Search

CN-122020567-A - Multi-condition adaptive residual service life prediction method for numerical control machine tool

CN122020567ACN 122020567 ACN122020567 ACN 122020567ACN-122020567-A

Abstract

The invention discloses a multi-working condition adaptation residual service life prediction method for a numerical control machine tool. The method comprises the steps of firstly carrying out degradation correlation screening and time sequence sample construction on multi-sensor monitoring signals, dividing a data set, then constructing a degradation trend perception feature extraction network to achieve unified feature coding of multi-source working condition data, designing a multi-branch migration prediction framework on the basis, achieving cross-working condition modeling through a shared feature coding and branch decoupling structure, further introducing an opposite type feature distribution alignment mechanism to reduce data distribution difference between different working conditions, simultaneously providing a multi-scale target degradation consistency constraint mechanism, effectively retaining specific degradation information of the target working conditions by establishing a comparison consistency relation between an input space and a feature space, constructing a source domain contribution self-adaptive collaborative optimization mechanism, carrying out dynamic weighting on contributions of different source working conditions in a prediction task, and finally achieving high-precision prediction of the residual service life of a cutter.

Inventors

  • JIN YI
  • ZHOU HAIGANG

Assignees

  • 南京震环智能装备有限公司

Dates

Publication Date
20260512
Application Date
20260410

Claims (9)

