Intelligent surface roughness classification system

Abstract

The invention discloses a surface roughness intelligent classification system, and mainly relates to the technical field of intelligent monitoring. Comprising the steps of collecting multi-channel vibration signals in lathe machining; capturing the multi-channel vibration signals generated in the machining process of the numerical control lathe in real time through a high-precision sensor; preprocessing the multichannel original vibration signals, wherein the preprocessing comprises denoising, filtering and normalization processing, and the denoising refers to removing noise interference in the signals through a digital filtering technology; the filtering is to select a proper frequency range and reserve vibration information related to the surface roughness; the normalization processing is to adjust the signal amplitude to a unified range so as to facilitate subsequent processing. The method provided by the invention has the beneficial effects that on-line monitoring is carried out by utilizing the vibration signal without additional measurement equipment, so that the measurement cost and the time cost are reduced. Meanwhile, due to the fact that real-time monitoring and feedback can be achieved through the method, machining parameters can be adjusted in time, and the rejection rate and the production cost are further reduced.

Description

Technical Field

[0001] The present invention relates to the field of intelligent monitoring technology, and specifically to an intelligent classification system for surface roughness. Background Art

[0002] The statements in this section merely provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] In the machinery manufacturing industry, surface roughness is one of the important parameters for measuring the surface quality of workpieces, directly affecting the mechanical properties, wear resistance, mating accuracy and service life of products. Although traditional surface roughness detection methods can provide accurate measurement results to a certain extent, they have many limitations. For example, the measurement speed is slow, and there is slight damage to the workpiece surface, which is not suitable for the measurement of high-precision or vulnerable workpieces.

[0004] With the development of intelligent manufacturing technology, the online monitoring method based on vibration signals has gradually become a research hotspot. Vibration signals are inherent signals generated during the machining process, which are closely related to the machining state of the workpiece surface. By analyzing the characteristic parameters such as the frequency and amplitude of vibration signals, the surface roughness of the workpiece can be indirectly reflected. However, most of the existing surface roughness monitoring methods based on vibration signals are only optimized for a single task. This single-task monitoring method results in limited model generalization ability and is difficult to adapt to complex and changeable machining environments.

[0005] In addition, in terms of feature extraction and classification algorithms, existing methods still have problems such as insufficient accuracy and poor robustness. Feature extraction is a key step in vibration signal analysis, directly related to the accuracy and stability of subsequent classification algorithms.

[0006] In order to overcome the deficiencies of the prior art, the present invention proposes a surface roughness classification method based on multi-task learning and vibration signals. This method collects multi-channel vibration signals during the machining process of a CNC lathe, uses advanced signal processing techniques to extract effective feature information, and constructs a multi-task learning network model for training and classification. Specifically, the present invention uses a Relative Position Matrix (RPM) to convert one-dimensional vibration signals into two-dimensional images, and then uses two-dimensional discrete wavelet transform (2D-DWT) to decompose and reconstruct the images to extract rich feature information. On this basis, the present invention introduces a Global Attention Mechanism (GAM) into the DenseNet model to construct a GAMDenseNet network as the encoder of the model to enhance the model's ability to extract key features. At the same time, a Channel Attention (CA) mechanism is incorporated into the decoder part to improve the model's ability to screen useful features for different tasks. This multi-task learning network model can simultaneously consider the monitoring of surface roughness and other related quality indicators, improving the model's generalization ability and robustness.

[0007] Furthermore, the present invention adopts the Gradient Normalization algorithm (GradNorm) to adaptively adjust the weight ratio of the loss functions of the two tasks of surface roughness recognition and dimensional accuracy prediction, so as to optimize the model training process. By dynamically adjusting the weight ratio, the model can better balance the learning effects of the two tasks during the training process, further improving the classification accuracy. Summary of the Invention

[0008] The object of the present invention is to provide an intelligent surface roughness classification system, which collects multi-channel vibration signals during the machining process of a numerical control lathe, extracts effective feature information by using advanced signal processing techniques, and constructs a multi-task learning network model for training and classification. Specifically, the present invention uses the Relative Position Matrix (RPM) to convert one-dimensional vibration signals into two-dimensional images, and then uses two-dimensional discrete wavelet transform (2D-DWT) to decompose and reconstruct the images to extract rich feature information. On this basis, the present invention introduces the Global Attention Mechanism (GAM) into the DenseNet model to construct the GAMDenseNet network as the encoder of the model to enhance the model's ability to extract key features. At the same time, the Channel Attention (CA) mechanism is incorporated into the decoder part to improve the model's ability to screen useful features for different tasks. This multi-task learning network model can simultaneously consider the monitoring of surface roughness and other related quality indicators, improving the generalization ability and robustness of the model.

