Numerically-controlled machine tool milling time sequence signal intelligent marking method and system based on Transform

Abstract

The invention provides a numerical control machine tool milling machining multi-process signal intelligent marking method based on Transform. Firstly, multi-mode time sequence signals (such as spindle vibration, cutting force and acoustic emission signals) are collected, denoising is performed through wavelet transform, time-frequency domain features are extracted, signal samples are classified according to process categories, and a high-quality training set is constructed by combining specialists to label the samples. Secondly, designing a multi-modal feature extraction framework based on a Transform model, introducing a multi-head attention mechanism, dynamically adjusting modal weights to generate joint feature representation, determining key modals and features through improved multi-modal feature sensitivity analysis, and accurately marking signals in combination with a self-supervised learning and physical constraint optimization model; finally, a sequential mark verification strategy is adopted, a high-quality mark data set is generated, and the accuracy and consistency of analysis in the machining process are improved. The method can be widely applied to machining quality analysis and intelligent optimization of the numerical control machine tool for complex precise parts, and has high efficiency and reliability.

Description

Technical Field

[0001] The invention relates to the field of high-precision intelligent manufacturing of high-precision CNC machine tools, and specifically to an intelligent marking technology for signals of different processes in the milling process of CNC machine tools. Background Art

[0002] During the milling process of CNC machine tools, a large amount of complex process signal data is generated along with workpiece processing, tool movement and equipment operation. These signals contain rich process status information and are the key to process optimization, fault diagnosis and quality control. However, unlabeled raw data usually has noise and redundancy, and direct use will lead to inefficient data analysis and modeling. How to accurately and efficiently label this data has become an important research topic in the field of intelligent manufacturing.

[0003] At present, the labeling of signal data of different processes in CNC machine tool processing mainly relies on manual work. Faced with the massive multimodal signal data in industrial production, manual labeling is time-consuming, labor-intensive, and extremely costly, and it is difficult to meet actual needs. In addition, due to the differences in the professional level of labelers and insufficient understanding of the processing process, manual labeling has problems such as high error rate and poor data consistency. This low-quality labeling will have an adverse effect on subsequent data utilization and analysis, reducing the prediction accuracy and decision reliability of the model. Although some rule-based automatic labeling methods and machine learning-assisted labeling methods have been proposed, these methods usually rely on preset rules or rely on a large amount of high-quality sample data for training, which is difficult to adapt to complex and changeable processing scenarios, and shows great limitations when processing signals of different processes.

[0004] The present invention proposes a Transform-based deep learning model to model and label the time series signal labeling task. This method uses the self-attention mechanism to extract signal features and dynamically adjusts the modal weights to achieve efficient labeling and accurate classification of multi-process signals, thereby improving the value density of manufacturing process data and providing high-quality data support for subsequent big data analysis and process optimization. Summary of the invention

[0005] In view of the problems of low marking efficiency, high cost and poor accuracy in the existing CNC machine tool processing signal marking relying on manual methods, the present invention proposes a deep learning model based on Transform to model and mark the time series signal marking task. It is difficult for the existing methods to effectively model the multimodal time series signals in the processing process, especially when the signal contains a lot of noise and nonlinear dynamic characteristics, the marking quality is difficult to meet the complex process requirements. At the same time, the existing marking technology shows insufficient robustness in processing long time series signals and capturing the dependencies between multimodal modes, and cannot adapt to the changing processing conditions in the manufacturing scene. The present invention introduces a deep learning model based on the Transform architecture, uses the self-attention mechanism to capture the long-range dependencies in the multimodal signals, and dynamically adjusts the weight distribution of different signal modes. The model integrates the constraints of physical laws in the feature extraction stage, taking into account the influence of multiple factors such as tools, workpieces, and equipment on the signal characteristics during the processing process. Combined with the dynamic weight adjustment strategy and the multi-head attention mechanism, efficient feature capture of multimodal signals can be achieved, and accurate marking of complex process signals can be achieved.

