A transform-based intelligent marking method and system for milling timing signals of a numerical control machine tool

Abstract

The application provides a kind of intelligent marking method for multi-process signal of milling machining of numerical control machine tool based on Transform. First, collect multi-modal time series signals (such as spindle vibration, cutting force, acoustic emission signal), denoise by wavelet transform and extract time-frequency domain features, classify signal samples according to process category, and construct high-quality training set combined with expert labeled samples. Second, design a multi-modal feature extraction framework based on Transform model, introduce multi-head attention mechanism, dynamically adjust modal weight to generate joint feature representation, determine key modal and feature through improved multi-modal feature sensitivity analysis, and combine self-supervised learning and physical constraint optimization model to accurately mark signals. Finally, use sequential labeling verification strategy to generate high-quality labeled dataset, improve the accuracy and consistency of process analysis. The application can be widely used in complex precision part numerical control machine tool machining quality analysis and intelligent optimization, 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, specifically to an intelligent marking technology for signals of different processes in the milling process of CNC machine tools. Background Technology

[0002] During CNC milling, a large amount of complex process signal data is generated along with workpiece machining, tool movement, and equipment operation. These signals contain rich process status information and are crucial for process optimization, fault diagnosis, and quality control. However, unlabeled raw data often contains noise and redundancy, leading to inefficient data analysis and modeling when used directly. How to accurately and efficiently label this data has become an important research topic in the field of intelligent manufacturing.

[0003] Currently, the labeling of signal data for different processes in CNC machine tool machining mainly relies on manual labor. Faced with the massive amounts of multimodal signal data in industrial production, manual labeling is time-consuming, labor-intensive, and extremely costly, making it difficult to meet actual needs. Furthermore, due to differences in the professional level of labelers and their insufficient understanding of the machining process, manual labeling suffers from high error rates and poor data consistency. This low-quality labeling negatively impacts subsequent data utilization and analysis, reducing the predictive accuracy and decision reliability of models. Although some rule-based automatic labeling methods and machine learning-assisted labeling methods have been proposed, these methods typically rely on pre-set rules or large amounts of high-quality sample data for training, making them difficult to adapt to complex and variable machining scenarios and exhibiting significant limitations when processing signals from different processes.

[0004] This invention proposes a Transform-based deep learning model for modeling and labeling time-series signal labeling tasks. This method utilizes a self-attention mechanism to extract signal features and dynamically adjusts modal weights to achieve efficient labeling and accurate classification of signals from multiple processes, 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] To address the problems of low efficiency, high cost, and poor accuracy associated with manual methods for marking CNC machine tool machining signals in existing technologies, this invention proposes a method based on...

[0006] This invention utilizes a deep learning model based on the Transform architecture to model and label time-series signals. Existing methods struggle to effectively model multimodal time-series signals during manufacturing processes, especially when signals contain significant noise and nonlinear dynamic characteristics, resulting in labeling quality that fails to meet the demands of complex processes. Furthermore, existing labeling techniques exhibit insufficient robustness in handling long-term signals and capturing dependencies between multiple modalities, making them unsuitable for the variable processing conditions in manufacturing scenarios. This invention introduces a deep learning model based on the Transform architecture, employing a self-attention mechanism to capture long-range dependencies in multimodal signals and dynamically adjusting the weight allocation of different signal modalities. The model incorporates physical constraints during the feature extraction stage, considering the influence of multiple factors such as tools, workpieces, and equipment on signal characteristics during processing. By combining a dynamic weight adjustment strategy and a multi-head attention mechanism, efficient feature capture of multimodal signals can be achieved, enabling accurate labeling of signals from complex processes.

[0007] This invention significantly improves the efficiency and quality of machining signal marking, solving the problem that marking results are easily affected by noise and parameter variations in existing technologies. Compared with traditional marking methods, this method exhibits higher robustness when processing high-noise signals and significantly improves computational efficiency. It is suitable for automated marking and classification of multimodal signals under different CNC machining conditions, providing high-quality basic data support for data-driven process optimization and prediction models in intelligent manufacturing systems.

