of intermediate feature map reflecting the degree of tooth surface concavity and convexity, transition fillet shape, and tooth profile undulation characteristics. A boundary feature encoding sub-network is constructed using the boundary input data as input. Convolutional kernels are placed along the same indexing direction as the tooth surface on the boundary constraint array to identify the effective grinding width, rim boundary position, and other features of the grinding wheel profile at each index position. The relative distance between tooth surfaces changes, generating a second type of intermediate feature map reflecting the grinding wheel's shape constraints and the distribution of physically accessible areas. A trajectory feature encoding sub-network is constructed using the trajectory input data as input. Convolutional kernels or one-dimensional convolutional kernels are arranged along the trajectory node index direction on the trajectory array to extract the coordinate change patterns of adjacent nodes in the historical motion trajectory, the variation law of node spacing, and the trajectory orientation features at different segments of the tooth tip, tooth vent, and tooth root, resulting in a third type of intermediate feature map. Feature alignment and fusion operations are set within or after the three types of feature encoding sub-networks. The intermediate features of the tooth surface, boundary, and trajectory at the same structural index position are spliced ​​or weighted in the channel dimension to form a fused feature vector that simultaneously contains tooth surface shape information, grinding wheel shape constraint information, and historical trajectory pattern information at each index position. Through nonlinear activation, normalization, and dimensionality compression, the fused feature vector is arranged into a continuous feature encoding matrix or feature encoding tensor throughout the entire tooth surface structural domain, serving as the feature encoding data output.

[0046] Specifically, a structural association layer is constructed to divide the fused features into tooth crest, tooth ventral, and tooth root partition features based on the feature encoding data, and arrange them into structural association data. This includes: First, based on the design parameters and discretization rules of the tooth surface structure in the tooth height direction, the boundary positions of the tooth tip region, tooth vent region, and tooth root region in the index dimension are determined. The fusion features within the corresponding index range are extracted from the feature encoding data to construct tooth tip partition feature sets, tooth vent partition feature sets, and tooth root partition feature sets, ensuring that the feature vectors within each set correspond to different discrete positions within the same structural region. Then, within each partition, the fusion features are rearranged according to the tooth direction index and processing order, reorganizing the feature encoding data, originally existing only as a two-dimensional matrix, into a feature sequence or feature submatrix with a clear order. This results in the tooth tip partition features forming a continuous feature distribution from the tooth tip edge to the tooth tip transition section within the tooth tip region; the tooth vent partition features forming a continuous feature distribution extending along the tooth direction within the tooth vent region; and the tooth root partition features forming a continuous feature distribution covering the root transition fillet and root connection curve within the tooth root region. Within each partition, one-dimensional convolutional units and local hinges are set... Connecting units or other network structures used to model sequence correlations perform local relation calculations on feature sequences within partitions, expressing the adjacent positional relationships and local change patterns between feature vectors within the network. This forms local correlation features in the tooth tip region within the tooth tip partition, local correlation features in the tooth belly region within the tooth belly partition, and local correlation features in the tooth root region within the tooth root partition. Based on this, cross-partition feature splicing units or cross-partition connection units connect the tooth tip partition features with the tooth belly partition features, and the tooth belly partition features with the tooth root partition features, according to the structural order along the tooth height direction. This allows features from different partitions to form transitional relationships at specific connection positions, thus establishing a step-by-step transition path between tooth tip, tooth belly, and tooth root in the feature space. Finally, the features calculated through intra-partition and inter-partition correlations are rearranged according to a preset index order, forming structurally correlated data with clear structural partition markings and continuous transitional relationships in both the tooth direction and tooth height direction.

[0047] In a preferred embodiment of the present invention, the method of constructing a deep learning model based on a convolutional neural network further includes: A chain-like fusion layer is constructed to combine the structural features corresponding to the tooth tip, tooth vent, and tooth root into multiple chain-like combinations according to the segment order based on the structural association data, thereby obtaining chain-like fused data; A control and adjustment layer is constructed to adjust the chained fusion data based on mega-entropy and chain amplitude. Mega-entropy is used to assign weights to each segment, and chain amplitude is used to limit the expansion range of each segment, thus obtaining control and adjustment data. An output layer is constructed to extract nodes of the motion trajectory based on the control and adjustment data, and sort the nodes according to the processing order to obtain the trajectory node output data.

