Surface roughness and size precision common prediction method based on relative position matrix

Abstract

The invention discloses a surface roughness and size precision common prediction method based on a relative position matrix, and mainly relates to the field of roughness detection. Comprising the steps that in the turning machining process, vibration signals are collected in real time through multi-channel acceleration sensors arranged at the positions of a main shaft, a tool and a workpiece clamp, and a roughness measuring instrument and a micrometer are synchronously used for measuring the surface roughness value and size data of a machined workpiece; normalization processing is carried out on the collected vibration signals, the vibration signals are converted into two-dimensional images to be expressed based on a relative position matrix algorithm, deep features are extracted through two-dimensional discrete wavelet transform, and the spatial expression ability of the signals is enhanced. The method has the beneficial effects that the vibration signal is converted into the two-dimensional RPM graph through the RPM, and the 2D-DWT is utilized to perform decomposition reconstruction and image fusion, so that the deep sharing features in the vibration signal are effectively extracted, and the method plays a key role in improving the prediction precision of the surface roughness and the size precision.

Description

Technical Field

[0001] The present invention relates to the field of roughness detection, and specifically to a method for jointly predicting surface roughness and dimensional accuracy based on a relative position matrix. Background Art

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

[0003] Under the background of the rapid development of intelligent manufacturing, surface processing quality has become one of the key indicators for measuring product performance, directly affecting the mechanical properties, durability, overall reliability and market competitiveness of products. As the core evaluation indicators of surface processing quality, the precise control of surface roughness and dimensional accuracy is crucial for meeting the strict requirements of the high-end manufacturing field. However, in the actual machining process, due to the complex interaction of various factors such as material properties, tool wear, and cutting parameters, the control of surface roughness and dimensional accuracy faces great challenges.

[0004] Traditionally, the prediction of surface roughness and dimensional accuracy mainly relies on a single model for a single evaluation index, and this method is inadequate when dealing with multiple evaluation indexes that need to be monitored simultaneously in the actual manufacturing process. Low efficiency: An independent prediction model needs to be constructed and trained for each evaluation index, resulting in a significant increase in computing resources and time costs. Limited accuracy: A single model often has difficulty in comprehensively capturing and utilizing the deep shared features in the vibration signal, which are crucial for improving the prediction accuracy. Lack of adaptability: In the actual machining process, factors such as working conditions and material properties may change, leading to a decline in the prediction performance of a single model.

[0005] To address these problems, the present invention proposes a method for jointly predicting surface roughness and dimensional accuracy based on a relative position matrix (RPM) and multi-task learning (MTL). This method converts the vibration signal into a two-dimensional image representation through RPM, and uses two-dimensional discrete wavelet transform (2D-DWT) for decomposition, reconstruction and image fusion to extract the deep shared features in the vibration signal. At the same time, an MTL network model is constructed by combining the hard parameter sharing (HPS) architecture and the attention mechanism to achieve the joint prediction of surface roughness and dimensional accuracy. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for jointly predicting surface roughness and dimensional accuracy based on a relative position matrix. The present invention converts vibration signals into two-dimensional RPM maps through RPM, and uses 2D-DWT for decomposition, reconstruction, and image fusion, effectively extracting deep shared features in the vibration signals. These features contain rich machining process information, which plays a key role in improving the prediction accuracy of surface roughness and dimensional accuracy. Experimental results show that compared with traditional single-model prediction methods, the method proposed in the present invention has significantly improved prediction accuracy.

[0007] To achieve the above object, the present invention is realized through the following technical solutions:

[0008] The method for jointly predicting surface roughness and dimensional accuracy based on a relative position matrix includes the following steps:

[0009] S1. During the turning process, multi-channel acceleration sensors arranged at the positions of the spindle, tool, and workpiece fixture are used to collect vibration signals in real time, and a roughness measuring instrument and a micrometer are synchronously used to measure the surface roughness value and dimensional data of the machined workpiece.

[0010] S2. The collected vibration signals are normalized, converted into a two-dimensional image representation based on the relative position matrix algorithm, and deep features are extracted through two-dimensional discrete wavelet transform to enhance the spatial expression ability of the signals.

