A multi-working-condition adaptive residual service life prediction method for a tool of a numerical control machine tool

Abstract

This invention discloses a multi-condition adaptive remaining service life prediction method for CNC machine tool tools. The method first performs degradation correlation screening and time-series sample construction on multi-sensor monitoring signals, and then divides the dataset. Subsequently, a degradation trend-aware feature extraction network is constructed to achieve unified feature encoding of multi-source working condition data. Based on this, a multi-branch migration prediction architecture is designed, achieving cross-working condition modeling through shared feature encoding and branch decoupling structures. Furthermore, an adversarial feature distribution alignment mechanism is introduced to reduce data distribution differences between different working conditions. Simultaneously, a multi-scale target degradation consistency constraint mechanism is proposed, effectively preserving the unique degradation information of the target working condition by establishing a contrast consistency relationship between the input space and the feature space. In addition, an adaptive collaborative optimization mechanism for source domain contributions is constructed to dynamically weight the contributions of different source working conditions in the prediction task. Finally, high-precision prediction of the remaining tool service life is achieved.

Description

Technical Field

[0001] This invention relates to the field of predicting the remaining service life of CNC machine tool tools, and in particular to a multi-condition adaptive method for predicting the remaining service life of CNC machine tool tools. Background Technology

[0002] In CNC machine tool machining, accurately predicting the remaining service life of cutting tools is crucial. This not only allows for the prediction of tool failure points, effectively preventing decreased machining accuracy and batch workpiece scrap, ensuring product quality and process stability, but also significantly reduces unplanned downtime, equipment damage, and safety accidents caused by sudden tool breakage, thereby directly lowering production and operating costs and maintenance risks. Simultaneously, this prediction supports the optimization of scientific tool replacement strategies, helps control resource waste caused by excessive tool stockpiling, and extends the effective service life of tools, forming the cornerstone of achieving efficient, safe, and low-cost intelligent manufacturing.

[0003] Traditional tool condition monitoring methods mainly rely on manual experience or periodic shutdowns for inspection, which suffers from high subjectivity and poor real-time performance, making it difficult to meet the demands of modern intelligent manufacturing for online monitoring and prediction. In recent years, deep learning methods have been widely applied in the field of remaining service life prediction. By automatically extracting complex degradation features, they have demonstrated superior performance in terms of modeling accuracy and generalization ability compared to traditional methods. For example, convolutional neural networks can effectively extract local degradation features, recurrent neural networks and their variants can model temporal dependencies, and attention-based models further enhance the ability to model long sequences. However, these methods are typically based on the assumption that the distribution of training and testing data is consistent and rely on sufficient labeled data for training. In real-world industrial scenarios, CNC machine tool processing conditions are complex and variable. Different cutting parameters and workpiece materials can lead to significant changes in the distribution of monitoring data, making it difficult to directly apply models trained under a single operating condition to new environments. Furthermore, acquiring labeled data covering the entire lifespan is costly, further limiting the practical application of these models.

[0004] To address the aforementioned issues, researchers have begun exploring domain-adaptive-based methods for predicting remaining service life to achieve knowledge transfer between different operating conditions. However, most existing methods are designed for single-source domain scenarios and still have significant shortcomings under multi-source operating conditions: Firstly, the correlation between data from different source operating conditions and the target operating condition varies. Existing methods typically employ equal-weight or fixed-weight strategies for fusion, making it difficult to fully utilize highly correlated source domain information and even introducing negative transfer effects from low-correlation data. Secondly, when aligning feature distributions, the preservation of degradation information specific to the target operating condition is often neglected, resulting in the weakening of key degradation features. Furthermore, existing feature extraction methods lack specialized modeling capabilities for the multi-scale temporal characteristics of tool degradation processes, making it difficult to fully characterize complex degradation patterns. Therefore, how to achieve effective knowledge transfer under multi-source operating conditions while simultaneously considering the differences in source domain contributions and the preservation of target degradation information has become an urgent technical problem to be solved.

[0005] A search revealed that Chinese invention patent CN111832624A discloses a method for predicting the remaining life of cutting tools based on adversarial transfer learning. The method includes: collecting data from the cutting processes of different types of cutting tools to determine historical and new types of cutting tool samples; using the data from the historical cutting tool samples to construct a feature extraction model and a nonlinear regression model for the historical cutting tool samples; performing adversarial domain adaptation on the data from the historical cutting tool samples and the data from the new cutting tool samples to construct a feature extraction model for the new cutting tool samples; using the feature extraction model for the new cutting tool samples constructed in step S13 to analyze and extract time-series signals from the new cutting tool samples, and transferring the nonlinear regression model from the historical cutting tool samples to the new cutting tool samples to achieve the prediction of the remaining life of the new cutting tool samples. This invention's method for predicting the remaining life of cutting tools based on adversarial transfer learning enables rapid construction of prediction models under new process conditions, reduces the requirement for a large number of new cutting tool samples, improves the accuracy of cutting tool life prediction, and has strong applicability.

[0006] 1. From an overall architectural perspective, this application addresses the problem of tool remaining life prediction under multi-source operating conditions. It constructs a multi-branch migration prediction architecture, where each source condition and target condition constitute an independent migration branch. A strategy of shared representations and hard parameter isolation is adopted. All branches share a degradation trend-aware feature extraction network, but the domain discrimination subnetwork and life regression prediction subnetwork of each branch are independent. This design enables the model to simultaneously handle knowledge transfer from multiple source conditions while preserving the specific information of each source domain. In contrast, the patent "A Tool Remaining Life Prediction Method Based on Adversarial Transfer Learning" only addresses the migration scenario from historical tool types to new tool types, employing a traditional adversarial domain adaptation architecture and failing to address the collaborative modeling problem of multi-source operating conditions.