[0008] To achieve the above objectives, this invention proposes an intelligent tagging technology for signals of different processes in CNC milling based on Transform, which includes the following steps:

[0009] 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, amplitude, and time span) for subsequent feature analysis and modeling.

[0010] S2: For different machining conditions, the signal samples are initially classified according to the process category (such as idling, acceleration, rough milling, semi-finish milling and finish milling). A small number of high-quality labeled samples annotated by experts are used as the training set, and multiple sets of experiments combining different modal signal features are designed to verify the importance of signal features in the classification task.

[0011] S3: Construct a multimodal signal feature extraction framework based on the Transform model. The model input includes the temporal features of multimodal signals. By introducing a multi-head attention mechanism, long-range dependencies between signals are captured, and the weight allocation of different modal signals is dynamically adjusted to generate joint feature representations. The multimodal features are further compressed and encoded by combining a feedforward neural network.

[0012] S4: Design an improved multimodal feature sensitivity analysis method. Through ablation experiments, analyze the influence of different modal signals and their feature combinations on the labeling results, identify the key modes and features that have the greatest impact on the labeling task, and provide a basis for model optimization.

[0013] S5: Using the joint feature representation obtained in steps S1 to S4 as input, a self-supervised learning method based on the Transformer architecture is adopted to model and label the time series signal; a physical constraint loss function is introduced into the model to embed prior physical information in the machining conditions (such as machining path and tool vibration mode) into the model to optimize the accuracy and consistency of the labeling results.

[0014] S6: Using a sequential labeling and validation strategy, the labeling results generated in step S5 are compared with the experimental labeling data to 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, ultimately generating a comprehensive and high-quality labeled dataset. The effectiveness of the labeled data for subsequent processing analysis is verified through experiments.

[0015] As a further preferred approach, the specific application steps for multimodal signal feature extraction based on the Transform model are as follows:

[0016] S31: Divide the multimodal time-series signal samples into segments of uniform length and construct the input matrix. Where N represents the number of samples, T represents the length of the time 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 (MFCC) and amplitude frequency components.

[0017]

[0018] Where Q = XW Q K = XW K , Let d represent the trainable weight matrix. k This represents the scaling factor for the feature dimension.

[0019] S33: Concatenate and linearly transform the output of multi-head attention to generate a multimodal joint feature representation F. joint :

[0020] F joint =Concat(Attention1,...,Attention H W O

[0021] in, This represents the projection matrix.

[0022] S34: Combining a feedforward neural network (FFN) with F joint Further nonlinear transformations and feature compression are performed to finally generate the feature representation F for the labeling task. final :

[0023] F final =FFN(F joint FFN consists of two fully connected layers, and the activation function is ReLU.

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

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

[0026]

[0027] The features of the mask location 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, y is the model's predicted value. i This is a true label.

[0031] S53: Physical constraint loss L phy Taking into account the physical laws governing the actual machining process, such as machining paths, tool vibration modes, and cutting force fluctuations:

[0032]

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

[0034] S54: Optimize the total loss function L total The labeling results are compared with the actual labeled data:

[0035] Ltotal =L pred +αL phy

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

[0037] The beneficial effects of this invention are mainly reflected in:

[0038] (1) A feature extraction method for multimodal time series signals 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 allocation 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 was designed. Through ablation experiments and sensitivity analysis, the influence of different modal signal features on the labeling results was accurately evaluated, providing a clear basis for model optimization. Compared with traditional feature sensitivity analysis techniques, the method of this invention has higher computational efficiency and stronger robustness.

[0040] (3) Combining physical constraints and a self-supervised learning mechanism, an optimization method for time-series signal labeling tasks is proposed. Prior physical information from the processing conditions is embedded into the model loss function, significantly improving the accuracy and consistency of the labeling task. Compared with existing technologies, this invention demonstrates higher labeling efficiency and quality when processing unlabeled signal samples. Attached Figure Description

[0041] Figure 1 This is the experimental data acquisition platform used in this invention for verifying the proposed method;

[0042] Figure 2 This is a flowchart of a CNC machine tool milling timing signal marking method based on Transform;

[0043] Figure 3 These are timing signal samples collected in the experiment, categorized by process type: idling, acceleration, rough milling, semi-finish milling, and finish milling, corresponding to five processes.

