Aluminum busbar machining parameter dynamic optimization system based on deep learning

Abstract

The invention relates to the field of aluminum bus processing intelligent control, in particular to an aluminum bus processing parameter dynamic optimization system based on deep learning, which comprises a server, a data preprocessing module, a pre-training module, a real-time optimization module, a data interaction module and a model iteration module. The data preprocessing module collects and processes related data and constructs a training data set; the pre-training module establishes a mapping relation between working conditions and processing parameters through a lightweight full-connection neural network, and outputs initial parameters; the real-time optimization module feeds back dynamic correction parameters in real time according to the numerical control system based on a deep reinforcement learning DQN algorithm; the data interaction module realizes two-way data transmission; and the model iteration module regularly fine-tunes model parameters. The system can adapt to dynamic working conditions, rapidly output optimal machining parameters, remarkably reduce machining deviation, improve production efficiency and reduce generation of waste products, and is adaptive to a mainstream numerical control system.

Description

Technical Field

[0001] This invention relates to the field of intelligent control for aluminum busbar processing, specifically to a deep learning-based dynamic optimization system for aluminum busbar processing parameters. Background Technology

[0002] As a core conductive component of power systems, the processing quality of aluminum busbars directly affects the stability of power transmission. During the processing of aluminum busbars, the settings of parameters such as cutting current and feed rate play a decisive role in dimensional accuracy and surface quality.

[0003] In existing technologies, the setting of aluminum busbar processing parameters suffers from the following shortcomings: First, the use of a fixed parameter table binds processing parameters to simple dimensions such as aluminum busbar thickness, failing to consider dynamic factors such as material fluctuations and changes in ambient temperature, leading to large fluctuations in processing deviations. Second, it relies on manual trial cutting and adjustment by technicians, requiring repeated testing for 1-2 hours under new working conditions, which is not only inefficient but also results in a significant waste of raw materials. Third, it lacks a real-time feedback mechanism; when encountering unexpected situations such as power grid voltage fluctuations or uneven material distribution during processing, parameters cannot be adjusted in a timely manner, easily leading to batch scrap. Therefore, there is an urgent need for a lightweight, rapidly deployable dynamic optimization system for processing parameters to solve the problems of fixed parameters, low efficiency, and poor adaptability in existing technologies. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide a deep learning-based dynamic optimization system for aluminum busbar processing parameters, solving the problems of fixed parameters, low efficiency, and poor adaptability in existing technologies, and enabling intelligent generation and dynamic correction of processing parameters.

[0005] The basic solution provided by this invention is a deep learning-based dynamic optimization system for aluminum busbar processing parameters, including a server. The server includes a data preprocessing module, a pre-training module, a real-time optimization module, a data interaction module, and a model iteration module. The data preprocessing module is used to collect aluminum busbar processing data, perform data cleaning, standardization, and labeling, and construct a training dataset. The aluminum busbar processing data includes raw operating condition data, raw processing parameter data, and quality data. The raw operating condition data includes the aluminum busbar's thermal expansion coefficient, thermal conductivity, thickness, length, surface roughness requirements, and ambient temperature, which are processed to form a operating condition feature vector. The raw processing parameter data includes processing current, feed rate, and processing time, which are processed to form a processing parameter vector. The quality data includes dimensional deviation data and surface roughness data. The labels are the optimal processing parameter vectors corresponding to each set of raw operating condition data, obtained by filtering the raw processing parameter data and quality data. The pre-training module is used to train a lightweight fully connected neural network based on a training dataset. The training dataset consists of working condition feature vectors and corresponding optimal processing parameter vectors obtained by processing historical data through a data preprocessing module. A mapping relationship between the working condition feature vectors and the processing parameter vectors is established, and the initial processing parameter vector is output after receiving the real-time working condition feature vector. The real-time optimization module is used to dynamically correct the initial machining parameter vector based on the real-time operating status data fed back by the CNC system using the deep reinforcement learning DQN algorithm, and output parameter fine-tuning instructions. The operating status data includes the pre-trained output initial machining parameter vector, the actual machining parameter deviation of the CNC system, the spindle load fluctuation of the equipment, and the sudden working condition indicator. The sudden working condition indicator includes voltage fluctuation indicator and material unevenness indicator. The data interaction module is used to realize bidirectional data transmission between the server and the CNC system, receive the original working condition data, original machining parameter data, quality data and operating status data uploaded by the CNC system, and issue the initial machining parameter vector and parameter fine-tuning instructions. The model iteration module is used to collect a complete dataset of real-time working condition feature vectors, corrected processing parameter vectors, and quality data for each processing operation, and periodically incrementally fine-tunes the model parameters of the pre-training module and the real-time optimization module to maintain mapping accuracy and correct response speed.