[0011] S3. A multi-task learning network model based on a hard parameter sharing architecture is constructed. The encoder uses an improved DenseNet network and integrates a global attention mechanism, and the decoder introduces a channel attention mechanism to realize the joint training of surface roughness recognition and dimensional accuracy prediction.

[0012] S4. The model is trained based on a joint loss function, and the task weights are dynamically adjusted through the GradNorm algorithm to optimize the multi-task cooperation effect.

[0013] S5. The trained model is used to realize the joint output of surface roughness classification and dimensional accuracy regression, supporting real-time machining process monitoring.

[0014] The acquisition of the multi-channel vibration signals includes: at least three independent acceleration sensors are used, which are respectively arranged at the positions of the spindle, tool, and workpiece fixture to ensure the comprehensiveness of the signal spatial distribution; the sampling frequency of the vibration signals is not less than 15 kHz, and the signal duration covers the complete machining cycle to capture dynamic change features; the surface roughness measurement uses the surface roughness value as a quantization index, and the measurement accuracy needs to reach 0.01 μm; the dimensional accuracy measurement uses single-point measurement values, and the measurement accuracy needs to reach 0.001 mm.

[0015] The construction of the relative position matrix includes: normalizing the multi-channel vibration signals to eliminate the dimension difference and ensure the comparability of signals in each channel; generating a two-dimensional matrix based on the signal time-series characteristics, where the number of rows and columns of the matrix respectively correspond to the number of sampling points and channels of the signal, and the matrix element value is the normalized signal amplitude; performing a standardization transformation on the matrix to limit the range of its element values within [0, 1] to avoid instability in model training caused by too large a numerical range.

[0016] The feature extraction of the two-dimensional discrete wavelet transform includes: selecting the Haar wavelet as the basis function, performing three-level decomposition on the relative position matrix, and extracting the low-frequency approximation sub-band and high-frequency detail sub-bands in the horizontal, vertical, and diagonal directions; reconstructing the image through the inverse wavelet transform to enhance the local feature expression ability of the signal; using the max-pooling operation to reduce the feature dimension, setting the pooling window size to 2×2 and the stride to 2 to reduce the computational complexity.

[0017] The improved DenseNet network encoder includes: integrating a global attention module after each dense block to enhance the expression of key features through the spatial attention mechanism; using a transition layer to reduce the feature map size, setting the reduction factor to 0.5 to balance feature expression and computational efficiency; the network depth is not less than 40 layers, including at least three dense blocks, and each dense block contains 6 convolutional layers with the convolutional kernel size set to 3×3.

[0018] The channel attention mechanism includes: performing global average pooling on the feature map output by the encoder to generate channel-level statistics, where the dimension of the statistics is the same as the number of channels; learning the channel weights through a two-layer fully connected network, using ReLU and Sigmoid as activation functions, the output dimension of the first fully connected layer is 1 / 4 of the number of channels, and the second layer restores to the number of channels; multiplying the weights with the feature map channel by channel to achieve feature enhancement or suppression and enhance the expression ability of task-related features.

[0019] The joint loss function includes: the cross-entropy loss for surface roughness recognition, which is used to evaluate the accuracy of the classification task; the mean squared error loss for dimensional accuracy prediction, which is used to evaluate the deviation of the regression task; dynamically adjusting the weights of the two losses through the GradNorm algorithm, with the weight update period set to once every 5 training epochs to ensure the stability of multi-task collaborative training.

[0020] The model training process includes: using the AdamW optimizer for parameter update, setting the initial learning rate to 0.001, the learning rate decay factor to 0.1, and the decay period to every 20 training epochs; introducing L2 regularization to prevent overfitting, setting the regularization coefficient to 0.0005 to limit the model complexity; the number of training epochs is not less than 200, and the early stopping mechanism is triggered based on the change in the validation set loss. If the validation set loss does not decrease for 10 consecutive epochs, the training stops.

[0021] The present invention includes the following steps:

[0022] Data acquisition: During the turning process, a multi-channel original vibration signal is collected by using an acceleration sensor and a vibration signal collector. Meanwhile, a roughness measuring instrument and a micrometer are used to measure the surface roughness value and dimensional data of the machined workpiece respectively.

