Deep learning-based processing surface roughness prediction method and system
By constructing a cross-fusion prediction model of deep learning, and utilizing the robot's internal joint torque and external vibration signals, the problem of predicting the surface roughness of robot machining was solved, and high-precision surface roughness prediction was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2025-05-26
- Publication Date
- 2026-06-23
AI Technical Summary
In robotic machining, traditional methods struggle to effectively predict surface roughness, and adding sensors increases costs.
A deep learning-based cross-fusion prediction model is constructed, which uses the robot's internal joint torque and external vibration signals to predict the surface roughness of the machined surface through feature dimensionality reduction, multi-channel feature enhancement, and cross-fusion modules.
It achieves high-precision prediction of surface roughness of robot processing under the condition of a small number of sensors, and adapts to the changes in surface roughness of processing under different postures and parameters.
Smart Images
Figure CN120493207B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and system for predicting the roughness of machined surfaces based on deep learning, belonging to the technical field of robotic machining. Background Technology
[0002] Robots are increasingly used in machining tasks due to their high flexibility and low cost. However, because of their low stiffness, robots are prone to significant vibrations during machining, which negatively impacts the surface roughness. Surface roughness is a key parameter for evaluating the morphology of machined surfaces. A significant correlation exists between surface roughness and machining parameters. Notably, due to the nonlinearity of robot performance, robots may exhibit different levels of surface roughness under the same machining parameters, making the prediction of surface roughness in robot-machined surfaces extremely challenging.
[0003] Traditional surface roughness prediction models consider the influence of various physical factors, such as tool runout and vibration. However, the factors affecting surface roughness are complex, with many mechanisms that are difficult to describe and model. Therefore, data-driven methods based on signal characteristics have been used in the task of predicting surface roughness. Patent CN119618146A discloses a surface roughness monitoring method with an embedded physical prediction model. This method trains the surface roughness prediction model by inputting theoretical cutting force, current signal, and machining parameters, and calculates the surface roughness value of the machined surface in real time. Patent CN119260492A proposes an on-machine prediction method and device for surface roughness in cylindrical grinding. This method acquires grinding images from any process, inputs them into an improved Swing Transformer model for prediction, and obtains the prediction results. These methods achieve surface roughness prediction by establishing an end-to-end module. However, robotic machining processes are particularly complex, making data-driven methods especially suitable for such complex tasks.
[0004] Data-driven approaches rely on inputs that effectively reflect data characteristics. In robotic machining tasks, machining status information can be reflected through vibration signals. To obtain more comprehensive and complementary information, more sensors are needed to complement the feature information; however, adding more sensors increases costs. Summary of the Invention
[0005] To address the challenge of improving the prediction of machined surface roughness while minimizing the use of sensors, this invention provides a deep learning-based method and system for predicting machined surface roughness.
[0006] The present invention provides a deep learning-based method for predicting the roughness of machined surfaces, comprising:
[0007] S1. Construct a dataset. The input data consists of the robot's internal joint torque and external vibration signals. The output data is the surface roughness of the machined surface corresponding to the input data.
[0008] S2. Construct a cross-fusion prediction model based on a deep learning network. The cross-fusion prediction model includes a feature dimensionality reduction module by channel, a multi-channel feature enhancement module, and a fusion module.
[0009] The robot's internal six-channel joint torque and external three-channel vibration signals are input to a channel-based feature reduction module. This module performs feature reduction on the internal six-channel joint torque and external three-channel vibration signals to obtain low-dimensional features for each channel. These low-dimensional features are then input to a parallel multi-channel feature enhancement module. This module performs dynamic interaction between the low-dimensional features of each channel to obtain multi-channel torque enhancement features and multi-channel vibration enhancement features corresponding to the internal six-channel joint torque and external three-channel vibration signals, respectively. These features are then input to a fusion module. This fusion module performs bidirectional fusion on the input multi-channel torque enhancement features and multi-channel vibration enhancement features to obtain fused features. The surface roughness of the machined surface is then predicted based on these fused features.
