CN-121901856-B - Robot joint module stage sensing fault diagnosis and life prediction method

CN121901856BCN 121901856 BCN121901856 BCN 121901856BCN-121901856-B

Abstract

The invention relates to a robot joint module stage-aware fault diagnosis and life prediction method, which comprises the steps of converting an original vibration signal of a robot joint module in a whole life cycle into a semantic health representation describing health state evolution, taking the semantic health representation as input, taking structural priori knowledge including a robot joint module degradation stage, a degradation trend and physical boundary constraint as output, constructing a large language model-based degradation reasoning model, taking the structural priori knowledge as input, degradation stage identification and residual life prediction as output, constructing a deep neural network-based predictor, dynamically adjusting training loss weights in different degradation stages and constructing physical heuristic losses based on the structural priori knowledge in a training process, fusing the training losses and physical information constraint losses into total losses, and carrying out joint training and optimization on model parameters. The invention enhances the prediction accuracy, physical consistency and robustness under complex working conditions.

Inventors

LIANG ZHONGWEI
LI ZHAO
Zhuang jichao
HUANG SIRUI
CHEN XIN
ZHANG CAIXIA

Assignees

广州大学

Dates

Publication Date: 20260512
Application Date: 20260320

Claims (10)

1. The life prediction method for the robot joint module stage perception is characterized by comprising the following steps of: Collecting an original vibration signal of the whole life cycle of a robot joint module, and converting the original vibration signal into semantic health representation describing health state evolution after signal processing; Constructing a big language model-based degradation reasoning model, which takes the semantic health representation as input to generate a structured priori knowledge containing degradation stages, degradation trends and physical boundary constraints of the robot joint module; Constructing a predictor based on a deep neural network, taking the structured priori knowledge as input, and generating degradation stage identification and residual life prediction results; The structure of the predictor comprises a GRU encoder, a degradation stage classification head and an RUL regression head, wherein the GRU encoder extracts time sequence degradation characteristics of vibration signals to obtain degradation embedding; The predictor is provided with a stage perception self-adaptive learning module which dynamically adjusts training loss function weights of the predictor in different degradation stages based on the degradation stages in the structured priori knowledge so as to balance stage classification loss and RUL regression loss in the training loss; the training loss and the physical heuristic loss are weighted and fused to form total loss, and model parameters are jointly trained and optimized by using a back propagation algorithm to form the predictor; The physical heuristic penalty consists of three constraints related to degradation characteristics, including: , wherein: physical heuristic penalty for time point t; Loss of constraint for monotonicity for each constraint strength , And Predicted remaining life values for time points t and t-1, respectively, degradation rate constraints , , Is the upper bound of degradation rate, degradation acceleration constraint , , For the predicted remaining life value at time point t-2, Representing the degradation phase Is a predicted degradation acceleration level of (a).
2. The method of claim 1, wherein the calculation of training loss function weights comprises: , In the formula, Training loss function weight for time point t, which is the first Probability of degradation of individual degradation phases Is used as a means for controlling the speed of the vehicle, Outputting by the degradation stage classification head; the method is characterized in that the method is a preset stage basic weight, and K is the total number of degradation stages.
3. The method according to claim 2, characterized in that the degradation phase The expected degradation acceleration level of (a) is defined as follows: , wherein early, middle, late represents the early, mid and late degradation stages of degradation, respectively.
4. A method according to claim 3, wherein when k=3, the preset phase basis weights are: , Wherein, the To characterize that as the degradation degree increases, the residual life regression analysis obtains higher weights, and k=1, 2,3 corresponds to early, late, respectively.
5. The method of claim 1, wherein the constructing of the semantic health representation comprises: performing time-frequency analysis on the acquired original vibration signal segment, and extracting a group of degradation sensitive characteristics from the time-frequency representation; extracting a plurality of degradation sensitive characteristics at the time t of an original vibration signal segment to form a multidimensional time-frequency characteristic vector, and normalizing; the multi-dimensional time-frequency characteristic vector is fused into a single health index by adopting principal component analysis To (3) pair Performing first order difference to obtain , 、 Respectively represent time 、 The health index value at the location, Representing the change rate of the health index, constructing a health index sequence 、 ; Extracting a set of statistical features from a vibration signal segment to provide supplemental degradation information, time Statistical feature vector at RMS, kurtosis, skewness, crest, entropy represents root mean square, kurtosis, skewness, crest factor and signal entropy respectively; Determining an operating condition vector , 、、 Respectively represent time Rotational speed, load level, temperature at; Constructing a semantic health representation based on the health index sequence, the statistical feature vector and the operation condition vector 。
6. The method of claim 5, wherein the degradation-sensitive characteristic comprises time-frequency energy, spectrum centroid, spectrum bandwidth, high frequency energy ratio.
7. The method of claim 1, wherein the constructing of the degenerate reasoning model comprises: Constructing a prompt template, converting the semantic health representation into a natural language description input pre-trained large language model, and outputting the parsed structured priori knowledge , Representing the inferred degradation phase early, middle, late represents the early, mid and late degradation phases of degradation, respectively; Shows degradation trend, stable, moderate, accelerating shows stable degradation trend, moderate degradation trend and accelerated degradation trend respectively, Represents an inferred constraint related to degradation characteristics that is embedded as a soft regularization term into the learning objective of the loss function.
8. The method of claim 7, wherein the upper bound of degradation rate Adaptively determining based on the degradation trend inferred by the large language model: 。
9. The method of claim 1, wherein the GRU encoder comprises an input projection layer, a time encoding layer and a global degradation embedding layer, wherein the input projection layer is an input end of the GRU encoder, the time encoding layer comprises a plurality of layers of GRU units, and the global degradation embedding layer takes a hidden state of a last layer of GRU units in a last time step as a degradation embedding vector.
10. The method of claim 9, wherein the RUL regression head performs regression based on the degenerate embedding vector to output a residual life prediction value: , in the training process, the predicted output value of the residual life Is constrained to be non-negative, ensuring physical rationality; Representing a hidden layer activation function; 、 representing the weight and bias of the RUL regression head output layer; representing the degenerate embedding vector, , Representing weights and offsets of the RUL regression header hidden layer; the degradation stage classification head includes a fully connected layer with a ReLU activation function and a softmax output layer.