[0023] Data processing: Preprocessing operations such as cleaning and cropping are performed on the collected vibration signal to remove noise and outliers. The preprocessed multi-channel vibration signal is converted into a two-dimensional RPM map by using RPM, and this process includes steps such as signal normalization, piecewise aggregate approximation, and RPM algorithm construction. The multi-channel RPM maps are fused and decomposed and reconstructed by using two-dimensional discrete wavelet transform (2D-DWT) to extract the deep shared features in the vibration signal and expand the data set.

[0024] Model construction: An MTL-GAMDenseNet-CA network model based on the hard parameter sharing (HPS) architecture is constructed. This model includes an encoder and two decoders. The encoder adopts an improved DenseNet network and introduces a GAM attention mechanism to form GAMDenseNet to improve the feature extraction ability. The decoder part introduces a CA attention mechanism to enhance the feature screening ability and realize the adaptability to different tasks (surface roughness recognition and dimensional accuracy prediction). The GradNorm algorithm is used to adaptively adjust the weights of the two task loss functions to achieve multi-task joint training and optimize the model performance.

[0025] Model training and evaluation: The processed data set is divided into a training set and a test set. The model is trained by using the training set data, the model parameters are optimized by using the Adam optimizer, and weight decay is adopted for regularization to prevent overfitting. After the training is completed, the performance of the model is evaluated by using the test set data, including the precision, recall, and F1 score of surface roughness recognition, and indexes such as MAE and RMSE of dimensional accuracy prediction.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] Improving the prediction accuracy: The present invention converts the vibration signal into a two-dimensional RPM map by using RPM, and performs decomposition, reconstruction, and image fusion by using 2D-DWT, effectively extracting the deep shared features in the vibration signal. These features contain rich machining process information and play a key role in improving the prediction accuracy of surface roughness and dimensional accuracy. The experimental results show that compared with the traditional single-model prediction method, the method proposed by the present invention has a significant improvement in the prediction accuracy.

[0028] Improve prediction efficiency: The multi-task learning network model constructed by combining the GAM and CA attention mechanisms realizes the joint prediction of surface roughness and dimensional accuracy. Compared with a single model, the multi-task learning model can share the underlying feature representation, reducing the consumption of computing resources and time costs. At the same time, by introducing the attention mechanism, the model can adaptively adjust the attention to different tasks, further improving the prediction efficiency.

[0029] Enhance model adaptability: During the actual machining process, factors such as working conditions and material properties may change, resulting in a decline in the prediction performance of the model. The method proposed in the present invention adaptively adjusts the weights of the two task loss functions by introducing the GradNorm algorithm, enabling the model to dynamically adjust the prediction strategy according to real-time data, thereby enhancing the adaptability and stability of the model. In addition, the application of RPM and 2D-DWT also enables the model to process different types of vibration signals, further improving the generalization ability of the model.

[0030] The method proposed in the present invention provides strong support for quality control and optimization in the intelligent manufacturing process through its advantages in improving prediction accuracy and efficiency, enhancing model adaptability, and providing effective solutions. Brief Description of the Drawings

[0031] Att Figure 1 is a data acquisition and processing flowchart of the method for jointly predicting surface roughness and dimensional accuracy based on the relative position matrix of the present invention.

[0032] Att Figure 2 is a model construction and training flowchart of the method for jointly predicting surface roughness and dimensional accuracy based on the relative position matrix in the present invention. Detailed Embodiments

[0033] 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.

[0034] The present invention relates to a method for jointly predicting surface roughness and dimensional accuracy based on a relative position matrix, and its main structure includes:

[0035] Data Acquisition and Preprocessing

[0036] Sensor Arrangement: Accelerometers are installed at the positions of the spindle, cutting tool, and workpiece fixture of the lathe machine. Ensure that the sensors can accurately capture the vibration signals during the machining process. The sensor at the spindle position should be installed on the top of the spindle box to monitor the radial and axial vibrations of the spindle; the sensor at the cutting tool position should be installed on the side of the tool post to monitor the cutting vibration of the tool; the sensor at the workpiece fixture position should be installed at the bottom of the fixture to monitor the clamping vibration of the workpiece.