[0006] The method of the present invention can significantly improve the efficiency and quality of processing signal marking, and solve the problem that the marking results in the prior art are easily affected by noise and parameter changes. Compared with traditional marking methods, this method shows higher robustness when processing high-noise signals, and the computational efficiency is significantly improved. It is suitable for the automatic marking and classification of multimodal signals under different working conditions of CNC machining, and provides high-quality basic data support for data-driven process optimization and prediction models in intelligent manufacturing systems.

[0007] To achieve the above object, the present invention proposes an intelligent marking technology for different process signals of milling processing of CNC machine tools based on Transform, which includes the following steps:

[0008] S1: Collect multimodal time series signals during CNC machining, including spindle vibration signals, cutting force signals and acoustic emission signals, use wavelet transform to denoise the original signals, and extract the time-frequency domain features of the signals (such as Mel-frequency cepstral coefficients, frequency components, amplitudes and time spans, etc.) for subsequent feature analysis and modeling.

[0009] S2: For different machining conditions, the signal samples are preliminarily classified according to the process categories (such as idling, acceleration, rough milling, semi-finishing milling and finishing milling). A small number of high-quality labeled samples annotated by experts are used as training sets, and multiple groups of different modal signal feature combination experiments are designed to verify the importance of signal features in the classification task.

[0010] S3: A multimodal signal feature extraction framework is constructed based on the Transform model. The model input includes the temporal features of the multimodal signal. By introducing a multi-head attention mechanism, the long-range dependencies between signals are captured, and the weight distribution of different modal signals is dynamically adjusted to generate a joint feature representation. The multimodal features are further compressed and encoded by combining a feedforward neural network.

[0011] S4: Design an improved multimodal feature sensitivity analysis method, analyze the influence of different modal signals and their feature combinations on the labeling results through ablation experiments, determine the key modalities and features that have the greatest impact on the labeling task, and provide a basis for model optimization.

[0012] S5: Taking the joint feature representation obtained in steps S1 to S4 as input, a self-supervised learning method based on the Transformer architecture is used to model and label the timing signal; a physical constraint loss function is introduced into the model, and the prior physical information in the processing condition (such as processing path, tool vibration mode) is embedded into the model to optimize the accuracy and consistency of the labeling results.

[0013] S6: Using a sequential labeling and verification strategy, the labeling results generated in step S5 are compared with the experimental labeling data, and the labeling accuracy and consistency are calculated; by dynamically adjusting the model parameters, the optimized model is used to predict and label unlabeled time series signal samples, and finally a comprehensive and high-quality labeled data set is generated, and the effectiveness of the labeled data for subsequent processing process analysis is verified through experiments.

[0014] As a further preferred embodiment, the specific application steps of constructing multimodal signal feature extraction based on the Transform model are as follows:

[0015] S31: Segment the multimodal time series signal samples into uniform lengths and construct the input matrix Where N represents the number of samples, T represents the length of the time series segment, and M represents the modal feature dimension. The initial signal features include the time-frequency domain features of spindle vibration, cutting force, and acoustic emission, such as Mel-frequency cepstral coefficients (MFCCs) and amplitude frequency components.

[0016]

[0017] Where Q = XW Q ,K=XW K ,V=XW V ,W Q ,W K , represents the trainable weight matrix, d k Represents the scaling factor of the feature dimension.

[0018] S33: Concatenate and linearly transform the output results of multi-head attention to generate a multimodal joint feature representation F joint :

[0019] F joint =Concat(Attention1,…,Attention H )W O

[0020] in, Represents the projection matrix.

[0021] S34: Combined with Feedforward Neural Network (FFN) for F joint Further nonlinear transformation and feature compression are performed to finally generate the feature representation F for the labeling task final :

[0022] F final =FFN(F joint )

[0023] FFN consists of two layers of fully connected networks, and the activation function uses ReLU.

