Intelligent design and manufacturing method and system for gear and spline machining

Abstract

The invention discloses an intelligent design and manufacturing method and system for gear and spline machining, and the method comprises the steps: building a hybrid intelligent model which comprises a large language model and a local small model; the hybrid intelligent model is used for designing technological parameters of the gear and the spline; constructing a digital twin system, and realizing real-time synchronization of the virtual model and the entity equipment; in a digital twinning system, optimizing technological parameters for machining gears and splines by utilizing a reinforcement learning technology; processing is carried out based on the optimized technological parameters; and performing quality detection by applying a machine learning technology. The invention further discloses a corresponding system. According to the method and system, advanced technologies such as artificial intelligence, digital twinning and machine learning are integrated, intelligent design and manufacturing of gear and spline machining are achieved, and the machining efficiency and the product quality are improved.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical processing technology, and specifically relates to an intelligent design and manufacturing method and system for gear and spline processing using artificial intelligence and digital twin technology. Background Art

[0002] Gears and splines are key components in mechanical transmission systems. Their processing accuracy and efficiency directly affect the performance and reliability of mechanical equipment. Traditional gear and spline processing methods mainly rely on manual experience to design and adjust process parameters, which has the following problems:

[0003] Low efficiency in process parameter design: Traditional methods require engineers to manually design process parameters based on experience, which is time-consuming and prone to errors.

[0004] Difficulty in optimizing the machining process: Due to the lack of real-time monitoring and intelligent optimization methods, it is difficult to dynamically adjust the process parameters during the machining process, resulting in difficulty in effectively improving the machining efficiency and quality.

[0005] Outdated quality inspection methods: Traditional quality inspections mostly use offline sampling methods, which cannot achieve real-time and comprehensive quality monitoring and easily result in the production of batches of unqualified products. Summary of the invention

[0006] In view of this, the present application provides an intelligent design and manufacturing method and system for gear and spline processing, which solves the problems in the prior art of low efficiency in process parameter design, difficulty in optimizing the processing process, lagging quality inspection methods, poor connection between design and manufacturing, and low data utilization.

[0007] The embodiment of the present application provides a method for intelligent design and manufacturing of gear and spline processing, including:

[0008] Establishing a hybrid intelligent model, wherein the hybrid intelligent model includes a large language model and a localized small model;

[0009] Designing process parameters of gears and splines using the hybrid intelligent model;

[0010] Build a digital twin system to achieve real-time synchronization between virtual models and physical devices;

[0011] In the digital twin system, the process parameters for machining the gears and splines are optimized using reinforcement learning technology;

[0012] Processing the gear and the spline based on the optimized process parameters of the gear and the spline;

[0013] Machine learning techniques are applied to perform quality inspection of the gears and splines.

[0014] The hybrid intelligent model is established, comprising:

[0015] Choose a large language model;

[0016] By using knowledge distillation technology, the large language model is distilled into a localized small model;

[0017] Alignment of the large language model and the localized small model is achieved.

[0018] The step of realizing the alignment of the large language model and the localized small model comprises the following steps:

[0019] Establishing a dynamic system model, and using a state space model to describe the knowledge conversion process from the large language model to the localized small model;

[0020] Gumbel-Softmax and direct estimator are used to deal with the discreteness problem in the knowledge conversion process;

[0021] Use multi-head attention mechanism and graph neural network to deal with the semantic coupling problem in the process of knowledge conversion;

[0022] Apply deep residual networks and nonlinear activation functions to handle nonlinear problems in the knowledge conversion process;

[0023] Use autoencoders and sparse attention mechanisms to handle the high-dimensionality problem of knowledge representation.

[0024] The construction of the digital twin system includes:

[0025] Establishing a virtual model of the gear and the spline;

[0026] Develop real-time data acquisition systems;

[0027] Realize real-time synchronization between virtual models and physical devices.

[0028] The method further includes: applying a general matrix multiplication (GEMM) optimization technique in the digital twin system, comprising the following steps:

[0029] Identify the matrix operations in the digital twin system, including geometric transformation, finite element analysis, dynamics simulation, and machine learning model reasoning;

[0030] Implementing cache optimization on the identified matrix operation links, including matrix blocking and data pre-fetching;

[0031] Apply SIMD instruction set optimization in the matrix operation link;

[0032] Implement multi-threaded parallelization for large-scale matrix operations;

[0033] GPU acceleration is implemented for finite element analysis tasks;

[0034] Develop an adaptive GEMM optimization strategy to dynamically select the optimal GEMM implementation based on the input data size and available hardware resources.

[0035] The step of optimizing the process parameters for machining the gears and splines using reinforcement learning technology comprises:

[0036] Establish a processing environment model, define the state space including process parameters, the action space of adjustable parameters of processing equipment, and the reward function based on processing quality and efficiency;

[0037] Select the DQN model as the reinforcement learning model;

[0038] Designing state-action value functions for estimating process parameter adjustments in a simulation environment;

[0039] In a simulation environment, training the reinforcement learning model based on the state-action value function;

[0040] The trained reinforcement learning model is used to optimize the process parameters for machining the gears and splines.

[0041] The method further comprises the following steps:

[0042] Apply target network technology to enhance the stability of the training process;

[0043] Develop a multi-agent reinforcement learning framework to optimize parameters of multiple processing equipment simultaneously;

[0044] Integrate expert knowledge and accelerate the learning process by introducing rule-based constraints.

[0045] Apply machine learning technology to perform quality inspection of the gears and splines, including:

[0046] Acquire images of the gear and spline in real time;

[0047] Using an image processing algorithm, extracting features from the images of the gears and splines;

[0048] Build and train convolutional neural networks;

[0049] The extracted features are input into the convolutional neural network to achieve defect detection and classification of the gears and splines.

[0050] The method further comprises the following steps:

[0051] Developing parametric design capabilities in the hybrid intelligent model to achieve automated customization of gears and splines;

[0052] Based on the parametric design using the hybrid intelligent model, the CAD / CAM system is integrated to achieve seamless connection from design to manufacturing;

[0053] Develop intelligent scheduling systems to optimize production processes and resource allocation.

[0054] The present application also provides a computer device, the computer device comprising:

[0055] at least one processor; and,

[0056] a memory communicatively connected to the at least one processor; wherein,

[0057] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned method for intelligent design and manufacturing of gear and spline processing.

[0058] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the above-mentioned method for intelligent design and manufacturing of gear and spline processing.

[0059] An embodiment of the present application also provides a computer program product, including computer instructions, characterized in that when the computer instructions are executed by a processor, the steps of the above-mentioned method for intelligent design and manufacturing of gear and spline processing are implemented.

[0060] The embodiment of the present application also provides a gear and spline processing intelligent design and manufacturing system, including:

[0061] An establishment module is used to establish a hybrid intelligent model, wherein the hybrid intelligent model includes a large language model and a localized small model;

[0062] A design module, used for designing process parameters of gears and splines using the hybrid intelligent model;

[0063] The simulation module is used to build a digital twin system and achieve real-time synchronization between the virtual model and the physical device;

[0064] An optimization module, used to optimize the process parameters of machining the gears and splines by using reinforcement learning technology in the digital twin system;

[0065] A processing module, used for processing the gear and the spline based on the optimized process parameters of the gear and the spline;

[0066] A detection module is used to apply machine learning technology to perform quality detection on the gears and splines.

[0067] This application has the following technical effects:

[0068] Improve the efficiency of process parameter design: Through the hybrid intelligent model, the intelligent design of gear and spline process parameters is realized, which greatly improves the design efficiency.

[0069] Real-time optimization of the machining process: The digital twin system and reinforcement learning technology are used to achieve real-time optimization of process parameters during the machining process, thereby improving machining efficiency and quality.

[0070] Improve the real-time and accuracy of quality inspection: Apply machine learning technology to perform real-time quality inspection, which significantly improves the efficiency and accuracy of inspection.

[0071] Achieve seamless integration of design and manufacturing: Through the integration of CAD / CAM system, seamless integration from design to manufacturing is achieved, thus improving production efficiency.

[0072] Improve data utilization: Through the digital twin system and machine learning technology, efficient use of processing process data is achieved, supporting continuous optimization and knowledge accumulation.

[0073] Enhance the adaptability and scalability of the system: Through the hybrid intelligent model and multi-agent reinforcement learning framework, the system has stronger adaptability and scalability, and can cope with different types of gear and spline processing requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0075] Figure 1 A flowchart of a method for intelligent design and manufacturing of gear and spline processing provided by an embodiment of the present invention;

[0076] Figure 2 This is the UI interface diagram of the automatic calculation of cylindrical gear substitute hobs;

[0077] Figure 3 This is the UI interface diagram for standard gear line length calculation;

[0078] Figure 4 This is the UI interface diagram for calculating the measured value of the chamfered teeth of a rectangular spline broach;

[0079] Figure 5 A module diagram of the intelligent design and manufacturing system for gear and spline processing provided by an embodiment of the present invention;

[0080] Figure 6 A structural diagram of the intelligent design and manufacturing system for gear and spline processing provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0081] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0082] It should be clear that the following embodiments of the present disclosure are described by specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other in the absence of conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0083] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0084] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout type is also more complicated.

[0085] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0086] Embodiment 1

[0087] like Figure 1 As shown, the embodiment of the present application provides a method for intelligent design and manufacturing of gear and spline processing, including:

[0088] [S1] Establishing a hybrid intelligent model, wherein the hybrid intelligent model includes a large language model and a localized small model;

[0089] In the present invention, the establishment of a hybrid intelligent model is the basis for realizing intelligent design and manufacturing of gear and spline processing. The hybrid intelligent model combines the powerful knowledge representation ability of the large language model and the efficient reasoning ability of the localized small model, providing intelligent support for subsequent process parameter design.

[0090] [S1.1] Select a large language model;

[0091] In this embodiment, a pre-trained large language model such as GPT-4, BERT or T5 can be selected as the basis. These models have accumulated rich general knowledge and language understanding capabilities through training on large-scale corpora. In the field of gear and spline processing, these models can understand and generate relevant professional terms, process flow descriptions, etc.

