Milling cutter wear prediction method based on parallel spatial-temporal feature extraction and physical description

Abstract

The invention relates to the technical field of machining cutter wear prediction, and provides a milling cutter wear prediction method based on parallel spatial-temporal feature extraction and physical description, and the method comprises the steps: collecting a cutter wear monitoring signal, carrying out the denoising of the signal through VMD decomposition in combination with a PCC coefficient, and constructing a data set; inputting data in the data set into a CNN-BiGRU parallel structure prediction model, performing joint modeling on local and global features of tool wear, and introducing an improved space-time double attention mechanism to enhance features; selecting a prediction model hyper-parameter and constructing a training unit to carry out performance evaluation to obtain an optimal prediction result; and constructing a staged physical degradation model based on tool wear characteristics, introducing a particle filter algorithm, and dynamically updating parameters of the physical degradation model by taking a prediction result of a prediction model as an observed quantity to realize residual service life estimation. According to the method, the prediction precision, the physical interpretability and the result credibility are improved.

Description

Technical Field

[0001] The embodiments of the present invention generally relate to the field of machining tool wear prediction technology, and particularly to a method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description. Background Technology

[0002] exist In milling of difficult-to-machine materials, cutting tools are prone to wear and failure due to high mechanical stress, thermal stress, and impact loads, directly leading to dimensional deviations and reduced surface quality of the workpiece. In traditional production, operators often subjectively determine tool change timing based on experience indicators such as machining time, spindle vibration, and chip color. This approach often results in assessment bias—conservative tool changes only achieve a tool life utilization rate of 50%-80%, while delayed tool changes may lead to tool failure and workpiece damage, causing approximately 20% downtime losses. This not only results in inefficient resource consumption but also falls significantly short of the digital and real-time requirements of intelligent manufacturing.

[0003] Traditional tool life prediction methods mainly include physical modeling and data-driven methods. Physical modeling is limited by the multi-field coupled nonlinear wear mechanism, typically based on idealized assumptions, making it difficult to accurately reflect the actual degradation process. Furthermore, parameter calibration relies heavily on offline experiments and simulations, resulting in high costs. Data-driven methods achieve efficient prediction with the support of multi-source sensors and deep learning; however, due to the "black box" nature of data-driven models, their decision logic and output results are difficult to correlate with the actual tool degradation process, leading to insufficient interpretability.

[0004] Given the shortcomings of single models, hybrid prediction methods that integrate the advantages of physical models and data-driven models have gradually become a research hotspot. However, in existing physical modeling techniques, tool wear modeling often uses logarithmic functions or polynomials to describe the wear degradation process. The model structure is relatively simple and it is difficult to characterize the nonlinear feature of accelerated degradation in the later stages of wear, failing to uniformly reflect the continuous evolution law of the tool's entire life cycle from slow wear to rapid failure. On the other hand, existing data-driven models mostly use single structures such as convolutional neural networks (CNN) or LSTM. Multi-structure models are usually based on a serial architecture of CNN and LSTM. As the model complexity increases, the computational storage and time costs increase significantly, training costs rise, and prediction timeliness is limited. At the same time, most methods rely on manually extracting wear state features, which can easily lead to a decrease in model efficiency and prediction accuracy under large-scale data conditions.

[0005] Therefore, there is an urgent need to develop a tool life prediction method that integrates the advantages of physical model constraints and data-driven approaches, effectively controlling model complexity while ensuring prediction accuracy. This method can accurately characterize the tool's entire life cycle degradation process and improve the model's timeliness and interpretability, thereby supporting the digital and green development needs of intelligent manufacturing. Summary of the Invention

[0006] To address the above issues, this invention constructs a parallel deep learning architecture to achieve collaborative modeling of CNN and BiGRU. This reduces model complexity while improving prediction timeliness. CNN extracts local features from time-series signals, while BiGRU characterizes global temporal degradation patterns, compensating for the shortcomings of a single structure. Simultaneously, a spatiotemporal dual attention mechanism is introduced to achieve end-to-end modeling, directly and adaptively extracting key features from the original signal, avoiding manual feature engineering and improving prediction efficiency. Furthermore, by combining the logarithmic-exponential composite function and particle filtering algorithm, dynamic fusion of the physical model and the data-driven model is achieved, improving prediction accuracy, physical interpretability, and result credibility, providing reliable technical support for tool remaining life prediction and predictive maintenance.

[0007] According to an embodiment of the present invention, a method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description is provided.

[0008] In a first aspect of the invention, a method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description is provided. The method includes: Step S01: Collect tool wear monitoring signals, use VMD decomposition combined with PCC coefficients to denoise the signals, and construct a dataset; Step S02: Input the data in the dataset into the CNN-BiGRU parallel structure prediction model to jointly model the local and global features of tool wear, and introduce an improved spatiotemporal dual attention mechanism to enhance the features; Step S03: Select the hyperparameters of the prediction model and construct training units to evaluate performance and obtain the optimal prediction results; Step S04: Construct a phased physical degradation model based on tool wear characteristics, introduce a particle filter algorithm, and use the prediction results of the prediction model as observations to dynamically update the parameters of the physical degradation model, thereby realizing the estimation of remaining service life.

