Tool wear prediction method based on environment robust causal representation, terminal and storage medium

Abstract

The application discloses a tool wear prediction method based on environment robust causal representation, a terminal and a storage medium. In the reasoning stage, only a multi-environment robust prediction branch is deployed, multi-modal monitoring data in the processing process needs to be collected, but it does not depend on complete modal data, and the modal missing state, noise and signal quality state are uniformly encoded as measurement environment variables to realize unified environment modeling. After modal encoding and fusion, the multi-modal fusion features are separated into causal representation and bias representation, so that the wear degradation information and environmental disturbance information are effectively separated in the representation space. Based on the environment prototype dictionary and the missing-noise debiasing module, the debiased causal representation is generated and corrected to obtain a robust debiased causal representation. After projecting the robust debiased causal representation, the tool wear prediction result is obtained. The application can adapt to low-cost sensor configuration and partial sensor failure scene, and realize robust and accurate tool wear prediction.

Description

Technical Field

[0001] This application relates to the fields of CNC machining and artificial intelligence technology, and in particular to a tool wear prediction method, terminal and storage medium based on environmentally robust causal characterization. Background Technology

[0002] In CNC machining, tool wear status typically requires comprehensive characterization using multi-source monitoring signals such as triaxial cutting force, triaxial vibration, and acoustic emission to capture tool degradation information from various perspectives, including cutting load, mechanical vibration, and high-frequency energy release. However, data quality degradation in actual industrial settings is not solely caused by the loss of a single mode; it can also be due to variations in noise type and intensity, decreased signal-to-noise ratio, signal drift, abnormal impacts, and changes in the measurement environment resulting from a combination of factors. These environmental changes can lead to the loss of key degradation information, damage to effective features, and enhanced environmental spurious correlations. This makes it difficult for models to distinguish between actual wear evolution and measurement disturbances, resulting in significant representational shifts when the environment changes, severely reducing the accuracy, stability, and reliability of tool wear prediction.

[0003] To address the problem of missing modalities, existing methods typically employ a process of completing the data before modeling, primarily including direct deletion, statistical imputation, machine learning imputation, and deep learning imputation. Direct deletion can lead to sample selection and distribution disruption when the missing proportion is high; statistical imputation and machine learning imputation methods struggle to characterize the complex nonlinear dynamic relationships in the tool wear process; while deep imputation methods based on self-attention mechanisms, generative models, or diffusion models utilize temporal dependencies and intervariate correlations for reconstruction, their primary goal remains restoring the missing data itself, making it difficult to guarantee consistency between the reconstructed signal and the actual tool degradation semantics.

[0004] To address noise pollution, existing methods typically employ a signal preprocessing approach followed by prediction, primarily including traditional filtering methods, time-frequency analysis methods, signal decomposition methods, and deep denoising methods. Traditional filtering and wavelet thresholding denoising tend to weaken subtle wear and degradation characteristics when suppressing noise; methods based on empirical mode decomposition or variational mode decomposition rely on manual rules and prior parameters, making it difficult to stably distinguish between real degradation information and environmental disturbances; while deep denoising methods possess strong nonlinear modeling capabilities, they are prone to overfitting or denoising bias when noise type, intensity, and signal quality vary with operating conditions.

[0005] Existing technologies typically treat modal missingness and noise contamination as independent problems, lacking a unified representation of the measurement environment comprised of missing states, noise types, noise intensity, and signal quality. This makes it difficult to characterize systematic shifts between different environments. Data completion or denoising methods, aiming to recover the observed signal, cannot guarantee that the recovered results retain the true causal semantics of wear. Instead, they may lose key features related to wear degradation while suppressing noise. Ordinary multimodal prediction and knowledge distillation methods easily transmit strong modal shortcuts or specific noise patterns from complete modes to the prediction branch, leading to the model incorrectly learning spurious correlations related to the environment. Furthermore, existing technologies lack prototype structures that can explicitly characterize typical missing noise environments, and mechanisms to constrain the consistency of the same degradation state across environments. Therefore, they are prone to failure under unseen combinations of missing states, mixed noise, or changes in signal quality, failing to meet the demand for high robustness in complex and variable measurement environments for tool wear prediction in industrial settings.

[0006] Therefore, existing technologies still need improvement and development. Summary of the Invention

[0007] The technical problem to be solved by this application is to provide a tool wear prediction method, terminal and storage medium based on environmental robust causal characterization, which addresses the above-mentioned deficiencies of the prior art. It aims to solve the problem that the quality of monitoring data is reduced due to the lack of strong modes and noise pollution in the actual machining environment, thereby affecting the accuracy of tool wear state characterization and prediction reliability.

[0008] The technical solution adopted in this application to solve the problem is as follows: In a first aspect, embodiments of this application provide a method for predicting tool wear based on environmentally robust causal characterization, the method comprising: During the inference phase, multimodal monitoring data and corresponding measurement environmental variables are acquired during the CNC machining process; the measurement environmental variables include modal missing state, noise and signal quality state. Robust causal projection features are generated based on the multimodal monitoring data and the measured environmental variables through a trained multi-environment robust prediction branch, and tool wear prediction results are output based on the robust causal projection features. The structure of the multi-environment robust prediction branch includes: A multi-modal encoder is used to encode the available modalities in the multi-modal monitoring data and to mark the missing modalities with a missing mask, thereby obtaining several modal coding features; A multimodal fusion module is used to perform multimodal fusion on the modal coding features to obtain fused features; The causal-bias decoupling module is used to perform causal-bias decoupling on the fused features to obtain the causal representation and bias representation corresponding to the multimodal monitoring data; The missing-noise debiasing module and the environmental prototype dictionary are used to generate a robust debiased causal representation of the multimodal monitoring data based on the causal representation, bias representation, and the measured environmental variables corresponding to the multimodal monitoring data. A projection head is used to project the robust debiased causal representation corresponding to the multimodal monitoring data to obtain the robust causal projection feature corresponding to the multimodal monitoring data. The prediction head is used to output tool wear prediction results based on the robust causal projection features corresponding to the multimodal monitoring data.

[0009] In one implementation, the multimodal monitoring data includes: a triaxial cutting force signal, a triaxial vibration signal, and an acoustic emission signal.

[0010] In one implementation, the missing-noise debiasing module and the environment prototype dictionary are specifically used for: Based on the bias representation corresponding to the multimodal monitoring data and the measured environmental variables, the causal representation corresponding to the multimodal monitoring data is corrected to obtain the debiased causal representation corresponding to the multimodal monitoring data. Calculate the distance between the bias representation corresponding to the multimodal monitoring data and each environmental prototype in the environmental prototype dictionary, and calculate the attention weight based on the distance value of each environmental prototype; An environment bias vector is generated based on the attention weights; the debiased causal representation corresponding to the multimodal monitoring data is corrected according to the environment bias vector to obtain a robust debiased causal representation corresponding to the multimodal monitoring data.