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

[0045] Figure 5 This invention provides the boundary partitioning results for different timing signals in CNC machine tool milling based on Transform.

[0046] Figure 6 This invention compares the actual time-series signal labels with the prediction results based on the Transform model.

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

[0048] Table 1

[0049]

[0050] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0051] Example 1: A method for modeling and labeling time-series signals using a Transform-based deep learning model. This method first extracts key feature information through various time-series signal acquisition and preprocessing steps; then, it constructs a Transform-based deep learning model to model the input signal; finally, through subsequent inference and evaluation steps, it achieves accurate and automated signal labeling. Specifically, it includes the following steps:

[0052] (1) Timing signal acquisition and preprocessing: For various timing signals (such as vibration signals, current signals, etc.), acquire data that meets the task requirements. Figure 1 The data acquisition platform for method implementation and verification of the present invention is illustrated. 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 employs a continuous sampling mode and is synchronized via a trigger signal.

[0053] After data acquisition is complete, the following preprocessing operations are performed on the raw signal:

[0054] 1. Noise reduction: Wavelet transform is used to decompose the signal into multiple scales and remove high-frequency noise components.

[0055] 2. Normalization: The signal amplitude is standardized to meet the zero mean and unit variance distribution.

[0056] 3. Segmentation: The signal is divided into multiple segments according to the process status to ensure that each segment contains a complete timing pattern.

[0057] In addition, time-domain features (such as mean and standard deviation), frequency-domain features (such as dominant frequency and bandwidth), and time-frequency-domain features (such as energy distribution of short-time Fourier transform) were extracted. These features will serve as input data, laying the foundation for subsequent model training.

[0058] (2) Constructing a 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). This feature representation not only includes information from a single modality, but also reflects the collaborative relationship between multiple modalities, providing rich information for model input.

[0059] (3) Transform-based temporal modeling: This invention employs a Transform-based deep learning model to model temporal signals. Specifically, a network architecture is designed that includes the following components:

[0060] • Embedding layer: Maps the input multimodal feature vectors to a high-dimensional space and adds positional information, thereby enhancing the temporal modeling capability through positional encoding.

[0061] • Multi-head self-attention mechanism: Captures long dependencies between different features in a time-series signal by using multiple attention heads.

[0062] • Feedforward neural networks: use non-linear activation functions to further process the features at each location.

[0063] • Residual connections and layer normalization: Improve network training stability and avoid gradient vanishing.

[0064] By combining the above modules, a deep learning model suitable for time-series signals is constructed, and parameter initialization is completed. Figure 2 A flowchart illustrating a Transform-based method for marking timing signals during CNC milling operations is presented.

[0065] (4) Model Training and Optimization: The constructed Transform model is trained. During training, supervised learning is used, and the labeled time-series signals are segmented as label data. Specific training details are as follows:

[0066] 1. Loss function: Cross-entropy loss is selected to measure the difference between the model's prediction and the true label.

[0067] 2. Algorithm optimization: The Adam optimizer is used, with an initial learning rate of 0.001, and a cosine annealing learning rate adjustment strategy is adopted to refine the model performance in the later stages of training.

[0068] 3. Regularization: To prevent overfitting, Dropout regularization is introduced with a dropout rate of 0.1. Weight decay (L2 regularization) is also added with a weight decay factor of 1e-5.

[0069] 4. Number of training rounds: The model is trained for a total of 100 rounds, with the complete training set used in each round, and the validation set is evaluated every 10 rounds to monitor model performance.

[0070] In addition, to ensure the stability and convergence speed of the model, a gradient clipping operation was added during training to limit the gradient norm to within 1. Figure 4 (a)-(d) show the training and testing process curves of the Transform-based CNC machine tool milling timing signal labeling model of the present invention, showing a trend of gradually decreasing training loss and steadily increasing verification accuracy.