[0092] For example, you can choose the llama 3 model, which is able to understand and generate complex descriptions related to gear and spline machining. By fine-tuning llama 3, it can be better adapted to specific tasks in the field of gear and spline machining.

[0093] [S1.2] Distilling the large language model into a localized small model through knowledge distillation technology;

[0094] The purpose of the knowledge distillation technique is to transfer the knowledge related to gear and spline processing from the large language model to a smaller and more efficient model. This step is achieved as follows:

[0095] Build a teacher-student model: Use the selected large language model as the teacher model and design a neural network with fewer parameters as the student model (i.e., a localized small model).

[0096] Prepare distillation dataset: Collect a large amount of text data related to gear and spline processing, including process parameter descriptions, processing flow instructions, etc.

[0097] Distillation process:

[0098] a. Input the distilled dataset into the teacher model (large language model) to obtain its output probability distribution.

[0099] b. Input the same data into the student model (localized small model) and obtain its output probability distribution.

[0100] c. Calculate the KL divergence between the output distributions of the teacher model and the student model as the distillation loss.

[0101] d. Also consider the prediction loss of the student model for the true label.

[0102] e. Combine the distillation loss and prediction loss to optimize the student model.

[0103] Iterative optimization: Repeat the above process until the performance of the student model reaches the preset target.

[0104] For example, you can use BERT-base as a teacher model and design a student model with a 3-layer Transformer structure. Through the above distillation process, you can get a localized small model with only 1 / 10 of the parameters of BERT-base, but its performance in the field of gear and spline processing is close to the original large model.

[0105] [S1.3] Align the large language model with the localized small model;

[0106] The purpose of model alignment is to ensure that the localized small model can accurately inherit the knowledge related to gear and spline processing in the large language model and achieve similar performance on specific tasks. The implementation of this step involves multiple technical points, which are explained in detail below:

[0107] [S1.3.1] Establish a dynamic system model and use a state space model to describe the knowledge conversion process from the large language model to the localized small model;

[0108] The establishment of a dynamic system model helps to better understand and control the knowledge conversion process. The specific implementation is as follows:

[0109] 1. Define the state vector: used to represent the state of the large language model and the localized small model during the knowledge conversion process, including model parameters, intermediate representation, etc.

[0110] 2. Define the transition function: Describe how the state evolves over time. A neural network can be used to parameterize this function.

[0111] 3. Define the observation function: map the state to an observable output, such as the performance of the model on a specific task.

[0112] For example, the following state-space model can be used:

[0113] x(t+1)=f(x(t),u(t))+w(t)

[0114] y(t)=h(x(t))+v(t)

[0115] Where x(t) is the state vector at time t, u(t) is the control input, y(t) is the observed output, w(t) and v(t) are the process noise and observation noise respectively. f and h are the transfer function and observation function respectively, which can be implemented using a neural network.

[0116] [S1.3.2] Using Gumbel-Softmax and straight-through estimators to deal with the discreteness problem in the knowledge conversion process

[0117] In the process of knowledge transfer, discrete decisions are involved, such as choosing which knowledge units to transfer. In order to make the whole process differentiable, the following techniques are used:

[0118] 1. Gumbel-Softmax: Used to approximate the discrete sampling process. The Gumbel-Max technique is used for discrete sampling in forward propagation, and the continuous approximation of the Softmax function is used in backward propagation.

[0119] 2. Straight-Through Estimator: Uses discrete values ​​during forward propagation, but assumes continuous functions during backpropagation.

[0120] Implementation example:

[0121] ```Python

[0122] def gumbel_softmax(logits,temperature=1.0):

[0123] gumbel_noise=-torch.log(-torch.log(torch.rand_like(logits)))

[0124] y_soft=F.softmax((logits+gumbel_noise) / temperature,dim=-1)

[0125] y_hard=torch.zeros_like(y_soft).scatter_(-1,y_soft.argmax(dim=-1,keepdim=True),1.0)

[0126] return(y_hard y_soft).detach()+y_soft

[0127] ```

[0128] Specifically, in the process of knowledge conversion, embodiments of the present invention often need to make discrete choices, such as deciding which knowledge units should be transferred. However, this discrete choice will bring about the problem of gradient calculation in deep learning, because discrete operations are usually not differentiable. To solve this problem, the present invention adopts two technologies: Gumbel-Softmax and straight-through estimator.

[0129] Gumbel-Softmax is a method that can generate approximate discrete samples while maintaining the differentiability of the operation. It works by adding random noise to the original probability distribution and then using a temperature parameter to control the discreteness of the output. When the temperature is close to zero, the output will be very close to the true discrete sample; when the temperature is higher, the output will be smoother.

[0130] The advantage of this approach is that it can produce nearly discrete outputs during forward propagation, while allowing smooth flow of gradients during backward propagation. This solves the gradient problem caused by discrete choices.

[0131] The straight-through estimator is a variant of Gumbel-Softmax. It uses true discrete values ​​during the forward pass (by taking the argmax), but pretends to use the continuous Gumbel-Softmax output during the backward pass. This "cheating" allows the gradient to flow through discrete operations that are not differentiable.

[0132] By combining the use of Gumbel-Softmax and the straight-through estimator, the embodiment of the present invention can perform discrete selection in the knowledge conversion process while maintaining the differentiability of the entire process, thereby achieving end-to-end training.

[0133] [S1.3.3] Use multi-head attention mechanism and graph neural network to deal with the semantic coupling problem in the process of knowledge conversion;

[0134] The semantic coupling problem refers to the complex dependencies between knowledge units in a large language model. In order to effectively deal with this problem, the following techniques are used:

[0135] Multi-head attention mechanism: allows the model to focus on different aspects of knowledge simultaneously, helping to capture complex semantic relationships.

[0136] Graph neural network: Represents knowledge as a graph structure, where nodes represent knowledge units and edges represent relationships between units. Through the message passing mechanism, the dependencies between knowledge units can be better handled.

[0137] Implementation example:

[0138] ```Python

[0139] class MultiHeadAttention(nn.Module):

[0140] def__init__(self,d_model,num_heads):

[0141] super().__init__()

[0142] self.num_heads = num_heads

[0143] self.attention=nn.MultiheadAttention(d_model,num_heads)

[0144] def forward(self,query,key,value):

[0145] return self.attention(query,key,value)[0]

[0146] class GraphConvolution(nn.Module):

[0147] def__init__(self,in_features,out_features):

[0148] super().__init__()

[0149] self.linear=nn.Linear(in_features,out_features)

[0150] def forward(self,x,adj):

[0151] return torch.matmul(adj,self.linear(x))

[0152] ```

[0153] Here, the embodiment of the present invention defines a multi-head attention module and a graph convolution module. In practical applications, these modules can be used in combination to capture the complex semantic relationship between knowledge units.

[0154] Specifically, in the process of knowledge conversion, there are often complex semantic associations and dependencies between different knowledge units. In order to effectively capture and process these relationships, the present invention combines the multi-head attention mechanism and the graph neural network.

[0155] The core idea of ​​the multi-head attention mechanism is to allow the model to focus on different parts of the input or different representation subspaces at the same time. It works through the following steps:

[0156] First, the input knowledge representation is projected into multiple different subspaces to form multiple "heads". Each head calculates attention independently.

[0157] Then, inside each head, the similarity between the query and the key is calculated and this similarity is used as the attention weight to weight the values.

[0158] Finally, the outputs of all heads are combined to obtain the final representation.

[0159] The advantage of this mechanism is that it allows the model to understand the input knowledge from multiple perspectives and capture semantic relationships at different levels.

[0160] Graph neural network is a deep learning method specifically used to process graph structured data. In the present invention, the embodiment of the present invention represents knowledge as a graph, where nodes represent knowledge units and edges represent the relationship between units. Graph neural network works in the following ways:

[0161] First, for each node in the graph, information about its neighboring nodes is collected.

[0162] This collected information is then used to update the representation of the current node.

[0163] This process is repeated many times, allowing information to propagate over long distances in the graph.

[0164] In this way, graph neural networks are able to effectively model complex dependencies between knowledge units.

[0165] The combined use of multi-head attention mechanism and graph neural network enables the model of the embodiment of the present invention to comprehensively and meticulously handle the semantic coupling problem in the knowledge conversion process. Multi-head attention provides the ability to focus on multiple semantic aspects, while graph neural network provides a framework for processing structured knowledge. This combination greatly enhances the model's ability to understand and convert complex knowledge, laying a solid foundation for knowledge transfer and application in gear and spline processing.

[0166] [S1.3.4] Apply deep residual networks and nonlinear activation functions to handle nonlinear problems in the knowledge conversion process

[0167] In the process of knowledge conversion, there is a complex nonlinear relationship between the large language model and the localized small model. In order to effectively deal with this problem, the present invention applies a deep residual network and a nonlinear activation function.

[0168] The deep residual network allows information to flow directly in the network by adding skip connections, which helps alleviate the gradient vanishing problem and enables deeper networks to learn effectively. In this embodiment, residual blocks can be introduced at various stages of knowledge conversion to enhance the expressiveness of the model.

[0169] The nonlinear activation function introduces nonlinear transformation capabilities to the model, enabling it to fit more complex functional relationships. Commonly used nonlinear activation functions include ReLU, LeakyReLU, GELU, etc. In the present invention, a suitable activation function can be selected according to the requirements of a specific task, or a parameterized activation function such as PReLU can be used to adapt to different nonlinear characteristics.

[0170] Through the combination of deep residual network and nonlinear activation function, the present invention can better capture the complex nonlinear relationship between the large language model and the localized small model, and improve the accuracy and efficiency of knowledge conversion.

[0171] For example, assume that the embodiment of the present invention is dealing with the problem of cutting force prediction in gear processing. There is a complex nonlinear relationship between cutting force and multiple processing parameters (such as cutting speed, feed rate, cutting depth, etc.).