[0009] Furthermore, the specific steps of step S01 are as follows: Step S011: Collect vibration and force signals during the milling process by using an accelerometer and force gauge installed on the machine tool spindle or worktable; Step S012: Perform VMD decomposition on the vibration and force signals; Step S013: Calculate the relationship between each IMF component obtained from VMD decomposition and the original signal. The Pearson correlation coefficient (PCC) between them; Step S014: Denoise the vibration and force signals in the X, Y, and Z directions in sequence, and combine them with the corresponding tool wear to construct a dataset; Step S015: Cut the data into sample sequences and divide them into training set, validation set and test set.

[0010] Furthermore, the specific steps of step S02 are as follows: Step S021: Input the training set data into the 1D-CNN unit to extract and compress the local features related to tool wear layer by layer; Step S022: Input the local features into the DAM module, perform channel attention and spatial attention calculations in sequence, and output the spatial feature vector; Step S023: Input the training set data into the 1D-CNN unit for spatial feature extraction, and simultaneously input it into the BiGRU unit for temporal feature extraction; Step S024: Input the temporal features into the DTA module, perform dimension recalibration and time step selection, and output the temporal feature vector; Step S025: Concatenate the spatial feature vector and the temporal feature vector along the Batchsize direction to obtain the fused features.

[0011] Furthermore, the DAM module described in step S022 is an improvement upon the CBAM module, and the steps are as follows: In the channel self-attention submodule, the local features are first divided into two branches, which perform convolution mapping and spatial dimension unfolding and rearrangement to obtain the first output feature and the second output feature. The self-attention weights between channels are calculated by the Softmax function for the first output feature and then weighted and fused with the second output feature to obtain the fused feature. The fused feature is then processed by convolution, LayerNorm and Sigmoid activation functions to generate channel attention weights, which are then weighted with the original input features channel by channel. In the spatial self-attention submodule, the input features are compressed through convolution and then multi-layer dilated convolution and normalization are introduced; after generating a spatial attention weight map through convolution, normalization and sigmoid activation, it is spatially weighted with the original input features; The weighted results of channel self-attention and spatial self-attention are fused to obtain the output of the DAM module.

[0012] Furthermore, the DTA module is improved based on the self-attention model, and the specific steps are as follows: Global average pooling and time step compression are performed on the temporal features output by BiGRU to obtain global tool wear state information; The importance of each dimension at each time step is learned by using two fully connected layers in the linear bottleneck layer, and the sigmoid function is used to generate dimension weights. The dimension weights are multiplied element-wise with the original features to obtain dimension-enhanced features, and the dimension-enhanced features are then non-linearly mapped using the tanh activation function. The importance of each time step is calculated using dynamic vectors, and the temporal attention weights are obtained by Softmax normalization. These temporal attention weights are then applied to the feature sequence to obtain the output of the DTA module.

[0013] Furthermore, the specific steps of step S04 are as follows: Step S041: Based on the nonlinear degradation characteristics of tool wear, construct a physical degradation model of tool wear based on the logarithmic-exponential form: , in, express Physically predicted tool wear value at any given time; , , , These are model parameters, used to characterize the degradation characteristics at different wear stages; Step S042: Physical degradation model of tool wear with respect to time Differentiating the equation, we obtain the ordinary differential degeneracy equation for tool wear: , in, This indicates the rate of degradation of tool wear over time. Step S043: Construct the state transition equation and observation equation for tool wear, that is, equation... Discretize and construct the state equation for tool wear: , in, , This is state transition noise; The Runge-Kutta method is used to obtain tool wear values: , in, Let the time step be [value], and: , Using the tool wear prediction values ​​output by the parallel data-driven model as observations, an observation equation is constructed: , in, For the first The parallel data at each time step drives the model output of the predicted wear value. To observe noise; Step S044: The particle filter algorithm is used to dynamically update the physical degradation model parameters as observations by using the prediction results of the parallel model, thereby realizing the estimation of remaining service life.

[0014] Furthermore, the particle filter algorithm flow is as follows: In the prediction phase, the numerical solution of the state equation obtained by Runge-Kutta is used to propagate the state of each particle; in the update phase, ... As an observation, the first is calculated according to the observation equation. Particle weight : , in, The variance representing the observation noise. Indicates the first Particles in The physical predicted wear value corresponding to the given time; The weights of all particles are normalized to obtain normalized weights. : , Based on the weighted particle resampling, the updated tool wear estimate is obtained: , By repeating the above process and recursively updating through particle filtering, a sequence of physical prediction values ​​for tool wear is continuously obtained, thus constructing the wear evolution trajectory throughout the tool's entire life cycle.

[0015] In a second aspect of the invention, an apparatus for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description is provided. The apparatus includes: Signal acquisition and denoising module: Acquires tool wear monitoring signals, uses VMD decomposition combined with PCC coefficients to denoise the signals, and constructs a dataset; Parallel feature extraction module: Input the data in the dataset into the CNN-BiGRU parallel structure prediction model, jointly model the local and global features of tool wear, and introduce an improved spatiotemporal dual attention mechanism to enhance the features; Parallel model training module: Selects prediction model hyperparameters and constructs training units for performance evaluation to obtain the optimal prediction results; The digital-analog fusion module constructs a phased physical degradation model based on tool wear characteristics, introduces a particle filtering algorithm, and uses the prediction results of the prediction model as observations to dynamically update the parameters of the physical degradation model, thereby realizing the estimation of remaining service life.