[0011] In one implementation, the training method for the multi-environment robust prediction branch includes: During the training phase, historical multimodal monitoring data, corresponding historical measurement environmental variables, and corresponding tool wear labels are acquired during the CNC machining process. Construct multi-environmental disturbance input samples based on the historical multimodal monitoring data and the historical measured environmental variables; A high-quality reference branch is used to generate reference causal projection features based on the historical multimodal monitoring data and the historical measured environmental variables, and the first tool wear prediction result is output based on the reference causal projection features; the structure of the high-quality reference branch is the same as the structure of the multi-environment robust prediction branch. Through the multi-environment robust prediction branch, a second tool wear prediction result is output based on the robust causal projection features corresponding to the multi-environment disturbance input samples. The robust causal projection features have the same dimension as the reference causal projection features. The robust causal projection features corresponding to the multi-environment perturbation input samples are distilled using the reference causal projection features as a knowledge source, and the multi-environment robust prediction branch is trained by combining multiple loss functions.

[0012] In one implementation, the steps of generating reference causal projection features based on the historical multimodal monitoring data and the historical measured environmental variables through a high-quality reference branch, and outputting a first tool wear prediction result based on the reference causal projection features, include: The historical multimodal monitoring data is encoded using encoders of multiple modes in the high-quality reference branch to obtain the modality coding features corresponding to the historical multimodal monitoring data. The multimodal fusion module in the high-quality reference branch performs multimodal fusion on the various modal coding features corresponding to the historical multimodal monitoring data to obtain reference fusion features. The causal-bias decoupling module in the high-quality reference branch is used to decouple the reference fusion features from causal to bias, thereby obtaining the causal representation and bias representation corresponding to the reference fusion features. Using the missing-noise debiasing module and the environmental prototype dictionary in the high-quality reference branch, a reference debiasing causal representation is generated based on the causal representation, bias representation, and historical measured environmental variables corresponding to the reference fusion features. The reference debiased causal representation is projected using the projection head in the high-quality reference branch to obtain the reference causal projection feature. The first tool wear prediction result is output based on the reference causal projection features through the prediction head in the high-quality reference branch.

[0013] In one implementation, the step of outputting a second tool wear prediction result based on the robust causal projection features corresponding to the multi-environment perturbation input samples through the multi-environment robust prediction branch, wherein the robust causal projection features have the same dimension as the reference causal projection features, includes: The available modes in the multi-environmental perturbation input samples are encoded by the encoder of the mode in the multi-environmental robust prediction branch, and the missing modes are marked by a missing mask to obtain several mode encoding features corresponding to the multi-environmental perturbation input samples. The multimodal fusion module in the multi-environment robust prediction branch performs multimodal fusion on the modal coding features corresponding to the multi-environment perturbation input samples to obtain the fused features corresponding to the multi-environment perturbation input samples. The causal-bias decoupling module in the multi-environment robust prediction branch performs causal-bias decoupling on the fusion features corresponding to the multi-environment perturbation input samples to obtain the causal representation and bias representation corresponding to the multi-environment perturbation input samples. The robust debiased causal representation of the multi-environmental perturbation input sample is generated by using the missing-noise debiasing module and the environment prototype dictionary in the multi-environmental robust prediction branch, based on the causal representation, bias representation, and historical measured environmental variables corresponding to the multi-environmental perturbation input sample. The robust causal projection features corresponding to the multi-environment perturbation input samples are obtained by projecting the robust debiased causal representations corresponding to the multi-environment perturbation input samples through the projection head in the multi-environment robust prediction branch. The second tool wear prediction result is output through the prediction head in the multi-environment robust prediction branch, based on the robust causal projection features corresponding to the multi-environment disturbance input samples.

[0014] In one implementation, the steps of distilling the robust causal projection features corresponding to the multi-environmental perturbation input samples using the reference causal projection features as a knowledge source, and training the multi-environmental robust prediction branch in conjunction with multiple loss functions include: Wear regression loss is calculated based on the tool wear label and the first tool wear prediction result; orthogonal constraint loss is calculated based on the causal characterization and bias characterization corresponding to the high-quality reference branch; bias prediction loss is calculated based on the bias characterization corresponding to the high-quality reference branch to predict the mode missing state, noise and signal quality state, and combined with the historical measurement environment variables; cross-environment invariance loss is calculated based on the causal characterization calculated by the high-quality reference branch based on the same sample under different measurement environments and the corresponding first tool wear prediction result. The high-quality reference branch is trained using a multi-objective joint optimization method, based on the wear regression loss, orthogonal constraint loss, bias prediction loss, and cross-environment invariance loss corresponding to the high-quality reference branch. Wear regression loss is calculated based on the tool wear label and the second tool wear prediction result; orthogonal constraint loss is calculated based on the causal characterization and bias characterization corresponding to the multi-environment robust prediction branch; bias prediction loss is calculated based on the bias characterization corresponding to the multi-environment robust prediction branch to predict mode missing state, noise and signal quality state, and combined with the historical measurement environmental variables; cross-environment invariance loss is calculated based on the causal characterization calculated by the multi-environment robust prediction branch based on the same sample under different measurement environments and the corresponding second tool wear prediction result; distillation loss is calculated based on the reference causal projection feature and the robust causal projection feature corresponding to the multi-environment perturbation input sample. The multi-environment robust prediction branch is trained using a multi-objective joint optimization approach, based on the wear regression loss, orthogonal constraint loss, bias prediction loss, cross-environment invariance loss, and distillation loss corresponding to the multi-environment robust prediction branch.

[0015] In one implementation, the distillation loss is the Kullback-Leibler divergence.

[0016] Secondly, embodiments of this application also provide a terminal, the terminal including a memory and one or more processors; the memory stores one or more programs; the programs include instructions for executing the tool wear prediction method based on environmental robust causal characterization as described above; the processor is used to execute the programs.

[0017] Thirdly, embodiments of this application also provide a computer-readable storage medium storing a plurality of instructions adapted to be loaded and executed by a processor to implement the steps of the tool wear prediction method based on environmental robust causal characterization as described above.

[0018] The beneficial effects of this application are as follows: In the inference phase, this application only deploys a multi-environment robust prediction branch. It requires collecting multi-modal monitoring data during the machining process but does not rely on complete modal data. It uniformly encodes missing modal states, noise, and signal quality states as measurement environmental variables to achieve unified environmental modeling. After modal encoding and fusion, the multi-modal fusion features are separated into causal representations and bias representations, effectively separating wear degradation information from environmental disturbance information in the representation space. Based on the environmental prototype dictionary and the missing-noise debiasing module, a debiased causal representation is generated and corrected to obtain a robust debiased causal representation. After projecting the robust debiased causal representation, the tool wear prediction result is obtained. This application can adapt to low-cost sensor configurations and scenarios with partial sensor failure, and achieves robust and accurate tool wear prediction. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart of the tool wear prediction method based on environmental robust causal characterization provided in the embodiments of this application.

[0021] Figure 2 This is a technical architecture diagram of model training provided in the embodiments of this application.

[0022] Figure 3 This is a schematic block diagram of the terminal provided in the embodiments of this application. Detailed Implementation

[0023] This application discloses a tool wear prediction method, terminal, and storage medium based on environmentally robust causal characterization. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit this application.

[0024] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0025] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0026] To address the aforementioned shortcomings of existing technologies, this application provides a tool wear prediction method based on environmentally robust causal characterization, such as... Figure 1 As shown, the method includes: Step S100: In the reasoning stage, acquire multimodal monitoring data and corresponding measurement environmental variables during the CNC machining process; the measurement environmental variables include modal missing state, noise and signal quality state.

