Methods, systems, terminals, and storage media for predicting remaining tool life

Abstract

This application discloses a method, system, terminal, and storage medium for predicting the remaining tool life, relating to the field of CNC machining technology. The method includes: acquiring multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted, wherein the multimodal sensing signals include cutting force signals, vibration signals, and acoustic emission signals; inputting the multimodal sensing signals and operating parameters into a trained end-to-end prediction model; extracting features from the multimodal sensing signals using the end-to-end prediction model to obtain the spatiotemporal features of each mode; obtaining a latent space vector of the operating parameters; decoupling the spatiotemporal features based on the latent space vector to obtain decoupling features decoupled from the operating conditions of the cutting tool to be predicted; and performing regression prediction based on the decoupling features to output the predicted remaining tool life value corresponding to the cutting tool to be predicted. This improves the accuracy of predicting the remaining tool life.

Description

Technical Field

[0001] This application relates to the field of CNC machining technology, and in particular to a method, system, terminal and storage medium for predicting the remaining service life of a cutting tool. Background Technology

[0002] In industrial automated production, cutting tools are the core actuators in the cutting process, and their condition directly determines machining accuracy, production efficiency, and even equipment safety. Therefore, accurately predicting the remaining service life of cutting tools is of paramount importance for achieving predictive maintenance, ensuring production continuity, and reducing operating costs.

[0003] In existing technologies, the remaining tool life is typically predicted using a model built on a single-condition modeling approach. This approach ensures data consistency by fixing cutting parameters (such as spindle speed, feed rate, and depth of cut), collecting sensor data throughout the tool's lifecycle, and establishing the remaining tool life prediction model. However, prediction models built on this single-condition modeling approach are only suitable for situations with a single task and stable environment. They cannot control the impact of changes in operating conditions on the prediction process. In actual production, however, actual operating conditions may change, affecting the accuracy of the prediction.

[0004] Therefore, the relevant technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this application is to provide a method, system, terminal, and storage medium for predicting the remaining service life of cutting tools. This aims to solve the technical problem that prediction models in related technologies are only applicable to situations with a single task and a stable environment, and cannot control the impact of changes in working conditions on the prediction process. This is not conducive to improving the accuracy of predictions when actual working conditions may change.

[0006] To achieve the above objectives, the first aspect of this application provides a method for predicting the remaining service life of a cutting tool, wherein the method includes: The multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted are obtained, wherein the multimodal sensing signals include cutting force signals, vibration signals and acoustic emission signals; The multimodal sensing signals and the operating parameters are input into a trained end-to-end prediction model. The end-to-end prediction model extracts features from the multimodal sensing signals to obtain the spatiotemporal features of each mode. The operating parameters are mapped to obtain the latent space vector of the operating conditions. The spatiotemporal features are decoupled based on the latent space vector of the operating conditions to obtain decoupling features that are decoupled from the operating conditions of the cutting tool to be predicted. Regression prediction is performed based on the decoupling features to output the predicted remaining tool life value of the cutting tool to be predicted.

[0007] Optionally, the cutting force signal is a three-channel cutting force signal, the vibration signal is a four-channel vibration signal, and the acoustic emission signal is a two-channel acoustic emission signal. The above operating parameters include at least one of cutting power, spindle speed, feed rate, and depth of cut.

[0008] Optionally, the above-mentioned end-to-end prediction model is used to extract features from the above-mentioned multimodal sensing signals to obtain the spatiotemporal features of each modality, including: Each mode of the multimodal sensing signal is taken as the signal to be processed, and the local temporal dynamic features corresponding to the signal to be processed are extracted by the multi-scale feature extraction network in the end-to-end prediction model. Calculate the statistical characteristics corresponding to the above-mentioned signal to be processed, wherein the above-mentioned statistical characteristics include the peak factor and / or root mean square. The aforementioned local temporal dynamic features and statistical features are aggregated and spliced ​​together along the channel dimension to obtain spliced ​​features. Based on the splicing characteristics described above, the spatiotemporal features corresponding to the signal to be processed are obtained through the long short-term memory network in the end-to-end prediction model.

[0009] Optionally, obtaining the latent space vector of the operating condition based on the above-mentioned operating condition parameter mapping includes: The above working condition parameters are input into the working condition encoder in the above end-to-end prediction model, and the above working condition latent space vector is obtained by mapping through the above working condition encoder. The aforementioned operating condition encoder is a multilayer sensor network; the aforementioned operating condition parameters include at least one of spindle speed, feed rate, and depth of cut.

[0010] Optionally, the above-mentioned decoupling processing of the spatiotemporal features based on the above-mentioned latent space vector of the working condition yields decoupling features that are decoupled from the working condition of the cutting tool to be predicted, including: Based on the above working condition latent space vector, the scaling factor and translation factor are obtained by mapping through the linear mapping fully connected network in the above end-to-end prediction model. For each of the above-mentioned spatiotemporal features, the mean and standard deviation corresponding to the above-mentioned spatiotemporal features are calculated in the spatial dimension, so as to standardize the above-mentioned spatiotemporal features based on the above-mentioned mean and standard deviation. Based on the scaling factor and the translation factor mentioned above, an affine transformation is performed on the standardized spatiotemporal features to obtain the decoupled features mentioned above.

[0011] Optionally, the regression prediction based on the decoupling features to output the predicted remaining tool life of the cutting tool to be predicted includes: For each of the above decoupling features, global average pooling is performed respectively, and the reliability confidence level corresponding to the above decoupling features is determined based on the features after global average pooling. Based on the aforementioned reliability confidence level, the aforementioned decoupling features are weighted to obtain the weighted features of each mode; The multi-head self-attention module in the above end-to-end prediction model is used to fuse the weighted features of each modality to obtain multi-modal fusion features. The multimodal fusion features are input into the regression module of the end-to-end prediction model to obtain the tool remaining life prediction value. The regression module is composed of a multi-layer nonlinear fully connected network.

[0012] Optionally, the above end-to-end prediction model is trained according to the following steps: Obtain a training dataset, wherein the training dataset is constructed based on at least one sample tool. For each sample tool, there are multiple training sample groups corresponding to different time steps in the training dataset. A training sample group includes a set of corresponding multimodal signal samples, working condition parameter samples, tool remaining life labels and working condition labels. Input the multimodal signal samples and working condition parameter samples from the above training dataset into the end-to-end prediction model to be trained, and obtain the multimodal fusion features and tool remaining life prediction values ​​obtained by the above end-to-end prediction model. Based on the aforementioned multimodal fusion features, the predicted working condition classification result is determined by a domain discriminator. Based on the above predicted working condition classification results and the above working condition labels, the domain alignment loss is determined; based on the above predicted tool remaining life value and the above predicted tool remaining life label, the regression loss is determined; based on the predicted tool remaining life value and the tool remaining life label corresponding to the training sample groups of the same sample tool at different time steps, the monotonic trend loss is determined. Among these, the above domain alignment loss is used to constrain the above multimodal fusion features to be independent of the working condition category; the above regression loss is used to constrain the consistency between the above predicted tool remaining life value and the above predicted tool remaining life label; and the above monotonic trend loss is used to constrain the change direction of the above predicted tool remaining life value over time to be consistent with the change direction of the above predicted tool remaining life label over time. Based on the aforementioned domain alignment loss, regression loss, and monotonic trend loss, the model parameters of the end-to-end prediction model are adjusted. Then, the process returns to the previous steps of inputting the multimodal signal samples and working condition parameter samples from the training dataset into the end-to-end prediction model to be trained, and obtaining the multimodal fusion features and tool remaining life prediction values ​​obtained by the end-to-end prediction model, until the preset training termination condition is met, and the trained end-to-end prediction model is obtained.