[0009] To achieve the above object, the present invention is realized through the following technical solutions:

[0010] An intelligent surface roughness classification system includes the following steps:

[0011] Collect multi-channel vibration signals during lathe machining: Real-time capture multi-channel vibration signals generated by a numerical control lathe during the machining process through high-precision sensors;

[0012] Preprocess the multi-channel original vibration signals: including denoising, filtering, and normalization processing. Denoising is to remove noise interference in the signal through digital filtering techniques; filtering is to select an appropriate frequency range to retain vibration information related to surface roughness; normalization processing is to adjust the signal amplitude to a unified range for subsequent processing;

[0013] Convert the preprocessed one-dimensional vibration signals into two-dimensional images using the relative position matrix: The relative position matrix method can effectively convert one-dimensional signals into two-dimensional images, thus more intuitively displaying the feature information in the signals;

[0014] The two-dimensional discrete wavelet transform is used to decompose and reconstruct the two-dimensional image to extract feature information: The two-dimensional discrete wavelet transform can decompose the image into wavelet coefficients of different frequencies and directions. By selecting appropriate wavelet coefficients for reconstruction, feature information closely related to surface roughness can be extracted;

[0015] Construct a multi-task learning network model: It includes a GAMDenseNet encoder and a decoder incorporating a channel attention mechanism. The GAMDenseNet encoder can efficiently extract feature information in the image, while the attention mechanism decoder can focus on the correlation between different channels and improve the prediction performance of the model;

[0016] The extracted feature information is input into the multi-task learning network model for training: The model simultaneously processes two tasks, surface roughness recognition and dimensional accuracy prediction. By sharing feature information and gradient information, collaborative optimization of the two tasks is achieved;

[0017] Use the model to predict the vibration signal and output the result: Input the new vibration signal into the trained model, and the model can quickly output the surface roughness classification result.

[0018] The GAMDenseNet encoder also includes a global average pooling layer and a fully connected layer for further extracting and integrating feature information; The global average pooling layer can reduce the dimension of the feature map while retaining key information; The fully connected layer can map the feature information to the classification result.

[0019] The decoder also includes an upsampling layer and a convolutional layer for converting the extracted feature information into a classification prediction result; The upsampling layer can restore the size of the feature map to the same size as the original image; The convolutional layer can convert the feature map into a classification prediction result through convolutional operations.

[0020] The preprocessing step also includes segmenting the vibration signal to facilitate subsequent feature extraction and classification prediction. Segmenting the vibration signal can divide the long-duration vibration signal into multiple short-duration segments.

[0021] The two-dimensional discrete wavelet transform is used to perform multi-level decomposition and reconstruction on the two-dimensional image to improve the richness of feature extraction. Multi-level decomposition and reconstruction can extract feature information of different frequencies and directions, thereby improving the model's ability to identify surface roughness.

[0022] The GAMDenseNet multi-task learning network model uses a gradient normalization algorithm to adaptively adjust the weight ratio of the loss functions of the two tasks of surface roughness recognition and dimensional accuracy prediction to optimize the model training process. The normalization algorithm can dynamically adjust the weight ratio according to the training progress and loss magnitude of the two tasks to achieve balanced optimization of the two tasks.

[0023] It also includes post-processing the classification results of the surface roughness. The post-processing steps include smoothing the classification results, removing outliers, or using an ensemble learning method to improve the classification accuracy.

[0024] The smoothing process in the post-processing steps can use methods such as moving average filtering or Gaussian filtering to remove noise and fluctuations in the classification results; removing outliers can be achieved by setting thresholds or using clustering algorithms; the ensemble learning method improves the classification accuracy by combining the prediction results of multiple models.

[0025] First, multi-channel vibration signals during the machining process of a CNC lathe are collected. These vibration signals contain rich machining process information and are the basis for subsequent feature extraction and classification.