[0027] Furthermore, the multimodal monitoring data includes: triaxial cutting force signal, triaxial vibration signal, and acoustic emission signal.

[0028] Specifically, the method in this embodiment is mainly divided into two stages: a training stage and an inference stage. The purpose of this method is not merely to train a model using historical data, but more importantly, to address the problem of how to stably and accurately output tool wear prediction results when faced with uncontrollable changes in the measurement environment, such as modal missingness, noise contamination, and signal quality degradation after actual deployment. Specifically, in the inference stage, multimodal monitoring data and corresponding measurement environment variables are acquired during the CNC machining process. This step is to enable the multi-environment robust prediction branch to accurately perceive the actual measurement conditions of the current machining site, thereby specifically eliminating environmental spurious correlations and characterization biases during feature extraction and prediction. The multimodal monitoring data mainly includes triaxial cutting force signals, triaxial vibration signals, and acoustic emission signals, comprehensively reflecting the tool wear degradation process from different physical dimensions such as cutting load, mechanical vibration, and high-frequency energy release. The measurement environment variables are used to uniformly describe the quality degradation status faced by the current multimodal monitoring data, specifically including modal missingness states (also known as modal missingness masks) and noise and signal quality states. Modal missing state indicates the current availability of various modes such as force, vibration, and acoustic emission; noise and signal quality state encompasses noise type, noise intensity, and signal-to-noise ratio (SNR), where noise type can include Gaussian noise, impulse noise, drift noise, periodic interference noise, and mixed noise formed by the superposition of multiple noises. This application encodes modal missing state, noise type, noise intensity, and signal quality state (SNR) into a unified measurement environment variable, thereby achieving unified environmental modeling of missing and noise. This allows for a structured and complete description of the complex measurement environment changes formed by the coupling of strong modal missing states and multiple types of noise. Through this unified adaptive approach, a single environment variable characterizes arbitrary combinations of modal missing states, noise types, and signal quality changes. This enables the model to simultaneously handle strong modal missing states, single noise, mixed noise, and different SNR conditions within the same framework, avoiding error accumulation caused by using multiple independent missing state and signal preprocessing procedures. This provides an accurate environmental context for the subsequent generation of robust causal projection features and consistent cross-environment tool wear prediction.

[0029] Step S200: Using the trained multi-environment robust prediction branch, generate robust causal projection features based on the multimodal monitoring data and the measured environmental variables, and output tool wear prediction results based on the robust causal projection features; The structure of the multi-environment robust prediction branch includes: A multi-modal encoder is used to encode the available modalities in the multi-modal monitoring data and to mark the missing modalities with a missing mask, thereby obtaining several modal coding features; A multimodal fusion module is used to perform multimodal fusion on the modal coding features to obtain fused features; The causal-bias decoupling module is used to perform causal-bias decoupling on the fused features to obtain the causal representation and bias representation corresponding to the multimodal monitoring data; The missing-noise debiasing module and the environmental prototype dictionary are used to generate a robust debiased causal representation of the multimodal monitoring data based on the causal representation, bias representation, and the measured environmental variables corresponding to the multimodal monitoring data. A projection head is used to project the robust debiased causal representation corresponding to the multimodal monitoring data to obtain the robust causal projection feature corresponding to the multimodal monitoring data. The prediction head is used to output tool wear prediction results based on the robust causal projection features corresponding to the multimodal monitoring data.

[0030] Specifically, during the inference phase, this application utilizes a trained multi-environment robust prediction branch to generate robust causal projection features based on acquired multimodal monitoring data and measurement environmental variables. Based on these features, it outputs tool wear prediction results, thereby achieving unified adaptation to complex measurement environments in industrial settings and robust wear state output. The multi-environment robust prediction branch is deployed independently during the inference phase and does not require continuous acquisition of complete modal data. It is applicable to scenarios with low-cost sensor configurations or partial sensor failures. Its specific structure and processing flow are as follows: First, the multimodal encoder receives input multimodal monitoring data that may contain missing modalities, noise contamination, or signal quality degradation. It encodes the monitoring data for each usable modality separately, while marking missing modalities using a missing mask, thus obtaining several modal coding features. Subsequently, the multimodal fusion module fuses these modal coding features to form a unified fusion feature, comprehensively expressing the overall state of the current sample under multi-source physical dimensions and measurement attenuation conditions. Next, the causal-bias decoupling module operates on this fused feature, decomposing it into a causal representation related to the tool degradation state and a bias representation related to measurement environment disturbances, thus separating wear degradation information from environmental disturbance information in the representation space. This decoupling operation can separate the essential degradation patterns formed during machining due to changes in cutting parameters and natural tool wear from the observation offsets caused by modal missingness, noise type and intensity, and signal drift. Then, the missing-noise debiasing module, combined with the environmental prototype dictionary, generates a robust debiased causal representation based on the causal representation, bias representation, and measurement environment variables of the current sample. The environmental prototype dictionary stores learnable prototype representations of typical missing and noisy environments, which can identify the typical or mixed environment in which the current sample is located, and correct the representation offset caused by environmental changes through prototype conditional missing-noise debiasing, thereby suppressing the interference of bias representation on causal representation, while not losing key information related to wear degradation. Subsequently, the projection head projects and maps this robust debiased causal representation to obtain robust causal projection features. Finally, the prediction head outputs the tool wear prediction result based on the robust causal projection feature, achieving an accurate and stable estimate of the tool wear value. This embodiment's method, through environmental modeling, causal-bias decoupling, prototype condition debiasing, and wear prediction, ensures that the model maintains cross-environmental prediction robustness and consistency under various environmental changes such as strong mode loss, mixed noise, and different signal-to-noise ratios.

[0031] In one implementation, the missing-noise debiasing module and the environment prototype dictionary are specifically used for: Based on the bias representation corresponding to the multimodal monitoring data and the measured environmental variables, the causal representation corresponding to the multimodal monitoring data is corrected to obtain the debiased causal representation corresponding to the multimodal monitoring data. Calculate the distance between the bias representation corresponding to the multimodal monitoring data and each environmental prototype in the environmental prototype dictionary, and calculate the attention weight based on the distance value of each environmental prototype; An environment bias vector is generated based on the attention weights; the debiased causal representation corresponding to the multimodal monitoring data is corrected according to the environment bias vector to obtain a robust debiased causal representation corresponding to the multimodal monitoring data.