[0024] More preferably, the specific application steps of time series signal modeling and labeling based on self-supervised learning are as follows:

[0025] S51: The joint feature representation F generated in step S3 is final As the model input, a self-supervised learning method based on the Transformer architecture is used to model and label the time series signal. Specifically, a mask mechanism is used to partially mask the input features and construct a mask sequence The shading ratio is p:

[0026]

[0027] The features at the mask positions are replaced by randomly initialized vectors.

[0028] S52: Define the loss function, including the prediction loss L pred and physical constraint loss L phy :

[0029]

[0030] in, is the model prediction value, y i For real mark.

[0031] S53: Physical constraint loss L phy Combined with the physical laws of the actual machining process, such as machining paths, tool vibration patterns and cutting force fluctuations:

[0032]

[0033] Among them, f pred,j represents the predicted physical quantity, f true,j represents the measured physical quantity, λ is the regularization coefficient, and Reg(g) represents the regularization constraint on the model output.

[0034] S54: Optimize the total loss function L total , compare the labeling results with the real labeling data:

[0035] L total =L pred +αL phy

[0036] Among them, α is the loss weight coefficient. The optimized model is used for automatic labeling of unlabeled signal samples.

[0037] The beneficial effects of the present invention are mainly manifested in:

[0038] (1) A multimodal time series signal feature extraction method based on the Transform model is proposed. By introducing a multi-head attention mechanism, it can effectively capture the long-range dependencies between multimodal signals and dynamically adjust the weight distribution of different modal signals. Compared with traditional feature extraction methods, this method has higher accuracy and generalization ability in modeling complex time series signals.

[0039] (2) An improved multimodal feature sensitivity analysis method is designed. Through ablation experiments and sensitivity analysis, the influence of different modal signal features on the labeling results is accurately evaluated, providing a clear basis for model optimization. Compared with traditional feature sensitivity analysis technology, the method of the present invention has higher computational efficiency and stronger robustness.

[0040] (3) Combining physical constraints and self-supervised learning mechanisms, an optimization method for time series signal labeling tasks is proposed, which embeds the prior physical information of the processing conditions into the model loss function, significantly improving the accuracy and consistency of the labeling task. Compared with the prior art, the present invention shows higher labeling efficiency and labeling quality when processing unlabeled signal samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is an experimental data acquisition platform used by the present invention to verify the proposed method;

[0042] Figure 2 It is a flow chart of the timing signal marking method for milling processing of CNC machine tools based on Transform;

[0043] Figure 3The time series signal samples corresponding to the five processes collected in the experiment are as follows: idling, acceleration, rough milling, semi-finishing milling and finishing milling;

[0044] Figure 4 It is the training and testing process curve of the CNC machine tool milling processing timing signal marking model based on Transform of the present invention;

[0045] Figure 5 It is the boundary division result of different timing signals of CNC machine tool milling processing based on Transform of the present invention;

[0046] Figure 6 It is a comparison between the real time series signal label of the present invention and the prediction result based on the Transform model;

[0047] Table 1 is a comparison of the labeling effect evaluation indicators of different models of the present invention; Table 1

[0048] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0049] Example 1, a method for modeling and labeling a time series signal labeling task based on a deep learning model of Transform. The method first extracts key feature information through a variety of time series signal acquisition and preprocessing steps; then builds a deep learning model based on Transform to model the input signal; and finally realizes the precision and automation of signal labeling through subsequent reasoning and evaluation steps. Specifically, the method includes the following steps:

[0050] (1) Time series signal acquisition and preprocessing: For various time series signals (such as vibration signals, current signals, etc.), collect data that meets the task requirements. Figure 1 The data acquisition platform used for implementing and verifying the method of the present invention is shown. The hardware configuration of the platform includes a high-precision vibration sensor (frequency response range of 0.1Hz to 20kHz), a multi-channel data acquisition card (sampling rate up to 100kHz) and an industrial computer (Intel Core i7 processor, 16GB memory, 1TB storage capacity). Signal acquisition adopts continuous sampling mode and is synchronized by trigger signals.