[0037] Data Acquisition: Set the sampling frequency of the sensors to not less than 15 kHz to ensure that the signal duration covers the entire machining cycle. At the same time, synchronously record the surface roughness value and dimensional data of the workpiece after machining. The surface roughness value is measured using a surface roughness measuring instrument, and the dimensional data is measured using a micrometer. During the measurement process, ensure the accuracy and stability of the measuring instrument to avoid inaccurate data caused by measurement errors.

[0038] Data Preprocessing: Normalize the collected vibration signals to eliminate the dimension difference. The minimum-maximum normalization method can be used for normalization to map the signal values to the interval [0,1]. Then, convert the vibration signals into a two-dimensional image representation based on the RPM algorithm. The number of rows and columns of the RPM matrix correspond to the number of sampling points and channels of the signal respectively, and the matrix element values are the normalized signal amplitudes. Perform a standardization transformation on the RPM matrix to limit the range of its element values to [0,1]. The Z-score standardization method can be used for the standardization transformation to make the mean of the matrix elements 0 and the standard deviation 1.

[0039] Feature Extraction

[0040] Select Wavelet Basis: Select the Haar wavelet as the basis function and perform three-level decomposition on the RPM matrix. Extract the low-frequency approximation sub-band and the high-frequency detail sub-bands in the horizontal, vertical, and diagonal directions. Appropriate decomposition levels and wavelet basis functions should be selected during the wavelet decomposition process to ensure the accuracy and effectiveness of the decomposition results.

[0041] Image Reconstruction: Reconstruct the image through inverse wavelet transform to enhance the local feature expression ability of the signal. Ensure the clarity and integrity of the reconstructed image during the inverse wavelet transform process to avoid information loss caused by reconstruction errors.

[0042] Dimensionality Reduction Processing: Use the max-pooling operation to reduce the feature dimension. Set the pooling window size to 2×2 and the stride to 2. The size of the final feature map is reduced to 1 / 8 of the original. Appropriate pooling window size and stride should be selected during the max-pooling operation to ensure that the dimensionality-reduced feature map can still retain the key features of the original signal.

[0043] Model Construction

[0044] Encoder Design: The encoder adopts an improved DenseNet network structure, including at least three DenseBlocks, each of which contains 6 convolutional layers with a convolutional kernel size of 3×3. A Global Attention Module (GAM) is integrated after each DenseBlock to enhance the expression of key features through a spatial attention mechanism (such as the SE module). A Transition Layer is used to reduce the size of the feature map, and the reduction factor is set to 0.5. During the encoder design process, the rationality and effectiveness of the network structure should be ensured to avoid performance issues caused by an overly complex or simple network structure.

[0045] Decoder Design: The decoder introduces a Channel Attention mechanism (CA). Global average pooling is performed on the feature map output by the encoder to generate channel-level statistics. Two fully connected networks are used to learn the channel weights, and the activation functions are ReLU and Sigmoid. The output dimension of the first fully connected layer is 1 / 4 of the number of channels, and the second layer restores to the number of channels. The weights are multiplied with the feature map channel by channel to achieve feature enhancement or suppression. During the decoder design process, the effectiveness and stability of the channel attention mechanism should be ensured to avoid performance degradation caused by the failure of the attention mechanism.

[0046] Multi-task Output Layer: The Softmax activation function is used for the surface roughness recognition task to output three probability values. The linear activation function is used for the dimensional accuracy prediction task to output continuous values. During the design of the multi-task output layer, the accuracy and reliability of the output results should be ensured to avoid misjudgment or missed judgment caused by inaccurate output results.

[0047] Model Training and Optimization

[0048] Loss Function Definition: The joint loss function is defined as the weighted sum of the cross-entropy loss for surface roughness recognition and the mean squared error loss for dimensional accuracy prediction. The weights of the two tasks are dynamically adjusted through the GradNorm algorithm to ensure the stability of multi-task collaborative training. During the loss function definition process, the rationality and effectiveness of the loss function should be ensured to avoid training instability or performance degradation caused by improper loss function design.

[0049] Optimizer Selection: The AdamW optimizer is used for parameter update. The initial learning rate is set to 0.001, the learning rate decay factor is 0.1, and the decay period is every 20 training epochs. During the optimizer selection process, the stability and convergence of the optimizer should be ensured to avoid training failure or poor performance caused by improper optimizer selection.