[0048] In this embodiment of the invention, a chain fusion layer is constructed to combine the structural features corresponding to the tooth tip, tooth vent, and tooth root into multiple chain combinations according to the segment order based on the structural association data, thereby obtaining chain fusion data. This avoids position mismatch or dimensional chaos of partitioned features after entering the deep network. A control adjustment layer is constructed to adjust the chain fusion data according to the mega-entropy and chain amplitude. The mega-entropy assigns weights to each segment, and the chain amplitude limits the expansion range of each segment, thereby obtaining control adjustment data. This avoids information loss or position inconsistency during the adjustment process. An output layer is constructed to extract nodes of the motion trajectory based on the control adjustment data and sort the nodes according to the processing order, thereby obtaining trajectory node output data. The trajectory nodes no longer remain in the implicit feature space but appear in the form of explicit numerical vectors, realizing a seamless data docking process from the deep learning model to the external control system.

[0049] Specifically, a chain-like fusion layer is constructed to combine the structural features corresponding to the tooth tip, tooth vent, and tooth root into multiple chain combinations according to segment order based on the structural association data, thereby obtaining chain-like fused data. This includes: First, the structural data is categorized and stored according to tooth tip, tooth side, and tooth root partitions, forming a basic feature tensor with partition identifier, segment index, and feature dimension as the index dimensions. Based on the segment division results of the tooth surface structure in the previous steps, a segment order index table is established. The index position of each segment in the tooth tip, tooth side, and tooth root partitions is written into this segment order index table, ensuring that each segment has a unique sequence number and feature position in each partition in the data structure. According to the segment order index table, the feature vectors of each segment in the tooth tip, tooth side, and tooth root partitions are sequentially read in the processing unit. The three types of feature vectors are concatenated along the feature dimension in a preset order to obtain a chained combined feature unit for that segment. All chained combined feature units corresponding to all segments are then processed. Arranged sequentially according to segment order, forming a chain-like combination sequence with segment order as a one-dimensional axis; one-dimensional convolution kernels or other local aggregation operators are set on the chain-like combination sequence, limiting the receptive field of each convolution kernel to the chain-like combination feature units of several adjacent segments. Weighted summation and nonlinear transformation operations are performed on the chain-like combination feature units within each receptive field, and the operation results are written into a new sequence feature tensor, so that the sequence feature tensor records the chain-like features that fuse the local context information of the tooth tip, tooth belly, and tooth root at each segment order position; after completing the convolution and nonlinear transformation of all segments, the sequence feature tensor is normalized and formatted, and the segment order index, the corresponding fused feature vector, and the partition identifier are packaged according to a predetermined data structure to obtain chain-like fused data.

[0050] Specifically, a control and adjustment layer is constructed to adjust the chained fusion data based on mega-entropy and chain amplitude. Mega-entropy assigns weights to each segment, and chain amplitude limits the expansion range of each segment, resulting in control and adjustment data, which includes: First, based on the segment order index recorded in the chain-fused data, each segment order index is mapped and matched with the corresponding mega-entropy and chain amplitude quantities, forming a one-dimensional mega-entropy sequence and a one-dimensional chain amplitude sequence indexed by the segment order. Then, based on the feature dimensions of the chain-fused data, a dimension expansion operation is performed on the mega-entropy and chain amplitude sequences. The mega-entropy quantity at each segment order position is copied or expanded through a linear transformation to the same feature dimension length as the feature vector of that segment, resulting in mega-entropy tensors and chain amplitude tensors corresponding element-wise in the segment order and feature dimensions of the chain-fused data. In the processing unit, a double adjustment coefficient is calculated for each segment order and each feature dimension. The value in the mega-entropy tensor is used as the complexity factor of that feature, and the value in the chain amplitude tensor is used as the expansion constraint factor of that feature. The complexity factor and expansion constraint factor are combined into an adjustment weight tensor according to a preset functional relationship. The adjustment weight tensor is structurally related to... The chained fusion data maintains index consistency. After constructing the adjustment weight tensor, element-wise operations are performed on the chained fusion data and the adjustment weight tensor at the corresponding segment order and feature dimension positions. The feature values ​​in the chained fusion data are multiplied or weighted and superimposed with the adjustment weights. The results are written into a new feature tensor, so that the new feature tensor numerically reflects the combined effect of the complexity information and expansion constraint information of each segment on the chained fusion features. Normalization and numerical range pruning are performed on the new feature tensor to control the norm of the feature vectors at each segment order position within a preset range, and feature values ​​exceeding the range are truncated by the upper or lower limit to avoid numerical overflow or gradient anomalies in subsequent network operations. After normalization, the new feature tensor is re-encapsulated, and the segment order index, the adjusted feature vector, and the additional index information related to adjustment are uniformly written into the output data structure to form control adjustment data.