[0032] Specifically, in the multi-environment robust prediction branch, the missing-noise debiasing module and the environmental prototype dictionary together constitute the key mechanism for eliminating environmental spurious correlations and reducing representational bias caused by environmental changes. First, the missing-noise debiasing module performs preliminary correction on the causal representation based on the bias representation corresponding to the multimodal monitoring data, and the measurement environmental variables consisting of modality missing masks, noise types, noise intensity, and signal-to-noise ratio, obtaining a debiased causal representation. This step utilizes environmental disturbance information extracted from the bias representation and the explicitly encoded modality missing state, noise state, and signal quality in the measurement environmental variables to conditionally suppress or adjust components in the causal representation contaminated by environmental factors, thereby weakening the direct interference of modality missingness and known noise types on the degraded feature expression. Subsequently, the missing-noise debiasing module performs prototype conditional debiasing correction in conjunction with the environmental prototype dictionary. The environmental prototype dictionary pre-establishes learnable prototype representations corresponding to typical and mixed measurement environments, including but not limited to complete modal environments, strong modality missing environments, Gaussian noise environments, impact noise environments, drift noise environments, periodic interference environments, and mixed environments formed by the coupling of multiple noises. During the inference phase, an environment prototype dictionary built upon bias representations is used to calculate the distance between the current sample's bias representation and each environment prototype in the dictionary. Based on these distances, an attention mechanism is then used to calculate the attention weights corresponding to each environment prototype. These attention weights reflect the similarity between the current sample's measurement environment and each typical environment prototype, effectively identifying whether the main environmental influence on the current sample is a single known environmental type or a mixed environment pattern formed by multiple interfering factors. Weighted aggregation of each environment prototype is performed based on the attention weights to generate an environment bias vector. This environment bias vector comprehensively reflects the overall deviation direction and magnitude of the current sample in the multidimensional environmental space. Finally, the missing-noise debiasing module performs further gating correction on the debiased causal representation based on the generated environment bias vector, obtaining a robust debiased causal representation. This correction process selectively compensates for or suppresses components in the causal representation influenced by environment prototype conditions through a gating mechanism, thereby effectively reducing environmental spurious correlations under the coupling effect of multi-factor environmental disturbances. This ensures that the final causal representation used for tool wear prediction maintains high semantic consistency and cross-environment robustness under different missing, noisy, and mixed environments.

[0033] In this embodiment, causal-bias decoupling separates stable tool wear degradation information from environmental disturbance information. Causal representation undertakes the wear prediction task, while bias representation undertakes the environmental identification task. The environmental prototype dictionary and its attention weights can reflect the main environmental influences on the current sample, taking into account both the interpretability of the model and the low-complexity deployment requirements under low-cost sensor configuration.

[0034] In one implementation, the training method for the multi-environment robust prediction branch includes: During the training phase, historical multimodal monitoring data, corresponding historical measurement environmental variables, and corresponding tool wear labels are acquired during the CNC machining process. Construct multi-environmental disturbance input samples based on the historical multimodal monitoring data and the historical measured environmental variables; A high-quality reference branch is used to generate reference causal projection features based on the historical multimodal monitoring data and the historical measured environmental variables, and the first tool wear prediction result is output based on the reference causal projection features; the structure of the high-quality reference branch is the same as the structure of the multi-environment robust prediction branch. Through the multi-environment robust prediction branch, a second tool wear prediction result is output based on the robust causal projection features corresponding to the multi-environment disturbance input samples. The robust causal projection features have the same dimension as the reference causal projection features. The robust causal projection features corresponding to the multi-environment perturbation input samples are distilled using the reference causal projection features as a knowledge source, and the multi-environment robust prediction branch is trained by combining multiple loss functions.

[0035] Specifically, the model training method used in this embodiment is collaborative training of a high-quality reference branch and a multi-environment robust prediction branch. The high-quality reference branch transmits stable causal knowledge after bias correction, enabling the multi-environment robust prediction branch to maintain consistent tool wear prediction capabilities under conditions of strong modal missingness and low-quality data, just as it does under high-quality data conditions. During the training phase, historical multimodal monitoring data, corresponding historical measurement environmental variables, and tool wear labels from the CNC machining process are first acquired. The historical measurement environmental variables are structured environmental descriptions formed by uniformly encoding modal missingness states, noise types, noise intensities, and signal quality states. Based on this historical data, multi-environment disturbance input samples are constructed. By actively applying different combinations of modal missingness, noise types, noise intensities, and signal-to-noise ratio changes to the original multimodal monitoring data, various measurement environments that may occur in the industrial field are simulated to expand the environmental coverage of the training samples.

[0036] The training framework in this embodiment simultaneously constructs a high-quality reference branch and a multi-environment robust prediction branch. Both have the same network structure but differ in input data quality and training objectives. The high-quality reference branch receives relatively complete and high-quality historical multimodal monitoring data and its corresponding measurement environmental variables. After processing such as encoding fusion, causal-bias decoupling, and missing-noise debiasing, it generates reference causal projection features and outputs a first tool wear prediction result based on these features. The multi-environment robust prediction branch receives multi-environment perturbation input samples constructed through active perturbation and their corresponding historical measurement environmental variables, generating robust causal projection features corresponding to these input samples. After projection mapping, the robust causal projection features have the same dimension as the reference causal projection features. Finally, the multi-environment robust prediction branch outputs a second tool wear prediction result based on these robust causal projection features.

[0037] In terms of loss function design, this embodiment uses reference causal projection features as the knowledge source. Projective causal feature distillation is performed on the robust causal projection features generated by the multi-environment robust prediction branch. The debiased causal representation of the high-quality reference branch is transferred as stable knowledge to the multi-environment robust prediction branch, guiding it to achieve accurate tool wear prediction under conditions of strong modal missingness and low-quality data. This avoids the multi-environment robust prediction branch directly imitating fused features containing environmental shortcuts, thus preventing the mislearning of specific noise patterns as degenerate features. Furthermore, a cross-environment causal invariance loss is introduced to constrain the causal representation and prediction results of the same sample to remain consistent under different measurement environments, ensuring that degenerate semantics remain stable under cross-environment conditions. By training the high-quality reference branch and the multi-environment robust prediction branch using multiple loss functions, the multi-environment robust prediction branch ultimately exhibits accurate and robust tool wear prediction performance and good engineering applicability when facing various missing, noisy, and mixed environmental changes.

[0038] In one implementation, the steps of generating reference causal projection features based on the historical multimodal monitoring data and the historical measured environmental variables through a high-quality reference branch, and outputting a first tool wear prediction result based on the reference causal projection features, include: The historical multimodal monitoring data is encoded using encoders of multiple modes in the high-quality reference branch to obtain the modality coding features corresponding to the historical multimodal monitoring data. The multimodal fusion module in the high-quality reference branch performs multimodal fusion on the various modal coding features corresponding to the historical multimodal monitoring data to obtain reference fusion features. The causal-bias decoupling module in the high-quality reference branch is used to decouple the reference fusion features from causal to bias, thereby obtaining the causal representation and bias representation corresponding to the reference fusion features. Using the missing-noise debiasing module and the environmental prototype dictionary in the high-quality reference branch, a reference debiasing causal representation is generated based on the causal representation, bias representation, and historical measured environmental variables corresponding to the reference fusion features. The reference debiased causal representation is projected using the projection head in the high-quality reference branch to obtain the reference causal projection feature. The first tool wear prediction result is output based on the reference causal projection features through the prediction head in the high-quality reference branch.