[0172] In this case, the embodiment of the present invention can construct a deep residual network to model this nonlinear relationship:

[0173] The input layer receives various processing parameters.

[0174] The network contains multiple residual blocks. Each residual block contains two convolutional layers and a skip connection. For example, the first residual block learns the interaction between cutting speed and feed rate, and the second residual block learns the further relationship between this interaction and cutting depth.

[0175] After each convolution layer, the embodiment of the present invention uses ReLU (Rectified Linear Unit) as a nonlinear activation function. ReLU can introduce nonlinear transformation while maintaining computational efficiency.

[0176] Skip connections allow the network to directly learn the identity mapping between parameters, which helps solve the gradient vanishing problem in deep networks.

[0177] The last layer uses a linear activation function and outputs the predicted cutting force.

[0178] In this way, the model of the embodiment of the present invention can capture the complex nonlinear relationship between cutting parameters and cutting forces while maintaining good training performance.

[0179] [S1.3.5] Using autoencoders and sparse attention mechanisms to handle high-dimensionality problems in knowledge representation

[0180] Large language models usually have high-dimensional knowledge representation, and directly converting them to localized small models will lead to information redundancy and low computational efficiency. To solve this problem, the present invention adopts autoencoders and sparse attention mechanisms.

[0181] Autoencoders can learn compact representations of data by compressing high-dimensional inputs into a low-dimensional latent space and then reconstructing the original input. In the knowledge conversion process, autoencoders can be used to compress the high-dimensional knowledge representation of a large language model into a lower-dimensional latent space, and then reconstruct the knowledge representation suitable for a localized small model from this latent space.

[0182] The sparse attention mechanism reduces the computational complexity by limiting the computational scope of attention, while retaining the ability to model long-distance dependencies. In the present invention, methods such as local sensitive hashing or learned sparse attention can be used to enable the model to efficiently process long sequence inputs, which is particularly useful for describing long process flows involved in gear and spline processing.

[0183] By combining the autoencoder and the sparse attention mechanism, the present invention can effectively handle the high-dimensional problem of knowledge representation and improve computational efficiency while retaining key information.

[0184] For example, consider the problem of transferring gear design knowledge. A large language model may contain a large amount of high-dimensional knowledge representation related to gear design, but not all knowledge is relevant to a specific gear processing task.

[0185] In this case, the embodiment of the present invention can use an autoencoder and a sparse attention mechanism to handle this problem:

[0186] First, the embodiment of the present invention uses an autoencoder to compress knowledge representation. The input may be a gear design knowledge vector containing thousands of features.

[0187] The encoder part may contain multiple fully connected layers, which progressively compress the input into a smaller latent space, such as only 100 dimensions.

[0188] The decoder part then attempts to reconstruct the original input from this compressed representation. By training the autoencoder to minimize the reconstruction error, the embodiment of the present invention can obtain a more compact knowledge representation.

[0189] Next, embodiments of the present invention apply a sparse attention mechanism on this compressed representation. For example, if embodiments of the present invention are processing a specific type of gear (such as a spur gear), embodiments of the present invention can design an attention mechanism that only focuses on the knowledge dimensions that are most relevant to this type of gear.

[0190] Specifically, embodiments of the present invention may calculate an attention score for each knowledge dimension, but force most scores to be close to zero. This can be achieved by adding an L1 regularization term to the attention calculation.

[0191] Finally, the embodiment of the present invention uses only those knowledge dimensions that obtain high attention scores to guide the gear processing process.

[0192] Through this method, the embodiment of the present invention can extract the most relevant parts from the high-dimensional general gear design knowledge, effectively reducing the dimension of the problem while retaining the most valuable information. This not only improves the computational efficiency, but also helps the model better focus on the most relevant knowledge for the current task.

[0193] [S2] Design of process parameters of gears and splines using the hybrid intelligent model

[0194] After the hybrid intelligent model is established, the present invention uses the model to design the process parameters of gears and splines. The purpose of this step is to automatically generate the optimal processing parameters according to the given gear and spline specifications.

[0195] The application of hybrid intelligent model in process parameter design is mainly reflected in the following aspects:

[0196] First, the large language model can understand and parse complex gear and spline specification descriptions and extract key information. For example, from a description such as "spur gear with module 2 and 20 teeth", the model can identify key parameters such as module, number of teeth, and gear type.

[0197] Secondly, the localized small model generates preliminary process parameter recommendations based on the extracted key information and the built-in professional knowledge base. These parameters include cutting speed, feed rate, cutting depth, etc.

[0198] The hybrid intelligent model then optimizes parameters by considering multiple factors, such as machining efficiency, surface quality, tool life, etc. The model can adjust and optimize the initially generated parameters based on historical data and expert rules.

[0199] Finally, the model generates a detailed process parameter report, including the specific parameter settings for each processing stage, as well as the basis for parameter selection and the expected results.

[0200] In this way, the hybrid intelligent model can quickly and accurately generate optimized process parameters for gears and splines of different specifications, greatly improving design efficiency and parameter quality.

[0201] Example: Designing the machining parameters of a spur gear

[0202] Assume that the embodiment of the present invention needs to design processing parameters for a spur gear with the following specifications:

[0203] Modulus: 3mm

[0204] Number of teeth: 30

[0205] Pressure angle: 20°

[0206] Material: 42CrMo steel

[0207] The application of hybrid intelligent model in this process is as follows:

[0208] 1. Input analysis:

[0209] The large language model component first parses the input gear specifications. It understands the meaning of "spur gear", identifies key parameters such as module, number of teeth, pressure angle, etc., and understands the material properties of 42CrMo steel.

[0210] 2. Preliminary parameter generation:

[0211] The localized small model generates preliminary process parameter recommendations based on the parsed information and the built-in professional knowledge base. For example:

[0212] Rough machining speed: 120m / min

[0213] Finishing cutting speed: 150m / min

[0214] Roughing feed rate: 0.2mm / rev

[0215] Finishing feed rate: 0.1mm / rev

[0216] Rough cutting depth: 2mm

[0217] Finishing cutting depth: 0.5mm

[0218] 3. Parameter optimization:

[0219] The hybrid intelligent model considers multiple factors for parameter optimization:

[0220] Processing efficiency: The model calculated that the processing could be completed in about 2 hours using these parameters.

[0221] Surface quality: The predicted surface roughness is Ra 1.6μm.

[0222] Tool life: It is estimated that the tool can machine about 50 gears at these parameters.

[0223] Processing accuracy: The predicted tooth profile error is within 6μm.

[0224] The model might suggest increasing the finishing cutting speed slightly to 160m / min to improve efficiency, while reducing the finishing feed rate to 0.08mm / rev to ensure better surface quality.

[0225] 4. Consider special requirements:

[0226] If there are special requirements, such as a higher surface quality, the model will further adjust the parameters. It may recommend adding a super finishing operation with the following parameters:

[0227] Super finishing cutting speed: 180m / min

[0228] Super finishing feed rate: 0.05mm / rev

[0229] Super finishing cutting depth: 0.1mm

[0230] 5. Generate process parameter report:

[0231] Finally, the model generates a detailed process parameter report including:

[0232] Specific parameter settings for each processing stage

[0233] Basis for parameter selection (such as material properties, processing efficiency requirements, etc.)

[0234] Expected results (such as expected processing time, surface quality, tool life, etc.)

[0235] Possible risks and precautions (such as processing steps that require special attention)

[0236] 6. Continuous optimization:

[0237] During the actual machining process, if the actual effect is found to deviate from the expected, the hybrid intelligent model can make real-time adjustments based on the feedback. For example, if the actual surface roughness is found to be higher than expected, the model may recommend further reducing the feed rate for finishing.

[0238] In this way, the hybrid intelligent model can comprehensively consider multiple factors such as gear specifications, material properties, and processing requirements, quickly generate optimized process parameters, and dynamically adjust them according to actual conditions, greatly improving the efficiency and quality of gear processing.

[0239] [S3] Build a digital twin system to achieve real-time synchronization between virtual models and physical devices

[0240] The construction of a digital twin system is the key to achieving real-time optimization of gear and spline machining processes. The system creates a virtual model of the machining equipment and process and keeps it synchronized with the physical equipment in real time, providing a basis for subsequent parameter optimization.

[0241] [S3.1] Establish a virtual model of the gear and spline

[0242] The establishment of the virtual model includes several key steps. First, the geometric features of the gears and splines need to be accurately modeled, which can be achieved through parametric CAD modeling technology. The model should include key parameters such as tooth shape, module, number of teeth, pressure angle, etc.

[0243] Secondly, it is necessary to establish a material model to describe the mechanical behavior of gear and spline materials during processing. This involves elastic-plastic models, thermodynamic models, etc., to accurately simulate the deformation and heat generation of materials during cutting.

[0244] Finally, it is necessary to establish a machining process model, including tool motion trajectory, cutting force model, heat conduction model, etc. These models together constitute a complete virtual machining process.

[0245] [S3.2] Develop a real-time data acquisition system

[0246] The real-time data acquisition system is the bridge between the virtual model and the physical equipment. The system needs to collect various real-time data from the physical processing equipment, such as cutting force, vibration, temperature, power, etc.

[0247] The present invention adopts multi-sensor fusion technology and comprehensively utilizes multiple sensors such as force sensors, acceleration sensors, thermocouples, power meters, etc. to obtain comprehensive processing status information. After preprocessing and feature extraction, the collected data is transmitted to the data processing center through a high-speed network.

[0248] [S3.3] Realize real-time synchronization between virtual models and physical devices

[0249] The core of the digital twin system is to achieve real-time synchronization between virtual models and physical devices. This process involves multiple aspects such as data synchronization, model update and status prediction.

[0250] In terms of data synchronization, the system needs to quickly map the real-time collected data to the virtual model and update the status of the virtual model. This requires the system to have low-latency data processing capabilities and efficient data mapping algorithms.