[0006] The principle of this invention is as follows: The data preprocessing module comprehensively collects working conditions, processing parameters, and quality data during the aluminum busbar processing. After standardization and cleaning, a high-quality training dataset is constructed, generating real-time working condition feature vectors that can be directly input into the model. The pre-training module uses historical data to train a lightweight fully connected neural network, learning the intrinsic mapping relationship between working conditions and optimal processing parameters, and realizing the rapid output of initial processing parameters under new working conditions. The real-time optimization module, based on the deep reinforcement learning DQN algorithm, continuously receives actual operating status data (including parameter deviations, equipment load, and sudden working conditions) from the CNC system, dynamically corrects the initial processing parameters, and ensures that the parameters always adapt to the actual working conditions during processing. The data interaction module acts as a communication bridge between the system and the CNC system, ensuring the real-time performance and reliability of data transmission. Finally, the model iteration module continuously collects complete data from each processing operation, periodically fine-tunes the model parameters, and continuously optimizes the model mapping accuracy and real-time correction capability as it is used, forming an adaptive closed loop.

[0007] The beneficial effects of this invention are as follows: By learning the dynamic mapping relationship between working conditions and processing parameters through a fully connected neural network, and combining it with the dynamic correction of the real-time optimization module, it replaces the fixed parameter table mode and can adapt to dynamic factors such as fluctuations in aluminum busbar material and changes in ambient temperature, significantly reducing processing deviation fluctuations; the pre-training module can directly output the initial processing parameters under new working conditions, eliminating the need for manual trial cutting and adjustment by technicians, shortening the debugging time for new working conditions from 1-2 hours to minutes, and reducing raw material waste; the real-time optimization module, based on real-time feedback data from the CNC system, quickly outputs parameter fine-tuning instructions, which can effectively cope with sudden working conditions such as power grid voltage fluctuations and material inhomogeneity, avoiding the generation of batch scrap.

[0008] Furthermore, the data preprocessing module includes a data cleaning unit, a feature standardization unit, and a label generation unit. The data cleaning unit is used to remove collected data corresponding to processing interruptions and quality inspection failures, and to use the mean to complete the missing original working condition data and original processing parameter data. The feature standardization unit is used to standardize the original working condition data, unify the data distribution range, and generate standardized working condition feature vectors. It also performs validity verification on the original processing parameter data, retaining processing parameter vectors that conform to the equipment's operating range. The label generation unit is used to select the optimal processing parameter vector from multiple sets of processing parameter vectors corresponding to each set of original working condition data. The optimal vector has a size deviation less than or equal to a size deviation threshold and a surface roughness less than or equal to a surface roughness threshold, and uses it as the training label for that set of working conditions.

[0009] The data cleaning unit ensures the integrity and reliability of the collected data by removing abnormal data and filling in missing data, thus avoiding interference from abnormal data on model training. The feature standardization unit eliminates the dimensional differences of data in different dimensions by unifying the data distribution range, thereby improving the convergence speed and fitting accuracy of model training. The label generation unit selects the optimal processing parameters that meet the quality threshold as training labels, enabling the model to learn the accurate mapping relationship between working conditions and optimal processing parameters.

[0010] Furthermore, the lightweight fully connected neural network of the pre-training module includes an input layer, three hidden layers, and an output layer; the input layer is a standardized working condition feature vector; the number of neurons in the hidden layer are 32, 16, and 8, respectively, and the activation function is ReLU, which is used to extract the nonlinear correlation between working condition features and processing parameters; the output layer is a processing parameter vector, and the activation function is Linear.

[0011] Employing a lightweight structure and a stepped neuron count design of 32, 16, and 8, the model minimizes computational and storage requirements while ensuring the ability to extract nonlinear correlations between working condition features and machining parameters. This allows it to adapt to the embedded hardware environment of CNC systems and ensures that the model inference latency meets real-time machining requirements.