[0039] Further, the step of outputting a second tool wear prediction result based on the robust causal projection features corresponding to the multi-environment perturbation input samples through the multi-environment robust prediction branch, wherein the robust causal projection features have the same dimension as the reference causal projection features, includes: The available modes in the multi-environmental perturbation input samples are encoded by the encoder of the mode in the multi-environmental robust prediction branch, and the missing modes are marked by a missing mask to obtain several mode encoding features corresponding to the multi-environmental perturbation input samples. The multimodal fusion module in the multi-environment robust prediction branch performs multimodal fusion on the modal coding features corresponding to the multi-environment perturbation input samples to obtain the fused features corresponding to the multi-environment perturbation input samples. The causal-bias decoupling module in the multi-environment robust prediction branch performs causal-bias decoupling on the fusion features corresponding to the multi-environment perturbation input samples to obtain the causal representation and bias representation corresponding to the multi-environment perturbation input samples. The robust debiased causal representation of the multi-environmental perturbation input sample is generated by using the missing-noise debiasing module and the environment prototype dictionary in the multi-environmental robust prediction branch, based on the causal representation, bias representation, and historical measured environmental variables corresponding to the multi-environmental perturbation input sample. The robust causal projection features corresponding to the multi-environment perturbation input samples are obtained by projecting the robust debiased causal representations corresponding to the multi-environment perturbation input samples through the projection head in the multi-environment robust prediction branch. The second tool wear prediction result is output through the prediction head in the multi-environment robust prediction branch, based on the robust causal projection features corresponding to the multi-environment disturbance input samples.

[0040] Specifically, multimodal monitoring data is collected during the CNC machining process, including triaxial vibration signals, triaxial cutting force signals, and acoustic emission signals. In this embodiment, the multimodal monitoring data used in the training phase is defined as historical multimodal monitoring data. It is understood that the elements contained in the complete multimodal monitoring data should be the same in both the inference and training phases.

[0041] For the The complete multimodal monitoring data for each sample is represented as follows: ; in, , , These represent the triaxial vibration signal, the triaxial cutting force signal, and the acoustic emission signal, respectively.

[0042] In addition to multimodal monitoring data, each sample also corresponds to a tool wear label. This is used to characterize the tool wear state during the CNC machining stage corresponding to the sample.

[0043] Modal missing state and noise and signal quality status Commonly defined as environmental variables for measurement : .

[0044] During the training phase, based on the modality missing state and noise and signal quality status Construct input samples with multiple environmental disturbances: ; in, This indicates missing mask and noise perturbation operations.

[0045] The multi-environment robust prediction branch is subjected to input coverage mode loss, noise type and intensity variation, signal-to-noise ratio decrease, signal drift, and environmental disturbances caused by a combination of factors.

[0046] The following steps are mainly performed during the high-quality reference representation construction phase: First, modal coding is performed: a high-quality reference branch is used to... Using the input as input, single-mode temporal features are extracted through a vibration encoder, a cutting force encoder, and an acoustic emission encoder, respectively, thus obtaining the modal coding features: ; Secondly, multimodal fusion and environmental debiasing are performed: various modal coding features are input into the multimodal fusion module of the high-quality reference branch, and a reference-debiased causal representation is obtained through causal-bias decoupling, environmental prototype matching, and prototype condition missing-noise debiasing. This reference-debiased causal representation eliminates strong modal shortcuts and environmental biases that may exist in the high-quality reference environment, and can be used for wear prediction, serving as a stable knowledge source for cross-environment causal feature distillation in multi-environment robust prediction branches.

[0047] ; in, Indicates reference fusion features; This indicates the multimodal fusion operation performed by the multimodal fusion module in the high-quality reference branch; This indicates a biased causal representation based on reference. This indicates the debiasing operation performed by the missing-noise debiasing module.

[0048] Next, wear prediction is performed: the high-quality reference branch is characterized based on the reference bias causal characterization. Output tool wear prediction value and wear regression loss Constraining prediction accuracy under high-quality reference environments, Mean squared error loss can be used.

[0049] In the multi-environment robust prediction branch modeling stage, the following steps are mainly performed: First, handle multi-environmental perturbation inputs: the multi-environment robust prediction branch... As input, the currently available modes are encoded, and the missing mask, noise type, noise intensity, and signal quality status are combined. For missing modes, a missing mask is used. Labeling is performed to prevent the model from mistaking zero values ​​for actual acquired signals.

[0050] Secondly, modality coding and fusion are performed: the coding and fusion process of the multi-environment robust prediction branch is represented as follows: ; in, This represents the multimodal fusion features obtained by the multi-environment robust prediction branch under various environmental perturbation conditions.

[0051] Next, causal-bias decoupling is performed: the causal-bias decoupling module will refer to the fused features. and fusion features It can be decomposed into two types of representations: one is the causal representation related to degradation. One type is used to represent information related to tool wear degradation stability; the other type is a bias characterization that is not related to degradation. , is used to represent non-degradation-related information caused by modal loss, noise pollution, signal quality degradation, and modal shortcuts.

[0052] For reference fusion features and fusion features Any one of the fusion features The decoupling process is represented as: ; in, and Decoupling can be achieved by using a two-branch multilayer perceptron network.

[0053] Subsequently, an environment prototype dictionary and prototype condition missing-noise debiasing are employed: the environment prototype dictionary-based missing-noise debiasing module is used to characterize the bias of the current sample. Modal missing state and noise and signal quality status , causal characterization After correction, a biased causal representation is obtained. Specifically, this module first, based on , and Estimate the environmental offset of the current sample : ; in, This represents the bias estimation network.

[0054] Then, a gating correction method is used to obtain a biased causal representation: ; in, This represents the correction coefficients generated by the gating network. This indicates element-wise multiplication.

[0055] The missing-noise debiasing module can reduce the characterization bias caused by strong mode missing, noise contamination and signal quality degradation, so that subsequent wear prediction mainly relies on stable degradation information, rather than on pseudo-related information in the measurement environment.

[0056] Typical environmental prototypes such as complete modalities, strong modalities missing, Gaussian noise, impact noise, and drift noise can be constructed in advance based on the bias representation of the training samples. Then, attention weights are calculated based on the distance between the current sample's bias representation and each environmental prototype, and an environmental bias vector is obtained to correct the causal representation.

[0057] Finally, cross-environment causal representation projection and wear prediction are performed: the reference debiased causal representation is sequentially input into the projection head (which can be defined as the reference projection head) and prediction head of the high-quality reference branch to obtain the first tool wear prediction result; and the robust debiased causal representation is sequentially input into the projection head (which can be defined as the robust projection head) and prediction head of the multi-environment robust prediction branch to obtain the second tool wear prediction result. After projection mapping, reference causal projection features and robust causal projection features of the same dimension can be obtained, providing a data foundation for subsequent model training.