[0013] A second aspect of this application provides a tool remaining service life prediction system, wherein the system includes: The data acquisition module is used to acquire the multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted, wherein the multimodal sensing signals include cutting force signals, vibration signals and acoustic emission signals; The prediction module is used to input the multimodal sensing signals and the operating parameters into a trained end-to-end prediction model, extract features from the multimodal sensing signals through the end-to-end prediction model to obtain the spatiotemporal features of each mode, obtain the operating condition latent space vector based on the operating parameters, decouple the spatiotemporal features based on the operating condition latent space vector to obtain decoupling features that are decoupled from the operating conditions of the cutting tool to be predicted, and perform regression prediction based on the decoupling features to output the predicted remaining tool life value corresponding to the cutting tool to be predicted.

[0014] A third aspect of this application provides a terminal, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of any of the above-described tool remaining service life prediction methods.

[0015] A fourth aspect of this application provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the above-described methods for predicting the remaining tool life.

[0016] As can be seen from the above, the present application provides a method for predicting the remaining service life of a cutting tool. The method includes: acquiring multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted, wherein the multimodal sensing signals include cutting force signals, vibration signals, and acoustic emission signals; inputting the multimodal sensing signals and operating parameters into a trained end-to-end prediction model; extracting features from the multimodal sensing signals through the end-to-end prediction model to obtain the spatiotemporal features of each mode; obtaining a latent space vector of the operating parameters; decoupling the spatiotemporal features based on the latent space vector to obtain decoupling features decoupled from the operating conditions of the cutting tool to be predicted; and performing regression prediction based on the decoupling features to output the predicted remaining service life value of the cutting tool to be predicted.

[0017] Thus, when predicting the remaining tool life, multimodal sensor signals and operating parameters are collected separately to perform predictions based on a trained end-to-end prediction model. During the prediction process, the trained end-to-end prediction model first extracts features from the multimodal sensor signals to obtain the spatiotemporal features of each mode. Then, it maps these features to the operating parameters to obtain a latent space vector. This latent space vector is then used to decouple the spatiotemporal features from the operating conditions of the cutting tool to be predicted, resulting in decoupled features. Regression prediction is then performed based on these decoupled features to output the predicted remaining tool life value for the cutting tool. In this way, decoupling features from operating conditions during the prediction process reduces the impact of changes in operating conditions on the prediction process, thereby improving the accuracy of the remaining tool life prediction. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, 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 of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for predicting the remaining service life of a cutting tool provided in an embodiment of this application; Figure 2 This is a schematic diagram of an end-to-end prediction model training process provided in an embodiment of this application; Figure 3 This is a schematic diagram of the constituent modules of a tool remaining service life prediction system provided in an embodiment of this application; Figure 4 This is a block diagram illustrating the internal structure of a terminal provided in an embodiment of this application. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0022] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to classification to [the described condition or event]."

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0026] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0027] The core of tool remaining life prediction lies in using historical and current cutting operation data to quantitatively assess the wear evolution trend of the tool over a future period, and then estimate the remaining effective service time of the tool before it can no longer continue normal cutting operations. For life prediction under variable operating conditions and multi-source sensing conditions, existing methods often handle this by restricting operating variables and static fusion, but their key bottleneck lies in the difficulty of dealing with the continuous uncertainty caused by the drift of multimodal signal quality.

[0028] Common strategies include single-condition modeling and modal feature cascade fusion. Single-condition modeling ensures data distribution consistency by fixing cutting parameters (such as spindle speed and feed rate), suitable for situations with simple tasks and stable environments. However, in actual production, frequent changes in operating conditions can be compounded by sensor link fluctuations and on-site interference, causing significant quality drift phenomena in multimodal signals, such as changes in signal-to-noise ratio, altered noise structure, channel distortion, and reliability degradation. This disrupts feature continuity and significantly weakens the model's generalization performance under new operating conditions. Therefore, prediction models built based on single-condition modeling are only suitable for situations with simple tasks and stable environments, and cannot control the impact of operating condition changes on the prediction process. In actual production, actual operating conditions may change, affecting the accuracy of predictions.

[0029] Modal feature cascade fusion method combines multi-source data such as cutting force, vibration, and acoustic emission by simply splicing or using a fixed fusion rate. While this can enrich the input information to some extent, it usually implies a static assumption that "the quality of each modal signal is constant and reliable". It is difficult to adapt to the dynamic drift of quality and the time-varying characteristics of reliability of multi-modal signals throughout their entire life cycle (for example, acoustic emission is more sensitive to early micro wear, while cutting force is more significant to later wear). This can induce fusion deviation when quality fluctuates or interference increases.

[0030] Furthermore, deep learning methods have been used for cross-condition adaptation and feature fusion, including unsupervised domain adaptation and attention-based weighted fusion. While these methods can align the distributional differences across different conditions and capture cross-parameter correlations, in unknown and complex transient conditions, the model may still struggle to fully decouple environmental noise from real wear information, leading to alignment bias. Simultaneously, complex field-collected multimodal data often exhibits significant quality drift and dynamic imbalance issues, with the reliability of different modes continuously changing with operating conditions, wear stages, and interference intensity. Since conventional attention mechanisms typically lack explicit and interpretable evaluation of underlying signal quality and signal-to-noise ratio (i.e., reliability assessment is lacking), the network is easily misled by high-amplitude, high-frequency noise "inferior modes" in quality drift scenarios, leading to the allocation of unreasonable fusion weights. This results in inferior mode dominance and generalization performance collapse, making it difficult to achieve stable and physically consistent tool life prediction throughout the entire lifecycle under varying operating conditions.

[0031] Specifically, existing single-condition modeling methods are prone to superimposed fluctuations in the field acquisition link and sensor interference when cutting parameters change frequently. This leads to significant signal quality drift in multimodal sensing signals (such as decreased signal-to-noise ratio, changes in noise spectrum structure, and attenuation of channel reliability), resulting in a drastic shift in feature distribution. Consequently, the model's cross-condition generalization ability collapses, leading to prediction failure. Conventional modal feature cascade fusion and numerical regression methods that rely solely on mean square error assume that the quality of each modal signal is stable and the feature correlation is constant. They struggle to perceive the dynamic drift of multimodal signal quality and the time-varying reliability under different operating conditions and wear stages. Under varying operating conditions, they are easily interfered with by high-frequency environmental noise or inferior modal signals, causing fusion bias and potentially disrupting the strictly monotonically increasing physical law of tool wear.

[0032] To address at least one of the aforementioned technical problems, this application proposes a method for predicting the remaining service life of a cutting tool. The method includes: acquiring multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted, wherein the multimodal sensing signals include cutting force signals, vibration signals, and acoustic emission signals; inputting the multimodal sensing signals and operating parameters into a trained end-to-end prediction model; extracting features from the multimodal sensing signals using the end-to-end prediction model to obtain the spatiotemporal features of each mode; mapping the operating parameters to obtain a latent space vector of the operating conditions; decoupling the spatiotemporal features based on the latent space vector of the operating conditions to obtain decoupling features decoupled from the operating conditions of the cutting tool to be predicted; and performing regression prediction based on the decoupling features to output a predicted value of the remaining tool life corresponding to the cutting tool to be predicted.