[0051] In a preferred embodiment of the present invention, a training sample set is constructed based on historical processing samples and their interference markers, and a deep learning model is trained based on the training sample set and the adjustment structure matrix to obtain a grinding trajectory model, including: By matching each historical processed sample with its corresponding interference marker, labeled sample data is obtained; Based on the labeled sample data, the continuous processing records are divided into training input sequences, and the corresponding interference marks are organized into training target sequences to obtain the training sample set; Predicted trajectory data is obtained by feeding the training input sequence into a deep learning model; Error data is obtained by comparing the predicted trajectory data with the training target sequence and calculating the degree of deviation between the two. Based on the error data and the adjustment structure matrix, the parameters of the deep learning model are updated until the error data meets the preset convergence condition, thus obtaining the grinding trajectory model.

[0052] In this embodiment of the invention, by pairing each historical processing sample with its corresponding interference marker, labeled sample data is obtained. This ensures that each trajectory node possesses both motion characteristics and interference state information simultaneously in the data structure, transforming the interference information from an independent detection result into labeled data that can be input into the neural network along with the trajectory data. Based on the labeled sample data, continuous processing records are divided into training input sequences, and the corresponding interference markers are organized into training target sequences, resulting in a training sample set. This transforms the original continuous processing records from a single long sequence into a fixed-length training sequence suitable for processing by convolutional neural networks or sequence models, ensuring that each training input sequence has a matching training target sequence. By inputting the training input sequence into a deep learning network... In the learning model, predicted trajectory data is obtained, ensuring a strict correspondence between the predicted trajectory data and the training target sequence in terms of the number of time steps, node index, and output dimension. This allows subsequent error calculations to compare the predicted value with the target value one by one according to the node index. By comparing the predicted trajectory data with the training target sequence, the degree of deviation between the two is calculated, resulting in error data. This multi-dimensional array reflects the error distribution between different tooth surface segments, laying the foundation for gradient calculation. Based on the error data and the adjustment structure matrix, the parameters of the deep learning model are updated until the error data meets the preset convergence condition, resulting in the grinding trajectory model. This allows the deep learning model to receive error information from different segments and regions in a structured manner during backpropagation and adjust the network parameters accordingly.

[0053] Specifically, based on the labeled sample data, the continuous processing records are divided into training input sequences, and the corresponding interference marks are organized into training target sequences to obtain the training sample set, which includes: First, the annotated historical processing sample data is read. Each record in the annotated sample data includes at least a processing timestamp, a trajectory node index, the tooth surface structure features corresponding to that trajectory node, grinding wheel boundary features, historical motion trajectory features, and interference markers for that trajectory node or trajectory segment. The data preprocessing module sorts multiple nodes in the same processing record according to the timestamp or trajectory node index, obtaining a sequence of continuous processing records arranged in processing order. The sorted continuous processing records are grouped by processing number, and abnormal records due to sensor failure, missing records, etc., are removed. Simultaneously, the feature dimensions between different processing batches are unified to ensure consistency in the numerical range and encoding format of tooth surface structure features, grinding wheel boundary features, and historical trajectory features. Then, sequence length and step parameters for training are set. For example, each training input sequence is set to contain a predetermined number of continuous trajectory nodes, and a sliding window method is used to sequentially truncate the continuous processing records according to a preset step size. Several trajectory nodes within the coverage area of ​​each sliding window are considered as a training input sequence. For each training input sequence, the multi-dimensional feature vectors corresponding to each node in the basic dataset are concatenated according to the trajectory node index order. The feature vectors at different time steps are sequentially superimposed to form a two-dimensional or three-dimensional training input tensor, and the time step dimension, feature dimension, and possible channel dimension are respectively identified on the tensor dimension. At the same time, interference mark information or target trajectory node information corresponding one-to-one with the above trajectory node indexes are extracted from the labeled sample data and arranged in the same time order to form a training target sequence with the same length as the training input sequence. Several training target sequences are numbered and stored according to processing records and window positions, and can be further processed by one-hot encoding or numerical normalization when needed. For tail trajectories that are less than a complete window length, they can be filled by padding with zero vectors at the input end or copying features from neighboring nodes, and the filling position is marked with a specific mask at the target end. These filler trajectories do not participate in error calculation in subsequent training stages. After the above division and organization, the training input sequences corresponding to all continuous processing records are associated and stored with the training target sequences to form a training sample set.