[0026] Next, the collected multi-channel vibration signals are preprocessed. The preprocessing steps include denoising, filtering, and normalization to eliminate noise interference and improve the signal quality. In addition, the vibration signals can also be segmented for subsequent feature extraction and classification prediction.

[0027] Then, the one-dimensional vibration signals are converted into two-dimensional images using the relative position matrix (RPM). The purpose of this step is to convert one-dimensional signals into two-dimensional images for subsequent feature extraction using two-dimensional discrete wavelet transform.

[0028] Next, two-dimensional discrete wavelet transform (2D-DWT) is used to decompose and reconstruct the two-dimensional images to extract feature information. 2D-DWT can decompose different frequency components in the image and extract features that have an important impact on classification. To improve the richness of feature extraction, a multi-level decomposition and reconstruction method can be used.

[0029] After feature extraction is completed, an MTL-GAMDenseNet-CA multi-task learning network model is constructed. This model includes a GAMDenseNet encoder and a decoder incorporating a channel attention mechanism. The GAMDenseNet encoder is used to extract feature information, and the decoder is used to perform classification prediction based on the extracted feature information. To improve the classification accuracy of the model, a gradient normalization algorithm can be used to adaptively adjust the weight ratio of the loss functions for the two tasks of surface roughness recognition and dimensional accuracy prediction.

[0030] Finally, the trained model is used to predict new vibration signals, and the surface roughness classification results are output. Based on the classification results, the surface quality of the workpiece can be evaluated and feedback can be provided.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] Improving Classification Accuracy: By introducing a multi-task learning framework and an attention mechanism, the classification accuracy of the model for surface roughness has been significantly improved. Multi-task learning enables the model to consider multiple related tasks simultaneously, thereby learning more comprehensive and in-depth feature representations. The introduction of the attention mechanism allows the model to pay more attention to the features that have an important impact on the classification results, further enhancing the classification performance. The advantage of this method is that it can not only accurately identify surface roughness but also predict other related quality indicators, such as dimensional accuracy, etc., thus improving the overall monitoring level of machining quality.

[0033] Enhancing Model Generalization Ability: Traditional methods often optimize only for a single task, resulting in limited model generalization ability. The method proposed in the present invention enables the model to maintain stable performance under different machining environments and conditions by considering multiple tasks simultaneously. In addition, by introducing advanced technologies such as the attention mechanism and the gradient normalization algorithm, the robustness and adaptability of the model are further enhanced. This enables the model to accurately classify surface roughness even in the face of complex and changing machining environments.

[0034] Reducing Measurement Costs: Traditional surface roughness measurement methods often require additional measurement equipment and professional personnel for operation, which are not only costly but also involve a cumbersome measurement process. The method proposed in the present invention uses vibration signals for on-line monitoring, eliminating the need for additional measurement equipment and reducing measurement costs and time costs. At the same time, since this method can achieve real-time monitoring and feedback, it helps to adjust machining parameters in a timely manner, further reducing the scrap rate and production costs.

[0035] Improving Production Efficiency: In the process of machining, real-time monitoring and feedback of the surface quality of workpieces are crucial. Traditional methods often require waiting until the machining is completed before measurement and evaluation, resulting in low production efficiency. The method proposed in the present invention can obtain surface roughness information in real time during the machining process and adjust machining parameters in a timely manner according to the classification results, thereby improving production efficiency and product quality. In addition, this method can also provide data support for subsequent machining processes, helping to optimize machining processes and improve overall production efficiency.

[0036] This method collects multi-channel vibration signals during the machining process of a CNC lathe, uses advanced signal processing techniques to extract effective feature information, and constructs a multi-task learning network model for training and classification. Specifically, the present invention uses a Relative Position Matrix (RPM) to convert one-dimensional vibration signals into two-dimensional images, and then uses two-dimensional discrete wavelet transform (2D-DWT) to decompose and reconstruct the images to extract rich feature information. On this basis, the present invention introduces a Global Attention Mechanism (GAM) into the DenseNet model to construct a GAMDenseNet network as the encoder of the model to enhance the model's ability to extract key features. At the same time, a Channel Attention (CA) mechanism is incorporated into the decoder part to improve the model's ability to screen useful features for different tasks. This multi-task learning network model can simultaneously consider the monitoring of surface roughness and other related quality indicators, improving the generalization ability and robustness of the model.