[0016] In a third aspect of the invention, an electronic device is provided. The electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the program to implement the method according to a first aspect of the invention.

[0017] In a fourth aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method according to a first aspect of the invention.

[0018] This invention constructs a parallel deep learning architecture to achieve collaborative modeling of CNN and BiGRU, improving prediction timeliness while reducing model complexity. CNN extracts local features of time series signals, while BiGRU characterizes global temporal degradation patterns, compensating for the shortcomings of a single structure. Simultaneously, a spatiotemporal dual attention mechanism is introduced to achieve end-to-end modeling, directly and adaptively extracting key features from the original signal, avoiding manual feature engineering and improving prediction efficiency. Furthermore, by combining the logarithmic-exponential composite function and particle filtering algorithm, dynamic fusion of the physical model and the data-driven model is achieved, improving prediction accuracy, physical interpretability, and result credibility, providing reliable technical support for tool remaining life prediction and predictive maintenance.

[0019] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description.

[0020] Beneficial effects: 1. A parallel deep learning architecture is adopted to improve prediction timeliness while reducing model complexity. CNN and BiGRU run independently and work together to capture key information for tool wear prediction. CNN units focus on extracting local features from time series signals, while stacked BiGRU units maintain global awareness of the complete sequence through a gating mechanism, retaining past information in the hidden layers to convey contextual degradation patterns, thus overcoming the limitations of a single network structure in feature representation and modeling capabilities; 2. By introducing a dual attention mechanism to achieve an end-to-end processing flow, the model can directly take the original monitoring signal as input and adaptively focus on key features that are highly correlated with tool wear in the time and space dimensions, thereby avoiding complex manual feature engineering and significantly improving modeling and prediction efficiency. 3. Construct a logarithmic-exponential composite function to characterize the tool life degradation process, improve physical interpretability, and combine it with a particle filter algorithm to achieve the complementary advantages of physical models and data-driven models. While ensuring prediction accuracy, the prediction credibility is enhanced, providing a reliable basis for accurate prediction of tool life and providing key technical support for achieving predictive maintenance and ensuring the stable operation of manufacturing systems. Attached Figure Description

[0021] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. Wherein: Figure 1A flowchart of a milling cutter wear prediction method based on a parallel spatiotemporal feature extraction and physical description according to an embodiment of the present invention is shown; Figure 2 A flowchart of a CNN-BiGRU parallel model according to an embodiment of the present invention is shown; Figure 3 (a) A dual attention module (DAM) according to an embodiment of the present invention is shown; Figure 3 (b) illustrates a Time Dimension Attention (DTA) module according to an embodiment of the present invention; Figure 4 A flowchart of a particle filter according to an embodiment of the present invention is shown; Figure 5 (a) shows the T1 tool wear prediction results according to an embodiment of the present invention; Figure 5 (b) shows the T2 tool wear prediction results according to an embodiment of the present invention; Figure 5 (c) shows the T3 tool wear prediction results according to an embodiment of the present invention; Figure 5 (d) shows the T1 tool prediction evaluation value according to an embodiment of the present invention; Figure 5 (e) shows the T2 tool prediction evaluation value according to an embodiment of the present invention; Figure 5 (f) shows the T3 tool prediction evaluation value according to an embodiment of the present invention; Figure 6 (a) illustrates the prediction results of a logarithmic polynomial model in the later stage of tool RUL according to an embodiment of the present invention; Figure 6 (b) shows the prediction results of the ResNet model in the later stage of tool RUL according to an embodiment of the present invention; Figure 6 (c) This illustrates the prediction results of the present method in the later stage of tool RUL according to an embodiment of the present invention; Figure 6 (d) Shows the statistical prediction indicators of the tool RUL for each model according to embodiments of the present invention; Figure 7 (a) Shows the full RUL prediction results for a T1 tool according to an embodiment of the present invention; Figure 7 (b) Shows the full RUL prediction results for the T2 tool according to an embodiment of the present invention; Figure 7 (c) Shows the full RUL prediction results for the T3 tool according to an embodiment of the present invention; Figure 8 A block diagram of a device for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description according to an embodiment of the present invention is shown; Figure 9 A schematic diagram of an apparatus for predicting milling cutter wear according to an embodiment of the present invention is shown. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] According to an embodiment of the present invention, a method for predicting milling cutter wear using parallel spatiotemporal feature extraction and physical description is proposed. By constructing a parallel deep learning architecture, collaborative modeling of CNN and BiGRU is achieved, which reduces model complexity while improving prediction timeliness. CNN extracts local features of time series signals, while BiGRU characterizes global temporal degradation patterns, compensating for the shortcomings of a single structure. At the same time, a spatiotemporal dual attention mechanism is introduced to achieve end-to-end modeling, directly and adaptively extracting key features from the original signal, avoiding manual feature engineering and improving prediction efficiency. Furthermore, by combining the logarithmic-exponential composite function and particle filtering algorithm, dynamic fusion of the physical model and the data-driven model is achieved, improving prediction accuracy, physical interpretability, and result credibility, providing reliable technical support for tool remaining life prediction and predictive maintenance.