[0007] 2. Regarding the core innovation mechanism, this application proposes several key technologies not found in the patent "A Tool Remaining Life Prediction Method Based on Adversarial Transfer Learning". First, this application designs a degradation trend-aware feature extraction network, which is composed of a cascaded local degradation pattern perception unit, a long sequence correlation modeling unit, and a feature compression mapping unit. This network can simultaneously capture short-term abrupt changes and long-term evolution trends in the tool wear process, specifically modeling the multi-scale temporal characteristics of tool degradation. In contrast, the patent "A Tool Remaining Life Prediction Method Based on Adversarial Transfer Learning" does not employ this structured design for the feature extraction network and lacks specialized modeling capabilities for the degradation process. Second, this application introduces a multi-scale target degradation consistency constraint mechanism. Through contrastive learning, it constructs consistency constraint losses between the target domain input and its deep features at multiple scales. During adversarial training, it explicitly preserves the degradation information specific to the target working condition, avoiding the loss of key degradation features due to distribution alignment. In contrast, the patent "A Tool Remaining Life Prediction Method Based on Adversarial Transfer Learning" only achieves distribution alignment through adversarial training and does not consider the preservation of target domain degradation information. Furthermore, this application constructs an adaptive collaborative optimization mechanism for source domain contributions, introduces learnable weight parameters for each source condition, and dynamically optimizes the weights using an improved fox search strategy. The prediction error on the validation set is used as the fitness function to achieve adaptive adjustment of the contribution of different source conditions, effectively avoiding the negative transfer problem caused by the equal weighting process in traditional methods. In contrast, the patent "A tool remaining life prediction method based on adversarial transfer learning" does not have the ability to perform differentiated modeling of contributions from multiple source domains.

[0008] 3. Regarding the loss function and optimization objective, this application jointly optimizes three types of losses: multi-scale target degradation consistency constraint loss, adversarial loss, and root mean square prediction loss. Through multi-objective collaborative optimization, it achieves the unity of distribution alignment, target degradation information preservation, and prediction accuracy improvement during the transfer process. In contrast, the patent "A tool remaining life prediction method based on adversarial transfer learning" only includes regression loss and adversarial loss, and lacks constraints on the preservation of target domain degradation information and the adaptive contribution of multi-source domains.

[0009] A search revealed that Chinese invention patent CN114398825A provides a method for predicting the remaining life of cutting tools under complex and variable working conditions. The specific steps are as follows: S1, collecting NC variable data from the CNC system and marking the remaining life of the cutting tool; S2, using compressed sensing to denoise the collected NC variable data; S3, dividing the NC variable data processed in S2 into domains based on the differences between working conditions, obtaining the source domain feature space and the target domain feature space of the cutting tool; S4, extracting sensitive features of the cutting tool life from the source domain feature space and the target domain feature space of the cutting tool using a one-dimensional residual block stacking method; S5, constructing a tool remaining life prediction transfer learning framework based on the sensitive features of the cutting tool life extracted in S4 using a domain adversarial method and a data distribution adaptive method. This invention not only improves the accuracy of tool remaining life prediction under a single working condition but also exhibits good generalization ability under complex and variable working conditions.

[0010] 1. From an overall architectural perspective, this application addresses the problem of tool remaining life prediction under multi-source operating conditions by constructing a multi-branch migration prediction architecture. Each source condition and target condition constitute an independent migration branch, employing a strategy of shared representations and hard parameter isolation. Each branch shares a degradation trend-aware feature extraction network, while also possessing its own independent domain discrimination sub-network and life regression prediction sub-network, thereby achieving collaborative modeling and differentiated utilization of multiple source conditions. In contrast, the patent "A Method for Predicting the Remaining Life of Cutting Tools Under Complex and Variable Operating Conditions," although also involving tool life prediction under complex and variable operating conditions, uses a single-source domain to target domain transfer learning framework. It extracts sensitive features by stacking identity residual blocks and decay residual blocks, and constructs a prediction model by combining domain adversarial and data distribution adaptive methods. It does not set up a multi-branch structure to handle multiple source domains, and therefore cannot decouple and differentiate the contribution modeling of different source conditions.

[0011] 2. Regarding the design of the feature extraction network, this application proposes a degradation trend-aware feature extraction network, which consists of three cascaded units: a local degradation pattern perception unit, a long sequence correlation modeling unit, and a feature compression mapping unit. The local degradation pattern perception unit uses a one-dimensional convolutional neural network to capture short-term wear mutation features; the long sequence correlation modeling unit uses a Transformer encoder to model degradation dependencies across time scales; and the feature compression mapping unit forms a compact feature representation through a fully connected network. These three units work together to jointly model local mutations and long-term evolution trends during tool degradation. The patent "Method for Predicting the Remaining Life of Cutting Tools Under Complex and Variable Working Conditions" uses a one-dimensional residual block stacking method to extract sensitive features, specifically including two structures: identity residual blocks and decaying residual blocks. The residual network is constructed through a pre-activation structure, batch regularization, random deactivation, and a direct connection mechanism. Although the residual network can extract deep features and alleviate the gradient vanishing problem, it is essentially still a general temporal feature extraction architecture. It has not been specifically designed for the multi-scale temporal characteristics of the tool degradation process and lacks the ability to jointly model local degradation mutations and long-range dependencies.