[0010] S3. Use the dataset to train the cross-fusion prediction model, and use the trained cross-fusion prediction model to predict the surface roughness of the machined surface.
[0011] Preferably, the mutual fusion module is implemented based on a dual multi-head cross-attention mechanism, which calculates the cross-modal attention vector between the multi-channel torque enhancement feature and the multi-channel vibration enhancement feature, and then fuses the obtained cross-modal attention vector with the multi-channel torque enhancement feature and the multi-channel vibration enhancement feature to obtain the fused feature.
[0012] Preferably, methods for calculating the cross-modal attention vector of the interaction between multi-channel torque enhancement features and multi-channel vibration enhancement features include:
[0013] The multi-channel torque enhancement feature and the multi-channel vibration enhancement feature are C1 and C2, respectively, and their corresponding cross-modal attention vectors are:
[0014]
[0015] W O Represents the trainable parameter matrix;
[0016] Concat() means concatenation. This represents the scaled dot product attention feature obtained by passing the cross-modal features from C1 to C2 through the m-th attention head. This represents the scaled dot product attention feature obtained by passing the m-th attention head through the cross-modal features from C2 to C1, where m = 1, ..., h, and h represents the number of attention heads. Let represent the trainable parameter matrix of the m-th attention head, and Softmax represent the activation function used to convert similarity into a probability distribution. These are the features assigned to the query vector Query, the key vector Key, and the value vector Value, respectively, with indices i = 1, 2, and d. k This represents the dimension of the key vector Key.
[0017] Preferably, the feature reduction module performs feature reduction on the internal joint torques of the robot's six channels and the vibration signals of the external three channels to obtain the low-dimensional features of each channel. The methods include:
[0018] A dataset X = [x1, x2, ..., x3] is constructed based on the internal joint torques or external vibration signals of each channel of the robot. n ], where the input feature x i i = 1, 2, ..., n, where n represents the number of input features in the dataset;
[0019] Construct a directed weighted graph based on dataset X.
[0020] According to the directed weighted graph Construct an undirected weighted graph G;
[0021] By effectively projecting the undirected weighted graph G, a low-dimensional feature map of the undirected weighted graph is obtained by minimizing the cross-entropy function. Based on the low-dimensional feature map, low-dimensional features of multi-channel internal joint torque or multi-channel external vibration signal are obtained.
[0022] As a preferred option, input feature x i satisfy:
[0023]
[0024] ρ i =min{d(x i x ij )|1≤j≤k,d(x i x ij )>0}
[0025] Where, x ij x represents i The j-th nearest neighbor, d(x) i x ij ) represents x i and x ij The distance between them, ρ i This represents a distance-indexed local connectivity constraint, σ.i Let x represent the local Riemannian metric at each point, and k represent x. i The number of neighbors under the metric.
[0026] As a preferred approach, a directed weighted graph is constructed based on the dataset X. The methods include:
[0027] Construct a directed weighted graph based on dataset X. V represents the simplified set of X, and E′ represents the directed edge formed: E′={(x i ,x ij The edge weights ω are set as follows: |1≤i≤n,1≤j≤k}
[0028]
[0029] ρ i =min{d(x i x ij )|1≤j≤k,d(x i x ij )>0}.
[0030] As a preferred option, based on the directed weighted graph Construct an undirected weighted graph G:
[0031] Obtaining a Directed Weighted Graph Given a weighted adjacency matrix A, obtain the symbolic metric B based on the weighted adjacency matrix A;
[0032]
[0033] in, The elements A of the weighted adjacency matrix A represent the Hadamard product. ij Indicates from x i To x j The probability of the existence of a directed edge in B is denoted by the symbolic measure of the element B. ij Let G represent the probability that at least one of the two directed edges exists. Construct an undirected weighted graph G based on the symbolic metric.
[0034] As a preferred option, the cross-entropy function is:
[0035]
[0036] Where, ω h (e) and ω l (e) represents two fuzzy sets: an undirected weighted graph and a low-dimensional feature graph, respectively, where e represents the weight of the edge.