[0054] Specifically, the predicted trajectory data is obtained by inputting the training input sequence into the deep learning model, including: First, several training input sequences and their corresponding training target sequences are read in batches from the training sample set. The training input sequences within the same batch are assembled into an input batch tensor according to a pre-defined tensor format. One dimension of the input batch tensor represents the sample number within the batch, another dimension represents the time step index, and the remaining dimensions represent the multi-channel feature vector at each time step. This input batch tensor is first fed into the input layer of the deep learning model. Based on the pre-defined network structure, the input layer maps data related to the tooth surface structure to the tooth surface input branch, data related to the grinding wheel boundary to the boundary input branch, and data related to the historical trajectory to the trajectory input branch, while keeping the time step index and sample index unchanged internally. Subsequently, the feature encoding layer performs one-dimensional or two-dimensional convolution operations, non-linear activation, and optional pooling operations on the time-series features of each branch, extracting shape change patterns, boundary constraint patterns, and trajectory evolution patterns within the local neighborhood. The encoded features from different branches are then concatenated or weighted and fused along the feature dimensions to form a fused feature representation for each time step. The fusion feature representation is achieved by rearranging or reorganizing the feature subspaces corresponding to the tooth tip, tooth belly, and tooth root through the structural association layer. This allows the network to explicitly distinguish different structural segments in the feature tensor. Then, the chain fusion layer concatenates and combines the features according to time and segment order, propagating upstream and downstream information to form a continuous chain relationship between features at adjacent time steps and adjacent segments in the feature space. Upon receiving the chain fusion features, the control adjustment layer reads the adjustment structure matrix corresponding to the current sample from the outside and applies numerical adjustments based on mega-entropy and chain amplitude to each time step and each feature channel, introducing complexity information and inter-segment expansion information into each element of the feature tensor. The processed feature tensor is then fed into the output layer. The output layer constructs a corresponding set of output neurons for each time step, performs a linear transformation on the adjusted feature vector, adds a bias term, and then generates the predicted trajectory node data corresponding to that time step through a predetermined activation function. The predicted trajectory node data may include the node's spatial coordinates, attitude parameters, or predicted values ​​related to the interference state. After performing the above forward propagation operation on all samples and all time steps within a batch, a set of predicted trajectory data that corresponds one-to-one with the positions of the training input sequence in the sample dimension and time dimension is obtained.

[0055] In a preferred embodiment of the present invention, error data is obtained by comparing the predicted trajectory data with the training target sequence and calculating the degree of deviation between the two, including: By matching the predicted trajectory data with the training target sequence on the trajectory node index, aligned prediction data and aligned target data are obtained; The node deviation data is obtained by calculating the numerical difference between each pair of trajectory nodes in the alignment prediction data and the alignment target data. By grouping and summarizing the node deviation data according to the processing order, grouped error data is obtained; By normalizing and aggregating the grouped error data, the errors of each group are converted into the overall deviation, thus obtaining the error data.

[0056] In this embodiment of the invention, by matching the predicted trajectory data with the training target sequence on the trajectory node index, aligned predicted data and aligned target data are obtained, so that the predicted data and target data have a unified comparison benchmark during neural network training. By calculating the numerical difference between each pair of trajectory nodes in the aligned predicted data and aligned target data, node deviation data is obtained, which can clearly record the degree and direction of the difference between the prediction and the target at each trajectory node. By grouping and summarizing the node deviation data according to the processing order, group error data is obtained, which improves the fine-grained error at the node level to the structured error expression at the segment level, making it convenient to process the error according to the importance of different segments during network training. By normalizing and aggregating the group error data, each group error is converted into the overall deviation degree, and error data is obtained, so that the error data becomes a comprehensive indicator reflecting the overall deviation degree of the trajectory, and can be used as a basis for judging whether the training has reached the preset convergence condition, forming a closed-loop error feedback mechanism in the neural network training cycle.

[0057] Specifically, node deviation data is obtained by calculating the numerical difference between each pair of trajectory nodes in the alignment prediction data and the alignment target data, including: First, after completing the one-to-one correspondence matching between the predicted trajectory data and the training target sequence in the trajectory node index dimension, for each trajectory node index, the trajectory nodes in the aligned predicted data and the trajectory nodes in the aligned target data are extracted as a pair of computational objects. Each trajectory node can be expressed in vector form, for example, including coordinate parameters describing the spatial position of the node, angle parameters describing the tool attitude, and process parameters describing the feed state; the system can establish a unified data structure for the above-mentioned parameters, so that the predicted nodes and target nodes under the same index are completely consistent in dimension and parameter type. Subsequently, the system calculates the numerical difference for each component of the predicted node and the target node corresponding to the index, such as the difference of the position coordinate component, the difference of the attitude angle component, and the difference of the feed component, and selects the absolute difference, square difference, or vector distance as the difference measurement method for the component as needed. In one embodiment, the differences of the above components can be weighted according to preset weights and combined into a node deviation scalar corresponding to the trajectory node; alternatively, without weighted synthesis, the differences of each component can be maintained as a multi-dimensional node deviation vector to retain finer-grained error information. The system repeats the above calculation process sequentially for all trajectory node indices, arranging the node deviation data corresponding to each node index according to the processing order of the trajectory nodes and storing it in the node deviation cache, thus forming a node deviation data sequence with the same length as the predicted trajectory. During the node deviation calculation process, obviously abnormal differences can be pruned or limited to prevent numerical overflow from affecting subsequent grouping statistics; different calculation precision or parameter selection strategies can also be set for node deviations in different regions based on the tooth surface region or processing segment to which the node belongs. Through the above processing, each predicted trajectory corresponds to a node-level error sequence, and each element in the sequence records the numerical difference between the predicted node and the target node under a unified trajectory node index, constituting the node deviation data.