[0058] In one implementation, the steps of distilling the robust causal projection features corresponding to the multi-environmental perturbation input samples using the reference causal projection features as a knowledge source, and training the multi-environmental robust prediction branch in conjunction with multiple loss functions include: Wear regression loss is calculated based on the tool wear label and the first tool wear prediction result; orthogonal constraint loss is calculated based on the causal characterization and bias characterization corresponding to the high-quality reference branch; bias prediction loss is calculated based on the bias characterization corresponding to the high-quality reference branch to predict the mode missing state, noise and signal quality state, and combined with the historical measurement environment variables; cross-environment invariance loss is calculated based on the causal characterization calculated by the high-quality reference branch based on the same sample under different measurement environments and the corresponding first tool wear prediction result. The high-quality reference branch is trained using a multi-objective joint optimization method, based on the wear regression loss, orthogonal constraint loss, bias prediction loss, and cross-environment invariance loss corresponding to the high-quality reference branch. Wear regression loss is calculated based on the tool wear label and the second tool wear prediction result; orthogonal constraint loss is calculated based on the causal characterization and bias characterization corresponding to the multi-environment robust prediction branch; bias prediction loss is calculated based on the bias characterization corresponding to the multi-environment robust prediction branch to predict mode missing state, noise and signal quality state, and combined with the historical measurement environmental variables; cross-environment invariance loss is calculated based on the causal characterization calculated by the multi-environment robust prediction branch based on the same sample under different measurement environments and the corresponding second tool wear prediction result; distillation loss is calculated based on the reference causal projection feature and the robust causal projection feature corresponding to the multi-environment perturbation input sample. The multi-environment robust prediction branch is trained using a multi-objective joint optimization approach, based on the wear regression loss, orthogonal constraint loss, bias prediction loss, cross-environment invariance loss, and distillation loss corresponding to the multi-environment robust prediction branch.

[0059] Furthermore, the distillation loss is the Kullback-Leibler divergence.

[0060] Specifically, during the training phase, the high-quality reference branch and the multi-environment robust prediction branch are jointly optimized through their respective multi-objective loss functions to achieve stable wear prediction capability under high-quality data and robust generalization capability under multi-environment perturbations, respectively.

[0061] The high-quality reference branch takes relatively complete and high-quality multimodal monitoring data as input. Its training objective is to establish accurate tool wear prediction capabilities while providing a semantically stable and bias-free causal representation knowledge source for subsequent knowledge distillation. The training of this branch involves the joint optimization of four loss terms: Wear regression loss: Calculated based on the first tool wear prediction result and tool wear label output by the high-quality reference branch, used to constrain the accuracy of the prediction value, so that the causal characterization can fully capture the essential information related to tool wear degradation.

[0062] Orthogonal constraint loss: calculated based on the causal and bias representations within the high-quality reference branch. By applying orthogonality constraints, the causal and bias representations are made independent or directionally separated in the representation space, thereby achieving effective decoupling of wear and degradation information from environmental disturbance information and avoiding the mixing of spurious environmental information in the causal representation.

[0063] Bias prediction loss: Based on the bias representation of the high-quality reference branch, this loss predicts environmental factors such as modal loss state, noise type, and signal quality state experienced by the current sample, and calculates the loss by comparing the prediction results with historical measurement environmental variables. This loss endows the bias representation with a clear environmental semantics, enabling it to undertake the task of environmental identification and ensuring that the decoupled bias representation truly encodes the changing information of the measurement environment.

[0064] Cross-environment invariance loss: For multiple perturbation versions of the same sample constructed under different measurement environments, each version is computed with its own causal representation and corresponding first tool wear prediction result through a high-quality reference branch, constraining these causal representations and prediction results to remain consistent under different environmental conditions. This loss, at the semantic level, forces samples in the same degradation state to have a cross-environment stable representation in the causal space, laying the foundation for the generalization ability of subsequent multi-environment robust prediction branches.

[0065] Through multi-objective joint optimization of the wear regression loss, orthogonal constraint loss, bias prediction loss and cross-environment invariance loss, the high-quality reference branch can achieve effective decoupling of causality and bias and semantic stability of cross-environment causal representation while maintaining prediction accuracy.

[0066] The multi-environment robust prediction branch takes actively perturbed multi-environment perturbation input samples as input. Its training objective is to achieve tool wear prediction capability consistent with the high-quality reference branch under conditions of strong modality loss and low-quality data. The training of this branch introduces an additional distillation loss on top of the four losses of the high-quality reference branch, forming a joint optimization of five losses.

[0067] Wear regression loss: Calculated based on the second tool wear prediction result output by the multi-environment robust prediction branch and the tool wear label. The constraint model can still accurately predict the wear value under disturbance conditions.

[0068] Orthogonal constraint loss: calculated based on the causal and bias representations within the multi-environment robust prediction branch, ensuring that the branch can also separate wear and degradation information from environmental disturbance information in the representation space, preventing environmental pseudo-correlation from interfering with the learning of causal representation.

[0069] Bias prediction loss: Based on the bias representation of the multi-environment robust prediction branch, the modal missing state, noise type and signal quality state of the current perturbation sample are predicted and calculated by comparing with historical measured environmental variables, so that the bias representation can maintain the ability to identify environmental factors.

[0070] Cross-environment invariance loss: For multiple perturbation versions of the same sample constructed under different measurement environments, causal representations and corresponding second tool wear prediction results are calculated through multi-environment robust prediction branches, constraining these causal representations and prediction results to remain consistent across environments. The weighted combination of environment prototypes can also describe mixed environments that did not appear alone during the training phase, thereby improving the model's adaptability to new missing combinations, new noise intensities, and signal quality fluctuations.

[0071] Distillation loss: Calculated based on the reference causal projection features generated by the high-quality reference branch and the robust causal projection features generated by the multi-environment robust prediction branch. This loss uses the debiased causal representation of the high-quality reference branch as a stable knowledge source to perform projection causal feature distillation on the multi-environment robust prediction branch. This guides the latter to learn cross-environment stable degenerate knowledge under low-quality input conditions, rather than directly imitating fused features containing environmental shortcuts, thereby avoiding mislearning specific noisy patterns or strong modal shortcuts as degenerate features.

[0072] Through multi-objective joint optimization of the five losses mentioned above, the multi-environment robust prediction branch is fully guided during the training phase to decouple causal and bias information, identify environmental states, maintain cross-environment consistency, and absorb high-quality debiased causal knowledge. After training, only this multi-environment robust prediction branch is deployed during the inference phase, which can complete environment-robust tool wear prediction based solely on the currently available modalities.

[0073] For example, regarding bias prediction supervision: to incorporate missing, noise, and signal quality-related information into the bias representation. Set the bias prediction head, based on Predicting missing modal states and noise and signal quality status The bias prediction loss is expressed as: ; in, and This represents the cross-entropy loss.

[0074] For orthogonality constraints: compute causal representations within a batch of samples. With bias representation The correlation between them, and the loss through orthogonal constraint. Reduce the statistical correlation between the two types of representations to separate wear and degradation information from missing noise bias information as much as possible.

[0075] For cross-environment causal invariance constraints and cross-environment causal knowledge transfer: the projection head is used to align the dimensions and scales of cross-environment stable degenerate knowledge, and the environment bias representation does not enter the distillation target. After completing the environment-debiased training of the high-quality reference branch, its parameters are frozen. The high-quality reference branch extracts only the debiased causal projection features under the high-quality reference environment, which serve as the supervision source for the multi-environment robust prediction branch to learn cross-environment stable degradation knowledge. When training the multi-environment robust prediction branch, the reference causal projection features are used as the knowledge source, and the robust causal projection features are distilled to make the multi-environment robust prediction branch approximate the stable degradation semantics provided by the reference under different measurement environments, rather than imitating the environment-related biases of the reference.

[0076] Distillation losses can be accounted for using Kullback-Leibler divergence: ; in, This represents the distillation temperature coefficient.