[0012] 3. Regarding transfer learning and distribution alignment mechanisms, this application adopts a multi-branch transfer prediction architecture and introduces adversarial training in each branch to achieve feature distribution alignment between the source and target domains. Simultaneously, through a multi-scale target degradation consistency constraint mechanism, it constructs a consistency relationship between the target domain input and its deep features at multiple scales, explicitly preserving the degradation information specific to the target operating condition during distribution alignment. The patent "Method for Predicting the Remaining Life of Cutting Tools Under Complex and Variable Operating Conditions" employs a strategy combining domain adversarial methods and data distribution adaptive methods to construct a transfer learning framework. On one hand, it uses a domain discriminant function and a life prediction function to form an adversarial learning mechanism; on the other hand, it uses the maximum mean difference of multiple kernels as a quantitative indicator, minimizing the difference in feature space distribution between the source and target domains through a linear combination of polynomial and Gaussian kernels. However, compared to the patent "Method for Predicting the Remaining Life of Cutting Tools Under Complex and Variable Operating Conditions," which lacks a dedicated target domain degradation information preservation mechanism, its distribution alignment process may lead to the weakening or loss of the target domain's unique degradation features. Furthermore, in contrast, the patent "Method for Predicting the Remaining Life of Cutting Tools Under Complex and Variable Working Conditions" treats the identity residual block, the decay residual block, and the fully connected layer as feature extraction functions, and constructs a domain discrimination function and a prediction function on this basis. Its adversarial mechanism is essentially a traditional adversarial domain adaptation method, and does not involve degradation information preservation strategies such as multi-scale consistency constraints.

[0013] 4. Regarding source domain contribution modeling, this application constructs an adaptive collaborative optimization mechanism for source domain contributions in multi-source operating scenarios. Learnable weight parameters are introduced for each source operating condition, and an improved fox search strategy is used to dynamically optimize the weights. The prediction error on the validation set is used as the fitness function to achieve adaptive adjustment of the contribution of different source operating conditions, effectively avoiding negative migration problems. The patent "Method for Predicting the Remaining Life of Cutting Tools Under Complex and Variable Operating Conditions" only addresses migration scenarios from a single source domain to the target domain and does not address the contribution modeling problem for multi-source operating conditions. Its technical solution does not set up any mechanism to distinguish or weight the contributions of different source domains.

[0014] 5. Regarding the loss function and optimization objective, this application jointly optimizes three types of losses: multi-scale objective degradation consistency constraint loss, adversarial loss, and root mean square prediction loss. Through multi-objective collaborative optimization, it achieves a unified approach to distribution alignment, preservation of degradation information, and improvement of prediction accuracy. In contrast, the optimization objective of the patent "Method for Predicting the Remaining Life of Cutting Tools Under Complex and Varied Working Conditions" consists of two parts: tool life prediction error and neighborhood discrimination error. The neighborhood discrimination error is achieved through adversarial learning, and the life prediction error is calculated through a prediction function. However, its loss function does not include a mechanism similar to multi-scale objective degradation consistency constraint, nor does it involve a constraint term for adaptive source domain contribution.

[0015] A search revealed a Chinese invention patent with publication number CN117910351A, which discloses a method for predicting the remaining life of a stochastically degraded tool using a digital-analog linkage approach. The method includes: S1: acquiring historical multi-source sensor data; S2: constructing a life prediction model; S3: constructing an optimization objective function for the remaining life prediction accuracy; S4: determining whether the prediction accuracy requirements are met; S5: outputting online composite health indicators; S6: adaptively updating the degradation model parameters using the EM method; S7: performing online prediction of the remaining life based on the network parameters and degradation model parameters; and S8: verifying the accuracy of the prediction results. The method uses a Transformer network to map multi-source sensor data to composite tool health indicators, and uses gated convolutional units to incorporate local features into an attention mechanism, mitigating information attenuation in deep neural network structures and improving the model's sensitivity to local information. By employing a digital-analog linkage method, the failure state, the failure threshold in the stochastic degradation model, and the composite health indicators at the time of failure are unified, thus unifying the health state and the composite health indicators.

[0016] 1. From an overall architectural perspective, this application addresses the problem of tool remaining life prediction under multi-source operating conditions. It constructs a multi-branch migration prediction architecture, where each source condition and target condition constitute an independent migration branch. A strategy of shared representation and hard parameter isolation is adopted, with each branch sharing a degradation trend perception feature extraction network and possessing its own independent domain discrimination sub-network and life regression prediction sub-network. This enables collaborative modeling and differentiated utilization of multiple source conditions. In contrast, the comparative application employs a "digital-model linkage" approach, constructing a life prediction model that includes a Transformer network and a stochastic degradation model. By combining the composite health index output by the neural network with the Wiener process, it uses an optimization objective function to force the network to output an index that conforms to the stochastic degradation law. Its overall architecture is a single model structure, lacking a multi-branch structure to handle multi-source conditions and failing to differentiate the contributions of different source domains.

[0017] 2. Regarding the design of the feature extraction network, this application proposes a degradation trend-aware feature extraction network, which consists of three cascaded units: a local degradation pattern perception unit, a long sequence correlation modeling unit, and a feature compression mapping unit. The local degradation pattern perception unit uses a one-dimensional convolutional neural network to capture short-term wear mutation features; the long sequence correlation modeling unit uses a Transformer encoder to model degradation dependencies across time scales; and the feature compression mapping unit forms a compact feature representation through a fully connected network. These three units work together to jointly model local mutations and long-term evolution trends during tool degradation. In contrast, the patent "A Method for Predicting the Remaining Life of a Randomly Degraded Tool Based on Digital-Analog Linkage" also uses a Transformer network as the core of feature extraction, but its structure is a combination of a local feature extraction layer, an encoding layer, and a regression layer. The local feature extraction layer introduces gated convolutional units, using reset and update gates to incorporate local features into the attention mechanism, thus alleviating the information decay problem in deep neural networks. However, the Transformer network in the patent "A Method for Predicting the Remaining Life of a Stochastically Degraded Tool with Digital-Analog Linkage" does not perform dedicated multi-scale temporal characteristic modeling for the tool degradation process. Its local feature extraction mainly relies on gating mechanisms to filter the information after convolution operations. In contrast, this application uses three cascaded units to perform structured design for short-term mutations, long-range dependencies, and feature compression, respectively. There are significant differences between the two in terms of the targeting and structuring of feature extraction.