[0037] Furthermore, the present invention uses the Gradient Normalization algorithm (GradNorm) to adaptively adjust the weight ratio of the loss functions of the two tasks of surface roughness recognition and dimensional accuracy prediction to optimize the model training process. By dynamically adjusting the weight ratio, the model can better balance the learning effects of the two tasks during the training process, further improving the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] FIG. Figure 1 is a data acquisition and processing flowchart of a surface roughness intelligent classification system according to the present invention.

[0039] FIG. Figure 2 is a model construction flowchart of a surface roughness intelligent classification system according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0040] The following further elaborates the present invention in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by this application.

[0041] The present invention relates to a surface roughness intelligent classification system, and its main structure includes:

[0042] Construction of Data Acquisition System: Sensor Selection and Arrangement: To achieve high-precision vibration signal acquisition, we selected acceleration sensors as the core components. These sensors have characteristics such as high sensitivity, low noise, and wide frequency response, and can accurately capture the weak vibration signals generated during the machining process. At the key parts of the CNC lathe, such as the spindle, worktable, and tool clamping device, we carefully arranged multiple groups of sensors according to the vibration transmission path and signal intensity distribution to form a sensor array. Each group of sensors contains three acceleration sensors in orthogonal directions to comprehensively capture the vibration information in three-dimensional space. In addition, to ensure the stability and reliability of the sensors, we used high-quality connecting wires and connectors and carried out strict testing and calibration.

[0043] Configuration of Data Acquisition System: The data acquisition system consists of a sensor array, a signal conditioning circuit, a high-speed data acquisition card, and a host computer software. The analog vibration signals output by the sensor array are amplified and filtered by the signal conditioning circuit and then digitized by the high-speed data acquisition card at a high sampling rate (such as above 20 kHz). The host computer software is responsible for real-time reception, storage, and preliminary processing of the data to ensure the integrity and accuracy of the data. In addition, the data acquisition system also has a synchronous triggering function, which can automatically start data acquisition at the beginning of machining to avoid data loss. To ensure the real-time performance and reliability of the data, we used high-performance computers and storage devices and carried out reasonable system optimization and configuration.

[0044] Data Preprocessing and Feature Extraction: Vibration Signal Denoising: The original vibration signals often contain useless information such as mechanical noise and electromagnetic interference, which will interfere with subsequent feature extraction and classification. Therefore, we first used adaptive filtering technology to remove the noise. The adaptive filter can dynamically adjust the filtering parameters according to the changes in the input signal, effectively suppressing the noise components while retaining the useful signals related to surface roughness. In addition, we also applied wavelet transform for multi-scale analysis to further separate the noise and the signal and improve the signal-to-noise ratio. To verify the denoising effect, we carried out a comparative experiment and found that the quality of the denoised signal was significantly improved.

[0045] Segmentation and Standardization of Vibration Signals: Considering the time-varying nature of vibration signals, we divided the long-term continuous signals into multiple short-term signal segments, each of which contains a fixed number of sampling points (such as 1024 points). This can not only ensure that the local features of the signal are not damaged but also reduce the computational amount and improve the processing efficiency. Subsequently, each signal segment is standardized, that is, the mean value of the signal is adjusted to 0 and the standard deviation is adjusted to 1 to eliminate the amplitude differences between different signal segments and provide a consistent data basis for subsequent feature extraction. To ensure the accuracy of the standardization process, we used a high-precision mathematical operation library and carried out strict testing.

[0046] Feature Extraction and Dimensionality Reduction: Feature extraction is a crucial step connecting signal processing and machine learning models. We use the Short-Time Fourier Transform (STFT) to convert the vibration signal from the time domain to the frequency domain, extracting spectral features such as the dominant frequency, bandwidth, energy distribution, etc. At the same time, combining time-domain features such as mean, variance, peak value, etc., we construct a multi-dimensional feature vector. To reduce feature redundancy and improve the model training efficiency, we use Principal Component Analysis (PCA) to perform dimensionality reduction on the feature vector, retaining the most representative feature components. To ensure the effectiveness of feature extraction and dimensionality reduction, we have conducted a large number of experiments and optimizations, and selected the best parameters and algorithms.