[0051] After data acquisition is completed, the following preprocessing operations are performed on the original signal:

[0052] 1. Denoising: Use wavelet transform to decompose the signal at multiple scales and remove high-frequency noise components.

[0053] 2. Normalization: Standardize the signal amplitude to satisfy the zero mean and unit variance distribution.

[0054] 3. Segment processing: Divide the signal into multiple segments according to the process status to ensure that each segment of data contains a complete timing pattern.

[0055] In addition, time domain features (such as mean, standard deviation), frequency domain features (such as main frequency, bandwidth) and time-frequency domain features (such as energy distribution of short-time Fourier transform) are also extracted. These features will serve as input data to lay the foundation for subsequent model training.

[0056] (2) Constructing multimodal feature representation: For time series signal features from different sources, a unified multimodal feature representation is constructed through feature fusion strategies (such as weighted fusion or dimensionality reduction processing). This feature representation not only includes the information of a single modality, but also reflects the synergistic relationship between multiple modalities, providing rich information for model input.

[0057] (3) Time series modeling based on Transform model: The present invention uses a deep learning model based on Transform to model time series signals. Specifically, a network architecture including the following components is designed:

[0058] Embedding layer: maps the input multimodal feature vector to a high-dimensional space and adds position information to enhance the temporal modeling capability through position encoding.

[0059] Multi-head self-attention mechanism: Capture the long-term dependencies between different features in the time series signal through multiple attention heads.

[0060] Feedforward Neural Network: The features at each position are further processed using a nonlinear activation function.

[0061] · Residual connection and layer normalization: Improve the training stability of the network and avoid gradient disappearance.

[0062] Through the combination of the above modules, a deep learning model suitable for time series signals is constructed and parameter initialization is completed. Figure 2 The flow chart of the Transform-based timing signal marking method for CNC machine tool milling processing is shown.

[0063] (4) Model training and optimization: Train the constructed Transform model. During the training process, supervised learning is used to segment the labeled time series signals as label data. The specific training details are as follows:

[0064] 1. Loss function: Cross entropy loss is used to measure the difference between the model prediction results and the true labels.

[0065] 2. Optimization algorithm: The Adam optimizer is used, the initial learning rate is set to 0.001, and the cosine annealing learning rate adjustment strategy is adopted to refine the model performance in the later stage of training.

[0066] 3. Regularization: To prevent overfitting, Dropout regularization is introduced, and the dropout rate is set to 0.1. At the same time, weight decay (L2 regularization) is added, and the weight decay factor is set to 1e-5.

[0067] 4. Number of training rounds: The model training was performed for a total of 100 rounds, using the full training set in each round, and the validation set was evaluated after every 10 rounds to monitor the model performance.

[0068] In addition, in order to ensure the stability and convergence speed of the model, a gradient clipping operation was added during the training process to limit the gradient norm to less than 1. Figure 4 (a)-(d) show the training and testing process curves of the Transform-based CNC machine tool milling processing timing signal labeling model of the present invention, showing a trend of gradually decreasing training loss and steadily improving verification accuracy.

[0069] (5) Model reasoning and signal labeling: The unlabeled time series signal is input into the trained Transform model to generate the corresponding state sequence and segment labels. The generated labeling results are used to further analyze the change pattern and important moments of the time series signal. Figure 5 (a)-(d) show the boundary division results of different timing signals of CNC machine tool milling processing based on Transform of the present invention.