[0071] (5) Model Inference and Signal Labeling: Input the unlabeled time series signal into the trained Transform model to generate the corresponding state sequence and segment labels. Use the generated labeling results to further analyze the change patterns and important moments of the time series signal. Figure 5 (a)-(d) show the boundary partitioning results of different timing signals in CNC machine tool milling based on Transform according to the present invention.

[0072] (6) Model Evaluation and Validation: Taking a 4mm milling cutter as an example, the model was validated using multiple sets of real signal datasets. Accuracy, recall, and F1 score were used to comprehensively evaluate the model's performance, and it was compared with existing methods. Specifically, a vibration signal dataset of industrial equipment was selected for the experiment, and the model performance was compared with traditional methods such as LSTM and HMM. In terms of accuracy, the Transform model achieved 95.4%, significantly higher than LSTM's 88.7%, Bi-LSTM's 90.2%, and HMM's 83.2%. In terms of recall, the Transform model achieved 94.8%, while LSTM, Bi-LSTM, and HMM achieved 87.3%, 91.4%, and 81.5%, respectively. In terms of F1 score, the Transform model achieved 95.1%, demonstrating its superior overall performance. Figure 6 The comparison between the actual time-series signal labels of the present invention and the prediction results based on the Transform model is shown, further supporting the above conclusions. The results are shown in Appendix Table 1.

[0073] Specifically, the implementation details of this method include:

[0074] S1: Acquire multiple sets of time-series signal data and perform denoising and normalization processing. For example, in a vibration signal task, the acquired signals need to be low-pass filtered first to remove high-frequency noise. Figure 3 The experiment showcases timing signal samples collected according to process categories (idling, acceleration, rough milling, semi-finish milling, and finish milling).

[0075] S2: Extract multimodal features (such as mean, standard deviation, frequency components, etc.), and reduce the dimensionality of high-dimensional features through principal component analysis (PCA) to obtain multimodal feature representations that can be used as input to the model.

[0076] S3: Construct a Transform model containing N layers of self-attention modules, where each layer includes a multi-head attention mechanism and a feedforward network. The model's hyperparameters (such as the number of attention heads, embedding dimension, etc.) are tuned through cross-validation.

[0077] S4: Supervised learning of the model is performed using the training dataset. Each set of time-series signal data is divided into several segments, labeled with the corresponding state sequence. During training, a regularization term (such as Dropout) is added to prevent overfitting.

[0078] S5: During the model inference phase, the Viterbi decoding algorithm is used to post-process the generated state sequence to ensure that the output time stamps conform to the continuity and logic of the task.

[0079] S6: Verify the model's labeling performance through experiments. For example, by comparing it with benchmark models (LSTM, Bi-LSTM, HMM) on a standard dataset, the results show that the Transform model of this invention outperforms the benchmark models in both long-term dependency handling and labeling accuracy.

[0080] Experimental results show that the method proposed in this invention can effectively solve the complex modeling and labeling problem of time-series signals, and has higher accuracy, robustness, and applicability. This invention collected time-series signals from various real-world scenarios (such as vibration signals from industrial equipment), and experimentally verified the effectiveness and universality of the model. The results show that its labeling effect is significantly better than traditional methods.

[0081] Finally, it should be noted that the above descriptions are merely some embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should be considered within the scope of protection of this invention.

Claims

1. A smart labeling method for multi-process signals in CNC milling machining based on Transform, characterized in that: The method includes the following steps: S1: Signal acquisition and preprocessing, acquiring multimodal 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 cepstral coefficients, frequency components, amplitude and time span; S2: Sample classification and feature combination analysis. For different processing conditions, the signal samples are initially classified according to the process categories of idle, acceleration, rough milling, semi-finish milling and finish milling. Using a small number of high-quality labeled samples annotated by experts as the training set, multiple sets of signal feature combination experiments with different modes 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 the multimodal signal are input. By introducing a multi-head attention mechanism, long-range dependencies between signals are captured, and the weight allocation of different modal signals is dynamically adjusted to generate joint feature representations. The joint features are compressed and encoded by a feedforward neural network. S4: Multimodal feature sensitivity analysis. An improved feature sensitivity analysis method is designed. Through ablation experiments, the influence of different modal signals and their feature combinations on the labeling results is analyzed to 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. It adopts a self-supervised learning method based on the Transformer architecture to model and label the joint feature representations obtained in steps S1 to S4. The model introduces a physical constraint loss function to embed prior physical information in the processing conditions into the model, thereby optimizing the accuracy and consistency of the labeling results. S6: Sequential labeling and validation. The labeling results generated in step S5 are compared with the experimental labeling data to 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, ultimately generating a comprehensive and high-quality labeled dataset and validating its effectiveness in subsequent processing analysis.