Description

Robot joint module stage sensing fault diagnosis and life prediction method Technical Field The invention relates to the technical field of machine learning-based mechanical key component state monitoring and health management, in particular to a robot joint module stage-aware fault diagnosis and service life prediction method. Background The robot joint module is used as a key component of industrial machinery, and the running state of the robot joint module directly determines the reliability and safety of the whole equipment system. The high-precision transmission and dynamic response capability of the robot joint module are the basis of flexible assembly of the robot, and once the component fails due to unplanned shutdown, the component not only causes high economic loss, but also can cause serious production safety accidents. Therefore, the development of the residual life prediction and fault degree diagnosis research of the robot joint module is a core task for realizing the construction of the equipment preventive maintenance and health management system. Along with the development of industrial big data technology, the existing method for predicting the residual life and diagnosing the fault degree of the robot joint module mainly focuses on the application of a deep learning model (such as CNN, transformer). Such methods aim to extract degradation features from a mass of vibration signals by constructing complex nonlinear mapping structures. However, the existing data driving method still faces the following bottlenecks in practical application: first, there is a lack of stage adaptability. The degradation process of the robot joint module typically exhibits significant stepwise heterogeneous characteristics, such as slow degradation in the early stages of operation and an exponentially accelerating degradation trend in the late stages of failure. The existing deep learning model mostly adopts a static training objective function, and is difficult to dynamically perceive different degradation stages in which the model is positioned in the model training process, so that the prediction precision in the key rapid degradation stage is reduced. Second, physical consistency is difficult to guarantee. The neural network model driven by pure data is regarded as a 'black box' model, and the prediction result of the neural network model is often only dependent on signal pattern matching, and the constraint of a physical rule is lacked. In the absence of explicit degenerate physical equations, the life prediction results output by the model may violate basic physical evolution logic, reducing the reliability of the prediction results. Third, high-level knowledge integration is difficult. Although domain experts have profound prior knowledge and reasoning logic for robot joint module degradation, how to effectively integrate these high-level, structured degradation reasoning knowledge into the underlying signal processing and end-to-end neural predictors is still lacking in efficient technical means. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a robot joint module stage-aware fault diagnosis and life prediction method, which aims to solve the problem that the prediction accuracy, physical consistency and robustness of the prior art under complex working conditions are limited. The technical scheme adopted by the invention is as follows: The invention provides a robot joint module stage-aware fault diagnosis and life prediction method, which comprises the following steps: Collecting an original vibration signal of the whole life cycle of a robot joint module, and converting the original vibration signal into semantic health representation describing health state evolution after signal processing; Constructing a big language model-based degradation reasoning model, which takes the semantic health representation as input to generate a structured priori knowledge containing degradation stages, degradation trends and physical boundary constraints of the robot joint module; Constructing a predictor based on a deep neural network, taking the structured priori knowledge as input, and generating degradation stage identification and residual life prediction results; The structure of the predictor comprises a GRU encoder, a degradation stage classification head and an RUL regression head, wherein the GRU encoder extracts time sequence degradation characteristics of vibration signals to obtain degradation embedding; The predictor is provided with a stage perception self-adaptive learning module which dynamically adjusts training loss function weights of the predictor in different degradation stages by utilizing the degradation stages in the structured priori knowledge so as to balance stage classification loss and RUL regression loss in the training loss; The training loss and the physical information constraint loss are weighted and fused to form total loss, and model paramete