[0018] 3. Regarding transfer learning strategies and domain adaptation mechanisms, this application constructs a multi-branch transfer prediction architecture for multi-source working conditions. Adversarial training is introduced into each branch to align the feature distributions of the source and target domains. Simultaneously, a multi-scale target degradation consistency constraint mechanism is used to construct a consistency relationship between the target domain input and its deep features at multiple scales, explicitly preserving the degradation information specific to the target working condition during the distribution alignment process. In contrast, the patent "A Digital-Model Linked Method for Predicting the Remaining Life of a Randomly Degraded Tool" does not address cross-working-condition transfer learning. Its technical solution lacks a distinction between the source and target domains, and also lacks an adversarial training mechanism or multi-scale consistency constraints. The key technical focus of the patent "A Digital-Model Linked Method for Predicting the Remaining Life of a Randomly Degraded Tool" lies in achieving "digital-model linkage" between the neural network and the random degradation model by optimizing the objective function. That is, it uses the tool remaining life prediction error as a loss function to force the Transformer network to output a composite health index that conforms to the linear Wiener process law, thereby solving the problem of difficulty in determining the failure threshold. Its core lies in the collaborative optimization of the deep learning model and the physical degradation model, rather than solving the cross-working-condition transfer problem.

[0019] 4. Regarding degradation modeling and parameter updating, this application uses a stochastic degradation model to model the tool degradation process and dynamically weights the contributions of different source domains through a source domain contribution adaptive collaborative optimization mechanism. However, it does not involve online adaptive updating of the degradation model parameters. In contrast, the patent "A Digital-Analog Linked Stochastic Degradation Tool Remaining Life Prediction Method" uses a linear Wiener process to model the degradation process of composite health indicators and uses a combination of expectation-maximization algorithm and Kalman filtering to adaptively update parameters such as the drift coefficient in the degradation model online, achieving bidirectional linkage between the neural network output and the degradation model parameters. Specifically, the patent "A Digital-Analog Linked Stochastic Degradation Tool Remaining Life Prediction Method" sets the drift coefficient as a random walk model, utilizes a state-space model framework, and recursively updates the degradation model parameters after obtaining new observation data at each monitoring time through E-step Kalman filtering and M-step parameter optimization, thereby enabling adaptive prediction of remaining life. Although this application achieves knowledge transfer from multiple source conditions during training through adversarial training and consistency constraints, it does not set up a similar online parameter adaptive update mechanism.

[0020] 5. Regarding the loss function and optimization objective, this application jointly optimizes three types of losses: multi-scale target degradation consistency constraint loss, adversarial loss, and root mean square prediction loss. Through multi-objective collaborative optimization, it achieves a unified approach to distribution alignment, preservation of degradation information, and improvement of prediction accuracy. In contrast, the optimization objective of the patent "A Digital-Analog Linked Method for Predicting the Remaining Life of a Randomly Degraded Tool" is a target function composed of the error value of the remaining life prediction accuracy. Its core is to achieve joint optimization of Transformer network parameters and failure thresholds by minimizing the mean square error between the predicted life and the actual life. The loss function design of the patent "A Digital-Analog Linked Method for Predicting the Remaining Life of a Randomly Degraded Tool" embodies the core idea of ​​"digital-analog linkage," that is, by optimizing the objective function, it simultaneously affects the construction of composite health indicators and the parameter estimation of the random degradation model. In contrast, the loss function of this application focuses more on domain adaptation and preservation of degradation information under multi-source working conditions. Summary of the Invention

[0021] To address the shortcomings of existing technologies in predicting the remaining service life of CNC machine tool tools under various operating conditions, such as insufficient cross-domain generalization ability, ineffective modeling of source domain contribution differences, and easy loss of target operating condition degradation information, this invention proposes a multi-operating-condition adaptive remaining service life prediction method for CNC machine tool tools that considers multi-source collaborative transfer and preservation of target degradation information. This method constructs a multi-source domain transfer learning framework, achieving cross-operating-condition knowledge transfer while preserving the unique degradation patterns of the target domain, thereby significantly improving the prediction accuracy and stability under complex operating conditions.

[0022] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0023] A multi-condition adaptive method for predicting the remaining service life of CNC machine tool cutting tools includes the following steps:

[0024] Step S1: Collect monitoring signals during the tool machining process through sensors and preprocess them to construct a multi-condition adaptive remaining service life prediction dataset for the tool.

[0025] Step S2: Construct a multi-branch migration prediction model based on a degradation trend-aware feature extraction network;

[0026] Step S3: Input the multi-condition adaptation remaining service life prediction dataset of the tool obtained in step S1 into the multi-branch migration prediction model;

[0027] Step S4: Train the multi-branch transfer prediction model by batch-wise jointly optimizing the multi-scale target degradation consistency constraint loss, adversarial loss, and root mean square prediction loss based on the source domain contribution adaptive collaborative optimization mechanism.

[0028] Step S5: After training, only the degradation trend perception feature extraction network and the life regression prediction sub-network in the multi-branch migration prediction model are retained for online tool remaining life prediction.

[0029] As a preferred technical solution of the present invention, step S1 is specifically as follows:

[0030] Step S11: Collect tool vibration signals, spindle current signals and acoustic emission signals during the tool machining process using sensors;

[0031] Step S12: Screen the degradation correlation of each monitoring signal and retain only the time-varying signals that change monotonically with the evolution of tool wear;

[0032] Step S13: Normalize the screened monitoring signals according to the processing conditions, and use the time series samples constructed by the sliding time window to take the remaining tool life corresponding to the end of each time window as the label of the window.