[0070] (6) Model evaluation and verification: Taking a 4mm milling tool as an example, the model was verified through multiple sets of real signal data sets, and the model performance was comprehensively evaluated using indicators such as accuracy, recall, and F1 score, and compared with existing methods. Specifically, the experiment selected a vibration signal data set of an industrial equipment and compared the model performance with traditional methods such as LSTM and HMM. In terms of accuracy, the Transform model reached 95.4%, significantly higher than 88.7% of LSTM, 90.2% of Bi-LSTM, and 83.2% of HMM; in terms of recall, the Transform model reached 94.8%, while LSTM, Bi-LSTM, and HMM were 87.3%, 91.4%, and 81.5%, respectively; in terms of F1 score, the Transform model reached 95.1%, showing its advantage in comprehensive performance. Figure 6 The comparison between the real time series signal label of the present invention and the prediction result based on the Transform model is shown, which further supports the above conclusion. The results are shown in Appendix 1.

[0071] Specifically, the implementation details of this method include:

[0072] S1: Collect multiple sets of time series signal data, and perform denoising and normalization. For example, in the vibration signal task, the collected signal needs to be low-pass filtered to remove high-frequency noise. Figure 3 The timing signal samples corresponding to the process categories (idling, acceleration, rough milling, semi-finishing milling and finishing milling) collected in the experiment are displayed.

[0073] S2: Extract multimodal features (such as mean, standard deviation, frequency components, etc.), reduce the dimensionality of high-dimensional features through principal component analysis (PCA), and obtain multimodal feature representations that can be used as model input.

[0074] S3: Build a Transform model with N layers of self-attention modules, where each layer contains a multi-head attention mechanism and a feed-forward network. The model's hyperparameters (such as the number of attention heads, embedding dimension, etc.) are tuned through cross-validation.

[0075] S4: Use the training data set to supervise the model. Each set of time series signal data is divided into several segments, and the labels are the corresponding state sequences. During the training process, regularization terms (such as Dropout) are added to prevent overfitting.

[0076] S5: In the model inference phase, the generated state sequence is post-processed using the Viterbi decoding algorithm to ensure that the output timing labels meet the continuity and logic of the task.

[0077] S6: Verify the labeling performance of the model through experiments. For example, a standard data set is selected to compare the benchmark models (LSTM, Bi-LSTM, HMM), and the results show that the Transform model of the present invention is superior to the benchmark models in terms of long-term dependency processing and labeling accuracy.

[0078] Experimental results show that the method proposed in this paper can effectively solve the complex modeling and labeling problems of time series signals, and has higher accuracy, robustness and applicability. This paper collects time series signals in various practical scenarios (such as industrial equipment vibration signals), and verifies the effectiveness and universality of the model through experiments. The results show that its labeling effect is significantly better than traditional methods.

[0079] Finally, it should be noted that the above are only some embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and there are many variations. All variations that can be directly derived or associated with the content disclosed by ordinary technicians in this field should be considered as the protection scope of the present invention.

Claims

1. An intelligent marking method for multi-process signals of CNC machine tool milling based on Transform, characterized in that: The method comprises the following steps: S1: Signal acquisition and preprocessing, collecting multi-modal time series signals during CNC machining, including spindle vibration signals, cutting force signals and acoustic emission signals; denoising the original signals through wavelet transform, and extracting the time-frequency domain features of the signals, including Mel-frequency cepstrum coefficients, frequency components, amplitudes and time spans; S2: Sample classification and feature combination analysis: for different machining conditions, the signal samples are preliminarily classified according to the process categories (such as idling, acceleration, rough milling, semi-finishing milling and finishing milling); a small number of high-quality labeled samples annotated by experts are used as training sets, and multiple groups of different modal signal feature combination experiments are designed to verify the importance of features in the classification task; S3: Multimodal feature extraction and joint modeling. A multimodal signal feature extraction framework is constructed based on the Transform model. The temporal features of multimodal signals are input. The long-range dependencies between signals are captured by introducing a multi-head attention mechanism. The weight distribution of different modal signals is dynamically adjusted to generate joint feature representations. The joint features are compressed and encoded by combining a feedforward neural network. S4: Multimodal feature sensitivity analysis: design an improved feature sensitivity analysis method, analyze the influence of different modal signals and their feature combinations on the labeling results through ablation experiments, and determine the key modalities and features that have the greatest impact on the labeling task; S5: A self-supervised learning labeling model based on physical constraints, which uses a self-supervised learning method based on the Transformer architecture to model and label the joint feature representation obtained in steps S1 to S4; the model introduces a physical constraint loss function to embed the prior physical information in the processing conditions into the model to optimize the accuracy and consistency of the labeling results; S6: Sequential labeling and verification, compare the labeling results generated in step S5 with the experimental labeling data, and calculate the labeling accuracy and consistency; by dynamically adjusting the model parameters, the optimized model is used to predict and label unlabeled time series signal samples, and finally a comprehensive and high-quality labeling data set is generated, and its effectiveness in subsequent processing analysis is verified.