[0251] In terms of model updating, the system needs to continuously adjust and optimize the parameters of the virtual model based on real-time data to ensure that the virtual model can accurately reflect the actual processing status. This can be achieved through online learning algorithms or adaptive filtering technology.

[0252] In terms of state prediction, the system also needs to predict the future processing state based on the current state and historical data. This can be achieved through time series analysis or machine learning models, providing a basis for real-time optimization of process parameters.

[0253] Through the above steps, the digital twin system constructed by the present invention can achieve real-time synchronization between the virtual model and the physical equipment, laying the foundation for subsequent process parameter optimization and quality control.

[0254] In addition, in this digital twin system, the digital twin system involves a large number of matrix operations when dealing with gear and spline design, especially in the following aspects:

[0255] Geometric transformation

[0256] Finite Element Analysis

[0257] Dynamics simulation

[0258] Machine Learning Model Inference

[0259] Therefore, general matrix multiplication (GEMM) can be adopted, and optimizing GEMM can significantly improve the performance of the entire system.

[0260] The general matrix multiplication GEMM optimization technology is applied in the digital twin system, specifically including:

[0261] 1. Identify the matrix operations in the digital twin system:

[0262] In the digital twin system, matrix operations are widely used in multiple links:

[0263] Geometric transformation: used to simulate operations such as rotation and translation of parts, usually involving the multiplication of 4x4 transformation matrices.

[0264] Finite element analysis: involves solving large-scale sparse matrices, such as operations on stiffness matrices.

[0265] Dynamic simulation: including the calculation and solution of mass matrix and damping matrix.

[0266] Machine learning model reasoning: such as matrix multiplication operations of fully connected layers.

[0267] 2. Implement cache optimization:

[0268] Matrix Blocking: Divide a large matrix into sub-matrices that fit the cache size. For example, for a system with 32KB L1 cache, you can choose a 32x32 float type sub-matrix (about 4KB).

[0269] Data prefetching: Use the __builtin_prefetch() function (GCC compiler) or the _mm_prefetch() instruction (Intel compiler) to preload data into the cache.

[0270] Example code snippet:

[0271] ```c

[0272] for(int i=0; i <N;i+=BLOCK_SIZE){

[0273] for(int j=0; j <N;j+=BLOCK_SIZE){

[0274] for(int k=0;k <N;k+=BLOCK_SIZE){

[0275] __builtin_prefetch(&C[i][j+BLOCK_SIZE],1,1);

[0276] / / Perform matrix multiplication of BLOCK_SIZE x BLOCK_SIZE

[0277] }

[0278] }

[0279] }

[0280] ```

[0281] 3. Apply SIMD instruction set optimization:

[0282] Use the AVX-512 instruction set to process multiple floating-point numbers in parallel.

[0283] Example code snippet:

[0284] ```c

[0285] #include<immintrin.h>

[0286] void gemm_avx512(float*A,float*B,float*C,int N){

[0287] for(int i=0; i <N;i++){

[0288] for(int j=0; j <N;j+=16){

[0289] __m512 c=_mm512_loadu_ps(&C[i*N+j]);

[0290] for(int k=0;k <N;k++){

[0291] __m512 a=_mm512_set1_ps(A[i*N+k]);

[0292] __m512 b=_mm512_loadu_ps(&B[k*N+j]);

[0293] c = _mm512_fmadd_ps(a,b,c);

[0294] }

[0295] _mm512_storeu_ps(&C[i*N+j],c);

[0296] }

[0297] }

[0298] }

[0299] ```

[0300] 4. Implement multi-threaded parallelization:

[0301] Use OpenMP for multi-threaded parallel processing.

[0302] Example code snippet:

[0303] ```c

[0304] #include<omp.h>

[0305] void gemm_parallel(float*A,float*B,float*C,int N){

[0306] #pragma omp parallel for collapse(2)

[0307] for(int i=0; i <N;i++){

[0308] for(int j=0; j <N;j++){

[0309] float sum = 0.0f;

[0310] for(int k=0;k <N;k++){

[0311] sum+=A[i*N+k]*B[k*N+j];

[0312] }

[0313] C[i*N+j]=sum;

[0314] }

[0315] }

[0316] }

[0317] ```

[0318] 5. GPU accelerated finite element analysis:

[0319] Use CUDA for GPU acceleration, especially for large-scale sparse matrix operations.

[0320] Example code snippet (CUDA kernel function):

[0321] ```cuda

[0322] __global__void sparse_matrix_vector_multiply(int*row_ptr,int*col_idx,float*values,float*x,float*y,int N){

[0323] int i=blockIdx.x*blockDim.x+threadIdx.x;

[0324] if(i <N){

[0325] float dot = 0;

[0326] int row_start = row_ptr[i];

[0327] int row_end = row_ptr[i+1];

[0328] for(int j = row_start; j <row_end;j++){

[0329] dot+=values[j]*x[col_idx[j]];

[0330] }

[0331] y[i] = dot;

[0332] }

[0333] }

[0334] ```

[0335] 6. Develop an adaptive GEMM optimization strategy:

[0336] Create a decision tree or lookup table to select the most appropriate GEMM implementation based on the size of the input matrix and the available hardware resources.

[0337] Example pseudocode:

[0338] ```Python

[0339] def select_optimal_gemm(M,N,K,available_memory,num_cores,has_gpu):

[0340] if has_gpu and M*N*K>GPU_THRESHOLD:

[0341] return gpu_gemm

[0342] elif M*N*K <L3_CACHE_SIZE:

[0343] return cache_optimized_gemm

[0344] elif num_cores>4and M*N*K>PARALLEL_THRESHOLD:

[0345] return parallel_gemm

[0346] else:

[0347] return basic_gemm

[0348] ```

[0349] This function will select the best GEMM implementation based on the dimensions of the input matrix (M, N, K), available memory, number of CPU cores, and whether a GPU is available. Thresholds (such as GPU_THRESHOLD, L3_CACHE_SIZE, PARALLEL_THRESHOLD) need to be determined through experiments and performance analysis.

[0350] For example, assuming that the embodiment of the present invention is simulating a complex gear processing process, which involves a large number of matrix operations, the embodiment of the present invention will take the finite element analysis of gears as an example to illustrate the GEMM optimization process.

[0351] Scenario: The embodiment of the present invention needs to perform stress analysis on a high-precision gear model, which contains 100,000 nodes.

[0352] 1. Identify the matrix operation link:

[0353] In finite element analysis, the embodiment of the present invention needs to solve the following equations:

[0354] K*u=f

[0355] Where K is the stiffness matrix (100,000 x 100,000), u is the displacement vector, and f is the force vector.

[0356] 2. Implement cache optimization:

[0357] Assuming that the CPU of the embodiment of the present invention has a 256KB L2 cache, the embodiment of the present invention can divide the stiffness matrix K into 512x512 sub-matrices, each of which occupies about 1MB (assuming double-precision floating point numbers are used).

[0358] Code example:

[0359] ```c

[0360] #define BLOCK_SIZE 512

[0361] for(int i=0; i <N;i+=BLOCK_SIZE){

[0362] for(int j=0; j <N;j+=BLOCK_SIZE){

[0363] __builtin_prefetch(&K[i][j+BLOCK_SIZE],1,1);

[0364] / / Process K[i:i+BLOCK_SIZE][j:j+BLOCK_SIZE] submatrix

[0365] }

[0366] }

[0367] ```

[0368] 3. Apply SIMD instruction set optimization:

[0369] Use the AVX-512 instruction set to process 8 double-precision floating-point numbers in parallel.

[0370] Code example:

[0371] ```c

[0372] #include<immintrin.h>

[0373] void matrix_vector_multiply_avx512(double*K,double*u,double*f,int N){

[0374] for(int i=0; i <N;i+=8){

[0375] __m512d sum=_mm512_setzero_pd();

[0376] for(int j=0; j <N;j++){

[0377] __m512d k=_mm512_loadu_pd(&K[i*N+j]);

[0378] __m512d u_vec=_mm512_set1_pd(u[j]);

[0379] sum=_mm512_fmadd_pd(k,u_vec,sum);

[0380] }

[0381] _mm512_storeu_pd(&f[i],sum);

[0382] }

[0383] }

[0384] ```

[0385] 4. Implement multi-threaded parallelization:

[0386] Use OpenMP to parallelize matrix-vector multiplication.

[0387] Code example:

[0388] ```c

[0389] #include<omp.h>

[0390] void parallel_matrix_vector_multiply(double*K,double*u,double*f,intN){

[0391] #pragma omp parallel for

[0392] for(int i=0; i <N;i++){

[0393] double sum = 0.0;

[0394] for(int j=0; j <N;j++){

[0395] sum+=K[i*N+j]*u[j];

[0396] }

[0397] f[i] = sum;

[0398] }

[0399] }

[0400] ```

[0401] 5. GPU acceleration:

[0402] Considering that the stiffness matrix K is usually sparse, the embodiment of the present invention can use the CUDA sparse matrix library cuSPARSE to accelerate the calculation.

[0403] Code example:

[0404] ```cuda

[0405] #include<cusparse.h>

[0406] cusparseStatus_t sparse_matrix_vector_multiply_gpu(

[0407] cusparseHandle_t handle,

[0408] int*rowPtr,

[0409] int*colIdx,

[0410] double*values,

[0411] int nnz,

[0412] double*u,

[0413] double*f,

[0414] int N

[0415] ){

[0416] cusparseMatDescr_t descr=0;

[0417] cusparseCreateMatDescr(&descr);

[0418] cusparseSetMatType(descr,CUSPARSE_MATRIX_TYPE_GENERAL);

[0419] cusparseSetMatIndexBase(descr,CUSPARSE_INDEX_BASE_ZERO);

[0420] const double alpha = 1.0;

[0421] const double beta = 0.0;

[0422] return cusparseDcsrmv(handle,CUSPARSE_OPERATION_NON_TRANSPOSE,

[0423] N,N,nnz,&alpha,descr,values,rowPtr,colIdx,

[0424] u,&beta,f);

[0425] }

[0426] ```

[0427] 6. Adaptive GEMM optimization strategy:

[0428] Create a decision function to choose the best implementation:

[0429] ```Python

[0430] def select_optimal_fem_solver(N,nnz,available_memory,num_cores,has_gpu):

[0431] density = nnz / (N*N)

[0432] if has_gpu and N>50000:

[0433] return gpu_sparse_solver

[0434] elif density<0.01:

[0435] return cpu_sparse_solver

[0436] elif num_cores>16and N>10000:

[0437] return parallel_dense_solver

[0438] else:

[0439] return basic_dense_solver

[0440] ```

[0441] Actual application process:

[0442] 1. First, the system identifies that matrix-vector multiplication in finite element analysis is the key computational bottleneck.