[0037] As a preferred embodiment, the modular multi-channel feature enhancement module includes two branch structures, one of which is configured with a 6×2 convolutional kernel to achieve cross-channel feature enhancement between the six channels inside the robot.
[0038] Another branch structure is configured with 3×2 convolutional kernels to achieve cross-channel feature enhancement between the three channels outside the robot.
[0039] The beneficial effects of this invention are as follows: The cross-fusion prediction model integrates the robot's internal joint torque signals and external vibration signals, realizing the interaction of signal features during the fusion process and better refining the processing state information. This invention reduces the dimensionality of features by channel to minimize interference from redundant features. The dimensionality-reduced features are constructed into a bi-branch structure, and dynamic interaction between features from different channels is achieved through a multi-channel feature enhancement module. Based on a dual multi-head cross-attention mechanism, this invention achieves collaborative interaction of information across modalities, thereby completing a bidirectional deep collaborative representation between the robot's internal and external signal features during the fusion process. Finally, by segmenting and aggregating the features, the surface roughness of the robot's machining is accurately predicted. By combining the robot's internal and external signals to predict the surface roughness of the machined surface, prediction of surface roughness under different postures and processing parameters can be achieved with high prediction accuracy. Attached Figure Description
[0040] Figure 1 This is a schematic diagram of the internal and external signal acquisition system during the robot processing of the present invention;
[0041] Figure 2 This is a schematic diagram of the structure of the cross-fusion prediction model in an embodiment of the present invention;
[0042] Figure 3 This is a schematic diagram of the structure of the multi-channel feature enhancement module in an example of the present invention;
[0043] Figure 4 This is a schematic diagram of the structure of the fusion module in an embodiment of the present invention;
[0044] Figure 5 This is a roughness prediction effect diagram of the cross-fusion prediction model in the embodiment of the present invention. Detailed Implementation
[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0046] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.
[0047] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.
[0048] To achieve accurate prediction of surface roughness in robot machining under different machining poses and parameters, this embodiment provides a surface roughness prediction method based on internal and external signal fusion and deep learning. By combining the robot's internal and external signals, the surface roughness of the robot machining is predicted, enabling prediction of surface roughness under different poses and machining parameters with high prediction accuracy. The deep learning-based surface roughness prediction method of this embodiment includes:
[0049] Step 1: Construct a dataset. The input data consists of the robot's internal joint torques and external vibration signals. The output data is the surface roughness of the machined surfaces corresponding to the input data.
[0050] A dataset of machined surface roughness was obtained through robot machining experiments with different workpiece placements and machining parameters. Simultaneously, torque signals from six internal channels and vibration signals from three external channels were acquired. Internal and external signal characteristics of the robot were extracted by channel in the time domain, frequency domain, and time-frequency domain.
[0051] Specifically, to acquire both internal and external signals from the robot, this embodiment obtains a dataset of machined surface roughness through robot machining experiments with different workpiece placements and machining parameters. This embodiment uses a KR500-MT industrial robot equipped with an electric spindle for machining operations. The tool holder is an HSK model HSK63A-SDC8-100L. During machining, joint torque signals from six internal channels and vibration signals from three external channels are simultaneously acquired. The robot joint torque signals are obtained through an interface in the robot controller, while a three-axis accelerometer is mounted on the electric spindle side to measure the vibration signals.
[0052] The surface roughness of robot-machined surfaces is closely related to the robot's posture and machining parameters. Considering this, machining experiments with different parameters were conducted at six different workpiece placement positions. The workpiece material was Al 6061, and the workpiece dimensions were 140×140×40mm. At the same workpiece placement position, multiple sets of machining experiments were conducted with varying spindle speed, feed rate, and axial depth of cut. Workpiece surface roughness information was collected using a TR200 surface roughness measuring instrument. After completing a certain axial depth of cut, the surface roughness of the stable machining section was measured three times using the surface roughness measuring instrument. During the measurement process, the workpiece was not disassembled to ensure in-situ measurement of the machined surface roughness. Figure 2 This is a schematic diagram of the internal and external signal acquisition system during the robot's processing.