[0024] The principles and spirit of the present invention will be explained in detail below with reference to several representative embodiments.

[0025] Figure 1 This is a schematic flowchart of a method for predicting milling cutter wear using parallel spatiotemporal feature extraction and physical description, according to an embodiment of the present invention. The method includes: Step S01: Collect tool wear monitoring signals, use VMD decomposition combined with PCC coefficients to denoise the signals, and construct a dataset; Step S02: Input the data in the dataset into the CNN-BiGRU parallel structure prediction model to jointly model the local and global features of tool wear, and introduce an improved spatiotemporal dual attention mechanism to enhance the features; Step S03: Select the hyperparameters of the prediction model and construct training units to evaluate performance and obtain the optimal prediction results; Step S04: Construct a phased physical degradation model based on tool wear characteristics, introduce a particle filter algorithm, and use the prediction results of the prediction model as observations to dynamically update the parameters of the physical degradation model, thereby realizing the estimation of remaining service life.

[0026] It should be noted that although the operation of the method of the present invention has been described in a specific order in the above embodiments and figures, this does not require or imply that the operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0027] To provide a clearer explanation of the above-mentioned method for predicting milling cutter wear using parallel spatiotemporal feature extraction and physical description, a specific embodiment is described below. However, it is worth noting that this embodiment is only for better illustrating the present invention and does not constitute an improper limitation of the present invention.

[0028] The following specific example will further illustrate the method for predicting milling cutter wear using parallel spatiotemporal feature extraction and physical description: Step S01: Collect tool wear monitoring signals, use VMD decomposition combined with PCC coefficients to denoise the signals, and construct a dataset.

[0029] The specific steps are as follows: Step S011: Collect vibration and force signals during the milling process by using an accelerometer and force gauge installed on the machine tool spindle or worktable.

[0030] Step S012: Perform variational mode decomposition (VMD) on the vibration and force signals. Set the number of modes. =8, penalty parameter =2000, tolerance error =1e-7. After decomposition, each signal is adaptively decomposed into 8 IMF components. The VMD decomposition formula is as follows:

[0031] in, It is the first One modal component, It is its center frequency. It is the Dirac function. By setting an appropriate number of modes... and penalty parameters It can effectively separate useful components from noise in a signal.

[0032] Step S013: Calculate the relationship between each IMF component obtained from VMD decomposition and the original signal. The Pearson correlation coefficient (PCC) between them. The formula for calculating PCC is:

[0033] in, Describing covariance, Indicates standard deviation; The threshold for PCC is set to 0.3, and IMF components with PCC values ​​greater than 0.3 are retained to obtain a denoised signal.

[0034] Step S014: Denoise the six-channel signals (vibration signals and force signals in the X, Y, and Z directions) sequentially, and combine them with the corresponding tool wear to construct a dataset.

[0035] Step S015: Divide the data into sample sequences of length 5000. All generated samples are randomly divided into training, validation, and test sets in a ratio of 7:2:1. Each sample sequence has a dimension of 6×5000, representing the readings of the 6 sensor channels at 5000 consecutive time points.

[0036] Step S02: Input the data in the dataset into the CNN-BiGRU parallel structure prediction model to jointly model the local and global features of tool wear, and introduce an improved spatiotemporal dual attention mechanism to enhance the features.

[0037] Step S021: Input the training set data into the 1D-CNN unit. This branch consists of six stages (layers 1 to 6). Each stage adopts a unified structure of "convolution-batch normalization-ReLU activation-Dropout-max pooling" to extract and compress local features related to tool wear layer by layer.

[0038] like Figure 2As shown, in the first layer, a convolution operation is performed on the six-channel input signal (the number of channels increases from 6 to 32, the kernel size is 1, and the stride is 1). Then, batch normalization, ReLU activation, and Dropout are performed sequentially. Initial downsampling is achieved through max pooling with a kernel size of 3 and a stride of 3 to obtain initial local feature representations. The second and third layers, while maintaining the number of channels at 32, all use a convolutional structure with a kernel size of 2 and a stride of 1 to model the features. Specifically, the second layer uses max pooling with a kernel size of 2 and a stride of 2; the third layer uses max pooling with a kernel size of 3 and a stride of 3 to extract higher-level local features. The fourth layer uses convolution to achieve channel dimensionality reduction (from 32 to 16), combined with max pooling with a kernel size of 2 and a stride of 2. The fifth and sixth layers, while maintaining 16 channels, both employ one-dimensional convolution with a kernel size of 2 and a stride of 1, and max pooling with a kernel size of 3 and a stride of 3 to further compress the features, ultimately obtaining a compact and highly discriminative 16×14 local feature set. .

[0039] Step S022: The local features (16×14) enter the dual attention module (DAM), and channel attention and spatial attention are calculated in sequence, finally outputting a spatial feature vector with a dimension of 1×64.