[0443] 2. The system detects that the stiffness matrix K is a 100,000 x 100,000 sparse matrix with approximately 5,000,000 non-zero elements (assuming that each node is connected to 50 other nodes on average).

[0444] 3. The system checks the available hardware resources: 64GB RAM, 32-core CPU, and 1 CUDA-supported GPU.

[0445] 4. Call the select_optimal_fem_solver function:

[0446] ```Python

[0447] optimal_solver=select_optimal_fem_solver(100000,5000000,64e9,32,True)

[0448] ```

[0449] 5. The function returns to gpu_sparse_solver because the problem size is large and a GPU is available.

[0450] 6. The system uses gpu_sparse_solver implemented by CUDA's cuSPARSE library to solve the finite element equations.

[0451] 7. During the calculation process, the system continuously monitors performance indicators. If low GPU utilization or memory bottleneck is detected, the system may switch to cpu_sparse_solver in the next iteration.

[0452] In this way, the digital twin system of the embodiment of the present invention can automatically select the optimal GEMM implementation method according to the specific problem scale and available resources, thereby greatly improving the efficiency of gear finite element analysis. This not only speeds up a single analysis, but also enables the system to perform more frequent real-time simulation and optimization during the actual processing process, improving the accuracy and efficiency of the entire gear processing process.

[0453] By implementing these optimization techniques, the embodiments of the present invention can significantly improve the performance of matrix operations in the digital twin system, thereby improving the response speed and efficiency of the entire system. This is crucial for real-time simulation and optimization of the machining process of gears and splines.

[0454] [S4] In the digital twin system, the process parameters of the gear and spline are optimized by using reinforcement learning technology.

[0455] After the digital twin system is built, the present invention uses reinforcement learning technology to dynamically optimize the processing parameters of gears and splines. The purpose of this step is to continuously adjust the process parameters according to real-time feedback during the actual processing process to achieve the best processing effect.

[0456] [S4.1] Establish a processing environment model, define the state space including process parameters, the action space of adjustable parameters of processing equipment, and the reward function based on processing quality and efficiency

[0457] The processing environment model is the basis of reinforcement learning. In the present invention, the state space includes the current process parameters (such as cutting speed, feed rate, cutting depth, etc.) and the real-time processing status (such as cutting force, vibration, temperature, etc.) obtained from the digital twin system. The action space defines the adjustable parameter range of the processing equipment, such as the adjustable range of cutting speed and feed rate.

[0458] The design of the reward function is crucial, as it directly affects the optimization direction of reinforcement learning. The reward function of the present invention comprehensively considers factors such as processing quality (such as surface roughness and geometric accuracy) and processing efficiency (such as material removal rate and processing time), as well as tool life. The reward function can be expressed as a weighted sum of these factors, and the weights can be adjusted according to specific processing requirements.

[0459] [S4.2] Select the DQN model as the reinforcement learning model

[0460] The present invention selects Deep Q-Network (DQN) as the reinforcement learning model. DQN is a reinforcement learning algorithm that combines Q-learning and deep neural networks, and is particularly suitable for processing problems in high-dimensional state spaces.

[0461] In this embodiment, the input of the DQN network is the current state (process parameters and processing state), and the output is the Q value of each action (parameter adjustment). The network structure can adopt a multi-layer fully connected network or a convolutional neural network, and the specific structure can be designed according to the characteristics of the input state.

[0462] [S4.3] Design state-action value functions for estimating process parameter adjustments in a simulation environment

[0463] The state-action value function (Q function) is the core of DQN, which estimates the long-term reward that can be obtained by taking a certain action in a given state. In the present invention, this function is used to evaluate the long-term benefits that can be brought by adjusting a certain process parameter under the current processing state.

[0464] The design of the function needs to take into account immediate rewards (such as current machining quality and efficiency) and future rewards (such as the impact on tool life). Through repeated trials and learning in a simulation environment, DQN is able to gradually optimize this state-action value function.

[0465] [S4.4] In a simulation environment, training the reinforcement learning model based on the state-action value function

[0466] The training process is carried out in the simulation environment provided by the digital twin system. This process includes the following key steps:

[0467] First, the parameters of the DQN network are initialized. Then, in each training round, the model selects an action (process parameter adjustment) based on the current state, observes the feedback from the environment (new state and reward), and stores this experience in the experience replay pool.

[0468] Next, a batch of experiences is randomly sampled from the experience replay pool and used to update the parameters of the DQN network. The update process uses the gradient descent method to minimize the mean square error between the predicted Q value and the target Q value.

[0469] In order to improve the stability of training, the present invention adopts the target network technology, that is, using a separate network to generate the target Q value and regularly copying the parameters of the DQN network to the target network.

[0470] [S4.5] Optimizing the process parameters for machining the gears and splines using the trained reinforcement learning model

[0471] After training, the reinforcement learning model is deployed to the actual processing system. During the processing, the model continuously adjusts the process parameters according to the real-time processing status. Specifically, the model receives real-time status information from the digital twin system and then selects the action with the largest Q value, that is, the optimal parameter adjustment solution.

[0472] This dynamic optimization process can adapt to various changes in the machining process, such as tool wear, changes in material properties, etc., so as to continuously maintain the optimal machining effect.

[0473] For example, the embodiment of the present invention will use a gear processing case to go through the entire process of S4.1-S4.5.

[0474] [S4.1] Establish a processing environment model, define the state space including process parameters, the action space of adjustable parameters of processing equipment, and the reward function based on processing quality and efficiency

[0475] Assume that the embodiment of the present invention is optimizing the machining process of a spur gear.

[0476] 1. State space definition:

[0477] Current cutting speed: 80-200m / min

[0478] Current feed rate: 0.05-0.3mm / rev

[0479] Current cutting depth: 0.1-3mm

[0480] Real-time cutting force: 0-2000N

[0481] Real-time vibration amplitude: 0-5mm / s

[0482] Real-time cutting temperature: 20-300℃

[0483] Processing time: 0-120min

[0484] 2. Action space definition:

[0485] Adjust cutting speed: -20, -10, 0, +10, +20m / min

[0486] Adjust feed rate: -0.05, -0.02, 0, +0.02, +0.05mm / rev

[0487] Adjust cutting depth: -0.5, -0.2, 0, +0.2, +0.5mm

[0488] 3. Reward function design:

[0489] R = w1*quality score + w2*efficiency score + w3*tool life score w4*energy consumption score

[0490] in:

[0491] Quality score: Based on surface roughness and tooth profile accuracy, ranging from 0-100

[0492] Efficiency score: Based on material removal rate, range 0-100

[0493] Tool Life Score: Based on estimated tool wear, range 0-100

[0494] Energy Score: Based on power consumption, range 0-100

[0495] w1, w2, w3, w4 are weight coefficients, which can be adjusted according to specific needs.

[0496] [S4.2] Select the DQN model as the reinforcement learning model

[0497] The embodiment of the present invention constructs a DQN model, whose structure is as follows:

[0498] 1. Input layer: 7 neurons, corresponding to 7 parameters of the state space

[0499] 2. Hidden layer 1: 64 neurons, ReLU activation function

[0500] 3. Hidden layer 2: 32 neurons, ReLU activation function

[0501] 4. Output layer: 15 neurons, corresponding to 15 possible operations in the action space (5 adjustment methods for each of the 3 parameters)

[0502] [S4.3] Design state-action value functions for estimating process parameter adjustments in a simulation environment

[0503] The state-action value function Q(s,a) of the embodiment of the present invention estimates the long-term reward that can be obtained by taking action a in the current state s. The specific implementation is as follows:

[0504] 1. Input the current state s (including current process parameters and processing status)

[0505] 2. For each possible action a (15 parameter adjustment methods), predict the Q value

[0506] 3. The calculation of Q value takes into account immediate rewards and possible future rewards:

[0507] Q(s,a) = immediate reward + γ * maximum future reward

[0508] Where γ is a discount factor used to balance the importance of immediate rewards and future rewards

[0509] For example, if the current state is a cutting speed of 120 m / min, the embodiment of the present invention predicts that the Q value of increasing by 10 m / min may be 80, which means that this action may bring certain long-term benefits.

[0510] [S4.4] In a simulation environment, training the reinforcement learning model based on the state-action value function

[0511] The embodiment of the present invention trains the DQN model in the simulation environment provided by the digital twin system. The training process is as follows:

[0512] 1. Initialization:

[0513] DQN network parameters are randomly initialized

[0514] The experience replay pool is initialized to empty

[0515] Set the number of training rounds to 10000

[0516] 2. For each training round:

[0517] a. Reset the simulation environment and get the initial state

[0518] b. For each step in the environment (up to 1000 steps):

[0519] Randomly select an action with probability ε, otherwise select the action with the largest Q value

[0520] Perform the selected action and observe the new state and reward

[0521] Store (state, action, reward, new state) into the experience replay pool

[0522] If the experience replay pool size exceeds 10,000, remove the oldest experience

[0523] Randomly sample 64 experiences from the experience replay pool for learning

[0524] Calculate the target Q value using the target network

[0525] Update the DQN network parameters to minimize the mean square error between the predicted Q value and the target Q value

[0526] Update the target network every 100 steps

[0527] c. If the termination condition is reached (such as processing is completed or a serious error occurs), end the current round

[0528] 3. During the training process, gradually reduce the ε value from 0.9 to 0.1 to balance exploration and exploitation

[0529] [S4.5] Optimizing the process parameters for machining the gears and splines using the trained reinforcement learning model

[0530] The trained DQN model is deployed to the actual gear processing system.