[0033] Step S14: Divide the preprocessed signal data into datasets, take the data under one working condition as the target domain, use 20% of it as the validation set for training, and use the remaining 80% as the test set for testing; use the data under the other working conditions as the multi-source domain for training.

[0034] As a preferred technical solution of the present invention, step S2 is specifically as follows:

[0035] Step S21: Construct a degradation trend perception feature extraction network:

[0036] It includes a local degradation pattern sensing unit for capturing short-term wear mutation features, a long-sequence correlation modeling unit for modeling degradation dependencies across time scales, and a feature compression mapping unit for reducing feature redundancy and forming a compact representation. The three units are cascaded together to extract a unified feature representation containing local degradation patterns and long-term evolution trends from the time series.

[0037] Step S22: Construct a multi-branch migration prediction model based on a degradation trend-aware feature extraction network:

[0038] Each source condition and target condition constitute an independent migration branch. All migration branches share the degradation trend perception feature extraction network described in step S21, and each includes an independent domain distribution discrimination subnetwork and a lifetime regression prediction subnetwork.

[0039] As a preferred technical solution of the present invention: in step S21:

[0040] The local degradation pattern perception unit adopts a one-dimensional convolutional neural network structure, including three one-dimensional convolutional layers. The kernel sizes of each convolutional layer are 16, 3, and 3, and the number of convolutional channels are 3, 64, and 64, respectively. Each convolutional layer is followed by a batch normalization layer and a non-linear activation function ReLU, and feature downsampling is achieved through a max pooling layer.

[0041] The long sequence correlation modeling unit adopts a Transformer encoder based on a self-attention mechanism, with 8 self-attention heads and a hidden layer dimension of 32-256;

[0042] The feature compression mapping unit adopts a fully connected network structure, including three fully connected layers, with 32, 128, and 256 neurons in each layer, and a non-linear activation function LeakyReLU is connected after each fully connected layer.

[0043] As a preferred technical solution of the present invention: the multi-branch migration prediction model in step S22 adopts a strategy of shared representation + hard parameter isolation, so that each migration branch shares the features extracted by the degradation trend perception feature extraction network, but the parameters of the domain discrimination sub-network and prediction sub-network unique to each migration branch are not shared, so as to maintain cross-domain consistency and source domain specificity at the same time.

[0044] As a preferred technical solution of the present invention: in step S22:

[0045] The domain distribution discriminant subnetwork includes three fully connected layers, with 256, 64, and 2 neurons per layer, and the output layer uses a sigmoid function to determine whether the input features come from the source domain or the target domain.

[0046] The lifespan regression prediction subnetwork includes two fully connected layers, each with 256 and 1 neurons, used to output the predicted value of the remaining tool lifespan.

[0047] As a preferred technical solution of the present invention, step S4 is specifically as follows:

[0048] Step S41: Perform adversarial training in each transfer branch to align the distribution of source domain features and target domain features in the discriminant space; at the same time, the feature updates generated during adversarial training are subject to the consistency constraints in step S42.

[0049] Step S42: In the feature distribution alignment process in step S41, a multi-scale target degradation consistency constraint loss is constructed. By calculating the consistency constraint loss between the target domain input and its deep feature representation at three scales, the loss of target domain-specific degradation information during the distribution alignment process is suppressed.

[0050] Step S43: To address the differences in contribution of different source conditions to the target prediction task, an adaptive collaborative optimization mechanism for source domain contribution is constructed. The consistency constraint loss, adversarial loss, and root mean square prediction loss obtained in step S42 are used together as optimization objectives. Weight parameters are introduced for each source condition, and a group search strategy is used to update the weights during the optimization process.

[0051] As a preferred technical solution of the present invention: the multi-scale target degradation consistency constraint loss in step S42 Based on contrastive learning, the following is the specific construction:

[0052] (1);

[0053] Among them, X u and X v Z represents the input of the u-th target domain and the feature of the v-th target domain, respectively. w This represents the data concatenated from the w-th target domain input and target domain features, where sim(·) represents the cosine similarity, B represents the batch size, and s represents the current scale. Let I be the scale set, and I(·) be the indicator function.

[0054] As a preferred embodiment of the present invention: the group search strategy in step S43 adopts an improved fox search strategy, as follows:

[0055] Step S431: Initialize the weighted population using Latin hypercube sampling;

[0056] Step S432: Iteratively update the weights using a dual-mode search strategy, including a jump-style global search and a random perturbation local search.

[0057] Step S433: Use a random reset strategy for out-of-bounds weight parameters to maintain the diversity of the solution space;

[0058] Step S434: Introduce a dynamic elite retention mechanism, with its elite ratio... satisfy:

[0059] (2);

[0060] in This represents the current iteration number. This represents the maximum number of iterations.

[0061] Step S435: Using the root mean square prediction loss on the validation data as the fitness function, the weights are updated and optimized to achieve adaptive adjustment of the contribution of multi-source information to the target task.

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0063] 1. This invention uses a multi-branch transfer modeling structure to decouple the modeling of different source conditions and target conditions, and combines a shared feature extraction network to achieve a unified representation, effectively improving the model's adaptability and prediction accuracy in multi-condition environments.

[0064] 2. This invention proposes a multi-scale target degradation consistency constraint mechanism, which constructs a consistency relationship between the input space and the feature space to effectively preserve the target degradation features during the migration process, thereby improving the prediction reliability.