[0047] Multi-Task Learning Model Construction and Training: Model Architecture Design: The multi-task learning model proposed in this invention is based on a deep learning framework, specifically adopting a hybrid architecture of Convolutional Neural Network (CNN) and Long Short-Term Memory Network (LSTM). CNN is responsible for extracting spatial features from the feature images of vibration signals, while LSTM is used to capture time-series features. The main body of the model consists of a shared feature extraction layer, two independent task branches (surface roughness classification branch and dimensional accuracy prediction branch), and a fusion layer. The shared feature extraction layer extracts low-level and high-level features of the vibration signal through multiple convolutional and pooling operations. The two task branches process their respective specific tasks, outputting classification labels and prediction values. The fusion layer comprehensively combines the information of the two tasks through weighted summation and other methods to achieve collaborative optimization between multiple tasks. To verify the effectiveness of the model architecture, we have conducted comparative experiments on multiple architectures and selected the best one.

[0048] Loss Function and Optimization Strategy: To balance the learning progress of the two tasks, we adopt the GradNorm algorithm, which dynamically adjusts the task weights according to the gradient magnitudes of the two task loss functions, ensuring that the two tasks maintain a relatively balanced progress during training. In terms of the loss function, the cross-entropy loss is used for the surface roughness classification task, and the mean squared error loss is used for the dimensional accuracy prediction task. The Adam optimizer is selected, and its adaptive learning rate adjustment mechanism helps the model quickly converge to the optimal solution. To further improve the training effect of the model, we have conducted comparative experiments on multiple optimization strategies and parameters and selected the best optimization strategy.

[0049] Data Augmentation and Model Training: To improve the generalization ability of the model, we performed data augmentation on the training dataset, including transformations such as rotation, scaling, and translation, to increase the diversity of samples. During the training process, the mini-batch gradient descent method was adopted, and a batch of data (such as 64 samples) was processed in each iteration to reduce memory occupancy and accelerate training. At the same time, an early stopping strategy was set. When the loss on the validation set no longer decreased, the training was stopped to prevent overfitting. To ensure the stability and reliability of model training, we adopted a variety of training techniques and strategies, such as learning rate decay, gradient clipping, etc., and conducted a large number of experiments and optimizations.

[0050] By introducing the multi-task learning framework and attention mechanism, the classification accuracy of the model for surface roughness has been significantly improved. Multi-task learning enables the model to consider multiple related tasks simultaneously, thereby learning more comprehensive and in-depth feature representations. The introduction of the attention mechanism enables the model to pay more attention to the features that have an important impact on the classification results, further enhancing the classification performance. The advantage of this method is that it can not only accurately identify surface roughness but also predict other related quality indicators, such as dimensional accuracy, etc., simultaneously, thus improving the overall monitoring level of machining quality. Traditional methods often optimize only for a single task, resulting in limited generalization ability of the model. The method proposed in the present invention enables the model to maintain stable performance under different machining environments and conditions by considering multiple tasks simultaneously. In addition, by introducing advanced technologies such as the attention mechanism and gradient normalization algorithm, the robustness and adaptability of the model are further enhanced. This enables the model to still accurately classify surface roughness when facing a complex and changing machining environment.

[0051] Traditional surface roughness measurement methods often require additional measurement equipment and professional personnel for operation, which are not only costly but also have a cumbersome measurement process. The method proposed in the present invention uses vibration signals for online monitoring, eliminating the need for additional measurement equipment, reducing the measurement cost and time cost. At the same time, since this method can achieve real-time monitoring and feedback, it helps to adjust the machining parameters in a timely manner, further reducing the scrap rate and production cost.

[0052] Improving production efficiency: During the machining process, real-time monitoring and feedback of the surface quality of workpieces are crucial. Traditional methods often require waiting until the machining is completed to measure and evaluate, resulting in low production efficiency. However, the method proposed in this invention can obtain surface roughness information in real time during the machining process and adjust the machining parameters in a timely manner according to the classification results, thereby improving production efficiency and product quality. In addition, this method can also provide data support for subsequent machining processes, helping to optimize the machining process and improve the overall production efficiency. This method collects multi-channel vibration signals during the CNC lathe machining process, uses advanced signal processing techniques to extract effective feature information, and constructs a multi-task learning network model for training and classification. Specifically, this invention uses the Relative Position Matrix (RPM) to convert one-dimensional vibration signals into two-dimensional images, and then uses two-dimensional discrete wavelet transform (2D-DWT) to decompose and reconstruct the images to extract rich feature information. On this basis, this invention introduces the Global Attention Mechanism (GAM) into the DenseNet model to construct the GAMDenseNet network as the encoder of the model to enhance the model's ability to extract key features. At the same time, the Channel Attention (CA) mechanism is incorporated into the decoder part to improve the model's ability to screen useful features for different tasks. This multi-task learning network model can simultaneously consider the monitoring of surface roughness and other related quality indicators, improving the generalization ability and robustness of the model.