[0077] like Figure 2 As shown, the model is trained using a multi-objective joint optimization approach. The complete training loss includes wear regression loss, causal feature distillation loss, bias prediction loss, orthogonal constraint loss, and cross-environment causal invariance loss. ; in, Used to monitor tool wear prediction results Used to constrain robust projection features to approximate reference projection features. Used to constrain bias characterization to identify modal missing states and noisy states. Used to reduce the correlation between causal representations and bias representations This is used to ensure that the causal characterization and prediction results of the same sample remain consistent under different measurement environments.

[0078] The high-quality reference branch training phase mainly optimizes , , and The training phase of the multi-environment robust prediction branch further incorporates the above-mentioned loss method. This allows for the reception of causal feature-level knowledge guidance from high-quality reference branches.

[0079] In summary, the technical advantages of this application include, but are not limited to: (1) Unified environment modeling: Modal missing mask, noise type, noise intensity and signal quality status are uniformly encoded as measurement environment variables, and a single structured description is used to cover any combination of strong modal missing, single noise, mixed noise and different signal-to-noise ratios. Multiple data quality degradation scenarios are handled simultaneously within a unified framework.

[0080] (2) Causal-bias decoupling to separate wear degradation information from environmental disturbance information: The causal-bias decoupling module decomposes the multimodal fusion features into degradation-related causal representations and environmental-related bias representations, and uses bias state supervision and orthogonal constraints to effectively separate the two in the representation space. The causal representation undertakes the task of tool wear prediction and retains stable degradation semantics; the bias representation undertakes the task of environmental recognition and encodes information on changes in measurement conditions, enabling the model to distinguish between real wear evolution and measurement disturbances.

[0081] (3) Environmental prototype dictionary and prototype condition bias correction to correct representation bias caused by multi-factor environmental disturbances: An environmental prototype dictionary covering complete modalities, strong modal missing, and different noise types is established. Attention weights are obtained by calculating the distance between the bias representation and each environmental prototype, generating an environmental bias vector, and gating correction is applied to the causal representation. The weighted combination of environmental prototypes can describe mixed environments that did not appear alone during the training phase, thereby improving the model's generalization ability to new missing combinations, new noise intensities, and signal quality fluctuations. The environmental attention weights can also explicitly explain the main environmental influences on the current sample, taking into account the interpretability of the model.

[0082] (4) Cross-environmental causal invariance constraint to ensure cross-environmental stability of degenerative semantics: constrain the causal representation and prediction results of the same degenerative sample to remain consistent under different measurement environments, and force the model to learn cross-environmental invariant tool wear degradation knowledge from the semantic level, rather than relying on strong modal shortcuts of specific modes or pseudo-correlation features related to the environment, thereby reducing the representational bias when the environment changes.

[0083] (5) Causal feature level knowledge distillation to transfer debiased stable knowledge: Using the debiased causal representation of the high-quality reference branch as the knowledge source, the projected causal feature distillation of the multi-environment robust prediction branch is performed to guide the representation learning under low-quality data conditions. This avoids the multi-environment robust prediction branch directly copying the strong mode shortcut or specific noise mode in the high-quality reference branch, thereby achieving accurate prediction under strong mode missing and low-quality data conditions.

[0084] (6) The inference phase relies only on currently available modalities, adapting to low-cost deployment scenarios: The inference phase only deploys multi-environment robust prediction branches, without requiring continuous acquisition of complete modal data, making it suitable for low-cost sensor configurations or situations where some sensors fail. After inputting multi-modal data with missing modalities, noise pollution, or signal quality degradation, the tool wear prediction value is directly output, meeting the industrial site's requirements for real-time performance and ease of deployment.

[0085] (7) Enhance generalization ability and engineering applicability in unseen environments: The cross-environmental causal invariance constraint combined with the weighted combination mechanism of the environmental prototype dictionary enables the model to handle missing combinations, mixed noise types and signal quality fluctuations that did not appear alone during the training phase, maintain stable prediction performance in complex and ever-changing industrial measurement environments, and have high engineering applicability and robustness.

[0086] Based on the above embodiments, this application also provides a terminal, the principle block diagram of which can be as follows: Figure 3 As shown, the terminal includes a processor, memory, network interface, and display screen connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a tool wear prediction method based on environmentally robust causal characterization. The display screen can be an LCD screen or an e-ink screen.

[0087] Those skilled in the art will understand that Figure 3 The block diagram shown is only a partial structural diagram related to the solution of this application and does not constitute a limitation on the terminal on which the solution of this application is applied. The specific terminal may include more or fewer components than shown in the figure, or combine some components, or have different component arrangements.

[0088] In one implementation, the terminal's memory stores one or more programs, and these programs are configured to be executed by one or more processors, and the programs contain instructions for performing a tool wear prediction method based on environmentally robust causal characterization.

[0089] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0090] In summary, this application discloses a tool wear prediction method, terminal, and storage medium based on robust environmental causal characterization. During the inference phase, this application deploys only a multi-environment robust prediction branch, requiring the collection of multi-modal monitoring data during the machining process, but not relying on complete modal data. It uniformly encodes missing modal states, noise, and signal quality states as measured environmental variables to achieve unified environmental modeling. After modal encoding and fusion, the multi-modal fusion features are separated into causal characterization and bias characterization, effectively separating wear degradation information from environmental disturbance information in the characterization space. Based on an environmental prototype dictionary and a missing-noise debiasing module, a debiased causal characterization is generated and corrected to obtain a robust debiased causal characterization. After projecting the robust debiased causal characterization, the tool wear prediction result is obtained. This application is adaptable to low-cost sensor configurations and scenarios with partial sensor failure, achieving robust and accurate tool wear prediction.

[0091] It should be understood that the application of this application is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for predicting tool wear based on environmentally robust causal characterization, characterized in that, The method includes: During the inference phase, multimodal monitoring data and corresponding measurement environmental variables are acquired during the CNC machining process; the measurement environmental variables include modal missing state, noise and signal quality state. Robust causal projection features are generated based on the multimodal monitoring data and the measured environmental variables through a trained multi-environment robust prediction branch, and tool wear prediction results are output based on the robust causal projection features. The structure of the multi-environment robust prediction branch includes: A multi-modal encoder is used to encode the available modalities in the multi-modal monitoring data and to mark the missing modalities with a missing mask, thereby obtaining several modal coding features; A multimodal fusion module is used to perform multimodal fusion on the modal coding features to obtain fused features; The causal-bias decoupling module is used to perform causal-bias decoupling on the fused features to obtain the causal representation and bias representation corresponding to the multimodal monitoring data; The missing-noise debiasing module and the environmental prototype dictionary are used to generate a robust debiased causal representation of the multimodal monitoring data based on the causal representation, bias representation, and the measured environmental variables corresponding to the multimodal monitoring data. A projection head is used to project the robust debiased causal representation corresponding to the multimodal monitoring data to obtain the robust causal projection feature corresponding to the multimodal monitoring data. The prediction head is used to output tool wear prediction results based on the robust causal projection features corresponding to the multimodal monitoring data.

2. The tool wear prediction method based on environmental robust causal characterization according to claim 1, characterized in that, The multimodal monitoring data includes: triaxial cutting force signal, triaxial vibration signal, and acoustic emission signal.