[0058] Specifically, by normalizing and aggregating the grouped error data, the errors of each group are converted into the overall deviation, resulting in error data, which includes: First, the errors of each group are normalized. This can be done by linearly mapping each group error to the maximum and minimum values ​​of all group error data, ensuring it falls within a preset standard value range; or by standardizing each group error based on the mean and standard deviation of all group errors, making the error distributions of different groups statistically comparable. During normalization, the system can apply the same or different scaling factors to different groups based on the tooth surface region, machining stage, or trajectory segment length corresponding to the group, to avoid a disproportionate impact on the overall deviation during subsequent aggregation due to a large original value in a certain local area. After normalization, the system combines all group errors according to predetermined aggregation rules. Weighting factors related to the importance, number of nodes, or corresponding structural complexity can be pre-set for each group, and the normalized errors of each group are aggregated using a weighted summation method to obtain error data representing the overall deviation between the entire predicted trajectory and the target trajectory; alternatively, arithmetic mean, geometric mean, or other aggregation functions can be used to compress the multidimensional group error data into a one-dimensional or low-dimensional overall error vector for loss calculation during subsequent training. Finally, the error data after normalization and aggregation is output to the training control module. This error data can be used as input to the loss function of the deep learning model, or as a basis for judging whether the current training round meets the preset convergence condition, thereby forming a stable and calculable global bias metric in the parameter update loop.

[0059] In a preferred embodiment of the present invention, the parameters of the deep learning model are updated based on error data and the adjustment structure matrix until the error data meets a preset convergence condition, thereby obtaining a grinding trajectory model, including: Based on the error data, determine the adjustment parameters in the deep learning model and obtain the parameter data to be updated; The error data is weighted according to the adjustment structure matrix, and different adjustment weights are assigned to the errors of different segments and different regions to obtain weighted error data. The updated parameter data is obtained by performing update calculations based on the weighted error data; The updated parameter data is applied to the deep learning model, and the error calculation and parameter update steps are repeated. When the error data meets the preset convergence condition, the parameter update step is stopped, and the grinding trajectory model is obtained.

[0060] In this embodiment of the invention, adjustment parameters in the deep learning model are determined based on error data to obtain parameter data to be updated. Parameters of different functional layers are recorded separately to enable differentiated updates of parameters at different layers in the future. The error data is weighted according to the adjustment structure matrix, assigning different adjustment weights to errors in different segments and regions to obtain weighted error data. This establishes a direct link between the strength of the error in the parameter space and the importance of the structure and the sensitivity of the segment. The parameter data to be updated is updated based on the weighted error data to obtain updated parameter data, realizing the mapping from structured weighted error to updated parameter data, providing a traceable data foundation for subsequent convergence judgment and parameter tuning records. The updated parameter data is applied to the deep learning model, and the error calculation and parameter update steps are repeated. When the error data meets the preset convergence conditions, the parameter update step is stopped, resulting in a grinding trajectory model. By repeating error calculation and parameter update in each iteration, the model parameters in the data space gradually approach a state that satisfies the distribution of the training samples.

[0061] Specifically, based on the error data, the adjustment parameters in the deep learning model are determined to obtain the parameter data to be updated, including: First, the predicted trajectory data and the corresponding training target sequence are aligned one by one on the trajectory node index. The difference between each pair of trajectory nodes is calculated to form an error data set including position error, direction error, and segment attribution information. Then, based on the network structure of the deep learning model, the error data is back-mapped along the network connections. First, the output layer directly connected to the trajectory node output is located, and the weight matrix and bias vector used to generate trajectory node coordinates, segment labels, or related control quantities in the output layer are marked as first-type adjustable parameters. Next, according to the network topology, the back-connection path of the output layer is traced sequentially to the chained fusion layer, structural association layer, feature encoding layer, and deep convolutional units adjacent to the input layer. The weight matrix, convolution kernel parameters, bias terms, and parameter units interacting with the adjustment structure matrix in these layers that participate in trajectory node calculation are marked as second-type adjustable parameters. During the marking process, a parameter index table is established for each type of adjustable parameter, recording the parameter's layer level, tensor shape, position in the overall parameter vector, and correspondence with the trajectory nodes involved in the error data. Finally, all the parameters marked as adjustable are organized according to hierarchical order, functional category, and data structure to generate parameter data to be updated.