[0033] Thus, when predicting the remaining tool life, multimodal sensor signals and operating parameters are collected separately to perform predictions based on a trained end-to-end prediction model. During the prediction process, the trained end-to-end prediction model first extracts features from the multimodal sensor signals to obtain the spatiotemporal features of each mode. Then, it maps these features to the operating parameters to obtain a latent space vector. This latent space vector is then used to decouple the spatiotemporal features from the operating conditions of the cutting tool to be predicted, resulting in decoupled features. Regression prediction is then performed based on these decoupled features to output the predicted remaining tool life value for the cutting tool. In this way, decoupling features from operating conditions during the prediction process reduces the impact of changes in operating conditions on the prediction process, thereby improving the accuracy of the remaining tool life prediction.

[0034] like Figure 1 As shown in the embodiment of this application, a method for predicting the remaining service life of a cutting tool is provided. Specifically, the method includes the following steps: Step S100: Obtain the multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted, wherein the multimodal sensing signals include cutting force signals, vibration signals and acoustic emission signals; Step S200: Input the multimodal sensing signals and the operating parameters into the trained end-to-end prediction model. Extract features from the multimodal sensing signals using the end-to-end prediction model to obtain the spatiotemporal features of each mode. Obtain the latent space vector of the operating parameters. Decouple the spatiotemporal features from the latent space vector to obtain decoupling features that are decoupled from the operating conditions of the cutting tool to be predicted. Perform regression prediction based on the decoupling features to output the predicted remaining tool life value corresponding to the cutting tool to be predicted.

[0035] The cutting tool to be predicted is one whose remaining life needs to be predicted. In this embodiment, the cutting force signal is a three-channel cutting force signal, the vibration signal is a four-channel vibration signal, and the acoustic emission signal is a two-channel acoustic emission signal. The above operating parameters include at least one of cutting power, spindle speed, feed rate, and depth of cut.

[0036] Specifically, in this embodiment, when making predictions based on a trained end-to-end prediction model, the model's input includes a multimodal sensing sensor dataset collected under multiple operating conditions and the corresponding current operating condition parameters. Specifically, this includes real-time collected three-channel cutting force signals. Four-channel vibration signal Two-channel acoustic emission signals and the corresponding current operating parameters. (Such as cutting power, spindle speed, feed rate, depth of cut, etc.).

[0037] Among them, the cutting force signal is acquired by a triaxial piezoelectric force gauge installed at the bottom of the machine tool table or inside the tool post base; the vibration signal is acquired in real time by three independent single-axis acceleration sensors physically arranged on the machine tool spindle housing; and the acoustic emission signal is acquired by two acoustic emission sensors closely arranged on the workpiece fixture.

[0038] These are forces in three orthogonal directions, used to characterize the interaction forces between the tool and the workpiece; It is acceleration in three orthogonal directions. The acceleration corresponding to the additionally added high-sensitivity, high-frequency chatter acceleration acquisition channel specifically along the machine tool cutting feed direction; and These are acoustic emission amplitudes obtained through acquisition from different channels; specifically... Acoustic emission data was acquired using the first channel of an acoustic emission acquisition system located in the machine tool cutting area (e.g., near the tool plate / workpiece fixture). Acoustic emission was acquired using the second channel, which is positioned in a symmetrical or complementary spatial orientation (e.g., on the base side away from the cutting interference zone).

[0039] It should be noted that the multimodal sensing signal corresponding to the current moment (time step) is acquired based on the set time window, that is, the multimodal sensing signal at the current moment includes the time sequence signal from a historical period up to the current moment. The operating parameters are the parameters corresponding to the current moment.

[0040] Furthermore, the aforementioned end-to-end prediction model is used to extract features from the multimodal sensing signals to obtain the spatiotemporal features of each modality, including: Each mode of the multimodal sensing signal is taken as the signal to be processed, and the local temporal dynamic features corresponding to the signal to be processed are extracted by the multi-scale feature extraction network in the end-to-end prediction model. Calculate the statistical characteristics corresponding to the above-mentioned signal to be processed, wherein the above-mentioned statistical characteristics include the peak factor and / or root mean square. The aforementioned local temporal dynamic features and statistical features are aggregated and spliced ​​together along the channel dimension to obtain spliced ​​features. Based on the splicing characteristics described above, the spatiotemporal features corresponding to the signal to be processed are obtained through the long short-term memory network in the end-to-end prediction model.

[0041] In this embodiment, after the data is input into the end-to-end prediction model, multi-scale spatiotemporal feature extraction is first performed through the model. Specifically, considering the characteristics of multimodal time-series signals including high-frequency transient impacts and low-frequency slow-varying wear, parallel condition-aware multi-scale feature extraction networks are constructed according to modal physical categories. Each point of the extraction network uses a parallel one-dimensional convolutional kernel containing coarse, medium, and fine receptive fields to capture signal representations of different frequency bands. It should be noted that in the above end-to-end prediction model, a separate multi-scale feature extraction network is set up for signals of different modalities, which can realize parallel processing of signals of different modalities. In this embodiment, the example of processing each modal signal as a separate signal to be processed is used for illustration, but it is not intended to be a specific limitation.

[0042] Specifically, let the current time be the first... The data slice (i.e., the slice corresponding to the current time) The first time window) The physical mode signal matrix is ​​as follows ,in These correspond to the combined cutting force signal matrix of the three-axis direction, the vibration signal matrix of the four-channel direction, and the acoustic emission signal matrix of the two-channel direction, respectively. For the independent first... Modal signals First, a multi-scale one-dimensional convolutional network is used. Deep spatial filtering is applied to extract local temporal dynamic features; simultaneously, a manually designed statistical feature mapping function is used. From the original In parallel computation, prior physical statistical feature matrices such as crest factor and root mean square (RMS) are calculated. The aforementioned statistical feature mapping function is a pre-set function used to calculate the statistical features of the input data, which can be set and adjusted according to actual needs. In this embodiment, crest factor and RMS are calculated simultaneously, but this is not a specific limitation. Subsequently, to fully integrate the spatial transient information and macroscopic physical degradation indicators of the signal, local temporal dynamic features and prior physical statistical features are aggregated and spliced ​​along the channel dimension. Finally, the spliced ​​features are input into a Long Short-Term Memory (LSTM) network to deeply mine the temporal dependencies of the cutting sequence, obtaining deep spatiotemporal features with complete representational capabilities. The complete mathematical expression of the above process is as follows: ; in, The parameter is Long Short-Term Memory Network Represents a multi-scale one-dimensional convolutional feature extraction network The corresponding set of learnable weights and bias parameters internally. This represents the set of learnable weights and bias parameters within a Long Short-Term Memory (LSTM) network. This represents a concatenation operation along the feature channel dimension. This represents the set of physical statistical features for the corresponding mode. This parallel independent extraction structure ensures that each heterogeneous physical sensor independently extracts the most representative initial feature sequence at its own scale.

[0043] The operating condition parameters are encoded to obtain the operating condition latent space vector. The process of obtaining the operating condition latent space vector based on the aforementioned operating condition parameter mapping includes: The above working condition parameters are input into the working condition encoder in the above end-to-end prediction model, and the above working condition latent space vector is obtained by mapping through the above working condition encoder. The aforementioned operating condition encoder is a multilayer sensor network; the aforementioned operating condition parameters include at least one of spindle speed, feed rate, and depth of cut. In this embodiment, the operating condition parameters including spindle speed, feed rate, and depth of cut are used as examples for specific explanation, but this is not intended as a specific limitation.