[0053] The combined internal and external signals of the robot characterize the processing state, and extracting discriminative features from continuous signals is a prerequisite for accurate prediction. This implementation method performs feature extraction through three-domain analysis: time domain, frequency domain, and time-frequency domain. The extraction process for internal and external robot signals is performed channel-wise. The six channels of internal robot signals consist of six torque signals in the joint space, while the three channels of external robot signals consist of three directional vibration signals in Cartesian space. Before signal processing begins, the incoming and outgoing signal segments are truncated, and only the intermediate stable segments of the signal are extracted. Furthermore, the start and end times of interception are calculated based on the roughness gauge's measurement length and signal sampling time. After processing, a total of 9×16 sets of time-domain features, 9×13 sets of frequency-domain features, and 9×16 sets of time-frequency signal features are extracted. A total of 6×45 sets of features are extracted from internal signals, and a total of 3×45 sets of features are extracted from external signals.
[0054] Step 2: Construct a cross-fusion prediction model based on deep learning modules, such as... Figure 2 As shown, it includes a feature dimensionality reduction module based on channels, a multi-channel feature enhancement module, and a mutual fusion module:
[0055] Step 21: The robot's internal six-channel joint torque and external three-channel vibration signals are input to the channel-based feature reduction module. The feature reduction module performs feature reduction on the internal six-channel joint torque and external three-channel vibration signals to obtain the low-dimensional features of each channel, which are then input to the parallel multi-channel feature enhancement module.
[0056] In robotic machining, a multi-sensor system is used to acquire signals, and complementary machining state responses can be derived through feature extraction. However, the extracted sensor signal features exhibit complex coupling relationships, and the presence of too many redundant features may reduce the model's learning ability. To address this issue, dimensionality reduction techniques can reduce data dimensionality while preserving inherent structural information, thereby reducing redundancy between signal features. The feature dimensionality reduction module in this implementation reduces redundant dimensions before feature fusion. It includes the following steps:
[0057] Step 211: Construct a dataset X = [x1, x2, ..., x] based on the internal joint torques or external vibration signals of each channel of the robot. n ], where the input feature x i i = 1, 2, ..., n, where n represents the number of input features in the dataset;
[0058] k represents the standard, therefore, the set is for each x i The closest distance ρ i and smoothing normalization factor σ i Calculated. Combined with the nearest neighbor descent method, it can be expressed as: ρ i =min{d(x i x ij )|1≤j≤k,d(x i x ij )>0}
[0059] Where, x ij x represents i The j-th nearest neighbor, d(x) i x ij ) represents x i and x ij The distance between them, ρ i This represents a distance-indexed local connectivity constraint that ensures x i Connect to at least one edge with weight 1. k represents x. i The number of neighbors under the metric.
[0060] Smoothing normalization factor σ i The local Riemann metric for the i-th input feature satisfies:
[0061]
[0062] Step 212: Construct a directed weighted graph based on dataset X.
[0063] A directed weighted graph was constructed while ensuring data connectivity. Where V represents the simplified set of X. E′ represents the directed edge formed E′={(x i,x ij The edge weights ω are set as follows: |1≤i≤n,1≤j≤k}
[0064]
[0065] Where ω((x) i ,x ij () represents the edge weight, which is the probability that a given edge exists. Directed weighted graph. It is asymmetric, meaning there may be two edges with unequal weights between two points. This construction method is used to capture the topological structure of the data to form a local graph, and then the local graphs are pieced together according to different data sets to form a global representation of the manifold.