2. The intelligent marking method for multi-process signals in CNC milling based on Transform as described in claim 1, characterized in that: The signal acquisition and preprocessing in step S1 includes the following sub-steps: S11: Signal denoising, using wavelet transform to denoise the acquired multimodal signals; S12: Feature extraction, extracting the time-frequency domain features of the signal, including Mel-frequency cepstral coefficients (MFCCs), frequency components, amplitude, and time span, for subsequent analysis and modeling; The Mel-frequency cepstral coefficients (MFCCs) are calculated based on the Short-Time Fourier Transform (STFT), using the following formula: in: S k The energy of the k-th Mel filter is represented by denoted by K; the total number of Mel filters is represented by K; and n is the order of the Mel frequency cepstral coefficients (MFCCs), which can be any positive integer. number.

3. The intelligent marking method for multi-process signals in CNC milling based on Transform as described in claim 1, characterized in that: Step S3, multimodal feature extraction and joint modeling, includes the following sub-steps: S31: Multimodal signal feature input, using the preprocessed multimodal signal as input to construct the embedding matrix of the Transform model; S32: The introduction of a multi-head attention mechanism calculates the long-range dependencies between signal features using the following formula: Where Q = XW Q K = XW K V = XW V Where Q is the query matrix, K is the key matrix, and V is the value matrix. For a trainable weight matrix, d k The scaling factor representing the feature dimension; S33: Modal weight adjustment, dynamically adjusting the weight allocation of different modal signals based on multi-head attention mechanism to generate joint feature representation; S34: Feedforward neural network encoding, which further compresses and encodes joint features through a feedforward neural network.

4. The intelligent marking method for multi-process signals in CNC milling based on Transform as described in claim 1, characterized in that: The multimodal feature sensitivity analysis in step S4 includes the following sub-steps: S41: Ablation experiment design, sequentially remove different modal signals or combinations of signal features, and compare and analyze the changes in the performance of the labeling task; S42: Feature sensitivity quantization calculates the contribution of different modal signals and their feature combinations to the performance of the labeling task, filters out the key modalities and features that have the greatest impact on the labeling task, and provides a basis for model optimization.

5. The intelligent marking method for multi-process signals in CNC milling based on Transform as described in claim 1, characterized in that: Step S5, the physical constraint-based self-supervised learning labeling model, includes the following sub-steps: S51: Physical constraint embedding. Design a physical constraint loss function to embed prior physical information such as machining path and tool vibration mode into the model. This physical constraint loss function measures the deviation between the labeling result and the ideal physical constraint in the form of mean square error by combining the prior physical information of machining path and tool vibration mode, thereby optimizing the accuracy and consistency of model labeling. The physical constraint loss function is designed as follows: in: L phys This represents the total physical constraint loss; The signal label represents the model's prediction; f path (x i ) represents the ideal labeling function based on the machining path, generated from the machining geometry trajectory; f vib (x i This represents the ideal signal representation based on the tool vibration mode, calculated from the tool modal characteristics; λ1 and λ2 represent weighting parameters used to balance machining path constraints and vibration mode constraints; N represents the number of samples; S52: Joint feature optimization optimizes the labeling accuracy and consistency of the model during self-supervised learning; S53: Labeling results generation. Based on the optimized model, unlabeled time series signal samples are predicted and labeled to generate a comprehensive labeled dataset.

Images