Claims

1. An intelligent classification system for surface roughness, characterized in that: Including the following steps: Collect multi-channel vibration signals during lathe machining: Real-time capture of multi-channel vibration signals generated by a CNC lathe during machining through high-precision sensors; Preprocess the multi-channel original vibration signals: including denoising, filtering, and normalization. Denoising is to remove noise interference in the signal through digital filtering technology; Filtering is to select an appropriate frequency range to retain vibration information related to surface roughness; Normalization processing is to adjust the signal amplitude to a unified range for subsequent processing; Convert the preprocessed one-dimensional vibration signal into a two-dimensional image using a relative position matrix: The relative position matrix method can effectively convert a one-dimensional signal into a two-dimensional image, thus more intuitively displaying the characteristic information in the signal; Use two-dimensional discrete wavelet transform to decompose and reconstruct the two-dimensional image to extract characteristic information: Two-dimensional discrete wavelet transform can decompose an image into wavelet coefficients of different frequencies and directions. By selecting appropriate wavelet coefficients for reconstruction, characteristic information closely related to surface roughness can be extracted; Construct a multi-task learning network model: including a GAMDenseNet encoder and a decoder incorporating a channel attention mechanism. The GAMDenseNet encoder can efficiently extract characteristic information in the image, while the attention mechanism decoder can pay attention to the correlation between different channels and improve the prediction performance of the model; Input the extracted characteristic information into the multi-task learning network model for training: The model simultaneously processes two tasks of surface roughness recognition and dimensional accuracy prediction. By sharing characteristic information and gradient information, collaborative optimization of the two tasks is achieved; Use the model to predict the vibration signal and output the result: Input a new vibration signal into the trained model, and the model can quickly output the surface roughness classification result.

2. The intelligent classification system for surface roughness according to claim 1, wherein: The GAMDenseNet encoder also includes a global average pooling layer and a fully connected layer for further extracting and integrating characteristic information; The global average pooling layer can reduce the dimension of the feature map while retaining key information; The fully connected layer can map the characteristic information to the classification result.

3. The intelligent surface roughness classification system according to claim 2, characterized in that: The decoder also includes an upsampling layer and a convolutional layer for converting the extracted characteristic information into a classification prediction result; The upsampling layer can restore the size of the feature map to the same size as the original image; The convolutional layer can convert the feature map into a classification prediction result through convolutional operations.

4. The intelligent classification system for surface roughness according to claim 3, wherein: The preprocessing step also includes segmenting the vibration signal to facilitate subsequent feature extraction and classification prediction. Segmenting the vibration signal can divide a long-period vibration signal into multiple short-period signals.

5. The intelligent classification system for surface roughness according to claim 4, wherein: Use two-dimensional discrete wavelet transform to perform multi-level decomposition and reconstruction on the two-dimensional image to improve the richness of feature extraction. Multi-level decomposition and reconstruction can extract characteristic information of different frequencies and directions, thereby improving the model's ability to recognize surface roughness.

6. The intelligent surface roughness classification system according to claim 5, characterized in that: The GAMDenseNet multi-task learning network model uses a gradient normalization algorithm to adaptively adjust the weight ratio of the loss functions of the two tasks of surface roughness recognition and dimensional accuracy prediction, so as to optimize the model training process. The normalization algorithm can dynamically adjust the weight ratio according to the training progress and loss of the two tasks, realizing the balanced optimization of the two tasks.

7. The intelligent surface roughness classification system according to claim 6, characterized in that: It also includes post-processing the surface roughness classification results. The post-processing steps include smoothing the classification results, removing outliers, or using an ensemble learning method to improve the classification accuracy.

8. The intelligent surface roughness classification system according to claim 7, characterized in that: The smoothing process in the post-processing steps can use methods such as moving average filtering or Gaussian filtering to remove noise and fluctuations in the classification results; removing outliers can be achieved by setting thresholds or using clustering algorithms; the ensemble learning method improves the classification accuracy by combining the prediction results of multiple models.

Images