2. According to claim 1, a method for intelligently marking multi-process signals of milling processing of CNC machine tools based on Transform, characterized in that: The signal acquisition and preprocessing in step S1 includes the following sub-steps: S11: signal denoising, using wavelet transform method to perform noise reduction on the collected multimodal signals; S12: Feature extraction, extracting the time-frequency domain features of the signal, including Mel-frequency cepstral coefficients (MFCCs), frequency components, amplitudes, and time spans, for subsequent analysis and modeling. MFCCs are calculated based on short-time Fourier transform (STFT). The specific formula is as follows: in: S k represents the energy of the kth Mel filter; K represents the total number of Mel filters; n is the MFCC order, and its value range is n=1,2L N.

3. According to claim 1, a method for intelligently marking multi-process signals of milling processing of CNC machine tools based on Transform, characterized in that: The multimodal feature extraction and joint modeling in step S3 includes the following sub-steps: S31: Multimodal signal feature input, taking the preprocessed multimodal signal as input to construct the embedding matrix of the Transform model; S32: The introduction of the multi-head attention mechanism calculates the long-range dependency between signal features through the following formula: in, Represented as query, key and value matrices, is the trainable weight matrix, d k Represents the scaling factor of the feature dimension. S33: Modality weight adjustment, dynamically adjusts the weight distribution of different modality signals based on the multi-head attention mechanism to generate joint feature representation; S34: Feedforward neural network encoding, the joint features are further compressed and encoded through the feedforward neural network.

4. According to claim 1, a method for intelligently marking multi-process signals of CNC machine tool milling based on Transform, characterized in that: The multimodal feature sensitivity analysis in step S4 includes the following sub-steps: S41: Ablation experiment design, remove different modal signals or signal feature combinations in turn, and compare and analyze the changes in labeling task performance; S42: Quantify feature sensitivity, calculate the contribution of different modal signals and their feature combinations to the labeling task performance, screen the key modes and features that have the greatest impact on the labeling task, and provide a basis for model optimization.

5. According to claim 1, a method for intelligently marking multi-process signals of CNC machine tool milling based on Transform, characterized in that: The physical constraint-based self-supervised learning labeling model in step S5 includes the following sub-steps: S51: Physical constraint embedding, a physical constraint loss function is designed to embed prior physical information such as machining path and tool vibration mode into the model. This formula combines the prior physical information of machining path and tool vibration mode to measure the deviation of the marking result from the ideal physical constraint in the form of mean square error, thereby optimizing the accuracy and consistency of the model marking. The physical constraint loss function is designed as follows: in: L phys represents the total physical constraint loss; A signal tag representing the model prediction; f path (x i ) represents the ideal marking function based on the machining path, which is generated by the machining geometry trajectory; f vib (x i ) represents the ideal signal representation based on the tool vibration mode, calculated from the tool modal characteristics; λ1 and λ2 represent weight parameters, which are used to balance the machining path constraints and vibration mode constraints; N represents the number of samples. S52: Joint feature optimization, optimizing the labeling accuracy and consistency of the model during self-supervised learning; S53: Generate labeling results, predict and label the unlabeled time series signal samples based on the optimized model to generate a comprehensive labeled data set.

Images