[0044] Specifically, to eliminate the "domain offset" problem of measurement signals caused by varying operating conditions, a condition encoder is introduced. This encoder uses a multi-layer perceptron (MLP) network to process the numerical operating parameters at the current moment. Mapped to a high-dimensional continuous latent space vector of working conditions : ; in, This represents the high-dimensional, continuous latent feature vector of the working condition obtained after network mapping. This represents the mapping function of a multilayer perceptron network. Indicates the current number The input numerical operating condition parameter feature vector under each time step (i.e., time slice sequence) specifically includes the spindle speed, feed rate, depth of cut, and other specific numerical dimensions set by the machine tool when acquiring the signal for that segment. This represents the set of all learnable weights and bias parameters within the multilayer perceptron network.

[0045] It should be noted that in the physical essence mapping, the working condition latent space vector obtained after nonlinear topological dimensionality enhancement... This aims to bridge the semantic gap between manually defined quantities and deep abstract features. This essentially characterizes the conditional constraint benchmark of the overall macroscopic cutting dynamics state of the machine tool machining system under this specific combination of cutting parameters, that is, it depicts a "congenital mechanical environment profile and parameter fingerprint" under the current specific speed and feed. This feature vector The nonlinear energy baseline shift and background interference intensity, which are independent of tool cumulative wear and purely caused by the set working conditions, were extracted. This provides high-dimensional conditional prior variables with clear discriminative properties for the subsequent network to accurately isolate and decouple these external working condition interferences.

[0046] Furthermore, conditional adaptive normalization decoupling is performed. The above-mentioned spatiotemporal features are decoupled based on the aforementioned latent space vector of the working condition to obtain decoupling features decoupled from the working condition of the cutting tool to be predicted, including: Based on the above working condition latent space vector, the scaling factor and translation factor are obtained by mapping through the linear mapping fully connected network in the above end-to-end prediction model. For each of the above-mentioned spatiotemporal features, the mean and standard deviation corresponding to the above-mentioned spatiotemporal features are calculated in the spatial dimension, so as to standardize the above-mentioned spatiotemporal features based on the above-mentioned mean and standard deviation. Based on the scaling factor and the translation factor mentioned above, an affine transformation is performed on the standardized spatiotemporal features to obtain the decoupled features mentioned above.

[0047] First, using the working condition vector Generate an adaptive scaling factor (Scale). Translation factor (Shift) Specifically, the end-to-end prediction model described above constructs two parallel linear mapping fully connected networks (i.e., a scaling network and a translation network) to map the latent vectors of the working conditions. As the common input of the two networks, the output is mapped to an adaptive scaling factor consistent with the dimension of the number of multimodal feature channels. Translation factor : ; ; in, and This represents the learnable weight matrix and bias terms of the scaled network. and The learnable weight matrix and bias terms represent the translation network; their specific values ​​are determined during model training. It should be noted that... Representative moment The corresponding scaling factor obtained from the data calculation, and The actual representation has the same meaning. Similarly, this will not be elaborated upon here.

[0048] The scaling and translation factors are obtained, and the spatiotemporal features of each modality are processed using the Conditional Adaptive Instance Normalization (Conditional AdaIN) module set in the end-to-end prediction model described above. Modulation is then performed. This step forces the features to decouple into a pure wear feature space that is independent of the operating conditions by eliminating the distribution offset caused by the mean and variance of the operating conditions in the original features.

[0049] Specifically, for the first Spatiotemporal characteristics of each modality Calculate its own mean in the spatial dimension. and standard deviation The spatiotemporal features are then standardized based on the mean and standard deviation to form dimensionless features with a mean of 0 and a variance of 1. Furthermore, the generated scaling factor is used... Translation factor Performing a dynamic affine transformation on the standardized features yields the decoupled feature representation, i.e., the decoupled features. : ; in, This indicates element-wise multiplication.

[0050] Considering that the sensitivity of different physical modal signals (cutting force, vibration, and acoustic emission sensor signals) to different cutting conditions and tool wear stages often varies over time, and that they are highly susceptible to sudden interference from non-stationary high-frequency environmental noise of the machine tool, resulting in severe fluctuations in the signal-to-noise ratio of a single type of signal, a modal reliability assessment module is introduced to calculate the reliability confidence level corresponding to each mode.

[0051] Specifically, the regression prediction based on the aforementioned decoupling features, to output the predicted remaining tool life value corresponding to the cutting tool to be predicted, includes: For each of the above decoupling features, global average pooling is performed respectively, and the reliability confidence level corresponding to the above decoupling features is determined based on the features after global average pooling. Based on the aforementioned reliability confidence level, the aforementioned decoupling features are weighted to obtain the weighted features of each mode; The multi-head self-attention module in the above end-to-end prediction model is used to fuse the weighted features of each modality to obtain multi-modal fusion features. The multimodal fusion features are input into the regression module of the end-to-end prediction model to obtain the tool remaining life prediction value. The regression module is composed of a multi-layer nonlinear fully connected network.

[0052] In this embodiment of the application, the multi-source signals containing each channel are aggregated into three independent modes (i.e. These correspond to the overall cutting force mode, the overall vibration mode, and the overall acoustic emission mode, respectively. For the current... Three decoupling features of the input at each time step Calculate their reliability confidence level independently at the current moment. The specific calculation formula is as follows: ; in, This represents a global average pooling operation performed along the temporal / spatial feature dimension. Its physical essence is to compress locally redundant high-frequency information, condensing and extracting a "global energy macroscopic state and waveform fingerprint compression vector" of the current segment of the sensor signal, which reflects the current overall quality level of that mode. This represents a multilayer perceptron network used for state evaluation (typically consisting of two fully connected layers and an activation function in between). This represents the Sigmoid activation function, which will... The arbitrary real-valued nonlinear compression of the output is mapped to Within the standard probability interval. That is, the final output of the first The dynamic reliability confidence level of each physical mode. The closer the value is to 1, the better the signal-to-noise ratio of the current sensor signal (such as acoustic emission) at that cutting speed or wear stage, and the more reliable the mapping relationship to tool decay. The closer the confidence level is to 0, the more it indicates that the current sensor (such as vibration sensor) is under strong external environmental interference or in a blind zone, and its performance is extremely degraded. This confidence level will be used as a hard gating weight for subsequent soft selection of features.

[0053] Condition-aware route allocation is performed based on the calculated reliability confidence score. Specifically, based on the calculated reliability confidence score, the decoupling features of each mode are weighted and activated to suppress inferior mode features containing a large amount of high-frequency abnormal noise and enhance the feature response of high-sensitivity modes. ; in, This represents the weighted feature obtained after weighting.

[0054] Furthermore, cross-modal attention fusion is performed: the weighted features obtained after weighting are fed into the multi-head self-attention module network in the end-to-end prediction model. The deep interaction weights between different physical modalities are calculated using the inner product mechanism of query, key, and value, and then these weights are accumulated and fused to obtain high-purity multimodal fusion features containing globally optimal information. .

[0055] Based on the above-mentioned multimodal fusion features Regression prediction of Remaining Useful Life (RUL) is then performed. Specifically, the aforementioned multimodal fusion features are used... The input end-to-end prediction model includes a regression module composed of a multi-layer nonlinear fully connected network. Output the normalized predicted tool life: ; in, This represents the predicted remaining tool life. Regression module The specific values ​​of the corresponding learnable parameters are determined based on the model training process.

