CN-121980346-A - Zero sample learning-based tool damage state monitoring method, equipment and medium

CN121980346ACN 121980346 ACN121980346 ACN 121980346ACN-121980346-A

Abstract

The invention provides a tool damage state monitoring method, equipment and medium based on zero sample learning, wherein the method comprises the steps of generating a high-dimensional semantic vector according to word vectors corresponding to tool-damage structured text descriptions and graph embedding vectors corresponding to tool-damage knowledge maps; the method comprises the steps of generating a multi-mode sensing feature vector according to data of a tool to be detected in a cutting process, mapping the multi-mode sensing feature vector into a semantic space corresponding to a high-dimensional semantic vector, generating a virtual sample under the condition that the high-dimensional semantic vector corresponding to a sample of an invisible damage type is used as a condition, training a zero sample classifier according to the virtual sample, determining a dynamic threshold according to the zero sample classifier, and carrying out real-time early warning monitoring on the tool to be detected based on the dynamic threshold, wherein the prior knowledge of the tool, namely semantic attribute and data of a known damage mode, can be utilized to realize high-precision identification of the unknown damage type, so that the adaptability of a monitoring system under a variable processing environment is improved.

Inventors

LUO CUI
ZHU BEIBEI
QIN LIN
XU YING
LIN HAI
LI HANYANG
LI YUAN
WU YUHANG

Assignees

上海航天控制技术研究所

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (10)

1. The tool breakage state monitoring method based on zero sample learning is characterized by comprising the following steps of: Step S100, generating a high-dimensional semantic vector according to a word vector corresponding to a preset cutter-damaged structured text description and a graph embedding vector corresponding to a preset cutter-damaged knowledge graph; Step 200, carrying out data preprocessing and feature fusion on the collected three-way vibration signal, three-way force signal and cutter rear face image of the cutter to be detected in the cutting process to generate a multi-mode sensing feature vector; Step S300, mapping the multi-mode sensing feature vector to a semantic space corresponding to the high-dimensional semantic vector; step S400, generating corresponding virtual samples based on a preset improved generation countermeasure network on the condition that high-dimensional semantic vectors corresponding to samples of a preset invisible damage class; s500, constructing a zero sample classifier based on a semantic self-encoder, and training the zero sample classifier according to the virtual samples; And step 600, determining a dynamic threshold according to the zero sample classifier, and carrying out real-time early warning monitoring on the tool to be detected based on the dynamic threshold.
2. The method according to claim 1, wherein step S100 comprises: step S110, generating a structured text description for each corresponding cutter-damage category according to the type of the cutter to be detected; step S120, extracting word vectors of the structured text description corresponding to each cutter-damage category through a preset deep bidirectional pre-training language model; step S130, constructing a cutter-damage knowledge graph comprising cutter attributes, damage mode entities and correlations; Step S140, extracting a graph embedding vector of a damaged mode node from the cutter-damaged knowledge graph through a preset graph rolling network; And step S150, splicing the word vector and the graph embedded vector, and performing principal component analysis dimension reduction processing to generate a high-dimensional semantic vector.
3. The method according to claim 2, wherein the step S150 includes: step S151, splicing the word vector corresponding to each cutter-damage category and the graph embedding vector in a first dimension to obtain a fusion vector corresponding to each cutter-damage category; Step S152, stacking fusion vectors corresponding to all cutter-damage categories into a data matrix, and carrying out standardization processing on each characteristic dimension of the data matrix to obtain a standardized data matrix; Step 153, calculating a covariance matrix of the standardized data matrix, and performing feature decomposition on the covariance matrix to obtain a corresponding feature value and a feature vector; Step S154, sorting the feature values according to descending order, and determining the feature values positioned in the preset target quantity as target feature values in the sorted feature values; and step S155, projecting the fusion vector corresponding to each cutter-damage category after the standardization processing to a subspace formed by a plurality of feature vectors corresponding to the target feature values, so as to obtain a high-dimensional semantic vector corresponding to each cutter-damage category.
4. A method according to claim 3, wherein said step S200 comprises: step S210, fixing a three-way acceleration vibration sensor on the surface of a workpiece or a workbench through a magnetic seat, and enabling three sensitive axes of the three-way acceleration vibration sensor to be aligned with three axes of a coordinate system of a processing machine tool of a tool to be detected; step S220, a multichannel force sensor is installed in a spindle taper hole of the processing machine tool, wherein a three-way force sensing unit is integrated inside the multichannel force sensor; Step S230, fixing an industrial camera on an external observation window of a processing cavity through a rigid support, so that a lens of the industrial camera is opposite to a tool to be detected, which is arranged on a main shaft of the processing machine tool; Step S240, in the cutting process of the tool to be detected, carrying out wavelet packet decomposition on the three-way vibration signals acquired by the three-way acceleration vibration sensor and the three-way force signals acquired by the multi-channel force sensor to obtain sensing feature vectors; Step S250, in the cutting process of the to-be-detected tool, global average pooling feature extraction is carried out on the tool rear surface image of the to-be-detected tool acquired by the industrial camera so as to obtain an image feature vector; and step S260, respectively carrying out maximum and minimum value normalization on the sensing feature vector and the image feature vector, and then splicing to obtain the multi-mode sensing feature vector.
5. The method according to claim 4, wherein the step S300 includes: Step S310, constructing a three-layer semantic mapping network which adopts a three-layer fully-connected neural network as a mapping network, wherein the dimension of an output layer of the three-layer semantic mapping network is the same as the dimension of a semantic space corresponding to the high-dimensional semantic vector; step S320, using a mean square error loss function to minimize and map Euclidean distance between the output vector of the three-layer semantic mapping network and the high-dimensional semantic vector; Step S330, training the three-layer semantic mapping network for a plurality of training periods by adopting a visible broken type sample; And step 340, mapping the multi-mode sensing feature vector to a semantic space corresponding to the high-dimensional semantic vector according to the trained three-layer semantic mapping network.
6. The method according to claim 5, wherein the step S400 includes: Step S410, constructing a generated countermeasure network, wherein the generated countermeasure network comprises a generator and a discriminator, the generator comprises four fully connected layers, and the discriminator comprises three fully connected layers; step S420, training the generator and the discriminator according to a preset improved loss function until the generated countermeasure network reaches a Nash equilibrium state; Step S430, inputting the high-dimensional semantic vector corresponding to the invisible damage type sample in the sample data set into the trained generation countermeasure network to generate a corresponding virtual sample.
7. The method according to claim 6, wherein the step S500 includes: step S510, training a semantic self-encoder on a high-dimensional semantic vector corresponding to a sample of a visible damage type in a sample data set; step S520, inputting the virtual sample into the three-layer semantic mapping network to obtain corresponding mapping characteristics; Step S530, encoding and decoding the mapping features through the semantic self-encoder to obtain a reconstructed semantic vector; And S540, calculating cosine similarity of the reconstructed semantic vector and high-dimensional semantic vectors corresponding to samples of all damage categories, and taking the damage category with the maximum cosine similarity as a prediction result so as to train a zero sample classifier of the semantic self-encoder.
8. The method according to claim 7, wherein the step S600 includes: Step S610, inputting a multi-mode sensing feature vector corresponding to a tool to be detected in a cutting process into the zero sample classifier to obtain a corresponding prediction confidence coefficient output by the zero sample classifier; And step 620, if the prediction confidence is lower than a preset dynamic threshold, determining that the cutter to be detected is abnormal or unknown damaged.
9. A non-transitory computer readable storage medium having stored therein at least one instruction or at least one program loaded and executed by a processor to implement the method of any one of claims 1-8.
10. An electronic device comprising a processor and the non-transitory computer readable storage medium of claim 9.