[0062] Specifically, the error data is weighted according to the adjustment structure matrix, assigning different adjustment weights to errors in different segments and regions to obtain weighted error data, which includes: First, based on the tooth surface region to which the trajectory node belongs, each node is labeled as a tooth tip node, tooth belly node, or tooth root node. Simultaneously, based on the local index in the structural feature matrix, each node is further assigned to its corresponding local segment and local region, and a segment identifier, region identifier, and structural index are added to each error value. Then, from the pre-constructed adjustment structure matrix, matrix elements corresponding one-to-one with each trajectory node are extracted according to the aforementioned segment identifier and local region identifier. These matrix elements serve as the adjustment weights for the node's error. The complex amplification coefficient obtained from mega-entropy and the extended balance coefficient obtained from chain amplitude are numerically combined to comprehensively characterize the node's importance in the overall structure and its constraint degree in inter-segment extended relationships. The complex amplification coefficient is first mapped to a weight component sensitive to local structural differences, and the extended balance coefficient is mapped to a weight component sensitive to inter-segment relationships. Then, numerical superposition or function synthesis is performed on the same node to obtain the final adjustment weight for that node. After determining the node-level weights, the basic error value of each trajectory node is multiplied by its corresponding adjustment weight to obtain the node-level weighted error value. Furthermore, in order to match the parameter grouping strategy in the parameter data to be updated, the weighted errors of nodes belonging to the same layer, the same parameter subset, or the same functional module can be aggregated according to the connection relationship between the node and each parameter subset in the network. The hierarchical weighted error vector is formed by summing, averaging, or other aggregation methods, so that each type of parameter subset corresponds to a set of weighted error data containing structural adjustment information, thus forming weighted error data.

[0063] Specifically, the updated parameter data is obtained by updating the parameter data based on the weighted error data, including: First, the weighted error is backpropagated according to the network structure. A matrix operation is performed between the weighted error associated with the output of a certain layer and the activation values ​​of the input features of that layer to calculate the gradient information of the weight matrix and bias vector of that layer. In convolutional layers, the weighted error is backpropagated within the feature map dimension using convolution kernels. A pointwise multiplication operation is performed on each convolution kernel parameter with its corresponding input feature segment, and the results are accumulated to obtain the gradient value of the convolution kernel weight. In fully connected layers or other linear transformation layers, the weighted error vector is multiplied by the input feature vector using inner or outer products to obtain the gradient matrix of the corresponding weight matrix. The gradient value of the bias term is then calculated based on the aggregated result of the weighted errors. During the calculation process, the same or different gradient scaling factors can be used for different parameter subsets to numerically scale the gradient matrix or gradient vector according to a preset step size factor, obtaining the update increment data for each parameter element. Subsequently, these update increments are added or subtracted from the current parameter values ​​in the parameter data to be updated in index order to generate new parameter values, which are then bound to the index information of the original parameters to form the updated parameter data. To facilitate subsequent iterations, the parameter values ​​before and after the update, along with the corresponding gradient magnitudes, can be recorded simultaneously in the updated parameter data. This allows for backtracking analysis of the parameter update trajectory during convergence assessment and debugging. Through these operations, the weight matrix, convolution kernel parameters, and bias vectors of each layer obtain updated results consistent with the weighted error data, forming an updated parameter data set.

[0064] The process involves applying updated parameter data to a deep learning model and repeatedly executing error calculation and parameter update steps. When the error data meets a preset convergence condition, the parameter update step is stopped, resulting in a grinding trajectory model. Specifically, this includes: After completing one round of parameter update calculations, each updated parameter value is first written to the corresponding parameter storage location within the deep learning model according to the parameter index, replacing the original weight matrix, convolution kernel parameters, and bias vector in the current model. This ensures the model structure remains unchanged while the parameter configuration is updated to the latest state. After writing the parameters, a new training input sequence is selected from the training sample set. The basic dataset representing the tooth surface structure, grinding wheel boundary, and historical trajectory is input into the model's input layer, causing the updated model to perform a new round of forward inference and generate new predicted trajectory data. Subsequently, the new predicted trajectory data is compared with the corresponding training target sequence, and the trajectory node-level error value is recalculated. New error data is formed through trajectory node index alignment, grouping, and normalization operations. This error data is then input again into the complete process of determining adjustment parameters based on the error, performing weighted processing using the adjustment structure matrix, and updating parameters based on the weighted error. To determine whether the training process has terminated, convergence conditions are pre-set, such as the overall value of the error data being lower than a certain threshold, the error change amplitude being lower than a certain proportion in consecutive iterations, or the error data no longer changing significantly within a preset number of iterations. After each round of parameter updates and error data recalculation, the current error data is compared with the convergence condition. If the error data does not meet the convergence condition, the next round of parameter update continues. When the error data meets the preset convergence condition, the parameter update step is stopped, and all parameter states of the deep learning model at this time are solidified as the final parameter configuration. The model is then saved as a grinding trajectory model for subsequent grinding trajectory generation, so that the basic dataset can be directly input into the model in practical applications to output grinding trajectory data.