[0056] It should be noted that the above end-to-end prediction model is trained according to the following steps: Obtain a training dataset, wherein the training dataset is constructed based on at least one sample tool. For each sample tool, there are multiple training sample groups corresponding to multiple different time steps (i.e. multiple different moments) in the training dataset. A training sample group includes a set of corresponding multimodal signal samples, working condition parameter samples, tool remaining life labels and working condition labels. Input the multimodal signal samples and working condition parameter samples from the above training dataset into the end-to-end prediction model to be trained, and obtain the multimodal fusion features and tool remaining life prediction values ​​obtained by the above end-to-end prediction model. Based on the aforementioned multimodal fusion features, the predicted working condition classification result is determined by a domain discriminator. Based on the above predicted working condition classification results and the above working condition labels, the domain alignment loss is determined; based on the above predicted tool remaining life value and the above predicted tool remaining life label, the regression loss is determined; based on the predicted tool remaining life value and the tool remaining life label corresponding to the training sample groups of the same sample tool at different time steps, the monotonic trend loss is determined. Among these, the above domain alignment loss is used to constrain the above multimodal fusion features to be independent of the working condition category; the above regression loss is used to constrain the consistency between the above predicted tool remaining life value and the above predicted tool remaining life label; and the above monotonic trend loss is used to constrain the change direction of the above predicted tool remaining life value over time to be consistent with the change direction of the above predicted tool remaining life label over time. Based on the aforementioned domain alignment loss, regression loss, and monotonic trend loss, the model parameters of the end-to-end prediction model are adjusted. Then, the process returns to the previous steps of inputting the multimodal signal samples and working condition parameter samples from the training dataset into the end-to-end prediction model to be trained, and obtaining the multimodal fusion features and tool remaining life prediction values ​​obtained by the end-to-end prediction model, until the preset training termination condition is met, and the trained end-to-end prediction model is obtained.

[0057] In this embodiment, based on the model training process, the learnable parameters of each module and network in the end-to-end prediction model are updated to obtain a trained end-to-end prediction model. It should be noted that the data input processing during model training can refer to the same processing process described above during model usage, and will not be repeated here. In this embodiment, the loss calculation during model training is further explained in detail.

[0058] Specifically, during model training, gradient reversal domain adversarial learning is performed. To ensure the model has strong generalization ability under new conditions, a gradient reversal layer (GRL) and a domain discriminator are introduced on the discriminative side of the network. Domain discriminator Receive fused features To classify which type of cutting condition the feature originates from. .

[0059] Then, the domain adversarial loss is calculated: in forward propagation, GRL behaves as an identity mapping; in backward propagation, GRL is calculated according to the coefficients. The gradient is inverted. The determiner attempts to minimize the classification error, while the feature extractor, due to the gradient inversion, is forced to learn generalized features that "cannot be distinguished by the work condition category." A work condition domain alignment loss is introduced. : ; in, This represents the total number of discrete operating condition types. For the unique heat-coded label of the actual working condition (i.e., the working condition label, as the supervision benchmark), if the first The sample belongs to the working condition category. The value is 1 if it is 1, otherwise it is 0. The representation domain discriminator predicts that the input feature belongs to the first domain. The probability value of a class of working conditions, that is, the probability value corresponding to the predicted working condition classification result, is the output of the model after soft normalization, indicating that the feature belongs to the first class. Probability distribution value of type of working condition (0 (real numbers of 1). This is a gradient inversion layer, which performs identity mapping during forward propagation. In backpropagation, the gradient is reversed. ( The adversarial learning rate constraint parameter is used to control the intensity of the zero-sum adversarial game between the feature extractor and the domain discriminator. This represents the identity matrix. The physical essence of this cross-entropy formula is to perform a logarithmic penalty calculation only on the predicted probability of the actual operating condition category, forcing the discriminator to continuously narrow the divergence between the predicted probability distribution and the one-hot distribution of the actual operating condition.

[0060] In this embodiment, domain alignment loss is introduced to "force" the feature extractor to remove traces of the cutting conditions during training. Specifically, there is severe feature coupling in variable cutting condition signals (e.g., high-amplitude physical vibrations may be induced by severe tool wear or simply by high spindle speed). Ordinary regression networks are prone to fitting such "spurious correlations." In this adversarial architecture, the domain discriminator uses real cutting condition labels as supervisory signals, acting as a "cutting condition detective" striving to minimize the classification error of the cutting environment; while the feature extraction network, driven by the negative penalty of backpropagation in the gradient inversion layer (GRL), is forced to engage in a "zero-sum game" with the domain discriminator, that is, the feature extractor is forced to update and optimize weights in the direction that maximizes the deterioration of the discriminator's classification accuracy. The final equilibrium result of this game training is that the front-end feature extraction network is forced to filter out and smooth out the heterogeneous baseline waveform deviation of the multimodal signal caused by different cutting parameters (spindle speed, feed) at the source. The feature tensor ultimately fed into the subsequent RUL regression module becomes a pure, condition-independent, generalized wear prior semantic representation that spans multiple operating conditions. This enables the subsequent regression prediction module to eliminate numerical overestimation or underestimation caused by different cutting environments, and to maintain high-precision life assessment performance with zero generalization degradation even when facing cutting conditions of new machine tools that are not seen in actual production or that are frequently changed.

[0061] To improve the model's robustness to outliers, a smooth Huber loss function is used to calculate the predicted values. With actual remaining tool life label Regression loss between : ; in, This represents the number of samples.

[0062] The complete piecewise form of the function is: ; in, The set smoothing threshold parameter can be set and adjusted according to actual needs to balance the smoothing derivative of mean square error (MSE) when the error is small and the robustness of mean absolute error (MAE) to outliers.

[0063] In this embodiment of the application, the physical monotonicity loss is also calculated. To constrain the model, the predicted remaining tool life over time must align with the actual physical laws governing tool wear; that is, the predicted value should monotonically decrease with machining time, without any local rebounds or oscillations that violate these physical laws. Tool wear exhibits irreversible and strictly monotonic characteristics in its physical mechanism. Existing numerical loss methods are prone to causing oscillations in the prediction curve. Therefore, a trend penalty term based on the difference between sample pairs is introduced to constrain the time step. and The sign of the predicted slope must be consistent with the sign of the true degradation slope: ; in, No penalty is imposed when the predicted evolution direction conforms to objective physical laws (product greater than 0), but a gradient penalty is imposed when there is a reverse fluctuation that violates physical laws.

[0064] Joint optimization based on the total loss function is performed: the entire network architecture model adopts an end-to-end deep learning training mechanism. The final joint objective loss function is composed of a weighted combination of the above-mentioned losses using hyperparameters. ; in, Represents the loss of a joint objective; and These are the balance coefficients for controlling the prior physical laws and the alignment strength of the working condition distribution, which can be set and adjusted according to actual needs.

[0065] It should be noted that the training termination condition can be set and adjusted according to actual needs, and is not specifically limited here. In the embodiments of this application, during the training iteration phase, the model solves the total loss function. The core optimization objective is to find the global or local minimum value. This is achieved by employing the gradient descent algorithm, based on the calculated... The error gradient is propagated backward layer by layer from back to front, updating all network weight parameters in the feature extraction, route weighting, and lifetime regression modules.