[0012] Furthermore, the dimensional deviation threshold is 0.05 mm, and the surface roughness threshold is 1.2 mm. The real-time optimization module includes a status acquisition unit, an action decision unit, and a reward calculation unit. The status acquisition unit collects real-time operating status data through a data interaction module, where the machining parameter deviation is the difference between the actual operating parameters of the CNC system and the corresponding parameters in the initial machining parameter vector. The action decision unit sets a discrete action space, with each dimension corresponding to five fine-tuning values ​​for one machining parameter: machining current fine-tuning values ​​are -5A, -3A, 0A, +3A, +5A; feed rate fine-tuning values ​​are -10mm / min, -5mm / min, 0mm / min, +5mm / min, +10mm / min; and machining time fine-tuning values ​​are -0.5s, -0.3s, 0s, +0.3s, +0.5s, forming a total of 125 action combinations. The reward calculation unit constructs a multi-dimensional reward function R = R1 + R2 + R3, where R1 is the quality reward: when the dimensional deviation is less than or equal to 0.05mm and the surface roughness is less than or equal to 1.2. When m, R1 = +10; when the dimensional deviation is greater than 0.05 mm and less than or equal to 0.1 mm, or the surface roughness is greater than 1.2 mm. m and less than or equal to 1.5 When the dimensional deviation is greater than 0.1 mm or the surface roughness is greater than 1.5 mm, R1 = +5; when the dimensional deviation is greater than 0.1 mm or the surface roughness is greater than 1.5 mm... When the feed rate is m, R1 = -20; R2 is the efficiency bonus: when the feed rate is greater than or equal to 90% of the feed rate in the initial machining parameter vector, R2 = +3; otherwise, R2 = 0; R3 is the penalty for sudden working conditions: when a sudden working condition occurs and no parameter fine-tuning command is output, R3 = -15; when a sudden working condition occurs and a parameter fine-tuning command is output, the dimensional deviation recovers to less than or equal to 0.05 mm and the surface roughness recovers to less than or equal to 1.5 mm. When m, R3 = +8.

[0013] The design of the discrete action space balances the accuracy and computational efficiency of parameter adjustment. The 125 action combinations can cover the parameter adjustment needs under most working conditions. The multi-dimensional reward function, through the synergistic effect of quality reward, efficiency reward and sudden working condition penalty, guides the model to prioritize ensuring the processing quality meets the standards while also taking into account processing efficiency. At the same time, it forces the model to respond to sudden working conditions in a timely manner, ensuring the comprehensiveness and effectiveness of parameter adjustment.

[0014] Furthermore, the incremental fine-tuning cycle of the model iteration module is that after processing 100 aluminum busbars, the newly collected complete dataset is automatically added to the training set after being processed by the data preprocessing module, and the weight parameters of the lightweight fully connected neural network and the Q network parameters of the DQN algorithm are fine-tuned simultaneously.

[0015] By setting a fixed iteration cycle, the newly collected complete dataset includes real-time operating conditions, corrected parameters, and quality results, which can reflect the dynamic changes of the actual processing scenario. Through incremental fine-tuning, the model can continuously adapt to the operating characteristics of specific processing equipment, tool wear patterns, and material differences of different batches of aluminum busbars, ensuring that the model mapping accuracy and correction response speed remain stable over the long term. Attached Figure Description

[0016] Figure 1 This is a system block diagram of an embodiment of the deep learning-based aluminum busbar processing parameter dynamic optimization system of the present invention. Detailed Implementation

[0017] The following detailed description illustrates the specific implementation method: The basic implementation examples are as follows: Figure 1 As shown: A deep learning-based dynamic optimization system for aluminum busbar processing parameters includes a server, which includes a data preprocessing module, a pre-training module, a real-time optimization module, a data interaction module, and a model iteration module. The data preprocessing module is used to collect aluminum busbar processing data, perform data cleaning, standardization, and labeling, and construct a training dataset. The aluminum busbar processing data includes raw operating condition data, raw processing parameter data, and quality data. The raw operating condition data includes the aluminum busbar's thermal expansion coefficient, thermal conductivity, thickness, length, surface roughness requirements, and ambient temperature, which are processed to form a operating condition feature vector. The raw processing parameter data includes processing current, feed rate, and processing time, which are processed to form a processing parameter vector. The quality data includes dimensional deviation data and surface roughness data. The labels are the optimal processing parameter vectors corresponding to each set of raw operating condition data, obtained by filtering the raw processing parameter data and quality data. The pre-training module is used to train a lightweight fully connected neural network based on a training dataset. The training dataset consists of working condition feature vectors and corresponding optimal processing parameter vectors obtained by processing historical data through a data preprocessing module. A mapping relationship between the working condition feature vectors and the processing parameter vectors is established, and the initial processing parameter vector is output after receiving the real-time working condition feature vector. The real-time optimization module is used to dynamically correct the initial machining parameter vector based on the real-time operating status data fed back by the CNC system using the deep reinforcement learning DQN algorithm, and output parameter fine-tuning instructions. The operating status data includes the pre-trained output initial machining parameter vector, the actual machining parameter deviation of the CNC system, the spindle load fluctuation of the equipment, and the sudden working condition indicator. The sudden working condition indicator includes voltage fluctuation indicator and material unevenness indicator. The data interaction module is used to realize bidirectional data transmission between the server and the CNC system, receive the original working condition data, original machining parameter data, quality data and operating status data uploaded by the CNC system, and issue the initial machining parameter vector and parameter fine-tuning instructions. The model iteration module is used to collect a complete dataset of real-time working condition feature vectors, corrected processing parameter vectors, and quality data for each processing operation, and periodically incrementally fine-tunes the model parameters of the pre-training module and the real-time optimization module to maintain mapping accuracy and correct response speed.