  1. 1. The multi-condition adaptive residual service life prediction method for the numerical control machine tool is characterized by comprising the following steps of: Step S1, monitoring signals in the cutter machining process are collected through a sensor and preprocessed, so that a multi-working-condition adaptive residual service life prediction data set of the cutter is constructed; s2, constructing a multi-branch migration prediction model based on a degradation trend perception feature extraction network; S3, inputting the multi-working condition adaptive residual service life prediction data set of the cutter obtained in the step S1 into a multi-branch migration prediction model; s4, training a multi-branch migration prediction model based on source domain contribution self-adaptive collaborative optimization mechanism in batches by combining and optimizing multi-scale target degradation consistency constraint loss, counterloss and root mean square prediction loss; And S5, after training is finished, only preserving a degradation trend perception feature extraction network and a life regression prediction sub-network in the multi-branch migration prediction model for predicting the residual life of the online tool.
  2. 2. The method for predicting the residual service life of the tool for the numerical control machine tool according to claim 1, wherein the step S1 is specifically as follows: S11, acquiring a cutter vibration signal, a main shaft current signal and an acoustic emission signal in the cutter machining process through a sensor; Step S12, carrying out degradation correlation screening on each monitoring signal, and only retaining time-varying signals which change monotonically along with the tool wear evolution; Step S13, respectively carrying out normalization processing on the screened monitoring signals according to the processing working conditions, adopting a time sequence sample constructed by a sliding time window, and taking the residual service life of the cutter corresponding to the tail end moment of each time window as a label of the window; step S14, carrying out data set division on the preprocessed signal data, taking the data under one working condition as a target domain, taking 20% of the data as a verification set to participate in training, taking the rest 80% as a test set to be used for testing, and taking the data under the rest working conditions as a multi-source domain to participate in training.
  3. 3. The method for predicting the residual service life of the tool for the numerical control machine tool according to claim 1, wherein the step S2 is specifically as follows: step S21, constructing a degradation trend perception feature extraction network: The method comprises a local degradation mode sensing unit for capturing short-time wear mutation characteristics, a long-sequence related modeling unit for modeling degradation dependency relationship across time scales, and a characteristic compression mapping unit for reducing characteristic redundancy and forming compact representation, wherein the three units are sequentially connected in cascade and are used for extracting unified characteristic representation containing local degradation modes and long-term evolution trend from a time sequence; Step S22, constructing a multi-branch migration prediction model based on a degradation trend perception feature extraction network: each source working condition and each target working condition form independent migration branches, and all the migration branches share the degradation trend perception characteristic extraction network in the step S21 and respectively comprise independent domain distribution discrimination sub-networks and life regression prediction sub-networks.
  4. 4. The method for predicting the remaining service life of multi-working condition adaptation for a tool of a numerically-controlled machine tool according to claim 3, wherein in the step S21: The local degradation mode sensing unit adopts a one-dimensional convolutional neural network structure and comprises 3 layers of one-dimensional convolutional layers, the convolution kernel size of each layer is 16, 3 and 3, the number of convolutional channels is 3, 64 and 64, each layer of convolutional is sequentially connected with a batch normalization layer and a nonlinear activation function ReLU, and feature downsampling is realized through a maximum pooling layer; The long sequence related modeling unit adopts a transducer encoder based on a self-attention mechanism, the self-attention head number is 8, and the hidden layer dimension is 32-256; the characteristic compression mapping unit adopts a full-connection network structure and comprises 3 full-connection layers, the number of neurons in each layer is 32, 128 and 256, and a nonlinear activation function LeakyReLU is connected after each full-connection layer.
  5. 5. The method for predicting the residual service life of the tool-oriented multi-working condition adaptation of the numerical control machine tool according to claim 3, wherein the multi-branch migration prediction model in the step S22 adopts a strategy of shared characterization and hard parameter isolation, so that each migration branch shares the characteristics extracted by the degradation trend perception characteristic extraction network, but the domain discrimination sub-network and the prediction sub-network parameters unique to each migration branch are not shared, so as to simultaneously maintain cross-domain consistency and source domain specificity.
  6. 6. The method for predicting the remaining service life of multi-working condition adaptation for a tool of a numerically-controlled machine tool according to claim 3, wherein in the step S22: The domain distribution discrimination sub-network comprises 3 layers of full-connection layers, the number of neurons in each layer is 256, 64 and 2, and an output layer adopts Sigmoid for discriminating whether input features come from a source domain or a target domain; The life regression prediction sub-network comprises 2 full-connection layers, the number of neurons in each layer is 256 and 1, and the life regression prediction sub-network is used for outputting the residual service life prediction value of the cutter.
  7. 7. The method for predicting the residual service life of the tool for the numerical control machine tool according to claim 1, wherein the step S4 is specifically as follows: Step S41, executing countermeasure training in each migration branch to align the distribution of the source domain features and the target domain features in the discrimination space, wherein the feature update generated in the countermeasure training process is limited by the consistency constraint in step S42; Step S42, constructing multi-scale target degradation consistency constraint loss in the characteristic distribution alignment process in step S41, and inhibiting the loss of specific degradation information of a target domain in the distribution alignment process by calculating the consistency constraint loss between the input of the target domain and the deep characteristic representation of the target domain on 3 scales; And S43, constructing a source domain contribution self-adaptive collaborative optimization mechanism aiming at the contribution difference of different source working conditions to the target prediction task, taking the consistency constraint loss, the antagonism loss and the root mean square prediction loss obtained in the step S42 together as optimization targets, introducing weight parameters for each source working condition, and updating the weights by adopting a group search strategy in the optimization process.
  8. 8. The multi-tool adaptive residual life prediction method for numerically controlled machine tools according to claim 7, wherein the multi-scale target degradation consistency constraint loss in step S42 is Based on contrast learning construction, the method concretely comprises the following steps: (1); wherein X u and X v represent the ith target domain input and the ith target domain feature, respectively, Z w represents data spliced by the ith target domain input and the ith target domain feature, sim (·) represents cosine similarity, B represents batch size, s represents current scale, For the set of scales, I (·) is the indication function.
  9. 9. The multi-working condition adaptive residual service life prediction method for the numerically-controlled machine tool according to claim 7, wherein the group search strategy in the step S43 adopts an improved fox search strategy, specifically comprising the following steps: Step S431, initializing a weight population by utilizing Latin hypercube sampling; step S432, carrying out iterative updating on weights through a dual-mode search strategy, wherein the iterative updating comprises jump global search and random disturbance local search; Step S433, adopting a random reset strategy to the out-of-range weight parameter to maintain the diversity of the solution space; step S434, introducing a dynamic elite retention mechanism, the elite ratio thereof The method meets the following conditions: (2); Wherein the method comprises the steps of For the current number of iterations, The maximum iteration number; and S435, updating and optimizing the weight by taking the root mean square prediction loss on the verification data as an adaptability function, so as to realize the self-adaptive adjustment of the contribution degree of the multi-source information to the target task.