Description

Zero sample learning-based tool damage state monitoring method, equipment and medium Technical Field The invention relates to the technical field of intelligent manufacturing and equipment state monitoring, in particular to a method, equipment and medium for monitoring the damage state of a cutter based on zero sample learning. Background In precision numerical control machining, a tool is used as a key component for directly executing a cutting task, and the health state of the tool directly influences machining quality, efficiency and cost. Tool breakage is one of the main failure modes, and if not detected in time, the tool can lead to workpiece rejection, machine tool damage and even safety accidents. Currently, the mainstream tool breakage monitoring methods can be divided into two types, namely a method based on a physical model and a method based on data driving. The latter, especially deep learning models such as convolutional neural networks CNN (Convolutional Neural Network), are of great interest due to their strong feature learning capabilities. However, these methods have inherent limitations as follows: The supervision learning model needs to provide hundreds of labeling samples for each damage mode, and in the actual industrial scene, some serious damage (such as cutter breakage) samples are easy to obtain, but micro damage (such as micro tipping and thermal cracking) samples are rare, so that a balanced training set is difficult to construct; The generalization capability is poor, namely, when the trained model faces to a tool type which does not appear in the training set or a brand new damage mode, the recognition performance is rapidly reduced, and the requirement of frequent replacement of the tool in the flexible manufacturing system cannot be met; the multi-mode information is not utilized enough, that most researches only adopt a single type of sensor signal, such as a vibration signal or a force signal, and the complementary value of the multi-source information such as vision, force, vibration and the like cannot be fully fused. Zero sample learning provides a solution to the above-described problem by introducing "semantic attributes" to migrate knowledge of known classes to unknown classes. However, in the field of the strong physical background of tool monitoring, how to construct an effective semantic space, how to fuse multi-mode sensing information and how to guarantee the real-time performance and the robustness of a model on an industrial site remain key technical problems to be solved. Disclosure of Invention Aiming at the technical problems, the invention adopts the following technical scheme: According to one aspect of the present application, there is provided a tool breakage state monitoring method based on zero sample learning, including: Step S100, generating a high-dimensional semantic vector according to a word vector corresponding to a preset cutter-damaged structured text description and a graph embedding vector corresponding to a preset cutter-damaged knowledge graph; Step 200, carrying out data preprocessing and feature fusion on the collected three-way vibration signal, three-way force signal and cutter rear face image of the cutter to be detected in the cutting process to generate a multi-mode sensing feature vector; step S300, mapping the multi-mode sensing feature vector into a semantic space corresponding to the high-dimensional semantic vector; step S400, generating corresponding virtual samples based on a preset improved generation countermeasure network on the condition that high-dimensional semantic vectors corresponding to samples of a preset invisible damage class; S500, constructing a zero sample classifier based on a semantic self-encoder, and training the zero sample classifier according to the virtual samples; And step S600, determining a dynamic threshold according to the zero sample classifier, and carrying out real-time early warning monitoring on the tool to be detected based on the dynamic threshold. According to another aspect of the present application, there is provided a non-transitory computer readable storage medium having stored therein at least one instruction or at least one program loaded and executed by a processor to implement the aforementioned zero sample learning based tool breakage condition monitoring method. According to yet another aspect of the present application, there is provided an electronic device comprising a processor and the aforementioned non-transitory computer-readable storage medium. The invention has at least the following beneficial effects: according to the zero sample learning-based tool breakage state monitoring method, firstly, a word vector corresponding to a tool-breakage structured text description and a graph embedding vector corresponding to a tool-breakage knowledge graph are used for generating a high-dimensional semantic vector, and data preprocessing and feature fusion are carried out on a three