[0040] like Figure 3 As shown in (a), the dual attention module is an improvement on the CBAM module. It replaces the channel attention based on global pooling in CBAM with a channel self-attention module, and replaces its single convolutional spatial attention with a multi-layer dilated convolutional spatial self-attention module. At the same time, it adopts a parallel fusion structure of channel and spatial attention to replace the original serial method, thereby enhancing the feature correlation modeling ability and spatial perception range, and further enhancing the feature modeling ability.

[0041] In the channel self-attention submodule, it is first divided into two branches to focus on local features. The output features are obtained by performing 1×1 convolution mapping and spatial unrolling and reshaping. and output features To characterize the global correlation between different channels: , , in and This indicates a convolution operation.

[0042] For output features The self-attention weights between channels are calculated using the Softmax function. and output features We perform weighted fusion to model the importance of each channel: , in This represents the weighted operation in matrix multiplication. Subsequently, Channel attention weights are generated using convolution, LayerNorm, and Sigmoid activation functions: , Finally, the data is weighted channel-by-channel with the original input features. Highlighting feature channels that are more sensitive to tool wear conditions, among which This indicates channel-by-channel multiplication.

[0043] In the spatial self-attention submodule, the input features are first processed... Channel compression is achieved through 1×1 convolution: , Subsequently, dilated convolutions and batch normalization are introduced to expand the receptive field, effectively capturing long-range dependencies between different spatial locations without significantly increasing computational cost. , The dilation rate of the dilated convolution is set to... And stack 3 layers to enhance the ability to extract multi-scale spatial features.

[0044] Subsequently, a spatial attention weight map is generated through convolution, normalization, and sigmoid activation, assigning higher weights to key spatial regions to enhance the expressive power of local key features: , Spatial weighting with the original input features .

[0045] Finally, the weighted results of channel self-attention and spatial self-attention are fused to obtain the output of the DAM module. ,in This indicates a feature fusion operation.

[0046] Step S023: While the training set data from step S021 is input into the 1D-CNN unit for spatial feature extraction, it is simultaneously input into the BiGRU unit for temporal feature extraction.

[0047] like Figure 2As shown, the BiGRU unit employs a two-layer stacked bidirectional GRU structure with an input feature dimension of 6, used to receive multi-channel time series signals. The hidden layer dimension of each GRU unit is set to 64, and the bidirectional structure ensures that the output feature dimension at each time step is 128 (2×64). Dropout layers are introduced between layers to alleviate overfitting, ultimately outputting time series features. It is 1×64.

[0048] Step S024: Input the temporal features (1×64) into the Time Dimension Attention Module (DTA) to perform dimension recalibration and time step selection, so as to enhance the model's ability to capture key temporal features.

[0049] like Figure 3 As shown in (b), this module is based on the self-attention model and improves upon it by replacing the feature weighting method of a single mapping with a dimensional attention mechanism. It introduces a feature compression-recalibration process that combines global average pooling with a linear bottleneck structure, and introduces a temporal self-attention mechanism to replace the single temporal weighting structure. It uses nonlinear mapping and learnable weight vectors to dynamically assign weights to each time step.

[0050] First, the temporal characteristics of the BiGRU output. ( Time step, Global average pooling (GAP) is performed on the time step dimension to compress the time step dimension and obtain global tool wear state information. , in, It is along Time step compression is used to extract dimensional GAP features.

[0051] Then, the importance of each dimension at each time step is learned using two fully connected layers in the Linear Bottleneck (LB) layer, and the sigmoid function is used to generate the dimension weights: , and The weights of the first and second layers of the linear bottleneck layer, ReLU activation function Multiply the dimensionality weights element-wise with the original features to obtain the dimensionality-enhanced features: , Features after dimensional enhancement Nonlinear mapping using the tanh activation function: , Through dynamic vectors Calculate the importance of each time step and obtain the temporal attention weights by Softmax normalization: , Applying temporal attention weights to the feature sequence emphasizes key time steps: , Step S025: Concatenate the 1×64 spatial feature vector output by DAM and the 1×64 temporal feature vector output by DTA along the Batchsize direction to obtain a 1×128 fused feature representation.

[0052] Step S03: Select the hyperparameters of the prediction model and construct training units to evaluate performance and obtain the optimal prediction results.

[0053] Step S031: Construct a model training unit based on the PyTorch 2.5.1 deep learning framework, and configure CUDA 11.8 and cuDNN 8.9.0 to achieve GPU-accelerated computing; set training hyperparameters according to the parallel model structure and data scale, adopt a mini-batch training strategy, set the batch size to 32, and initialize the learning rate to 0.002. Step S032: Iteratively train the tool wear prediction model using the training set, evaluate the model's prediction performance on the validation set, and optimize and update the parallel model parameters based on the evaluation results.

[0054] Specifically, the parallel model parameter optimization adopts the Adam optimization algorithm and introduces L2 regularization constraints to suppress overfitting. At the same time, the learning rate is dynamically adjusted by combining an exponential decay learning rate scheduling strategy to improve training stability and convergence efficiency.

[0055] Step S033: Based on the MSE loss value, select the model parameters with excellent performance to obtain the optimal parallel model prediction results.

[0056] Step S04: Construct a phased physical degradation model based on tool wear characteristics, introduce a particle filter algorithm, and use the prediction results of the prediction model as observations to dynamically update the parameters of the physical degradation model, thereby realizing the estimation of remaining service life.