[0531] 1. Initialization: Set the initial process parameters (cutting speed 120m / min, feed rate 0.1mm / rev, cutting depth 1mm)

[0532] 2. During the processing, every 5 seconds:

[0533] a. Obtain the current status (including process parameters and processing status) from the digital twin system

[0534] b. Input the state into the DQN model to obtain the Q value of each possible action

[0535] c. Choose the action with the highest Q value

[0536] d. Execute the selected action and adjust the corresponding process parameters

[0537] 3. Practical application cases:

[0538] Initial state: cutting speed 120m / min, feed rate 0.1mm / rev, cutting depth 1mm, cutting force 1000N, vibration amplitude 2mm / s, temperature 150℃

[0539] DQN model prediction: The Q value is the highest when the cutting speed is increased by 10m / min

[0540] System execution: Adjust the cutting speed to 130m / min

[0541] After 5 seconds, the new state: the cutting force decreased to 950N, the vibration amplitude slightly increased to 2.2mm / s, and the temperature increased to 160℃

[0542] The DQN model predicts again: Keep the Q value of the current parameters as high as possible

[0543] System execution: Keep current parameters unchanged

[0544] This process will continue throughout the entire machining cycle, continuously optimizing the process parameters to achieve the best machining results. In this way, the reinforcement learning model can dynamically adjust the process parameters according to the real-time machining status and adapt to various changes, such as tool wear, changes in material properties, etc., so as to continuously maintain the best machining results.

[0545] [S5] Processing the gear and the spline based on the optimized process parameters of the gear and the spline;

[0546] After obtaining the optimized process parameters, the system applies these parameters to the actual processing process. This process involves the control and monitoring of processing equipment.

[0547] First, the optimized process parameters are transmitted to the CNC system to update the machining program, including updating parameters such as cutting speed, feed rate, and cutting depth.

[0548] Secondly, during the machining process, the digital twin system continues to monitor the actual machining status, including cutting force, vibration, temperature, etc. If a significant deviation is detected, the system will trigger a new round of parameter optimization.

[0549] Finally, after the processing is completed, the system will record the data of the entire processing process, including process parameters, processing status, final product quality, etc. These data will be used for subsequent analysis and model optimization.

[0550] In addition, the embodiment of the present invention further includes the following steps:

[0551] [S6] Apply machine learning technology to perform quality inspection on the gears and splines

[0552] Quality inspection is the last key link in the processing process. The present invention applies machine learning technology to achieve automatic and high-precision quality inspection.

[0553] [S6.1] Real-time acquisition of images of the gear and spline

[0554] The present invention uses high-resolution industrial cameras to capture images of gears and splines. The cameras are strategically placed at key locations on the processing equipment to capture the key features of gears and splines. During the image acquisition process, the control of lighting conditions is considered to ensure the consistency of image quality.

[0555] [S6.2] Extracting features from the image of the gear and spline using an image processing algorithm

[0556] Image processing is the basis of quality inspection. The present invention uses a variety of image processing techniques to extract key features of gears and splines. This includes edge detection algorithms to identify tooth profiles, Hough transforms to detect roundness, and texture analysis to assess surface quality.

[0557] The results of feature extraction include geometric dimensions (such as module, number of teeth, pressure angle, etc.), surface features (such as roughness, smoothness, etc.) and existing defects (such as cracks, notches, etc.).

[0558] [S6.3] Build and train a convolutional neural network

[0559] The present invention uses a convolutional neural network (CNN) to detect and classify defects in gears and splines. The structure of the CNN includes multiple convolutional layers, pooling layers, and fully connected layers.

[0560] The network's training data includes a large number of labeled gear and spline images, covering a variety of defect types and normal samples. The training process uses the cross entropy loss function and Adam optimizer, and uses data enhancement techniques (such as rotation, scaling, adding noise, etc.) to increase the generalization ability of the model.

[0561] Specifically, a CNN model dedicated to gear and spline defect detection is built and trained. The following is the detailed process:

[0562] Network architecture design: The embodiment of the present invention designs a multi-layer CNN, and the specific structure is as follows:

[0563] Input layer: accepts RGB images of size 224x224x3

[0564] Convolutional layer 1: 32 3x3 convolution kernels, stride 1, ReLU activation function

[0565] Max pooling layer 1: 2x2 pooling with a stride of 2

[0566] Convolutional layer 2: 64 3x3 convolution kernels, stride 1, ReLU activation function

[0567] Max pooling layer 2: 2x2 pooling with a stride of 2

[0568] Convolutional layer 3: 128 3x3 convolution kernels, stride 1, ReLU activation function

[0569] Max pooling layer 3: 2x2 pooling with a stride of 2

[0570] Fully connected layer 1: 512 neurons, ReLU activation function

[0571] Dropout layer: dropout rate is 0.5

[0572] Fully connected layer 2: 128 neurons, ReLU activation function

[0573] Output layer: The number of neurons is determined according to the number of defect types, and the Softmax activation function is used

[0574] Data preparation:

[0575] Collect a large number of gear and spline images, including all possible defect types (such as cracks, nicks, surface scratches, etc.) and normal samples

[0576] Annotate images to indicate defect type and location

[0577] The dataset is divided into training set (70%), validation set (15%) and test set (15%)

[0578] Data enhancement: To increase the generalization ability of the model, the embodiment of the present invention enhances the training data:

[0579] Random rotation (-15° to 15°)

[0580] Random horizontal and vertical flips

[0581] Randomly adjust brightness and contrast (±10%)

[0582] Add random noise

[0583] Model training:

[0584] Choose cross entropy as the loss function

[0585] The Adam optimizer is used and the learning rate is set to 0.001

[0586] The batch size is set to 32

[0587] The number of training rounds is set to 100, but an early stopping mechanism is set. If the performance of the validation set does not improve after 10 consecutive rounds, the training is stopped.

[0588] Use the learning rate decay strategy to reduce the learning rate to half of its original value every 20 rounds.

[0589] Model evaluation: Evaluate the model performance on the test set and calculate the following metrics:

[0590] Accuracy

[0591] Precision

[0592] Recall

[0593] F1 score

[0594] Model optimization: Based on the evaluation results, the following optimizations may be required:

[0595] Adjust the network structure (such as increasing or decreasing the number of layers)

[0596] Modify hyperparameters (such as learning rate, batch size)

[0597] Increase training data or further enhance data

[0598] [S6.4] Inputting the extracted features into the convolutional neural network to achieve defect detection and classification of the gear and spline

[0599] In actual inspection, the system first preprocesses and extracts features from the captured images, and then inputs the extracted features into the trained CNN. The CNN outputs the type and location of the defect, or determines it as a qualified product.

[0600] The system also integrates an interpretability module that can generate heat maps to show the areas of network attention, helping operators understand the detection results.

[0601] In this way, the present invention realizes efficient and accurate gear and spline quality detection, greatly improving detection efficiency and accuracy.

[0602] In addition, the embodiment of the present invention further includes the following steps:

[0603] A1. Apply target network technology to enhance the stability of the training process

[0604] The target network technique is designed to address the instability problem that may occur in Q-learning when using neural networks as function approximators.

[0605] Specifically include:

[0606] a. Network structure:

[0607] Main network (Q network): used for selecting actions and updating

[0608] Target network: used to calculate the target Q value

[0609] b. Update mechanism:

[0610] The main network is updated at every step

[0611] The target network parameters θ' are periodically copied from the master network parameters θ, with a frequency of every C steps:

[0612] θ'←θ

[0613] c. Loss function:

[0614] L=(r+γ*max_a'Q'(s',a';θ')Q(s,a;θ))^2

[0615] Where Q' is the target network and Q is the main network

[0616] d. Implementation example:

[0617] ```Python

[0618] class DQNAgent:

[0619] def __init__(self):

[0620] self.q_network=self.build_network()

[0621] self.target_network=self.build_network()

[0622] self.update_target_network()

[0623] def update_target_network(self):

[0624] self.target_network.set_weights(self.q_network.get_weights())

[0625] def train_step(self,state,action,reward,next_state,done):

[0626] target=reward

[0627] if not done:

[0628] target+=self.gamma*np.amax(self.target_network.predict(next_state)[0])

[0629] target_f=self.q_network.predict(state)

[0630] target_f[0][action]=targetself.q_network.fit(state,target_f,epochs=1,verbose=0)

[0631] def update_target_if_needed(self,step):

[0632] if step%self.target_update_frequency==0:

[0633] self.update_target_network()

[0634] ```

[0635] A2. Develop a multi-agent reinforcement learning framework to optimize the parameters of multiple processing equipment at the same time. Multi-agent reinforcement learning allows embodiments of the present invention to optimize multiple processing equipment at the same time, taking into account the mutual influence between the equipment. Specifically, it includes:

[0636] a. Agent Design:

[0637] Each processing equipment acts as an independent intelligent entity

[0638] Each agent has its own observation space, action space and reward function

[0639] b.Environmental design:

[0640] The global state contains the state of all devices

[0641] Global rewards take into account the performance of the entire production line

[0642] c. Learning algorithm: Use Multi-Agent DDPG (MADDPG)

[0643] d. Network structure:

[0644] Each agent has its own Actor network and Critic network

[0645] Actor Network: Input local observations, output actions

[0646] Critic network: inputs the global state and the actions of all agents, and outputs the Q value

[0647] e. Implementation example:

[0648] ```Python

[0649] class MADDPGAgent:

[0650] def__init__(self,state_size,action_size,agent_num):

[0651] self.actor=self.build_actor(state_size,action_size)

[0652] self.critic=self.build_critic(state_size*agent_num,action_size*agent_num)

[0653] self.target_actor=self.build_actor(state_size,action_size)

[0654] self.target_critic=self.build_critic(state_size*agent_num, action_size*agent_num)

[0655] def build_actor(self,state_size,action_size):

[0656] #Build Actor network

[0657] def build_critic(self,state_size,action_size):

[0658] #Build Critic network

[0659] def train(self,states,actions,rewards,next_states,dones):

[0660] #Training process

[0661] class MultiAgentEnvironment:

[0662] def__init__(self,agent_num):

[0663] self.agents=[MADDPGAgent(...)for_in range(agent_num)]

[0664] def step(self,actions):

[0665] #Execute actions and return new status, rewards, etc.