[0065] 3. This invention uses a source domain contribution adaptive collaborative optimization mechanism to dynamically model the role of different source conditions in the prediction task, avoiding information redundancy or negative transfer problems caused by equal weighting of each source domain in traditional methods, thereby improving the overall performance of the model. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the present invention;

[0067] Figure 2 This is a graph showing the predicted remaining service life of CNC machine tool cutting tools according to the present invention. Detailed Implementation

[0068] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0069] like Figure 1-2 As shown, the present invention proposes a multi-condition adaptive remaining service life prediction method for CNC machine tool tools, which includes the following steps:

[0070] Step S1: Collect monitoring signals during the tool machining process through sensors and preprocess them to construct a multi-condition adaptive remaining service life prediction dataset for the tool.

[0071] Specifically as follows:

[0072] Step S11: Acquire tool vibration signals, spindle current signals, and acoustic emission signals during the tool machining process using sensors. Specifically, multiple sensors are arranged in the CNC machine tool spindle system. Vibration signals are acquired using a three-axis accelerometer mounted on the outer wall of the spindle box. The accelerometer is located no more than 50 mm from the tool mounting position to ensure effective detection of vibration changes caused by tool wear. Acoustic emission signals are acquired using a broadband acoustic emission sensor coupled to the surface of the tool holder or spindle housing. Spindle current signals are acquired using a Hall current sensor installed in the machine tool spindle drive circuit. All sensor signals are synchronously acquired via a data acquisition card, with a sampling frequency set to 20.48 kHz to meet the requirements for high-frequency degradation feature extraction. Simultaneously, the acquisition system is equipped with an anti-aliasing filter to reduce noise interference.

[0073] Step S12: Use the Pearson correlation coefficient to screen the degradation correlation of the monitoring signals monitored by each sensor, and retain only the time-varying signals that change monotonicly with the evolution of tool wear.

[0074] Step S13: The filtered signals are processed using maximum and minimum normalization to the [0, 1] interval according to the machining conditions. Considering the varying lifespans of different tools, for ease of training, the lifespan of each tool is also normalized to the [0, 1] interval, meaning the remaining lifespan when fully healthy is 1, and the remaining lifespan when failed is 0. A time series sample is constructed using a sliding time window, with the remaining tool lifespan at the end of each time window serving as the label for that window. Each time window has an equal length of at least 2048 and a step size of at least 256. The sample for each time window can be represented as... C × T The matrix, where C Indicates the number of sensor channels. T Indicates the length of the time window;

[0075] Step S14: Divide the preprocessed signal data into datasets, take the data under one working condition as the target domain, use 20% of it as the validation set for training, and use the remaining 80% as the test set for testing; use the data under the other working conditions as the multi-source domain for training.

[0076] Step S2: Construct a multi-branch migration prediction model based on a degradation trend-aware feature extraction network;

[0077] Specifically as follows:

[0078] Step S21: Construct a degradation trend perception feature extraction network:

[0079] It includes a local degradation pattern sensing unit for capturing short-term wear mutation features, a long-sequence correlation modeling unit for modeling degradation dependencies across time scales, and a feature compression mapping unit for reducing feature redundancy and forming a compact representation. The three units are cascaded together to extract a unified feature representation containing local degradation patterns and long-term evolution trends from the time series.

[0080] Among them, the local degradation pattern perception unit adopts a one-dimensional convolutional neural network structure, including three one-dimensional convolutional layers, with kernel sizes of 16, 3, and 3 respectively, and the number of convolutional channels of 3, 64, and 64 respectively. After each convolutional layer, a batch normalization layer and a non-linear activation function ReLU are connected in sequence, and feature downsampling is achieved through a max pooling layer with a pooling kernel size of 2.

[0081] The long sequence correlation modeling unit adopts a Transformer encoder based on self-attention mechanism, with 8 self-attention heads and a hidden layer dimension of 32-256. It uses residual connection and layer normalization structure. The feature sequence input to the Transformer encoder is first positionally encoded using sine-cosine position encoding method, and then fed into the multi-head self-attention layer.

[0082] The feature compression mapping unit adopts a fully connected network structure, including three fully connected layers with 32, 128, and 256 neurons in each layer. Each fully connected layer is followed by a non-linear activation function LeakyReLU (with a negative slope of 0.2) to flatten the output features of the Transformer encoder before inputting them into the fully connected layer.

[0083] Step S22: Construct a multi-branch migration prediction model based on a degradation trend-aware feature extraction network:

[0084] The multi-branch migration prediction model employs a strategy of shared representations and hard parameter isolation. This allows each migration branch to share features extracted by the degradation trend-aware feature extraction network, but the parameters of the domain-discriminative subnetwork and prediction subnetwork, unique to each migration branch, are not shared. This approach maintains both cross-domain consistency and source domain specificity.

[0085] Each source condition and target condition constitute an independent migration branch. The number of branches is set to 3, which is the number of source domains. All migration branches share the degradation trend perception feature extraction network described in step S21, and each includes an independent domain distribution discrimination subnetwork and a lifetime regression prediction subnetwork.

[0086] Wherein: the domain distribution discriminant subnetwork includes 3 fully connected layers, with 256, 64 and 2 neurons in each layer, and the output layer uses Sigmoid to distinguish whether the input features come from the source domain or the target domain;

[0087] The lifespan regression prediction subnetwork includes two fully connected layers, each with 256 and 1 neurons, used to output the predicted value of the remaining tool lifespan.

[0088] Step S3: Input the multi-condition adaptation remaining service life prediction dataset of the tool obtained in step S1 into the multi-branch migration prediction model in batches. Each training batch contains both source domain samples and target domain samples, and the number of source domain samples and target domain samples is kept in a 1:1 ratio.

[0089] Step S4: Train the multi-branch transfer prediction model by batch-wise jointly optimizing the multi-scale target degradation consistency constraint loss, adversarial loss, and root mean square prediction loss based on the source domain contribution adaptive collaborative optimization mechanism.