[0066] Step 213: Based on the directed weighted graph Construct an undirected weighted graph G:
[0067] Let A be The weighted adjacency matrix represents the relationships between nodes in the graph, taking into account the symbolic metric B;
[0068]
[0069] in The elements A of the weighted adjacency matrix A represent the Hadamard (or pointwise) product. ij Indicates from x i To x j The probability of the existence of a directed edge in B is denoted by the symbolic measure of the element B. ij Let G represent the probability that at least one of the two directed edges exists. Therefore, the high-dimensional graph representing G forms an undirected weighted graph given by the symbolic metric B, which approximates the assumed local connectivity and uniform distribution of the original manifold, capturing the underlying structure of the data.
[0070] Step 214: Perform effective projection on the undirected weighted graph G, and obtain the low-dimensional feature map of the undirected weighted graph by minimizing the cross-entropy function. Based on the low-dimensional feature map, obtain the low-dimensional features of the joint torque of the six internal channels or the vibration signal of the three external channels:
[0071] Based on graph G, an effective projection is performed to calculate the low-dimensional layout. This part optimizes the cross-entropy between the weighted edges of the undirected weighted graph and the low-dimensional feature map, making the low-dimensional feature map composed of these points as close as possible to the undirected weighted graph. The cross-entropy function can be expressed as:
[0072]
[0073] Where CE(ω) h ,ω l It can measure two probability distributions, ω h (e) and ω l(e) represents two fuzzy sets, an undirected weighted graph and a low-dimensional feature graph, respectively, ω h (e) and ω l The difference information between (e) represents the similarity between the high-dimensional space and the transformed-dimensional space, respectively, where e represents the edge weight. This formula represents the computational principle of the force-directed graph layout algorithm. The first term represents the attractive force between point spans, and therefore has a greater weight in high-dimensional conditions because the distance between points should be as small as possible. The second term represents the repulsive force between the two ends, which should also be as small as possible. Finally, the optimal weight of the edges in the low-dimensional space is found by minimizing the cross-entropy function. This process, combined with stochastic gradient descent to optimize the mapping, yields the low-dimensional data representation, N. D This represents the data dimension after dimensionality reduction. Finally, low-dimensional features of multi-channel internal joint torque or multi-channel external vibration signals are obtained based on the low-dimensional feature map.
[0074] Step 22: The multi-channel feature enhancement module performs dynamic interaction between different channel features on the low-dimensional features of the corresponding channels to obtain the multi-channel torque enhancement features and multi-channel vibration enhancement features corresponding to the joint torque of the six internal channels and the vibration signals of the three external channels, respectively, and simultaneously inputs them into the mutual fusion module.
[0075] Dimensionality reduction techniques provide a low-dimensional, efficient, and information-preserving feature representation. However, channel-level processing mechanisms have inherent limitations, failing to effectively utilize the dynamic correlations and potential complementary information between cross-channel features. To address this issue, this implementation proposes a parallel multi-channel feature enhancement module that achieves complementarity of feature information from different channels through a dynamic feature interaction mechanism based on cross-channel convolution kernels. The input is N obtained after dimensionality reduction. D The module addresses the quantitative characteristics by designing cross-channel convolutional kernels with differentiated sensing regions to adapt to the heterogeneous characteristics of the robot's internal and external signal channels. Then, based on the spatial distribution characteristics of the internal and external signal channels, a parallel dual-branch convolutional submodule is constructed to establish a nonlinear mapping relationship between channels in the depth dimension. This implementation's multi-channel feature enhancement module includes two branch structures, such as... Figure 3 As shown, one branch structure is configured with a 6×2 convolutional kernel to achieve cross-channel feature enhancement between the six channels inside the robot; the other branch structure is configured with a 3×2 convolutional kernel to achieve cross-channel feature enhancement between the three channels outside the robot. This enables dynamic interaction between features from different channels, allowing features to gradually develop during training, thereby enhancing the expressive power of information.