[0018] Furthermore, the data preprocessing module includes a data cleaning unit, a feature standardization unit, and a label generation unit. The data cleaning unit is used to remove collected data corresponding to processing interruptions and quality inspection failures, and to use the mean to complete the missing original working condition data and original processing parameter data. The feature standardization unit is used to standardize the original working condition data, unify the data distribution range, and generate standardized working condition feature vectors. It also performs validity verification on the original processing parameter data, retaining processing parameter vectors that conform to the equipment's operating range. The label generation unit is used to select the optimal processing parameter vector from multiple sets of processing parameter vectors corresponding to each set of original working condition data. The optimal vector has a size deviation less than or equal to a size deviation threshold and a surface roughness less than or equal to a surface roughness threshold, and uses it as the training label for that set of working conditions.

[0019] Furthermore, the lightweight fully connected neural network of the pre-training module includes an input layer, three hidden layers, and an output layer; the input layer is a standardized working condition feature vector; the number of neurons in the hidden layer are 32, 16, and 8, respectively, and the activation function is ReLU, which is used to extract the nonlinear correlation between working condition features and processing parameters; the output layer is a processing parameter vector, and the activation function is Linear.

[0020] Furthermore, the dimensional deviation threshold is 0.05 mm, and the surface roughness threshold is 1.2 mm. The real-time optimization module includes a status acquisition unit, an action decision unit, and a reward calculation unit. The status acquisition unit collects real-time operating status data through a data interaction module, where the machining parameter deviation is the difference between the actual operating parameters of the CNC system and the corresponding parameters in the initial machining parameter vector. The action decision unit sets a discrete action space, with each dimension corresponding to five fine-tuning values ​​for one machining parameter: machining current fine-tuning values ​​are -5A, -3A, 0A, +3A, +5A; feed rate fine-tuning values ​​are -10mm / min, -5mm / min, 0mm / min, +5mm / min, +10mm / min; and machining time fine-tuning values ​​are -0.5s, -0.3s, 0s, +0.3s, +0.5s, forming a total of 125 action combinations. The reward calculation unit constructs a multi-dimensional reward function R = R1 + R2 + R3, where R1 is the quality reward: when the dimensional deviation is less than or equal to 0.05mm and the surface roughness is less than or equal to 1.2. When m, R1 = +10; when the dimensional deviation is greater than 0.05 mm and less than or equal to 0.1 mm, or the surface roughness is greater than 1.2 mm. m and less than or equal to 1.5 When the dimensional deviation is greater than 0.1 mm or the surface roughness is greater than 1.5 mm, R1 = +5; when the dimensional deviation is greater than 0.1 mm or the surface roughness is greater than 1.5 mm... When the feed rate is m, R1 = -20; R2 is the efficiency bonus: when the feed rate is greater than or equal to 90% of the feed rate in the initial machining parameter vector, R2 = +3; otherwise, R2 = 0; R3 is the penalty for sudden working conditions: when a sudden working condition occurs and no parameter fine-tuning command is output, R3 = -15; when a sudden working condition occurs and a parameter fine-tuning command is output, the dimensional deviation recovers to less than or equal to 0.05 mm and the surface roughness recovers to less than or equal to 1.5 mm. When m, R3 = +8.