[0066] To ensure that the model converges sufficiently while avoiding overfitting to historical working conditions, the training process in this embodiment does not rely on a single fixed error threshold, but instead employs a multi-dimensional, composite training termination strategy: Validation set early stopping mechanism: A separate cutting validation set is created in addition to the full training set. After each round of full data training, the total loss of the model on the validation set is calculated. If the validation set loss falls below the threshold for early stopping, the model is stopped early. If no significant decrease occurs within 20 consecutive epochs (i.e., the set tolerance parameter, such as 20 consecutive epochs), or even if a rebound occurs, it is determined that the model's generalization performance has reached its limit and has begun to overfit. At this point, the system will forcibly trigger an early termination command and automatically backtrack to call the historically optimal set of network weight configurations as the final saved model; Maximum Epochs (Max Epochs): As a fallback mechanism for the algorithm, a global maximum number of training epochs (e.g., 500) is pre-defined physically to forcibly terminate endless gradient wandering and limit computational overhead. In actual engineering implementation, training immediately terminates when any of the above conditions are met.

[0067] Figure 2 This is a schematic diagram of an end-to-end prediction model training process provided in an embodiment of this application. It should be noted that... Figure 2 In this context, r1, r2, and r3 represent the reliability confidence levels of the corresponding modes, while v1, v2, and v3 represent the weighted features obtained after weighting the decoupled features through the condition-aware router. These weighted features can then be fed into a multi-head self-attention module for further processing. Specifically, Figure 2 The calculation process for the three types of losses is shown in the diagram. The specific calculation methods can be found in the description above and will not be repeated here. The corresponding domain alignment loss, regression loss, and monotonic trend loss are calculated. After calculating these three losses, the joint objective loss is further calculated to update the model parameters.

[0068] This application provides a method for predicting the remaining service life of a cutting tool. The method includes: acquiring multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted, wherein the multimodal sensing signals include cutting force signals, vibration signals, and acoustic emission signals; inputting the multimodal sensing signals and operating parameters into a trained end-to-end prediction model; extracting features from the multimodal sensing signals using the end-to-end prediction model to obtain the spatiotemporal features of each mode; mapping the operating parameters to obtain a latent space vector of the operating conditions; decoupling the spatiotemporal features based on the latent space vector of the operating conditions to obtain decoupling features decoupled from the operating conditions of the cutting tool to be predicted; and performing regression prediction based on the decoupling features to output the predicted remaining service life value of the cutting tool to be predicted.

[0069] Thus, when predicting the remaining tool life, multimodal sensor signals and operating parameters are collected separately to perform predictions based on a trained end-to-end prediction model. During the prediction process, the trained end-to-end prediction model first extracts features from the multimodal sensor signals to obtain the spatiotemporal features of each mode. Then, it maps these features to the operating parameters to obtain a latent space vector. This latent space vector is then used to decouple the spatiotemporal features from the operating conditions of the cutting tool to be predicted, resulting in decoupled features. Regression prediction is then performed based on these decoupled features to output the predicted remaining tool life value for the cutting tool. In this way, decoupling features from operating conditions during the prediction process reduces the impact of changes in operating conditions on the prediction process, thereby improving the accuracy of the remaining tool life prediction.

[0070] Specifically, the proposed method for predicting remaining tool life can achieve multi-condition adaptive operation considering the quality drift of multimodal signals, thereby solving the problem of degraded life prediction performance caused by cross-condition feature distribution shift and fixed multimodal fusion weights. In actual production environments, due to factors such as variable machining requirements and frequent adjustments to cutting parameters (such as rotational speed and feed rate), significant drift in the quality of multimodal sensor signals often occurs, manifested as decreased signal-to-noise ratio, increased noise, channel distortion, or reliability degradation, which leads to domain shift at the feature level, resulting in decreased generalization ability and significantly reduced prediction accuracy of the model under new operating conditions. At the same time, since different physical modes (such as force, vibration, and acoustic emission) have different sensitivities to specific operating conditions and different wear stages, the signal quality and reliability of each modal signal dynamically change with time / operating conditions during operation. If the fusion strategy still adopts fixed weights or a static mechanism, the model will have difficulty perceiving quality drift and will be easily misled by low-quality or low-sensitivity modes, resulting in fusion bias and further deteriorating life prediction performance. This application's solution achieves effective decoupling of operating environment features and global alignment of cross-domain wear feature distributions by introducing conditional adaptive normalization and gradient inversion domain adversarial mechanisms into the feature extraction and representation network. Simultaneously, it employs an adaptive weighting strategy of operating condition-aware routing and dynamic evaluation of modal signal quality, combined with monotonic trend loss constraints consistent with the physical nature of irreversible tool decay, to suppress interference from low-quality, low-signal-to-noise-ratio, and insensitive modes, thereby enhancing the feature contribution of high-reliability modes under the current operating condition. Through the synergistic effect of these mechanisms, accurate and physically consistent prediction of remaining tool life can be achieved even under conditions of frequent changing operating conditions and multimodal signal quality drift.

[0071] Specifically, conditional adaptive instance normalization and gradient inversion domain adversarial learning are introduced into the feature extraction network, combined with a working condition-aware routing and modal signal quality / reliability dynamic evaluation mechanism. This achieves effective decoupling of environmental interference features under different working conditions and adaptive allocation of high-reliability, high-sensitivity modal information, thereby achieving efficient global alignment of the underlying feature distribution in the context of multimodal signal quality drift. Simultaneously, a monotonic trend loss mechanism consistent with the physical nature of irreversible tool decay and a cross-modal attention fusion mechanism are introduced to suppress interference from low signal-to-noise ratio and low-reliability modes and correct irrational prediction curve oscillations, ensuring accurate and physically reasonable tool remaining life prediction results even under frequent changing working conditions and dynamic drift of multimodal signal quality.

[0072] In the processing, the input is first constructed using cutting operation data containing multiple working condition parameters and multimodal sensing signals (cutting force, vibration, acoustic emission, etc.). After the multimodal data is processed by a multi-scale convolutional network to extract deep spatiotemporal features, the multimodal features are decoupled from working condition interference through a conditionally adaptive instance normalization module and used to train a unified feature representation network. At the modality fusion end, in response to the signal quality drift of each physical mode under dynamic cutting conditions caused by changes in time and working conditions (such as signal-to-noise ratio fluctuations, noise enhancement, channel reliability attenuation, and sensitivity changes), the real-time effective weights of each mode are adaptively calculated using a working condition-aware routing and modal signal quality / reliability dynamic evaluation mechanism. These weights are then fed into a cross-modal attention module to suppress the inferior noise interference introduced by quality drift, resulting in high-purity adaptive multimodal fusion features. Subsequently, a domain adversarial cross-learning method based on a gradient inversion layer is used to globally align the low-level features of different working conditions to alleviate the distribution shift caused by the superposition of quality drift and working condition changes. Then, a monotonic trend loss conforming to the irreversible tool decay mechanism and mean square error are combined to impose physical constraints on the fused temporal features. Finally, the results are input into a multi-layer regression prediction module to output the remaining tool life, thus achieving high-precision and physically consistent life prediction under complex conditions of frequent changing working conditions and multi-modal signal quality drift. The proposed solution includes the following three key processes: The process of decoupling features across different operating conditions and achieving global adversarial alignment involves introducing conditional adaptive instance normalization and a gradient inversion layer as an adaptive medium between multi-scale feature extraction and deep representation. This is achieved through forward feature modulation and reverse gradient adversarial mechanisms to completely eliminate operating condition interference. After encoding the current cutting parameters into a latent space vector, a scaling factor and a translation factor are generated using the conditional adaptive instance normalization module to progressively adaptively normalize the multimodal features, thus eliminating distribution biases caused by specific operating conditions. Simultaneously, the extracted features are continuously manipulated by the gradient inversion layer during backpropagation to restore the pure wear semantic structure. This process not only eliminates the severe "domain shift" problem caused by changing operating conditions but also achieves global generalization alignment of the underlying wear features across multiple operating conditions through an adversarial mechanism, thereby greatly improving the model's predictive generalization ability under conditions where no new operating conditions are encountered or where operating conditions frequently change.