[0076] Step 23: The mutual fusion module performs bidirectional fusion of the input multi-channel torque enhancement features and multi-channel vibration enhancement features to obtain fused features, and predicts the surface roughness of the machined surface based on the fused features;
[0077] To achieve information coordination and interaction during the fusion process, this invention implements a mutual fusion module based on a dual multi-head cross-attention mechanism. This module calculates the attention weights between two different sequences to simulate intramodal relationships. In this invention, these two sequences are represented as internal signal features of the robot. and external signal characteristics These features are assigned to the Query(Q), Key(K), and Value(V) components of the attention mechanism to achieve feature fusion. The fusion process consists of two parts: (1) Internal to external feature fusion (C1→C2): C2 is assigned as K and V, while C1 is assigned as Q. (2) External to internal feature fusion (C2→C1): C1 is assigned as K and V, while C2 is assigned as Q. By fusing these two parts to better complete the feature state information and realize the interaction between features, the mutual fusion module of this embodiment is based on the dual multi-head cross-attention mechanism. It calculates the cross-modal attention vector of the interaction between the multi-channel torque enhancement feature and the multi-channel vibration enhancement feature, and then fuses the obtained cross-modal attention vector with the multi-channel torque enhancement feature and the multi-channel vibration enhancement feature to obtain the fused feature. Specifically, the method for calculating the cross-modal attention vector of the interaction between the multi-channel torque enhancement feature and the multi-channel vibration enhancement feature includes:
[0078] The multi-channel torque enhancement feature and the multi-channel vibration enhancement feature are C1 and C2, respectively, and their corresponding cross-modal attention vectors are:
[0079]
[0080] W O Represents the trainable parameter matrix;
[0081] Concat() means concatenation. This represents the scaled dot product attention feature obtained by passing the cross-modal features from C1 to C2 through the m-th attention head. This represents the scaled dot product attention feature obtained by passing the m-th attention head through the cross-modal features from C2 to C1, where m = 1, ..., h, and h represents the number of attention heads. Let represent the trainable parameter matrix of the m-th attention head, and Softmax represent the activation function used to convert similarity into a probability distribution. These are the features assigned to the query vector Query, the key vector Key, and the value vector Value, respectively, with indices i = 1, 2, and d. k This represents the dimension of the key vector Key.
[0082] For the output cross-modal attention vector and This module chooses to add these two parts to the previous enhanced features, and then aggregate them together in the depth direction to form the final fused features.
[0083] Step 3: Train the cross-fusion prediction model using the dataset, and use the trained cross-fusion prediction model to predict the surface roughness of the machined surface.
[0084] Step 3 involves cross-referencing and selecting hyperparameters for the prediction models during training: 9 channels for dimensionality reduction, 6 batch sizes, and 300 epochs. The training and test set ratio is 8:2.
[0085] This embodiment also provides a deep learning-based surface roughness prediction system, including a storage device, a processor, and a computer program stored in the storage device and executable on the processor. The processor executes the computer program to implement the steps of the deep learning-based surface roughness prediction method described above.
[0086] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.
Claims
1. A deep learning-based method for predicting the roughness of machined surfaces, characterized in that, include: S1. Construct a dataset. The input data consists of the robot's internal joint torque and external vibration signals. The output data is the surface roughness of the machined surface corresponding to the input data. S2. Construct a cross-fusion prediction model based on a deep learning network. The cross-fusion prediction model includes a feature dimensionality reduction module by channel, a multi-channel feature enhancement module, and a fusion module. The robot's internal six-channel joint torque and external three-channel vibration signals are input to a channel-based feature reduction module. This module performs feature reduction on the internal six-channel joint torque and external three-channel vibration signals to obtain low-dimensional features for each channel. These low-dimensional features are then input to a parallel multi-channel feature enhancement module. This module performs dynamic interaction between the low-dimensional features of each channel to obtain multi-channel torque enhancement features and multi-channel vibration enhancement features corresponding to the internal six-channel joint torque and external three-channel vibration signals, respectively. These features are then input to a fusion module. This fusion module performs bidirectional fusion on the input multi-channel torque