[0021] Furthermore, the incremental fine-tuning cycle of the model iteration module is that after processing 100 aluminum busbars, the newly collected complete dataset is automatically added to the training set after being processed by the data preprocessing module, and the weight parameters of the lightweight fully connected neural network and the Q network parameters of the DQN algorithm are fine-tuned simultaneously.

[0022] In this embodiment, the system is deployed on an industrial-grade server, and the data interaction module uses the Profinet protocol to achieve bidirectional communication with the CNC system. The data preprocessing module collects 3000 sets of historical processing data (covering working conditions such as aluminum busbar thickness 5-50mm, ambient temperature 15-35℃, and processing current 80-150A). The data cleaning unit removes 280 sets of abnormal data due to processing interruptions and quality inspection failures, and uses the mean to complete 120 sets of missing data. The feature standardization unit processes the original working condition data using the Z-score standardization method, ensuring that the mean of each dimension is 0 and the standard deviation is 1. Simultaneously, the validity of the original processing parameter data is verified, removing parameter vectors that exceed the equipment's operating range (processing current 50-200A, feed speed 100-500mm / min, processing time 0.5-5s). The label generation unit filters out parameters with dimensional deviations less than or equal to 0.05mm and surface roughness less than or equal to 1.2mm. The optimal processing parameter vector of m is used as the label, and finally 2840 sets of effective training datasets are constructed.

[0023] The pre-training module uses the Python+PyTorch framework to train a lightweight fully connected neural network, setting the number of iterations to 200 rounds and the learning rate to 0.001 (decreasing by 10% every 50 rounds). After training, the model is converted to TensorFlow Lite format. The real-time optimization module's state acquisition unit collects CNC system operating status data every 500ms. When a sudden voltage fluctuation is detected (the actual processing current decreases by 10% compared to the initial parameters), the action decision unit selects a fine-tuning instruction of +5A processing current from 125 action combinations. Since the adjusted dimensional deviation is restored to 0.04mm, the reward calculation unit outputs a total reward of R1=+10, R2=+3, and R3=+8, guiding the model to strengthen the adjustment strategy.

[0024] After each 100 aluminum busbars are machined, the model iteration module automatically adds the newly collected complete dataset to the training set for incremental fine-tuning. After three iterations, the model's prediction accuracy for the initial parameters under new working conditions improves by 15%, and the real-time corrected machining dimensional deviation stabilizes at less than or equal to 0.04 mm, with a surface roughness less than or equal to 1.2 mm. m.

[0025] This embodiment achieves intelligent dynamic optimization of aluminum busbar processing parameters through complete module collaboration and parameter design, which not only ensures the stability of processing quality but also improves production efficiency. Furthermore, it has low deployment cost, strong adaptability, and can be widely applied to various CNC processing scenarios for aluminum busbars.

[0026] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A deep learning-based dynamic optimization system for aluminum busbar processing parameters, characterized in that: The server includes a data preprocessing module, a pre-training module, a real-time optimization module, a data interaction module, and a model iteration module. The data preprocessing module is used to collect aluminum busbar processing data, perform data cleaning, standardization, and labeling, and construct a training dataset. The aluminum busbar processing data includes raw operating condition data, raw processing parameter data, and quality data. The raw operating condition data includes the aluminum busbar's thermal expansion coefficient, thermal conductivity, thickness, length, surface roughness requirements, and ambient temperature, which are processed to form a operating condition feature vector. The raw processing parameter data includes processing current, feed rate, and processing time, which are processed to form a processing parameter vector. The quality data includes dimensional deviation data and surface roughness data. The labels are the optimal processing parameter vectors corresponding to each set of raw operating condition data, obtained by filtering the raw processing parameter data and quality data. The pre-training module is used to train a lightweight fully connected neural network based on a training dataset. The training dataset consists of working condition feature vectors and corresponding optimal processing parameter vectors obtained by processing historical data through a data preprocessing module. A mapping relationship between the working condition feature vectors and the processing parameter vectors is established, and the initial processing parameter vector is output after receiving the real-time working condition feature vector. The real-time optimization module is used to dynamically correct the initial machining parameter vector based on the real-time operating status data fed back by the CNC system using the deep reinforcement learning DQN algorithm, and output parameter fine-tuning instructions. The operating status data includes the pre-trained output initial machining parameter vector, the actual machining parameter deviation of the CNC system, the spindle load fluctuation of the equipment, and the sudden working condition indicator. The sudden working condition indicator includes voltage fluctuation indicator and material unevenness indicator. The data interaction module is used to realize bidirectional data transmission between the server and the CNC system, receive the original working condition data, original machining parameter data, quality data and running status data uploaded by the CNC system, and issue the initial machining parameter vector and parameter fine-tuning instructions. The model iteration module is used to collect a complete dataset of real-time working condition feature vectors, corrected processing parameter vectors, and quality data for each processing operation, and periodically incrementally fine-tunes the model parameters of the pre-training module and the real-time optimization module to maintain mapping accuracy and correct response speed.