[0057] Step S041: Based on the nonlinear degradation characteristic of tool wear exhibiting "slow initial wear and accelerated later wear" with machining time, a physical degradation model of tool wear based on logarithmic-exponential form is constructed to describe the overall evolution trend of tool wear. Its expression is: , in, express Physically predicted tool wear value at any given time; , , , These are model parameters, used to characterize the degradation characteristics at different wear stages.

[0058] Step S042: To describe the dynamic evolution of tool wear state, the equation is... Regarding time Differentiating the equation, we obtain the ordinary differential degeneracy equation for tool wear: , in, This indicates the rate of degradation of tool wear over time.

[0059] Step S043: Construct the state transition equation and observation equation for tool wear.

[0060] Specifically, the formula Discretize and construct the state equation for tool wear: , in, , This is state transition noise.

[0061] To obtain a numerical solution to the state equations, the Runge-Kutta method is used to obtain the tool wear values: , in, Let the time step be [value], and: , Simultaneously, using the tool wear prediction value output by the parallel data-driven model as the observation, an observation equation is constructed: , in, For the first The parallel data at each time step drives the model output of the predicted wear value. To observe noise.

[0062] Step S044: The particle filter algorithm is used to dynamically update the physical degradation model parameters as observations by using the prediction results of the parallel model, thereby realizing the estimation of remaining service life.

[0063] The particle filter algorithm flow is as follows: Figure 4 As shown, set the number of particles. =1000, initialize the particle set To characterize tool wear Probability distribution; in the prediction phase, the state of each particle is propagated using the numerical solution of the state equation obtained by Runge-Kutta; in the update phase, ... As an observation, the first is calculated according to the observation equation. Particle weight : , in, The variance representing the observation noise. Indicates the first Particles in The physical predicted wear value corresponding to the given time.

[0064] To ensure the comparability of weights, the weights of all particles are normalized to obtain normalized weights. : , Subsequently, the particles are resampled according to their weights to avoid particle degradation, resulting in an updated estimate of tool wear. , By repeating the above process and recursively updating through particle filtering, a sequence of physical prediction values ​​for tool wear is continuously obtained, thus constructing the wear evolution trajectory throughout the tool's entire life cycle.

[0065] Furthermore, let the maximum permissible wear threshold of the tool be... When the predicted wear evolution trajectory reaches this threshold, the corresponding time is... Then, the remaining tool life at the current moment is: .

[0066] In addition, in this embodiment, to verify the milling cutter wear prediction method for parallel spatiotemporal feature extraction and physical description proposed in this invention, the following experimental examples were also conducted: This method was implemented on a self-built dataset for γ-TiAl milling. The experiment used a CAMUMC850 ultrasonic-assisted three-axis machining center, but is not limited to this; a triaxial accelerometer was mounted on the fixture for synchronous data acquisition. Vibration acceleration signals in three directions ( , , The data is monitored by an IGTech A26F100T01C system; a Kistler 5167A triaxial force gauge is installed on the CNC milling machine table for synchronous data acquisition. Cutting force signals in three directions ( , , All signals were sampled at a frequency of 20kHz, and the original signal... The data is collected online in real time using a data acquisition card and stored in a computer.

[0067] The cutting tool used was an EH1301L four-flute carbide end mill with a cutting diameter of 10mm, a radius of 0.5mm, and a cutting length of 25mm. Vertical milling was performed, but not limited to this type. The workpiece used in the experiment was a γ-TiAl alloy with a diameter of 74.8mm and a length of 199mm, but not limited to this type. Specific milling parameters are detailed in Table 1. During the experiment, the workpiece was installed in a worktable fixture, and wet milling was performed using cutting fluid. Each machining stroke was 100mm long, and a total of 138 machining strokes were performed.

[0068] Table 1

[0069] After each machining stroke, the tool is removed from the machine tool, and the width of the wear band on the tool's flank is measured using a KEYENCE VHX-6000 ultra-depth-of-field microscope according to GB / T 16460-2016 standard. (Value). Three measurement points are taken in the uniform wear area of ​​each cutting tooth, and the maximum value is used as the label of the tool wear amount for that stroke.

[0070] The collected multi-source sensor signals were paired with corresponding tool wear values ​​to construct a tool wear monitoring dataset, which was then input into the proposed model to verify the effectiveness of the method. The performance of the proposed method in tool wear modeling and remaining service life prediction was evaluated by comparing the model predictions with actual values. The experimental results are as follows: Figure 5 (a), (b), and (c) demonstrate the parallel data-driven model of this method in... , , The wear prediction effect on the cutting tool shows that the predicted values ​​of this model have a high degree of consistency with the actual wear values, the prediction error is small, and the prediction curve basically remains within the range of actual wear. Within the error deviation band, it indicates that the model can accurately characterize the evolution trend of tool wear. Furthermore, as... Figure 5 As shown in (d), (e), and (f), this paper also selects five representative benchmark models for comparison: SVR based on feature engineering, CNN with automatic feature extraction capability, BiGRU with bidirectional temporal modeling characteristics, CNN-BiGRU that integrates convolutional and recurrent features, and ResNet with deep residual and multi-scale representation capabilities. Quantitative evaluation results show that the proposed method achieves the best performance on multiple evaluation metrics, with the lowest MAE at 3.98, the lowest RMSE at 5.44, and the highest R² determination at 0.982, fully verifying the accuracy and superiority of this model in tool wear prediction tasks.