[0065] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for generating a grinding trajectory model for a hobbing grinder based on deep learning, characterized in that, The method includes: Obtain the hob tooth surface structure, grinding wheel boundary dimensions, and historical machining samples to obtain the basic dataset; A multidimensional feature channel is established based on the basic dataset, and features of the tooth surface region, grinding wheel boundary region and historical trajectory region in the basic dataset are fused based on the multidimensional feature channel to obtain the structural feature matrix; Based on the structural feature matrix, the tooth is divided into segments according to the structural features of the tooth tip, tooth vent, and tooth root, and the complexity of the tooth surface structure in the local region is calculated to obtain the mega-entropy. Construct the chain amplitude structure matrix based on the mega-entropy and calculate the degree of extension of the hob between different segments to obtain the chain amplitude. A dual-adjustment structure is established based on the mega-entropy and chain amplitude, and the chain amplitude structure matrix is ​​adjusted and fused according to the dual-adjustment structure to obtain the adjustment structure matrix. A deep learning model is constructed based on a convolutional neural network. The deep learning model includes an input layer, a feature encoding layer, a structural association layer, a chained fusion layer, a control and adjustment layer, and an output layer. A training sample set is constructed based on historical processing samples and their interference markers. The deep learning model is then trained based on the training sample set and the adjustment structure matrix to obtain the grinding trajectory model. Input the basic dataset into the grinding trajectory model to generate grinding trajectory data.

2. The method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 1, characterized in that, A multi-dimensional feature channel is established based on the basic dataset, and features from the tooth surface region, grinding wheel boundary region, and historical trajectory region in the basic dataset are fused using the multi-dimensional feature channel to obtain a structural feature matrix, including: By mapping tooth surface structure data, grinding wheel boundary size data, and historical trajectory sample data to tooth surface region dataset, boundary region dataset, and trajectory region dataset, respectively, region classification data is obtained; Based on the regional classification data, the tooth surface region dataset, boundary region dataset, and trajectory region dataset are mapped to the first feature channel, the second feature channel, and the third feature channel to obtain channel mapping data. Based on the channel mapping data, the data in the three types of feature channels are aligned according to the structure index order to ensure that the data at the same index position form a corresponding relationship, thus obtaining the aligned channel data; Based on the alignment channel data, the data in the three feature channels at each index position are combined to obtain the combined feature data; Based on the combined feature data, the combined features at different index positions are arranged according to the row and column correspondence to obtain the structural feature matrix.

3. The method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 2, characterized in that, Based on the structural feature matrix, segments are divided according to the structural features of the tooth tip, tooth vent, and tooth root. The complexity of the tooth surface structure in local regions is calculated to obtain the mega-entropy, including: By extracting the data from the structural feature matrix according to the positions of the tooth tip, tooth vent, and tooth root, and dividing them into several local segments, segment division data is obtained. Based on the segment division data, the strength of the change in tooth surface structure in each local segment is calculated to obtain the local difference term; the overall complexity within different local segments is calculated to obtain the regional complexity term; the overall change intensity of complexity between each local segment is calculated to obtain the regional jump term; and the degree of correlation between different tooth surface structures in the same local segment is calculated to obtain the segment correlation term. By fusing the local difference term, the regional complexity term, the regional jump term, and the segment correlation term, we obtain the mega-entropy quantity.

4. The method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 3, characterized in that, Construct the chain amplitude structure matrix based on the mega-entropy and calculate the extent of the hob's extension between different segments to obtain the chain amplitude, including: Based on the segment division data and mega-entropy, the characteristic changes between different local segments are combined in pairs, and the expansion relationship between each segment pair is sequentially filled into the data array to obtain the chain amplitude structure matrix. Based on the chain length structure matrix, the degree of cumulative difference between adjacent segments is calculated, the expansion trend caused by local structural differences between different segments of the hob is identified, and the inter-segment expansion term is obtained; the relative position of each segment in the direction from tooth tip to tooth root is calculated, the degree of influence of different segment positions on the overall expansion is identified, and the position modulation term is obtained. The overall complexity is calculated to amplify or compress the expansion results of each segment, resulting in a complexity coupling term. The degree of mutual influence between expansion results of different segments is calculated, resulting in a chain amplitude coupling term. The inter-segment expansion term, position modulation term, complexity coupling term, and chain amplitude coupling term are fused together to obtain the chain amplitude quantity.