[0073] Modal reliability dynamic assessment and intelligent routing process: In the multimodal feature fusion stage, an adaptive dynamic routing and cross-modal attention dual mechanism is designed to balance and filter the contributions of each modality at both the feature allocation and deep interaction levels, ensuring accurate lifetime prediction even under high-concurrency environmental noise or degradation of the perception quality of a certain modality. In the shallow routing stage, the reliability confidence is dynamically calculated based on the global state of each single-modal feature. The weights of modalities with high reliability (sensitive to the current wear stage) are amplified, while the weights of inferior modalities containing a large amount of high-frequency abnormal noise are forcibly suppressed. In the deep interaction stage, the filtered and weighted features are fed into a multi-head self-attention mechanism to further adaptively extract the cooperative dependencies across the modal space. To avoid the problem of inferior modality dominance caused by traditional simple splicing fusion, this mechanism effectively overcomes the technical defects of traditional feature cascading methods that are susceptible to low signal-to-noise ratio modal interference. Through adaptive screening and deep purification of multi-source heterogeneous features, the effectiveness of multimodal fusion representation and the noise robustness of the model in complex cutting environments are greatly enhanced.

[0074] Physical Monotonic Trend Constraint and Joint Optimization Process: A multi-level loss constraint mechanism combining numerical smoothing and physical priors is employed in the life regression and test output stages. By calculating the correlation between the predicted slope and the actual decay slope between time series sample pairs, a strict match between the model output and the physical laws of tool decay is achieved, reducing the prediction oscillation problem caused by the lack of physical common sense in purely data-driven models. Specifically, a unique monotonic trend loss is introduced on top of the basic Huber regression loss. A strong gradient penalty is applied only when the predicted temporal evolution direction of the model undergoes a reverse jump that violates the physical essence of "irreversible monotonically increasing decay" of the tool. This positive guidance mechanism not only eliminates the frequent irrational spike jumps in conventional regression networks but also constrains the network to focus more on fitting the long-term cumulative wear trend, further improving the smoothness and physical rationality of the final life prediction curve in practical industrial applications.

[0075] This approach enables various technical effects, such as improved cross-domain generalization to address multimodal signal quality drift. In actual CNC machining, frequent changes in cutting parameters (such as spindle speed and feed rate) can be compounded by on-site interference and sensor link fluctuations, leading to significant signal quality drift in multimodal sensing signals (e.g., decreased signal-to-noise ratio, altered noise spectrum structure, and reduced channel reliability). This further induces shifts in the underlying feature distribution, causing a sharp decline in the generalization performance of traditional single-condition models under new conditions. This application introduces Conditional Adaptive Instance Normalization (Conditional AdaIN) and Gradient Reversal Layer (GRL) domain adversarial learning mechanisms. At the feature level, this decouples the quality drift-related interference components from effective wear information and globally aligns the latent layer representations for different working conditions. Even when encountering unseen new working conditions or environments with frequent changes in working conditions accompanied by signal quality drift during testing, the network can still suppress representational biases caused by quality drift, extract stable and effective tool wear representations, and reduce the risk of performance collapse in cross-domain applications.

[0076] Adaptive Reliability Balancing and Fusion Robustness under Multimodal Signal Quality Drift: Since different physical modes exhibit time-varying sensitivities to processing conditions and wear stages, and the signal-to-noise ratio (SNR) and reliability of each mode undergo significant dynamic drift and imbalance under sudden cutting interference, simple splicing or static fusion methods are easily dominated by locally high-amplitude inferior noise modes, suppressing effective weak modes and leading to decreased anti-interference capability. To suppress the fusion bias of "inferior mode dominance," this application employs a condition-aware routing and modal signal quality / reliability dynamic evaluation mechanism. Based on the global characterization of each mode, the reliability confidence level at the current moment is adaptively estimated, and the fusion weights are dynamically adjusted through cross-modal attention. This amplifies the contribution of high-reliability and high-sensitivity modes while suppressing low SNR and noise-contaminated modes under quality drift scenarios, achieving dynamic screening and deep purification of multi-source signals.

[0077] Lifetime evolution constraints that maintain physical consistency under signal quality drift and operating condition fluctuations: Data-driven models that rely solely on mean square regression (MSE) typically only focus on numerical fitting at discrete moments. In complex environments, they are easily affected by multimodal signal quality drift and noise fluctuations, resulting in oscillations in the predicted curve, which is inconsistent with the physical principle that tool degradation is "irreversible and monotonically degraded." To narrow this gap between pure data fitting and physical laws, this application introduces a monotonic trend loss based on the sign difference of the temporal state slope. This constraint ensures that the predicted evolution direction is consistent with the actual wear evolution slope over the time span of continuous sample pairs. Furthermore, an asymmetric penalty is applied to reverse jumps that violate degradation laws. Therefore, even when severe operating condition fluctuations and multimodal signal quality drift coexist, it can still output a smooth and physically consistent high-precision remaining life prediction.

[0078] like Figure 3 As shown in the figure, corresponding to the above-described tool remaining service life prediction method, this application embodiment also provides a tool remaining service life prediction system, the tool remaining service life prediction system comprising: The data acquisition module 310 is used to acquire the multimodal sensing signals and working parameters corresponding to the cutting tool to be predicted, wherein the multimodal sensing signals include cutting force signals, vibration signals and acoustic emission signals; The prediction module 320 is used to input the multimodal sensing signals and the operating parameters into a trained end-to-end prediction model, extract features from the multimodal sensing signals through the end-to-end prediction model to obtain the spatiotemporal features of each mode, obtain the operating condition latent space vector based on the operating parameters, decouple the spatiotemporal features based on the operating condition latent space vector to obtain decoupling features decoupled from the operating conditions of the cutting tool to be predicted, and perform regression prediction based on the decoupling features to output the predicted remaining tool life value corresponding to the cutting tool to be predicted.

[0079] Thus, when predicting the remaining tool life, multimodal sensor signals and operating parameters are collected separately to perform predictions based on a trained end-to-end prediction model. During the prediction process, the trained end-to-end prediction model first extracts features from the multimodal sensor signals to obtain the spatiotemporal features of each mode. Then, it maps these features to the operating parameters to obtain a latent space vector. This latent space vector is then used to decouple the spatiotemporal features from the operating conditions of the cutting tool to be predicted, resulting in decoupled features. Regression prediction is then performed based on these decoupled features to output the predicted remaining tool life value for the cutting tool. In this way, decoupling features from operating conditions during the prediction process reduces the impact of changes in operating conditions on the prediction process, thereby improving the accuracy of the remaining tool life prediction.

[0080] It should be noted that the specific structure and implementation of the above-mentioned tool remaining service life prediction system and its various modules or units can be referred to the corresponding descriptions in the above method embodiments, and will not be repeated here.

[0081] It should be noted that the division of the various modules in the above-mentioned tool remaining service life prediction system is not unique and is not intended as a specific limitation.