[0090] Specifically as follows:

[0091] Step S41: In each transfer branch, an alternating optimization strategy is used to perform adversarial training to align the distribution of source domain features and target domain features in the discriminant space. That is, the parameters of the feature extraction network are fixed, the parameters of the domain distribution discriminant sub-network are updated, and then the parameters of the domain distribution discriminant sub-network are fixed, and the feature extraction network is updated through a gradient reversal layer. The adversarial loss function is binary classification cross-entropy, the source domain label is set to 1, and the target domain label is set to 0. At the same time, the feature updates generated during the adversarial training process are restricted by the consistency constraint in step S42 to avoid destroying the original degenerate structure of the target condition during the distribution alignment process.

[0092] Step S42: In the feature distribution alignment process in step S41, a multi-scale target degradation consistency constraint loss is constructed. The consistency constraint loss between the target domain input and its deep feature representation is calculated at three scales: the original input layer at the shallow c-scale, the output layer of the local degradation pattern perception unit at the middle scale, and the output layer of the long sequence correlation modeling unit at the deep scale. This suppresses the loss of target domain-specific degradation information during the distribution alignment process.

[0093] Multi-scale target degradation consistency constraint loss Based on contrastive learning, the following is the specific construction:

[0094] (1);

[0095] Among them, X u and X v Z represents the input of the u-th target domain and the feature of the v-th target domain, respectively. w This represents the data concatenated from the w-th target domain input and target domain features, where sim(·) represents the cosine similarity, B represents the batch size, and s represents the current scale. Let I be the scale set, and I(·) be the indicator function.

[0096] Step S43: To address the differences in contribution of different source conditions to the target prediction task, an adaptive collaborative optimization mechanism for source domain contribution is constructed. The consistency constraint loss, adversarial loss, and root mean square prediction loss obtained in step S42 are used together as optimization objectives. Weight parameters are introduced for each source condition, and a group search strategy is used to update the weights during the optimization process.

[0097] Specifically, the group search strategy adopts an improved fox search strategy, as follows:

[0098] Step S431: Initialize the weighted population using Latin hypercube sampling;

[0099] Step S432: Iteratively update the weights using a dual-mode search strategy, including a jump-style global search and a random perturbation local search.

[0100] Step S433: Use a random reset strategy for out-of-bounds weight parameters to maintain the diversity of the solution space;

[0101] Step S434: Introduce a dynamic elite retention mechanism, with its elite ratio... satisfy:

[0102] (2);

[0103] in This represents the current iteration number. This represents the maximum number of iterations.

[0104] Step S435: Using the root mean square prediction loss on the validation data as the fitness function, the weights are updated and optimized to achieve adaptive adjustment of the contribution of multi-source information to the target task.

[0105] The improved fox search strategy is implemented as follows:

[0106] (1) Set the population size to 10; set the maximum number of iterations to 100; set the weight vector dimension to equal the number of source domains (3), and limit the weight values ​​to the range of [0, 1]. To ensure strong global exploration capability in the early stages of the search, use Latin hypercube sampling to initialize the weight population to improve the search space coverage;

[0107] (2) The weights are iteratively updated through a dual-mode search strategy. The global search phase is achieved by a jump update strategy based on the current global best individual, while the local search phase searches within the neighborhood of the current individual through random perturbation. The selection probability of both is set to 0.5 to achieve a balance between global exploration and local development.

[0108] (3) A random reset strategy is adopted for the out-of-bounds weight parameters, and they are remapped to random positions in the interval [0, 1] to maintain the diversity of the solution space;

[0109] (4) Introduce a dynamic elite retention mechanism, and its elite ratio satisfy:

[0110] (2);

[0111] in This represents the current iteration number. This represents the maximum number of iterations.

[0112] (5) The root mean square prediction loss on the validation data is used as the fitness function to update and optimize the weights, thereby realizing the adaptive adjustment of the contribution of multi-source information to the target task.

[0113] Regarding network training parameter settings, this embodiment uses the Adam optimizer to train the model, with an initial learning rate of 0.001, and employs an exponential decay strategy to gradually reduce the learning rate during training. The batch size is set to 64, and the number of training epochs is set to 150. An early stopping strategy is introduced during training: the model is considered converged and training is terminated when the root mean square error of the validation set does not decrease within 10 consecutive epochs.

[0114] Step S5: After training, only the degradation trend perception feature extraction network and the life regression prediction subnetwork in the multi-branch migration prediction model are retained for online tool remaining life prediction. The real-time monitoring signal collected by the tool to be predicted under the target working condition is input into the degradation trend perception feature extraction network, and the remaining life prediction value at the current moment is output through the life regression prediction subnetwork. The specific process includes:

[0115] Perform the same preprocessing on the real-time signal as during the training phase;

[0116] Input samples are continuously generated using a sliding window method;

[0117] A moving average is applied to the prediction results of multiple consecutive windows to reduce prediction fluctuations.

[0118] Based on the above technical embodiments, the present invention constructs a multi-source domain transfer learning framework, which not only realizes cross-condition knowledge transfer but also maintains the unique degradation mode of the target domain, thereby significantly improving the prediction accuracy and stability under complex conditions.