5. The method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 4, characterized in that, A dual-adjustment structure is established based on the mega-entropy and chain amplitude, and the chain amplitude structure matrix is ​​adjusted and fused according to the dual-adjustment structure to obtain the adjustment structure matrix, including: By converting mega-entropy into complex amplification factors and chain amplitude into extended balance factors, the adjustment data is obtained; Based on the adjustment data, a complex amplification coefficient and an extended balance coefficient are simultaneously superimposed numerically to ensure that the two form a corresponding relationship within the same adjustment framework, thus obtaining dual-adjustment structure data. Based on the dual-adjustment structure data, the adjustment effect of complex amplification coefficient and extended balance coefficient is applied simultaneously to each matrix element of the chain axis structure matrix to obtain the correction matrix data; Based on the correction matrix data, the data related to the tooth tip, tooth vent, and tooth root are summarized and arranged in the processing order to obtain the adjustment structure matrix.

6. The method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 5, characterized in that, Building deep learning models based on convolutional neural networks includes: Construct an input layer to receive the basic dataset, divide the basic dataset into tooth surface input data, boundary input data, and trajectory input data, and align them to obtain the input data; A feature coding layer is constructed to extract features of tooth surface shape change, grinding wheel shape constraint, and historical motion trajectory based on the input data, and combine them into fusion features to obtain feature coding data; A structural association layer is constructed to divide the fused features into tooth crest partition features, tooth ventral partition features, and tooth root partition features based on the feature encoding data, and arrange them into structural association data.

7. The method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 6, characterized in that, Building deep learning models based on convolutional neural networks also includes: A chain-like fusion layer is constructed to combine the structural features corresponding to the tooth tip, tooth vent, and tooth root into multiple chain-like combinations according to the segment order based on the structural association data, thereby obtaining chain-like fused data; A control and adjustment layer is constructed to adjust the chained fusion data based on mega-entropy and chain amplitude. Mega-entropy is used to assign weights to each segment, and chain amplitude is used to limit the expansion range of each segment, thus obtaining control and adjustment data. An output layer is constructed to extract nodes of the motion trajectory based on the control and adjustment data, and sort the nodes according to the processing order to obtain the trajectory node output data.

8. The method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 7, characterized in that, A training sample set is constructed based on historical processing samples and their interference markers. The deep learning model is then trained using the training sample set and the adjustment structure matrix to obtain the grinding trajectory model, including: By matching each historical processed sample with its corresponding interference marker, labeled sample data is obtained; Based on the labeled sample data, the continuous processing records are divided into training input sequences, and the corresponding interference marks are organized into training target sequences to obtain the training sample set; Predicted trajectory data is obtained by feeding the training input sequence into a deep learning model; Error data is obtained by comparing the predicted trajectory data with the training target sequence and calculating the degree of deviation between the two. Based on the error data and the adjustment structure matrix, the parameters of the deep learning model are updated until the error data meets the preset convergence condition, thus obtaining the grinding trajectory model.

9. The method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 8, characterized in that, Error data is obtained by comparing the predicted trajectory data with the training target sequence and calculating the degree of deviation between the two, including: By matching the predicted trajectory data with the training target sequence on the trajectory node index, aligned prediction data and aligned target data are obtained; The node deviation data is obtained by calculating the numerical difference between each pair of trajectory nodes in the alignment prediction data and the alignment target data. By grouping and summarizing the node deviation data according to the processing order, grouped error data is obtained; By normalizing and aggregating the grouped error data, the errors of each group are converted into the overall deviation, thus obtaining the error data.

10. A method for generating a grinding trajectory model for a hobbing grinder based on deep learning according to claim 9, characterized in that, Based on the error data and the adjustment structure matrix, the parameters of the deep learning model are updated until the error data meets the preset convergence condition, resulting in a grinding trajectory model, including: Based on the error data, determine the adjustment parameters in the deep learning model and obtain the parameter data to be updated; The error data is weighted according to the adjustment structure matrix, and different adjustment weights are assigned to the errors of different segments and different regions to obtain weighted error data. The updated parameter data is obtained by performing update calculations based on the weighted error data; The updated parameter data is applied to the deep learning model, and the error calculation and parameter update steps are repeated. When the error data meets the preset convergence condition, the parameter update step is stopped, and the grinding trajectory model is obtained.

Images