[0666] def reset(self):

[0667] #Reset the environment

[0668] ```

[0669] A3. Integrate expert knowledge and accelerate the learning process by introducing rule-based constraints

[0670] By integrating expert knowledge, the embodiments of the present invention can guide the reinforcement learning algorithm to converge to an effective strategy more quickly. Specifically, it includes:

[0671] a. Knowledge Representation:

[0672] Representing expert knowledge using if-then rules

[0673] Converting rules into a form that can be interpreted by neural networks

[0674] b. Constrained integration method:

[0675] Action space constraints: restricting the available actions in certain states Reward shaping: adjusting the reward function based on expert rules

[0676] Policy initialization: Initialize the policy network using expert rules

[0677] c. Implementation example:

[0678] ```Python

[0679] class KnowledgeBasedDQNAgent(DQNAgent):

[0680] def __init__(self):

[0681] super().__init__()

[0682] self.expert_rules=self.load_expert_rules()

[0683] def load_expert_rules(self):

[0684] # Load expert rules

[0685] return {

[0686] 'safe_speed_range':(50,200),

[0687] 'max_feed_rate':0.5,

[0688] 'tool_change_interval':1000

[0689] }

[0690] def select_action(self,state):

[0691] action=super().select_action(state)

[0692] return self.apply_expert_constraints(state,action)

[0693] def apply_expert_constraints(self,state,action):

[0694] if state['tool_wear']>0.8:

[0695] return 'change_tool'

[0696] if action['speed'] < self.expert_rules['safe_speed_range'][0]:

[0697] action['speed'] = self.expert_rules['safe_speed_range'][0]

[0698] if action['speed'] > self.expert_rules['safe_speed_range'][1]:

[0699] action['speed'] = self.expert_rules['safe_speed_range'][1]

[0700] if action['feed_rate'] > self.expert_rules['max_feed_rate']:

[0701] action['feed_rate'] = self.expert_rules['max_feed_rate']

[0702] return action

[0703] def compute_reward(self, state, action, next_state):

[0704] base_reward = super().compute_reward(state, action, next_state)

[0705] if action == 'change_tool' and state['tool_wear'] > 0.8:

[0706] base_reward += 10 # 奖励及时更换刀具

[0707] return base_reward

[0708] def initialize_network(self):

[0709] # Initialize network weights using expert rules

[0710] # This may require some complex mapping logic

[0711] ```

[0712] By implementing these technical details, the embodiments of the present invention can:

[0713] 1. Use the target network to stabilize the DQN training process and avoid drastic fluctuations in Q-value estimation.

[0714] 2. Use a multi-agent framework to optimize multiple processing equipment simultaneously, taking into account the mutual influence between equipment and the performance of the overall production line.

[0715] 3. Integrate expert knowledge to accelerate the learning process and prevent the agent from making obviously wrong decisions in the early stages, thereby accelerating convergence and improving final performance. The combination of these technologies will significantly improve the performance and efficiency of the gear and spline processing intelligent design and manufacturing system of the embodiment of the present invention.

[0716] The embodiment of the present invention also includes:

[0717] B1. Develop parametric design capabilities in hybrid intelligent models to achieve automated customization of gears and splines

[0718] Specifically include:

[0719] a. Parametric model construction:

[0720] Use geometric constraints and mathematical equations to define key parameters of gears and splines, such as module, number of teeth, pressure angle, helix angle, and more.

[0721] Establish the relationship model between parameters, such as the relationship between the tooth addendum coefficient and the module, the relationship between the tooth root height coefficient and the module, etc.

[0722] b. Knowledge graph integration:

[0723] Build a knowledge graph for gear and spline design, including information on design specifications, material properties, processing techniques, etc.

[0724] Use graph databases (such as Neo4j) to store knowledge graphs for fast query and reasoning.

[0725] c. Application of hybrid intelligent model:

[0726] Use large language models to understand and parse user demand descriptions.

[0727] Use localized small models for specific parameter calculation and optimization.

[0728] d. Optimization algorithm integration:

[0729] Integration of evolutionary algorithms such as genetic algorithm and particle swarm optimization is used to optimize the parameters of gears and splines under given constraints.

[0730] e. Implementation example:

[0731] ```Python

[0732] class ParametricDesignSystem:

[0733] def __init__(self):

[0734] self.llm=load_large_language_model()

[0735] self.local_model=load_local_model()

[0736] self.knowledge_graph=load_knowledge_graph()

[0737] self.optimization_engine=OptimizationEngine()

[0738] def parse_user_requirements(self,user_input):

[0739] return self.llm.extract_parameters(user_input)

[0740] def generate_initial_design(self,parameters):

[0741] initial_design=self.local_model.calculate_initial_parameters(parameters)

[0742] constraints=self.knowledge_graph.query_constraints(parameters)

[0743] return self.optimization_engine.optimize(initial_design,constraints)

[0744] def visualize_design(self,design_parameters):

[0745] #Generate 3D model or 2D drawing

[0746] ```

[0747] like Figure 2 As shown, Figure 2 It is the UI interface for the automatic calculation of the substitute hob for cylindrical gears. The hybrid intelligent model can generate an automatic program. The operation method is as follows: 1. Input the gear parameters (module, number of teeth, pressure angle, helix angle, pitch arc tooth thickness) and the diameter at the distortion position. 2. Input the hob parameters (module, pressure angle). 3. Click "Calculate" to get the calculation results (substitute roller pitch circle diameter, pitch arc tooth thickness and distortion, etc.). 4. Click "Clear" to clear all input parameters and calculation results, and the key automatically jumps to the input parameter "module" position. 5. Click "Exit" to exit the program and end the program.

[0748] Figure 3 It is the UI interface for calculating the length of standard gear line method. Figure 3 Operation method: 1. Input the standard gear parameters (module, number of teeth, normal pressure angle, helix angle, normal tooth thickness). 2. Input the measured number of teeth. 3. Click "Calculate" to get the calculation result (normal length). 4. Click "Clear" to clear all input parameters and calculation results, and the key automatically jumps to the input parameter "module" position. 5. Click "Exit" to exit the program and end the program operation.

[0749] Figure 4 This is the UI interface for calculating the measured value of the chamfered teeth of a rectangular spline broach. The same calculation method as above can automatically obtain the measured value.

[0750] B2. Based on the use of hybrid intelligent models for parametric design, the CAD / CAM system is integrated to achieve seamless connection from design to manufacturing.

[0751] Specifically include:

[0752] a.CAD system integration:

[0753] Develop API interfaces to automatically transfer parametric design results to CAD systems (such as SolidWorks, AutoCAD, etc.).

[0754] Use the secondary development interface of the CAD system to automatically generate 3D models and 2D engineering drawings.

[0755] b.CAM system integration:

[0756] Automatically import CAD models into CAM systems (such as Mastercam, CATIA, etc.).

[0757] Based on the machining parameters generated by the hybrid intelligent model, the NC machining program is automatically generated.

[0758] c. Digital Twin Integration:

[0759] Synchronize design and processing parameters to the digital twin system to achieve virtual simulation and optimization.

[0760] d.Data exchange standards:

[0761] Use standard formats such as STEP and IGES to ensure data exchange compatibility between different systems.

[0762] e. Implementation example:

[0763] ```Python

[0764] class IntegratedDesignManufactureSystem:

[0765] def __init__(self):

[0766] self.parametric_design_system=ParametricDesignSystem()

[0767] self.cad_interface=CADSystemInterface()

[0768] self.cam_interface=CAMSystemInterface()

[0769] self.digital_twin=DigitalTwinSystem()

[0770] def design_to_manufacture(self,user_requirements):

[0771] #Parametric Design

[0772] design_params=self.parametric_design_system.generate_design(user_requirements)

[0773] #CAD Modeling

[0774] cad_model=self.cad_interface.create_model(design_params)

[0775] #CAM program generation

[0776] cam_program=self.cam_interface.generate_program(cad_model,design_params)

[0777] #DigitalTwinSynchronization

[0778] self.digital_twin.update_model(design_params,cam_program)

[0779] return cam_program

[0780] class CADSystemInterface:

[0781] def create_model(self,params):

[0782] #Call the CAD system API to create a model

[0783] class CAMSystemInterface:

[0784] def generate_program(self,cad_model,params):

[0785] #Call the CAM system API to generate the machining program

[0786] ```

[0787] B3. Develop intelligent scheduling system to optimize production process and resource allocation

[0788] Specifically include:

[0789] a. Production Modeling:

[0790] Represent the production process using discrete event simulation models.

[0791] Build resource models, including machines, workers, tools, etc.

[0792] b. Scheduling algorithm:

[0793] Implements a mixed integer programming (MIP) solver for global optimization.

[0794] Integrated heuristic algorithms (such as simulated annealing and ant colony algorithm) are used to quickly solve large-scale problems.

[0795] c. Real-time data integration:

[0796] Collect production data in real time through Internet of Things (IoT) devices.

[0797] Use a stream processing framework such as Apache Flink to process real-time data streams.

[0798] d. Prediction model:

[0799] Integrate machine learning models to predict equipment failures, production bottlenecks, and more.

[0800] e. Dynamic adjustment:

[0801] Implement a dynamic rescheduling mechanism to respond to emergencies and plan changes.

[0802] f. Visual interface:

[0803] Develop a web-based visualization interface to display production plans and real-time progress.