[0071] This experiment also compared the computational efficiency of each model, as shown in Table 2. The total time of the parallel data-driven model of our method is 325.2 ms, which is moderately better than the comparison models. Although its inference time is not absolutely superior, for example, the inference time of SVR is only 81.6 ms, our model completely avoids the featureization step. In intelligent manufacturing environments, tool wear prediction often requires online processing, and the total time (including preprocessing and inference) is a key indicator. The total time of our model is shorter than CNN, BiGRU, CNN-BiGRU, and ResNet, and only slightly slower than SVR. However, SVR relies on featureization processing and has poor prediction performance, which may introduce latency and instability in dynamic industrial environments.

[0072] Table 2

[0073] Figure 6 (a), (b), and (c) show the tool RUL prediction results of multiple models. Through comparative experiments, the values ​​of three indicators of the fusion model proposed in this invention and other models on the test set were recorded and analyzed. The results show that the method of this invention is significantly lower than the comparative model in RMSE and MAE, and is closest to 1 in R², thus verifying the superiority of this invention in terms of prediction accuracy and stability from multiple perspectives. Figure 7 (a), (b), and (c) represent the invention in the three knives ( , , The remaining tool life prediction results on the model show that, with the accumulation of observation data, the posterior estimation accuracy of the tool degradation model parameters gradually improves, the prediction bias band continuously narrows, and the later prediction results exhibit convergence characteristics. Therefore, the prediction results of this model meet the actual needs of industrial production and can effectively support preventive maintenance decisions.

[0074] Based on the same inventive concept, this invention also proposes a device for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description. The implementation of this device can be found in the implementation of the method described above; repeated details will not be repeated. Figure 8 As shown, the device 100 includes: Signal acquisition and denoising module 101: Acquires tool wear monitoring signals, uses VMD decomposition combined with PCC coefficients to denoise the signals, and constructs a dataset; Parallel feature extraction module 102: Inputs the data in the dataset into the CNN-BiGRU parallel structure prediction model, jointly models the local and global features of tool wear, and introduces an improved spatiotemporal dual attention mechanism to enhance the features; Parallel model training module 103: Selects prediction model hyperparameters and constructs training units for performance evaluation to obtain the optimal prediction results; The digital-analog fusion module 104 constructs a phased physical degradation model based on tool wear characteristics, introduces a particle filtering algorithm, and uses the prediction results of the prediction model as observations to dynamically update the parameters of the physical degradation model, thereby realizing the estimation of remaining service life.

[0075] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0076] like Figure 9 The device includes a central processing unit (CPU), which can perform various appropriate actions and processes based on computer program instructions stored in read-only memory (ROM) or loaded from storage units into random access memory (RAM). RAM can also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0077] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0078] The processing unit executes the various methods and processes described above, such as method steps S01 to S04. For example, in some embodiments, method steps S01 to S04 may be implemented as a computer software program tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of method steps S01 to S04 described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute method steps S01 to S04 by any other suitable means (e.g., by means of firmware).

[0079] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload programmable logic devices (CPLDs), and so on.

[0080] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0081] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0082] Furthermore, although the operations are described in a specific order, this should be understood as requiring that such operations be performed in the specific order shown or in sequential order, or requiring that all illustrated operations be performed to achieve the desired result. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the invention. Certain features described in the context of individual embodiments may also be implemented in combination in a single implementation. Conversely, various features described in the context of a single implementation may also be implemented individually or in any suitable sub-combination in multiple implementations.

[0083] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for predicting milling cutter wear using parallel spatiotemporal feature extraction and physical description, characterized in that, The method includes: Step S01: Collect tool wear monitoring signals, use VMD decomposition combined with PCC coefficients to denoise the signals, and construct a dataset; Step S02: Input the data in the dataset into the CNN-BiGRU parallel structure prediction model to jointly model the local and global features of tool wear, and introduce an improved spatiotemporal dual attention mechanism to enhance the features; Step S03: Select the hyperparameters of the prediction model and construct training units to evaluate performance and obtain the optimal prediction results; Step S04: Construct a phased physical degradation model based on tool wear characteristics, introduce a particle filter algorithm, and use the prediction results of the prediction model as observations to dynamically update the parameters of the physical degradation model, thereby realizing the estimation of remaining service life.

2. The method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description according to claim 1, characterized in that, The specific steps of step S01 are as follows: Step S011: Collect vibration and force signals during the milling process by using an accelerometer and force gauge installed on the machine tool spindle or worktable; Step S012: Perform VMD decomposition on the vibration and force signals; Step S013: Calculate the relationship between each IMF component obtained from VMD decomposition and the original signal. The Pearson correlation coefficient (PCC) between them; Step S014: Denoise the vibration and force signals in the X, Y, and Z directions in sequence, and combine them with the corresponding tool wear to construct a dataset; Step S015: Cut the data into sample sequences and divide them into training set, validation set and test set.