[0050] Regularization Processing: L2 regularization is introduced to prevent overfitting. The regularization coefficient is set to 0.0005 to limit the model complexity. During the regularization processing, the rationality and effectiveness of the regularization coefficient should be ensured to avoid performance issues caused by an overly large or small regularization coefficient.

[0051] Training strategy: The number of training rounds shall be no less than 200 rounds. An early stopping mechanism is adopted and triggered based on the change of the validation set loss. If the validation set loss does not decrease for 10 consecutive rounds, the training will stop. During the formulation of the training strategy, the stability and efficiency of the training process should be ensured, and the situation of too long training time or poor performance caused by improper training strategy should be avoided.

[0052] Prediction and output

[0053] Model evaluation: Precision, Recall, and F1-score are used to evaluate the performance of the surface roughness recognition task; Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Maximum Absolute Error (MaxAE) are used to evaluate the accuracy of the dimensional accuracy prediction task. The comprehensive evaluation adopts a multi-task weighted score, and the weight coefficients are assigned according to the importance of the tasks (for example, roughness recognition accounts for 0.6, and dimensional prediction accounts for 0.4). During the model evaluation process, the accuracy and reliability of the evaluation indicators should be ensured, and the inaccurate evaluation results caused by unreasonable evaluation indicators should be avoided.

[0054] Real-time prediction: The trained model is deployed to the turning processing site. During the processing, vibration signals are collected in real time and input into the model for prediction. The surface roughness classification results and dimensional accuracy prediction values are output to support the real-time monitoring of the processing process. During the real-time prediction process, the timeliness and accuracy of the prediction results should be ensured, and the out-of-control processing process or product quality problems caused by delayed or incorrect prediction results should be avoided.

[0055] Improving prediction accuracy: In the present invention, the vibration signal is converted into a two-dimensional RPM map through RPM, and 2D-DWT is used for decomposition, reconstruction, and image fusion, effectively extracting the deep shared features in the vibration signal. These features contain rich processing process information, which plays a key role in improving the prediction accuracy of surface roughness and dimensional accuracy. The experimental results show that compared with the traditional single-model prediction method, the method proposed in the present invention has a significant improvement in prediction accuracy.

[0056] Enhancing prediction efficiency: A multi-task learning network model constructed by combining GAM and CA attention mechanisms realizes the joint prediction of surface roughness and dimensional accuracy. Compared with a single model, the multi-task learning model can share the underlying feature representation, reducing the consumption of computing resources and time costs. At the same time, by introducing the attention mechanism, the model can adaptively adjust the attention to different tasks, further improving the prediction efficiency.

[0057] Enhance model adaptability: During the actual machining process, factors such as working conditions and material properties may change, leading to a decline in the prediction performance of the model. The method proposed in the present invention adaptively adjusts the weights of the two task loss functions by introducing the GradNorm algorithm, enabling the model to dynamically adjust the prediction strategy according to real-time data, thereby enhancing the adaptability and stability of the model. In addition, the application of RPM and 2D-DWT also enables the model to process different types of vibration signals, further improving the generalization ability of the model.

Claims

1. A method for jointly predicting surface roughness and dimensional accuracy based on a relative position matrix, characterized in that: Including the following steps: During the turning process, vibration signals are collected in real time by multi-channel acceleration sensors arranged at the positions of the spindle, tool, and workpiece fixture. At the same time, a roughness measuring instrument and a micrometer are used to measure the surface roughness value and dimensional data of the machined workpiece; The collected vibration signals are normalized, converted into a two-dimensional image representation based on the relative position matrix algorithm, and deep features are extracted through two-dimensional discrete wavelet transform to enhance the spatial expression ability of the signals; A multi-task learning network model based on a hard parameter sharing architecture is constructed. The encoder uses an improved DenseNet network and integrates a global attention mechanism, and the decoder introduces a channel attention mechanism to realize the joint training of surface roughness recognition and dimensional accuracy prediction; The model is trained based on the joint loss function, and the task weights are dynamically adjusted through the GradNorm algorithm to optimize the multi-task cooperation effect; The trained model is used to achieve the joint output of surface roughness classification and dimensional accuracy regression, supporting real-time monitoring of the machining process.