enhancement features and multi-channel vibration enhancement features to obtain fused features. The surface roughness of the machined surface is then predicted based on these fused features. The feature reduction module performs feature reduction on the internal joint torques of the robot's six channels and the vibration signals of the external three channels to obtain the low-dimensional features of each channel. Methods include: A dataset is constructed based on the internal joint torques or external vibration signals of each channel of the robot. Input features , , This indicates the number of input features in the dataset; Construct a directed weighted graph based on dataset X. ; According to the directed weighted graph Construct an undirected weighted graph G; By effectively projecting the undirected weighted graph G, the low-dimensional feature map of the undirected weighted graph is obtained by minimizing the cross-entropy function. Based on the low-dimensional feature map, the low-dimensional features of the multi-channel internal joint torque or multi-channel external vibration signal are obtained. Input features satisfy: ； ； in, x represents i The j-th nearest neighbor, x represents i and The distance between them This represents a local connectivity constraint based on a distance index. Let x represent the local Riemannian metric at each point, and k represent x. i The number of neighbors it has under the metric; Construct a directed weighted graph based on dataset X. The methods include: Construct a directed weighted graph based on dataset X. (V, , V represents the simplified set of X. Represents the formed directed edges Edge weights Set to: ； ； According to the directed weighted graph Construct an undirected weighted graph G: Obtaining a Directed Weighted Graph (V, , The weighted adjacency matrix A is used to obtain the symbolic metric B. ； in, The elements A of the weighted adjacency matrix A represent the Hadamard product. ij Indicates from x i To x j The probability of the existence of a directed edge in B is denoted by the symbolic measure of the element B. ij Let G represent the probability that at least one of the two directed edges exists. Construct an undirected weighted graph G based on the symbolic metric. S3. Use the dataset to train the cross-fusion prediction model, and use the trained cross-fusion prediction model to predict the surface roughness of the machined surface.
2. The deep learning-based method for predicting the roughness of machined surfaces according to claim 1, characterized in that, The mutual fusion module is implemented based on a dual multi-head cross-attention mechanism. It calculates the cross-modal attention vector between the multi-channel torque enhancement feature and the multi-channel vibration enhancement feature, and then fuses the obtained cross-modal attention vector with the multi-channel torque enhancement feature and the multi-channel vibration enhancement feature to obtain the fused feature.
3. The deep learning-based method for predicting the roughness of machined surfaces according to claim 2, characterized in that, Methods for calculating the cross-modal attention vector of the interaction between multi-channel torque enhancement features and multi-channel vibration enhancement features include: The multi-channel torque enhancement feature and the multi-channel vibration enhancement feature are C1 and C2, respectively, and their corresponding cross-modal attention vectors are: ； ； ； ； Represents the trainable parameter matrix; Indicates splicing, This represents the scaled dot product attention feature obtained by passing the cross-modal features from C1 to C2 through the m-th attention head. This represents the scaled dot product attention feature obtained by passing the cross-modal features from C2 to C1 through the m-th attention head. h represents the number of attention heads; , , Let represent the trainable parameter matrix of the m-th attention head, and Softmax represent the activation function used to convert similarity into a probability distribution. , , These are the features assigned to the query vector Query, the key vector Key, and the value vector Value, respectively, with indices i=1,2. This represents the dimension of the key vector Key.
4. The deep learning-based method for predicting the roughness of machined surfaces according to claim 1, characterized in that, The cross-entropy function is: ； in, and Let represent two fuzzy sets: an undirected weighted graph and a low-dimensional feature graph, respectively, and let e represent the weight of the edge.
5. The deep learning-based method for predicting the roughness of machined surfaces according to claim 1, characterized in that, The multi-channel feature enhancement module includes two branch structures, one of which is configured with a 6×2 convolutional kernel to achieve cross-channel feature enhancement between the six channels inside the robot; Another branch structure is configured with 3×2 convolutional kernels to achieve cross-channel feature enhancement between the three channels outside the robot.
6. A deep learning-based surface roughness prediction device, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that, The processor executes the computer program to implement the steps of the deep learning-based method for predicting the roughness of machined surfaces as described in any one of claims 1 to 5.