2. The deep learning-based dynamic optimization system for aluminum busbar processing parameters according to claim 1, characterized in that: The data preprocessing module includes a data cleaning unit, a feature standardization unit, and a label generation unit. The data cleaning unit removes collected data corresponding to processing interruptions and quality inspection failures, and uses the mean to complete missing original operating condition data and original processing parameter data. The feature standardization unit standardizes the original operating condition data, unifies the data distribution range, and generates standardized operating condition feature vectors. It also verifies the validity of the original processing parameter data, retaining processing parameter vectors that conform to the equipment's operating range. The label generation unit selects the optimal processing parameter vector from multiple sets of processing parameter vectors corresponding to each set of original operating condition data. This optimal vector has a size deviation less than or equal to a size deviation threshold and a surface roughness less than or equal to a surface roughness threshold, and uses it as the training label for that set of operating conditions.

3. The deep learning-based dynamic optimization system for aluminum busbar processing parameters according to claim 2, characterized in that: The lightweight fully connected neural network of the pre-training module includes an input layer, three hidden layers, and an output layer. The input layer is a standardized working condition feature vector. The number of neurons in the hidden layers are 32, 16, and 8, respectively, and the activation function is ReLU, which is used to extract the nonlinear correlation between working condition features and processing parameters. The output layer is a processing parameter vector, and the activation function is Linear.

4. The deep learning-based dynamic optimization system for aluminum busbar processing parameters according to claim 3, characterized in that: The dimensional deviation threshold is 0.05 mm, and the surface roughness threshold is 1.2 mm. The real-time optimization module includes a status acquisition unit, an action decision unit, and a reward calculation unit. The status acquisition unit collects real-time operating status data through a data interaction module, where the machining parameter deviation is the difference between the actual operating parameters of the CNC system and the corresponding parameters in the initial machining parameter vector. The action decision unit sets a discrete action space, with each dimension corresponding to five fine-tuning values ​​for one machining parameter: machining current fine-tuning values ​​are -5A, -3A, 0A, +3A, +5A; feed rate fine-tuning values ​​are -10mm / min, -5mm / min, 0mm / min, +5mm / min, +10mm / min; and machining time fine-tuning values ​​are -0.5s, -0.3s, 0s, +0.3s, +0.5s, forming a total of 125 action combinations. The reward calculation unit constructs a multi-dimensional reward function R = R1 + R2 + R3, where R1 is the quality reward: when the dimensional deviation is less than or equal to 0.05mm and the surface roughness is less than or equal to 1.

2. When m, R1 = +10; when the dimensional deviation is greater than 0.05 mm and less than or equal to 0.1 mm, or the surface roughness is greater than 1.2 mm. m and less than or equal to 1.5 When the dimensional deviation is greater than 0.1 mm or the surface roughness is greater than 1.5 mm, R1 = +5; when the dimensional deviation is greater than 0.1 mm or the surface roughness is greater than 1.5 mm... When the feed rate is m, R1 = -20; R2 is the efficiency bonus: when the feed rate is greater than or equal to 90% of the feed rate in the initial machining parameter vector, R2 = +3; otherwise, R2 = 0; R3 is the penalty for sudden working conditions: when a sudden working condition occurs and no parameter fine-tuning command is output, R3 = -15; when a sudden working condition occurs and a parameter fine-tuning command is output, the dimensional deviation recovers to less than or equal to 0.05 mm and the surface roughness recovers to less than or equal to 1.5 mm. When m, R3 = +8.

5. The deep learning-based dynamic optimization system for aluminum busbar processing parameters according to claim 4, characterized in that: The incremental fine-tuning cycle of the model iteration module is as follows: after processing 100 aluminum busbars, the newly collected complete dataset is automatically added to the training set after being processed by the data preprocessing module, and the weight parameters of the lightweight fully connected neural network and the Q network parameters of the DQN algorithm are fine-tuned simultaneously.

Images