2. The method for jointly predicting surface roughness and dimensional accuracy based on the relative position matrix according to claim 1, wherein: The acquisition of the multi-channel vibration signals includes: at least three independent acceleration sensors are used and arranged at the positions of the spindle, tool, and workpiece fixture respectively to ensure the comprehensiveness of the signal spatial distribution; the sampling frequency of the vibration signals is not less than 15 kHz, and the signal duration covers the complete machining cycle to capture the dynamic change characteristics; the surface roughness is measured using the surface roughness value as a quantization index, and the measurement accuracy needs to reach 0.01 μm; the dimensional accuracy is measured using a single-point measurement value, and the measurement accuracy needs to reach 0.001 mm.

3. The method for jointly predicting surface roughness and dimensional accuracy based on the relative position matrix according to claim 2, characterized in that: The construction of the relative position matrix includes: normalizing the multi-channel vibration signals to eliminate the dimension difference and ensure the comparability of the signals in each channel; generating a two-dimensional matrix based on the signal time series characteristics, where the number of rows and columns of the matrix correspond to the number of sampling points and the number of channels of the signal respectively, and the matrix element values are the normalized signal amplitudes; performing a standardization transformation on the matrix to limit the range of its element values within [0,1] to avoid instability in model training caused by too large a numerical range.

4. The method for jointly predicting surface roughness and dimensional accuracy based on a relative position matrix according to claim 3, wherein: The feature extraction of the two-dimensional discrete wavelet transform includes: selecting the Haar wavelet as the basis function, performing three-level decomposition on the relative position matrix, and extracting the low-frequency approximation sub-band and high-frequency detail sub-bands in the horizontal, vertical, and diagonal directions; reconstructing the image through the inverse wavelet transform to enhance the local feature expression ability of the signal; using the maximum pooling operation to reduce the feature dimension, and setting the pooling window size to 2×2 and the stride to 2 to reduce the computational complexity.

5. The method for jointly predicting surface roughness and dimensional accuracy based on the relative position matrix according to claim 4, characterized in that: The improved DenseNet network encoder includes: integrating a global attention module after each dense block to enhance the expression of key features through the spatial attention mechanism; using a transition layer to reduce the feature map size, and setting the reduction factor to 0.5 to balance the feature expression and computational efficiency; the network depth is not less than 40 layers, including at least three dense blocks, and each dense block contains 6 convolutional layers with the convolutional kernel size set to 3×3.

6. The method for jointly predicting surface roughness and dimensional accuracy based on a relative position matrix according to claim 5, wherein: The described channel attention mechanism includes: performing global average pooling on the feature map output by the encoder to generate channel-level statistics, where the dimension of the statistics is the same as the number of channels; learning channel weights through a two-layer fully connected network, with ReLU and Sigmoid as activation functions, the output dimension of the first fully connected layer being 1 / 4 of the number of channels, and the second layer restoring to the number of channels; multiplying the weights with the feature map channel by channel to achieve feature enhancement or suppression, enhancing the expression ability of task-related features.

7. The method for jointly predicting surface roughness and dimensional accuracy based on the relative position matrix according to claim 6, wherein: The described joint loss function includes: the cross-entropy loss for surface roughness recognition, which is used to evaluate the accuracy of the classification task; the mean squared error loss for dimensional accuracy prediction, which is used to evaluate the deviation of the regression task; dynamically adjusting the weights of the two losses through the GradNorm algorithm, with the weight update period set to once every 5 training epochs to ensure the stability of multi-task collaborative training.

8. The method for jointly predicting surface roughness and dimensional accuracy based on the relative position matrix according to claim 7, wherein: The described model training process includes: using the AdamW optimizer for parameter update, with the initial learning rate set to 0.001, the learning rate decay factor to 0.1, and the decay period to every 20 training epochs; introducing L2 regularization to prevent overfitting, with the regularization coefficient set to 0.0005 to limit the model complexity; the number of training epochs is not less than 200, and the early stopping mechanism is triggered based on the change in the validation set loss. If the validation set loss does not decrease for 10 consecutive epochs, the training stops.

Images