Description

Multi-condition adaptive residual service life prediction method for numerical control machine tool Technical Field The invention relates to the field of prediction of residual service life of a cutter of a numerical control machine tool, in particular to a multi-working condition adaptive residual service life prediction method for the cutter of the numerical control machine tool. Background In the machining process of a numerical control machine tool, accurate prediction of the residual service life of a cutter is important. The method can not only pre-judge the failure point of the cutter, effectively avoid the reduction of machining precision and the scrapping of batch workpieces, ensure the quality and the process stability of products, but also remarkably reduce the unplanned shutdown, equipment damage and safety accidents caused by the breakage of the sudden cutter, thereby directly reducing the production operation cost and the maintenance risk. Meanwhile, the prediction supports scientific tool replacement strategy optimization, is beneficial to controlling resource waste caused by excessive warehouse preparation, prolongs the effective service cycle of the tools, and is a foundation stone for realizing efficient, safe and low-cost intelligent manufacturing. The traditional cutter state monitoring method mainly relies on manual experience judgment or regular shutdown detection, has the problems of strong subjectivity, poor real-time performance and the like, and is difficult to meet the requirements of modern intelligent manufacturing on online monitoring and prediction. In recent years, the deep learning method is widely applied in the field of residual service life prediction, and by automatically extracting complex degradation characteristics, the deep learning method shows better performance than the traditional method in terms of modeling accuracy and generalization capability. For example, the convolutional neural network can effectively extract local degradation characteristics, the cyclic neural network and the variant thereof can model time sequence dependency, and the model based on the attention mechanism further improves long-sequence modeling capability. However, the above-described methods are typically based on the assumption that the training data is consistent with the test data distribution, and rely on sufficient annotation data for training. In an actual industrial scene, the machining working condition of the numerical control machine tool is complex and changeable, and different cutting parameters and workpiece materials can cause the distribution of the monitoring data to change remarkably, so that a model trained under a single working condition is difficult to be directly applied to a new working condition environment. Meanwhile, the cost for acquiring the annotation data covering the complete life cycle is high, and the practical application of the model is further limited. In order to solve the problems, researchers begin to explore a residual service life prediction method based on domain adaptation so as to realize knowledge migration among different working conditions. However, most of the existing methods are designed aiming at single-source field scenes, and have obvious defects under multi-source working conditions, namely, on one hand, the correlation between different source working condition data and target working conditions is different, the existing methods generally adopt equal-weight or fixed-weight strategies for fusion, so that the high-correlation source field information is difficult to fully utilize, even negative migration influence caused by low-correlation data is introduced, on the other hand, when characteristic distribution alignment is carried out, the maintenance of special degradation information of the target working conditions is often ignored, so that key degradation characteristics are weakened, and in addition, the existing characteristic extraction method lacks special modeling capability aiming at multi-scale time sequence characteristics of a tool degradation process, so that the complex degradation mode is difficult to fully describe. Therefore, how to realize effective knowledge migration under the multi-source working condition and simultaneously consider the source domain contribution difference and the target degradation information retention becomes a technical problem to be solved urgently. The Chinese patent with the publication number of CN111832624A discloses a tool residual life prediction method based on anti-migration learning, which comprises the steps of collecting data of cutting processes of different types of tools, determining historical type tool samples and new type tool samples, constructing a historical type feature extraction model and a nonlinear regression model by using the data of the historical type tool samples, performing contrast domain adaptation on the data of the historical type tool samples and the da