3. The tool wear prediction method based on environmental robust causal characterization according to claim 1, characterized in that, The missing-noise debiasing module and the environmental prototype dictionary are specifically used for: Based on the bias representation corresponding to the multimodal monitoring data and the measured environmental variables, the causal representation corresponding to the multimodal monitoring data is corrected to obtain the debiased causal representation corresponding to the multimodal monitoring data. Calculate the distance between the bias representation corresponding to the multimodal monitoring data and each environmental prototype in the environmental prototype dictionary, and calculate the attention weight based on the distance value of each environmental prototype; An environment bias vector is generated based on the attention weights; the debiased causal representation corresponding to the multimodal monitoring data is corrected according to the environment bias vector to obtain a robust debiased causal representation corresponding to the multimodal monitoring data.

4. The tool wear prediction method based on environmental robust causal characterization according to claim 1, characterized in that, The training method for the multi-environment robust prediction branch includes: During the training phase, historical multimodal monitoring data, corresponding historical measurement environmental variables, and corresponding tool wear labels are acquired during the CNC machining process. Construct multi-environmental disturbance input samples based on the historical multimodal monitoring data and the historical measured environmental variables; A high-quality reference branch is used to generate reference causal projection features based on the historical multimodal monitoring data and the historical measured environmental variables, and the first tool wear prediction result is output based on the reference causal projection features; the structure of the high-quality reference branch is the same as the structure of the multi-environment robust prediction branch. Through the multi-environment robust prediction branch, a second tool wear prediction result is output based on the robust causal projection features corresponding to the multi-environment disturbance input samples. The robust causal projection features have the same dimension as the reference causal projection features. The robust causal projection features corresponding to the multi-environment perturbation input samples are distilled using the reference causal projection features as a knowledge source, and the multi-environment robust prediction branch is trained by combining multiple loss functions.

5. The tool wear prediction method based on environmental robust causal characterization according to claim 4, characterized in that, The steps of generating reference causal projection features based on the historical multimodal monitoring data and the historical measured environmental variables through a high-quality reference branch, and outputting the first tool wear prediction result based on the reference causal projection features, include: The historical multimodal monitoring data is encoded using encoders of multiple modes in the high-quality reference branch to obtain the modality coding features corresponding to the historical multimodal monitoring data. The multimodal fusion module in the high-quality reference branch performs multimodal fusion on the various modal coding features corresponding to the historical multimodal monitoring data to obtain reference fusion features. The causal-bias decoupling module in the high-quality reference branch is used to decouple the reference fusion features from causal to bias, thereby obtaining the causal representation and bias representation corresponding to the reference fusion features. Using the missing-noise debiasing module and the environmental prototype dictionary in the high-quality reference branch, a reference debiasing causal representation is generated based on the causal representation, bias representation, and historical measured environmental variables corresponding to the reference fusion features. The reference debiased causal representation is projected using the projection head in the high-quality reference branch to obtain the reference causal projection feature. The first tool wear prediction result is output based on the reference causal projection features through the prediction head in the high-quality reference branch.

6. The tool wear prediction method based on environmental robust causal characterization according to claim 4, characterized in that, The step of outputting the second tool wear prediction result based on the robust causal projection features corresponding to the multi-environment perturbation input samples through the multi-environment robust prediction branch includes: The available modes in the multi-environmental perturbation input samples are encoded by the encoder of the mode in the multi-environmental robust prediction branch, and the missing modes are marked by a missing mask to obtain several mode encoding features corresponding to the multi-environmental perturbation input samples. The multimodal fusion module in the multi-environment robust prediction branch performs multimodal fusion on the modal coding features corresponding to the multi-environment perturbation input samples to obtain the fused features corresponding to the multi-environment perturbation input samples. The causal-bias decoupling module in the multi-environment robust prediction branch performs causal-bias decoupling on the fusion features corresponding to the multi-environment perturbation input samples to obtain the causal representation and bias representation corresponding to the multi-environment perturbation input samples. The robust debiased causal representation of the multi-environmental perturbation input sample is generated by using the missing-noise debiasing module and the environment prototype dictionary in the multi-environmental robust prediction branch, based on the causal representation, bias representation, and historical measured environmental variables corresponding to the multi-environmental perturbation input sample. The robust causal projection features corresponding to the multi-environment perturbation input samples are obtained by projecting the robust debiased causal representations corresponding to the multi-environment perturbation input samples through the projection head in the multi-environment robust prediction branch. The second tool wear prediction result is output through the prediction head in the multi-environment robust prediction branch, based on the robust causal projection features corresponding to the multi-environment disturbance input samples.

7. The tool wear prediction method based on environmental robust causal characterization according to claim 4, characterized in that, The steps of distilling the robust causal projection features corresponding to the multi-environment perturbation input samples using the reference causal projection features as a knowledge source, and training the multi-environment robust prediction branch in conjunction with multiple loss functions include: Wear regression loss is calculated based on the tool wear label and the first tool wear prediction result; orthogonal constraint loss is calculated based on the causal characterization and bias characterization corresponding to the high-quality reference branch; bias prediction loss is calculated based on the bias characterization corresponding to the high-quality reference branch to predict the mode missing state, noise and signal quality state, and combined with the historical measurement environment variables; cross-environment invariance loss is calculated based on the causal characterization calculated by the high-quality reference branch based on the same sample under different measurement environments and the corresponding first tool wear prediction result. The high-quality reference branch is trained using a multi-objective joint optimization method, based on the wear regression loss, orthogonal constraint loss, bias prediction loss, and cross-environment invariance loss corresponding to the high-quality reference branch. Wear regression loss is calculated based on the tool wear label and the second tool wear prediction result; orthogonal constraint loss is calculated based on the causal characterization and bias characterization corresponding to the multi-environment robust prediction branch; bias prediction loss is calculated based on the bias characterization corresponding to the multi-environment robust prediction branch to predict mode missing state, noise and signal quality state, and combined with the historical measurement environmental variables; cross-environment invariance loss is calculated based on the causal characterization calculated by the multi-environment robust prediction branch based on the same sample under different measurement environments and the corresponding second tool wear prediction result; distillation loss is calculated based on the reference causal projection feature and the robust causal projection feature corresponding to the multi-environment perturbation input sample. The multi-environment robust prediction branch is trained using a multi-objective joint optimization approach, based on the wear regression loss, orthogonal constraint loss, bias prediction loss, cross-environment invariance loss, and distillation loss corresponding to the multi-environment robust prediction branch.

8. The tool wear prediction method based on environmental robust causal characterization according to claim 7, characterized in that, The distillation loss is the Kullback-Leibler divergence.

9. A terminal, characterized in that, The terminal includes a memory and one or more processors; the memory stores one or more programs; the programs contain instructions for executing the tool wear prediction method based on environmental robust causal characterization as described in any one of claims 1 to 8; the processors are used to execute the programs.

10. A computer-readable storage medium storing a plurality of instructions thereon, characterized in that, The instructions are applicable to be loaded and executed by a processor to implement the steps of the tool wear prediction method based on environmentally robust causal characterization as described in any one of claims 1 to 8.

Images