3. The method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description according to claim 1, characterized in that, The specific steps of step S02 are as follows: Step S021: Input the training set data into the 1D-CNN unit to extract and compress the local features related to tool wear layer by layer; Step S022: Input the local features into the DAM module, perform channel attention and spatial attention calculations in sequence, and output the spatial feature vector; Step S023: Input the training set data into the 1D-CNN unit for spatial feature extraction, and simultaneously input it into the BiGRU unit for temporal feature extraction; Step S024: Input the temporal features into the DTA module, perform dimension recalibration and time step selection, and output the temporal feature vector; Step S025: Concatenate the spatial feature vector and the temporal feature vector along the Batchsize direction to obtain the fused features.

4. The method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description according to claim 3, characterized in that, The DAM module described in step S022 is an improvement on the CBAM module, and the steps are as follows: In the channel self-attention submodule, the local features are first divided into two branches to perform convolutional mapping and spatial expansion and rearrangement to obtain the first output feature and the second output feature. The self-attention weights between channels are calculated by the Softmax function for the first output feature and then weighted and fused with the second output feature to obtain the fused feature. The fused features are processed by convolution, LayerNorm and Sigmoid activation functions to generate channel attention weights, which are then weighted channel-wise with the original input features. In the spatial self-attention submodule, the input features are compressed through convolution and then multi-layer dilated convolution and normalization are introduced; after generating a spatial attention weight map through convolution, normalization and sigmoid activation, it is spatially weighted with the original input features; The weighted results of channel self-attention and spatial self-attention are fused to obtain the output of the DAM module.

5. The method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description according to claim 3, characterized in that, The DTA module described is an improvement based on the self-attention model, and the specific steps are as follows: Global average pooling and time step compression are performed on the temporal features output by BiGRU to obtain global tool wear state information; The importance of each dimension at each time step is learned by using two fully connected layers in the linear bottleneck layer, and the sigmoid function is used to generate dimension weights. The dimension weights are multiplied element-wise with the original features to obtain dimension-enhanced features, and the dimension-enhanced features are then non-linearly mapped using the tanh activation function. The importance of each time step is calculated using dynamic vectors, and the temporal attention weights are obtained by Softmax normalization. These temporal attention weights are then applied to the feature sequence to obtain the output of the DTA module.

6. The method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description according to claim 1, characterized in that, The specific steps of step S04 are as follows: Step S041: Based on the nonlinear degradation characteristics of tool wear, construct a physical degradation model of tool wear based on the logarithmic-exponential form: ， in, express Physically predicted tool wear value at any given time; , , , These are model parameters, used to characterize the degradation characteristics at different wear stages; Step S042: Physical degradation model of tool wear with respect to time Differentiating the equation, we obtain the ordinary differential degeneracy equation for tool wear: ， in, This indicates the rate of degradation of tool wear over time. Step S043: Construct the state transition equation and observation equation for tool wear, that is, equation... Discretize and construct the state equation for tool wear: ， in, , This is state transition noise; The Runge-Kutta method is used to obtain tool wear values: ， in, Let the time step be [value], and: ， Using the tool wear prediction values ​​output by the parallel data-driven model as observations, an observation equation is constructed: ， in, For the first The parallel data at each time step drives the model output of the predicted wear value. To observe noise; Step S044: The particle filter algorithm is used to dynamically update the physical degradation model parameters as observations by using the prediction results of the parallel model, thereby realizing the estimation of remaining service life.

7. The method for predicting milling cutter wear by parallel spatiotemporal feature extraction and physical description according to claim 6, characterized in that, The specific process of the particle filter algorithm is as follows: In the prediction phase, the numerical solution of the state equation obtained by Runge-Kutta is used to propagate the state of each particle; in the update phase, ... As an observation, the first is calculated according to the observation equation. Particle weight : ， in, The variance representing the observation noise. Indicates the first Particles in The physical predicted wear value corresponding to the given time; The weights of all particles are normalized to obtain normalized weights. : ， Based on the weighted particle resampling, the updated tool wear estimate is obtained: ， By repeating the above process and recursively updating through particle filtering, a sequence of physical prediction values ​​for tool wear is continuously obtained, thus constructing the wear evolution trajectory throughout the tool's entire life cycle.

8. A device for predicting milling cutter wear through parallel spatiotemporal feature extraction and physical description, characterized in that, The device implements the method as described in any one of claims 1 to 7, comprising: Signal acquisition and denoising module: Acquires tool wear monitoring signals, uses VMD decomposition combined with PCC coefficients to denoise the signals, and constructs a dataset; Parallel feature extraction module: Input the data in the dataset into the CNN-BiGRU parallel structure prediction model, jointly model the local and global features of tool wear, and introduce an improved spatiotemporal dual attention mechanism to enhance the features; Parallel model training module: Selects prediction model hyperparameters and constructs training units for performance evaluation to obtain the optimal prediction results; The digital-analog fusion module constructs a phased physical degradation model based on tool wear characteristics, introduces a particle filtering algorithm, and uses the prediction results of the prediction model as observations to dynamically update the parameters of the physical degradation model, thereby realizing the estimation of remaining service life.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images