[0804] g. Implementation example:

[0805] ```Python

[0806] class IntelligentSchedulingSystem:

[0807] def __init__(self):

[0808] self.simulation_model=DiscreteEventSimulation()

[0809] self.mip_solver=MixedIntegerProgrammingSolver()

[0810] self.heuristic_solver=HeuristicSolver()

[0811] self.iot_interface=IoTInterface()

[0812] self.ml_predictor=MachineLearningPredictor()

[0813] def generate_schedule(self, orders, resources):

[0814] initial_schedule = self.mip_solver.solve(orders, resources)

[0815] return self.heuristic_solver.refine(initial_schedule)

[0816] def update_schedule(self, new_data):

[0817] predictions = self.ml_predictor.predict(new_data)

[0818] if self.need_reschedule(predictions):

[0819] return self.generate_schedule(self.current_orders, self.current_resources)

[0820] return self.current_schedule

[0821] def run(self):

[0822] while True:

[0823] new_data = self.iot_interface.get_real_time_data()

[0824] self.current_schedule = self.update_schedule(new_data)

[0825] self.visualization.update(self.current_schedule)

[0826] class IoTInterface:

[0827] def get_real_time_data(self):

[0828] # Get real-time data from IoT devices

[0829] class MachineLearningPredictor:

[0830] def predict(self,data):

[0831] #Use machine learning model to make predictions

[0832] Class Visualization:

[0833] def update(self,schedule):

[0834] #Update the Web interface display

[0835] ```

[0836] By implementing these technical details, the embodiments of the present invention can:

[0837] 1. Realize the automated parametric design of gears and splines to quickly respond to customers' customized needs.

[0838] 2. Seamlessly integrate design and manufacturing processes, reduce manual intervention and improve production efficiency.

[0839] 3. Intelligently schedule production resources, optimize production processes, and improve the efficiency and flexibility of the overall production line.

[0840] Embodiment 2

[0841] like Figure 5 As shown, the embodiment of the present application also provides a gear and spline processing intelligent design and manufacturing system 50, including:

[0842] Establishing module 51, used to establish a hybrid intelligent model, wherein the hybrid intelligent model includes a large language model and a localized small model;

[0843] A design module 52, for designing process parameters of gears and splines using the hybrid intelligent model;

[0844] The simulation module 53 is used to build a digital twin system and realize real-time synchronization between the virtual model and the physical device;

[0845] An optimization module 54, used to optimize the process parameters of machining the gears and splines by using reinforcement learning technology in the digital twin system;

[0846] A processing module 55, for processing the gear and the spline based on the optimized process parameters of the gear and the spline;

[0847] The detection module 56 is used to apply machine learning technology to perform quality detection of the gears and splines.

[0848] Embodiment 3

[0849] The present application also provides a computer device, such as Figure 6 As shown, the computer device comprises:

[0850] at least one processor; and,

[0851] a memory communicatively connected to the at least one processor; wherein,

[0852] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned intelligent design and manufacturing method for gear and spline processing.

[0853] An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions are used to enable a computer to execute the above-mentioned intelligent design and manufacturing method for gear and spline processing.

[0854] An embodiment of the present application also provides a computer program product, including computer instructions, characterized in that when the computer instructions are executed by a processor, the steps of the above-mentioned intelligent design and manufacturing method for gear and spline processing are implemented.

[0855] Each module in the above gear and spline processing intelligent design and manufacturing system can be implemented in whole or in part by software, hardware and their combination. Each module can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the corresponding operations of each module.

[0856] In one embodiment, a computer device is also provided. The computer device is the gear and spline processing intelligent design and manufacturing system mentioned in the above method embodiment. Its internal structure diagram can be as follows: Figure 6 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface.

[0857] Among them, the processor of the computer device is used to provide computing and control capabilities, and can be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto. The processor may include one or more processors, for example, one or more central processing units (CPUs). When the processor is a CPU, the CPU may be a single-core CPU or a multi-core CPU. The processor may also include one or more special-purpose processors, and the special-purpose processor may include a GPU, an FPGA, etc., for accelerated processing. The processor is used to call the program code and data in the memory and execute the steps in the above method embodiment. For details, please refer to the description in the method embodiment, which will not be repeated here.

[0858] The memory of the computer device includes, but is not limited to, non-volatile storage media and internal memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium.

[0859] The input / output interface of the computer device is used to exchange information between the processor and external devices.

[0860] The communication interface of the computer device is used to communicate with an external terminal via a network connection.

[0861] When the computer program is executed by a processor, a method for seafloor geological imaging is implemented.

[0862] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the division of each unit / module is only a logical function division, and there may be other division methods in actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. The mutual coupling, direct coupling, or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of systems or units, which can be electrical, mechanical or other forms.

[0863] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0864] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable systems. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from a website site, computer, server or data center to another website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (digital subscriber line, DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media integrations. The available medium may be a read-only memory (ROM), or a random access memory (RAM), or a magnetic medium, such as a floppy disk, a hard disk, a tape, a disk, or an optical medium, such as a digital versatile disc (DVD), or a semiconductor medium, such as a solid state disk (SSD), etc.

[0865] The system of the embodiments of the present disclosure can execute the method provided by the embodiments of the present disclosure, and the implementation principles are similar. The actions performed by each module in the system of each embodiment of the present disclosure correspond to the steps in the method of each embodiment of the present disclosure. For the detailed functional description of each module of the system, please refer to the description in the corresponding method shown in the previous text, which will not be repeated here.

[0866] The above is only an optional implementation method for some implementation scenarios of the present disclosure. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical concept of the scheme of the present disclosure, other similar implementation methods based on the technical ideas of the present disclosure are also within the protection scope of the embodiments of the present disclosure.

Claims

1. An intelligent design and manufacturing method for gear and spline processing, characterized in that: The following steps are involved: Establishing a hybrid intelligent model, wherein the hybrid intelligent model includes a large language model and a localized small model; Designing process parameters of gears and splines using the hybrid intelligent model; Build a digital twin system to achieve real-time synchronization between virtual models and physical devices; In the digital twin system, the process parameters for machining the gears and splines are optimized using reinforcement learning technology; Processing the gear and the spline based on the optimized process parameters of the gear and the spline; Machine learning techniques are applied to perform quality inspection of the gears and splines.

2. The method according to claim 1, characterized in that: The hybrid intelligent model is established, comprising: Choose a large language model; By using knowledge distillation technology, the large language model is distilled into a localized small model; Alignment of the large language model and the localized small model is achieved.

3. The method according to claim 2, characterized in that The step of realizing the alignment of the large language model and the localized small model comprises the following steps: Establishing a dynamic system model, and using a state space model to describe the knowledge conversion process from the large language model to the localized small model; Gumbel-Softmax and direct estimator are used to deal with the discreteness problem in the knowledge conversion process; Use multi-head attention mechanism and graph neural network to deal with the semantic coupling problem in the process of knowledge conversion; Apply deep residual networks and nonlinear activation functions to handle nonlinear problems in the knowledge conversion process; Use autoencoders and sparse attention mechanisms to handle the high-dimensionality problem of knowledge representation.

4. The method according to claim 1, characterized in that: The construction of the digital twin system includes: Establishing a virtual model of the gear and the spline; Develop real-time data acquisition systems; Realize real-time synchronization between virtual models and physical devices.

5. The method according to claim 4, characterized in that The method further includes: applying a general matrix multiplication (GEMM) optimization technique in the digital twin system, comprising the following steps: Identify the matrix operations in the digital twin system, including geometric transformation, finite element analysis, dynamics simulation, and machine learning model reasoning; Implementing cache optimization on the identified matrix operation links, including matrix blocking and data pre-fetching; Apply SIMD instruction set optimization in the matrix operation link; Implement multi-threaded parallelization for large-scale matrix operations; GPU acceleration is implemented for finite element analysis tasks; Develop an adaptive GEMM optimization strategy to dynamically select the optimal GEMM implementation based on the input data size and available hardware resources.

6. The method according to claim 1, characterized in that The step of optimizing the process parameters for machining the gears and splines using reinforcement learning technology comprises: Establish a processing environment model, define the state space including process parameters, the action space of adjustable parameters of processing equipment, and the reward function based on processing quality and efficiency; Select the DQN model as the reinforcement learning model; Designing state-action value functions for estimating process parameter adjustments in a simulation environment; In a simulation environment, training the reinforcement learning model based on the state-action value function; The trained reinforcement learning model is used to optimize the process parameters for machining the gears and splines.

7. The method according to claim 6, characterized in that The method further comprises the following steps: Apply target network technology to enhance the stability of the training process; Develop a multi-agent reinforcement learning framework to optimize parameters of multiple processing equipment simultaneously; Integrate expert knowledge and accelerate the learning process by introducing rule-based constraints.

8. The method according to claim 1, characterized in that Apply machine learning technology to perform quality inspection of the gears and splines, including: Acquire images of the gear and spline in real time; Using an image processing algorithm, extracting features from the images of the gears and splines; Build and train convolutional neural networks; The extracted features are input into the convolutional neural network to achieve defect detection and classification of the gears and splines.

9. The method according to claim 1, characterized in that: The following steps are also included: Developing parametric design capabilities in the hybrid intelligent model to achieve automated customization of gears and splines; Based on the parametric design using the hybrid intelligent model, the CAD / CAM system is integrated to achieve seamless connection from design to manufacturing; Develop intelligent scheduling systems to optimize production processes and resource allocation.

10. An intelligent design and manufacturing system for gear and spline processing, characterized in that: include: An establishment module is used to establish a hybrid intelligent model, wherein the hybrid intelligent model includes a large language model and a localized small model; A design module, used for designing process parameters of gears and splines using the hybrid intelligent model; The simulation module is used to build a digital twin system and achieve real-time synchronization between the virtual model and the physical device; An optimization module, used to optimize the process parameters of machining the gears and splines by using reinforcement learning technology in the digital twin system; A processing module, used for processing the gear and the spline based on the optimized process parameters of the gear and the spline; A detection module is used to apply machine learning technology to perform quality detection on the gears and splines.

Images