[0119] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A multi-condition adaptive remaining service life prediction method for CNC machine tool cutting tools, characterized in that, Includes the following steps: Step S1: Collect monitoring signals during the tool machining process through sensors and preprocess them to construct a multi-condition adaptive remaining service life prediction dataset for the tool. Step S2: Construct a multi-branch migration prediction model based on a degradation trend-aware feature extraction network; Step S2 is as follows: Step S21: Construct a degradation trend perception feature extraction network: It includes a local degradation pattern sensing unit for capturing short-term wear mutation features, a long-sequence correlation modeling unit for modeling degradation dependencies across time scales, and a feature compression mapping unit for reducing feature redundancy and forming a compact representation. The three units are cascaded together to extract a unified feature representation containing local degradation patterns and long-term evolution trends from the time series. Step S22: Construct a multi-branch migration prediction model based on a degradation trend-aware feature extraction network: Each source condition and target condition constitute an independent migration branch. All migration branches share the degradation trend perception feature extraction network described in step S21, and each includes an independent domain distribution discrimination subnetwork and a lifetime regression prediction subnetwork. Step S3: Input the multi-condition adaptation remaining service life prediction dataset of the tool obtained in step S1 into the multi-branch migration prediction model; Step S4: Train the multi-branch transfer prediction model by batch-wise jointly optimizing the multi-scale target degradation consistency constraint loss, adversarial loss, and root mean square prediction loss based on the source domain contribution adaptive collaborative optimization mechanism. Step S4 is as follows: Step S41: Perform adversarial training in each transfer branch to align the distribution of source domain features and target domain features in the discriminant space; at the same time, the feature updates generated during adversarial training are subject to the consistency constraints in step S42. Step S42: In the feature distribution alignment process in step S41, a multi-scale target degradation consistency constraint loss is constructed. By calculating the consistency constraint loss between the target domain input and its deep feature representation at three scales, the loss of target domain-specific degradation information during the distribution alignment process is suppressed. The multi-scale target degradation consistency constraint loss in step S42 Based on contrastive learning, the following is the specific construction: （1）； Among them, X u and X v Z represents the input of the u-th target domain and the feature of the v-th target domain, respectively. w This represents the data concatenated from the w-th target domain input and target domain features, where sim(·) represents the cosine similarity, B represents the batch size, and s represents the current scale. Let I be the scale set, and I(·) be the indicator function; Step S43: To address the differences in contribution of different source conditions to the target prediction task, an adaptive collaborative optimization mechanism for source domain contribution is constructed. The consistency constraint loss, adversarial loss, and root mean square prediction loss obtained in step S42 are used together as optimization objectives. Weight parameters are introduced for each source condition, and a group search strategy is used to update the weights during the optimization process. Step S5: After training, only the degradation trend perception feature extraction network and the life regression prediction sub-network in the multi-branch migration prediction model are retained for online tool remaining life prediction.

2. The multi-condition adaptive remaining service life prediction method for CNC machine tool cutting tools according to claim 1, characterized in that, The specific steps of S1 are as follows: Step S11: Collect tool vibration signals, spindle current signals and acoustic emission signals during the tool machining process using sensors; Step S12: Screen the degradation correlation of each monitoring signal and retain only the time-varying signals that change monotonically with the evolution of tool wear; Step S13: Normalize the screened monitoring signals according to the processing conditions, and use the time series samples constructed by the sliding time window to take the remaining tool life corresponding to the end of each time window as the label of the window. Step S14: Divide the preprocessed signal data into datasets, take the data under one working condition as the target domain, use 20% of it as the validation set for training, and use the remaining 80% as the test set for testing; use the data under the other working conditions as the multi-source domain for training.

3. The multi-condition adaptive remaining service life prediction method for CNC machine tool tools according to claim 1, characterized in that, In step S21: The local degradation pattern perception unit adopts a one-dimensional convolutional neural network structure, including three one-dimensional convolutional layers. The kernel sizes of each convolutional layer are 16, 3, and 3, and the number of convolutional channels are 3, 64, and 64, respectively. Each convolutional layer is followed by a batch normalization layer and a non-linear activation function ReLU, and feature downsampling is achieved through a max pooling layer. The long sequence correlation modeling unit adopts a Transformer encoder based on a self-attention mechanism, with 8 self-attention heads and a hidden layer dimension of 32-256; The feature compression mapping unit adopts a fully connected network structure, including three fully connected layers, with 32, 128, and 256 neurons in each layer, and a non-linear activation function LeakyReLU is connected after each fully connected layer.

4. The multi-condition adaptive remaining service life prediction method for CNC machine tool cutting tools according to claim 3, characterized in that, In step S22, the multi-branch migration prediction model adopts a strategy of shared representation + hard parameter isolation, so that each migration branch shares the features extracted by the degradation trend perception feature extraction network, but the parameters of the domain discrimination sub-network and prediction sub-network unique to each migration branch are not shared, so as to maintain cross-domain consistency and source domain specificity at the same time.

5. The multi-condition adaptive remaining service life prediction method for CNC machine tool cutting tools according to claim 3, characterized in that, In step S22: The domain distribution discriminant subnetwork includes three fully connected layers, with 256, 64, and 2 neurons per layer, and the output layer uses a sigmoid function to determine whether the input features come from the source domain or the target domain. The lifespan regression prediction subnetwork includes two fully connected layers, each with 256 and 1 neurons, used to output the predicted value of the remaining tool lifespan.

6. The multi-condition adaptive remaining service life prediction method for CNC machine tool cutting tools according to claim 1, characterized in that, The group search strategy in step S43 adopts an improved fox search strategy, as follows: Step S431: Initialize the weighted population using Latin hypercube sampling; Step S432: Iteratively update the weights using a dual-mode search strategy, including a jump-style global search and a random perturbation local search. Step S433: Use a random reset strategy for out-of-bounds weight parameters to maintain the diversity of the solution space; Step S434: Introduce a dynamic elite retention mechanism, with its elite ratio... satisfy: （2）； in This represents the current iteration number. This represents the maximum number of iterations. Step S435: Using the root mean square prediction loss on the validation data as the fitness function, the weights are updated and optimized to achieve adaptive adjustment of the contribution of multi-source information to the target task.

Images