[0082] Based on the above embodiments, this application also provides a terminal, the principle block diagram of which can be as follows: Figure 4As 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 a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps of any of the aforementioned tool remaining service life prediction methods. The display screen can be an LCD screen or an e-ink screen.

[0083] Those skilled in the art will understand that Figure 4 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.

[0084] In one embodiment, a terminal is provided, the terminal including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of any of the tool remaining service life prediction methods provided in the embodiments of this application.

[0085] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the tool remaining service life prediction methods provided in this application.

[0086] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the above device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above device can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0088] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0089] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0090] In the embodiments provided in this application, it should be understood that the disclosed systems / terminal devices and methods can be implemented in other ways. For example, the system / terminal device embodiments described above are merely illustrative. For instance, the division of modules or units described above is merely a logical functional division, and in actual implementation, it can be divided in other ways. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0091] If the integrated modules / units described above are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, and software distribution media, etc. It should be noted that the content included in the computer-readable storage medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction.

[0092] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions are not in essence a departure from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for predicting the remaining service life of a cutting tool, characterized in that, The method includes: The multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted are obtained, wherein the multimodal sensing signals include cutting force signals, vibration signals and acoustic emission signals; The multimodal sensing signals and the operating parameters are input into a trained end-to-end prediction model. The end-to-end prediction model extracts features from the multimodal sensing signals to obtain the spatiotemporal features of each mode. The operating parameters are mapped to obtain the latent space vector of the operating conditions. The spatiotemporal features are decoupled based on the latent space vector of the operating conditions to obtain decoupling features that are decoupled from the operating conditions of the cutting tool to be predicted. Regression prediction is performed based on the decoupling features to output the predicted remaining tool life value corresponding to the cutting tool to be predicted.

2. The method for predicting the remaining service life of a cutting tool according to claim 1, characterized in that, The cutting force signal is a three-channel cutting force signal, the vibration signal is a four-channel vibration signal, and the acoustic emission signal is a two-channel acoustic emission signal; The operating parameters include at least one of cutting power, spindle speed, feed rate, and depth of cut.

3. The method for predicting the remaining service life of a cutting tool according to claim 1, characterized in that, The step of extracting features from the multimodal sensing signals using the end-to-end prediction model to obtain the spatiotemporal features of each modality includes: Each mode of the multimodal sensing signal is taken as the signal to be processed, and the local temporal dynamic features corresponding to the signal to be processed are extracted by the multi-scale feature extraction network in the end-to-end prediction model. Calculate the statistical features corresponding to the signal to be processed, wherein the statistical features include the peak factor and / or root mean square. The local temporal dynamic features and the statistical features are aggregated and spliced ​​together along the channel dimension to obtain spliced ​​features; Based on the splicing features, the spatiotemporal features corresponding to the signal to be processed are obtained through the long short-term memory network in the end-to-end prediction model.

4. The method for predicting the remaining service life of a cutting tool according to claim 1, characterized in that, The step of obtaining the latent space vector of the operating condition based on the operating condition parameters includes: The operating condition parameters are input into the operating condition encoder in the end-to-end prediction model, and the operating condition latent space vector is obtained by mapping through the operating condition encoder. The operating condition encoder is a multilayer sensor network; the operating condition parameters include at least one of spindle speed, feed rate, and depth of cut.

5. The method for predicting the remaining service life of a cutting tool according to claim 1, characterized in that, The step of decoupling the spatiotemporal features based on the latent space vector of the working condition to obtain decoupling features decoupled from the working condition of the cutting tool to be predicted includes: Based on the working condition latent space vector, the scaling factor and translation factor are obtained by mapping through the linear mapping fully connected network in the end-to-end prediction model. For each of the aforementioned spatiotemporal features, the mean and standard deviation corresponding to the spatiotemporal features are calculated in the spatial dimension, so as to standardize the spatiotemporal features based on the mean and standard deviation. Based on the scaling factor and the translation factor, an affine transformation is performed on the standardized spatiotemporal features to obtain the decoupled features.

6. The method for predicting the remaining service life of a cutting tool according to claim 1, characterized in that, The regression prediction based on the decoupling features, to output the predicted remaining tool life value corresponding to the cutting tool to be predicted, includes: For each of the decoupling features, global average pooling is performed, and the reliability confidence level corresponding to the decoupling feature is determined based on the features after global average pooling. Based on the reliability confidence level, the decoupling features are weighted to obtain the weighted features of each mode; The multi-head self-attention module in the end-to-end prediction model is used to fuse the weighted features of each modality to obtain multi-modal fusion features. The multimodal fusion features are input into the regression module in the end-to-end prediction model to obtain the tool remaining life prediction value, wherein the regression module is composed of a multi-layer nonlinear fully connected network.

7. The method for predicting the remaining service life of a cutting tool according to claim 6, characterized in that, The end-to-end prediction model is trained according to the following steps: Obtain a training dataset, wherein the training dataset is constructed based on at least one sample tool, and for each sample tool, there are multiple training sample groups corresponding to different time steps in the training dataset. A training sample group includes a set of corresponding multimodal signal samples, working condition parameter samples, tool remaining life labels and working condition labels. Input the multimodal signal samples and working condition parameter samples from the training dataset into the end-to-end prediction model to be trained, and obtain the multimodal fusion features and tool remaining life prediction values ​​obtained by the end-to-end prediction model. Based on the multimodal fusion features, the predicted working condition classification result is determined by a domain discriminator; The domain alignment loss is determined based on the predicted working condition classification result and the working condition label. The regression loss is determined based on the predicted tool remaining life value and the tool remaining life label. The monotonic trend loss is determined based on the predicted tool remaining life value and the tool remaining life label corresponding to the training sample groups of the same tool at different time steps. The domain alignment loss is used to constrain the multimodal fusion feature from being independent of the working condition category. The regression loss is used to constrain the consistency between the predicted tool remaining life value and the tool remaining life label. The monotonic trend loss is used to constrain the direction of change of the predicted tool remaining life value over time to be consistent with the direction of change of the tool remaining life label over time. The model parameters of the end-to-end prediction model are adjusted according to the domain alignment loss, the regression loss, and the monotonic trend loss. Then, the process of inputting the multimodal signal samples and working condition parameter samples from the training dataset into the end-to-end prediction model to be trained, and obtaining the multimodal fusion features and tool remaining life prediction values ​​obtained by the end-to-end prediction model is returned to the end until the preset training termination condition is met, and the trained end-to-end prediction model is obtained.

8. A tool remaining service life prediction system, characterized in that, The system includes: The data acquisition module is used to acquire the multimodal sensing signals and operating parameters corresponding to the cutting tool to be predicted, wherein the multimodal sensing signals include cutting force signals, vibration signals and acoustic emission signals; The prediction module is used to input the multimodal sensing signals and the operating conditions parameters into a trained end-to-end prediction model, extract features from the multimodal sensing signals through the end-to-end prediction model to obtain the spatiotemporal features of each mode, obtain the operating condition latent space vector based on the operating conditions parameters, decouple the spatiotemporal features based on the operating condition latent space vector to obtain decoupling features decoupled from the operating conditions of the cutting tool to be predicted, and perform regression prediction based on the decoupling features to output the predicted remaining tool life value corresponding to the cutting tool to be predicted.

9. A terminal, characterized in that, The terminal includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the tool remaining service life prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the tool remaining service life prediction method as described in any one of claims 1 to 7.

Images