CN-121997255-A - Model training method, state and residual life prediction method, device and medium

CN121997255ACN 121997255 ACN121997255 ACN 121997255ACN-121997255-A

Abstract

The present disclosure provides a model training method, a state and remaining life prediction method, a device and a medium, comprising obtaining multi-sensor time sequence data; the method comprises the steps of carrying out feature extraction on multi-sensor time sequence data through a first branch to obtain space-frequency fusion features, carrying out bidirectional state capturing on the space-frequency fusion features through a second branch to obtain a bidirectional hidden state sequence, carrying out historical state association on the bidirectional hidden state sequence through a third branch to obtain an attention weight set, distributing attention weights to all states in the bidirectional hidden state sequence based on the attention weight set, carrying out feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through a fourth branch to predict the belonging state and the residual service life of sample equipment, and training a large language model based on the belonging state and the residual service life of the sample equipment by adopting a preset loss function to obtain an equipment state and residual service life prediction model. Therefore, equipment fault detection and service life prediction are effectively realized.

Inventors

ZHANG XIAOWEI
CHEN TONG
Luan Xinrui
LIU YING
ZHANG YIMENG
Luo Jiashuo
DONG YUCAI
YIN FEI
DONG WENTAO
Kong Zining
XIAO LONGBIN
DU JIAN
CUI WEI
LIN YUANYUAN
ZHANG SHITAI

Assignees

中国电子科技集团公司第十五研究所

Dates

Publication Date: 20260508
Application Date: 20260112

Claims (10)

1. The device state and residual life prediction model training method is characterized by being applied to a large language model, wherein the large language model comprises a first branch, a second branch, a third branch and a fourth branch, wherein the first branch is used for extracting spatial features and frequency domain features of a device, and the second branch is used for capturing a forward hidden state and a backward hidden state in the running process of the device through bidirectional time sequence modeling; The method comprises the following steps: Acquiring multi-sensor time sequence data of sample equipment; extracting characteristics of multi-sensor time sequence data of the sample equipment through the first branch to obtain space-frequency fusion characteristics; the second branch is used for capturing the bidirectional state of the space-frequency fusion characteristic to obtain a bidirectional hidden state sequence; historical state association is carried out on the bidirectional hidden state sequence through the third branch, so that an attention weight set for representing different states is obtained; Distributing attention weights to all states in the bidirectional hidden state sequence based on the attention weight set so as to update the weight of the bidirectional hidden state sequence; performing characteristic compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch, and predicting the belonging state and the residual service life of the sample equipment; And training the large language model by adopting the preset loss function based on the belonging state and the residual service life of the sample equipment to obtain an equipment state and residual service life prediction model.
2. The method according to claim 1, wherein the feature extraction of the multi-sensor time series data of the sample device by the first branch to obtain a space-frequency fusion feature comprises: performing multi-scale feature extraction on multi-sensor time sequence data of the sample equipment through depth separable convolution; Capturing a local transformation mode in the multi-scale features through a spatial feature extraction unit to obtain the spatial features, and extracting features of different frequency components in the multi-scale features through a multi-channel convolution kernel in a frequency domain feature extraction unit to obtain the frequency domain features; And fusing the spatial features and the frequency domain features to obtain the space-frequency fusion features.
3. The method according to claim 1, wherein said performing bidirectional state capturing on the space-frequency fusion feature by the second branch to obtain a bidirectional hidden state sequence includes: Capturing performance degradation trend of the sample equipment in the history operation process from the space-frequency fusion characteristic through a forward LSTM to obtain a forward hidden state; identifying hysteresis response and abnormal mutation of the sample equipment in the history operation process from the space-frequency fusion characteristics through a backward LSTM to obtain a backward hidden state; And splicing the forward hiding state and the backward hiding state to obtain the bidirectional hiding state sequence.
4. The method according to claim 1, wherein said feature compressing and nonlinear transforming the updated bi-directional hidden state sequence through the fourth branch, predicting the state and remaining service life of the sample device, comprises: Performing feature compression and nonlinear transformation on the updated bidirectional hidden state sequence through a full connection layer to obtain nonlinear transformation features; Performing classification judgment on the nonlinear transformation characteristics to obtain the state of the sample equipment; And carrying out regression prediction on the nonlinear transformation characteristics to obtain the residual service life of the sample equipment.
5. The method of claim 1, wherein the acquiring multi-sensor timing data of the sample device comprises: Collecting high-frequency time sequence data through a plurality of types of sensors deployed on the sample equipment; and performing wavelet transformation and feature standardization processing on the high-frequency time sequence data to obtain multi-sensor time sequence data of the sample equipment.
6. A device state and remaining life prediction method, comprising: Acquiring multi-sensor time sequence data of target equipment; inputting the multi-sensor time sequence data of the target equipment into an equipment state and residual service life prediction model to obtain the belonging state and residual service life of the target equipment; Wherein the device state and remaining life prediction model is generated by training the method of any one of claims 1-5.
7. The method as recited in claim 6, further comprising: Determining a fault damage level of the target equipment based on the state of the target equipment and the residual service life; and generating health state evaluation and equipment maintenance suggestions of the target equipment based on the fault damage level of the target equipment.
8. The device state and residual life prediction model training device is characterized by being applied to a large language model, wherein the large language model comprises a first branch, a second branch, a third branch and a fourth branch, wherein the first branch is used for extracting spatial features and frequency domain features of equipment, and the second branch is used for capturing a forward hidden state and a backward hidden state in the running process of the equipment through bidirectional time sequence modeling; The device comprises: The first acquisition module is used for acquiring multi-sensor time sequence data of the sample equipment; the extraction module is used for extracting characteristics of multi-sensor time sequence data of the sample equipment through the first branch to obtain space-frequency fusion characteristics; the capturing module is used for capturing the bidirectional state of the space-frequency fusion characteristic through the second branch to obtain a bidirectional hidden state sequence; The association module is used for carrying out historical state association on the bidirectional hidden state sequence through the third branch to obtain attention weight sets for representing different states; The distribution module is used for distributing attention weights to all states in the bidirectional hidden state sequence based on the attention weight set so as to update the weight of the bidirectional hidden state sequence; The prediction module is used for carrying out characteristic compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch, and predicting the belonging state and the residual service life of the sample equipment; And training the large language model by adopting the preset loss function based on the belonging state and the residual service life of the sample equipment to obtain an equipment state and residual service life prediction model.
9. A device state and remaining life prediction apparatus, comprising: the second acquisition module is used for acquiring multi-sensor time sequence data of the target equipment; The determining module is used for inputting the multi-sensor time sequence data of the target equipment into the equipment state and residual service life prediction model to obtain the belonging state and residual service life of the target equipment; Wherein the device state and remaining life prediction model is generated by training the method of any one of claims 1-5.
10. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when executed by a processor, implements the device state and remaining life prediction model training method according to any one of claims 1 to 5, or implements the device state and remaining life prediction method according to claim 6 or 7.

Description

Model training method, state and residual life prediction method, device and medium Technical Field The embodiment of the disclosure relates to the technical field of artificial intelligence, in particular to a model training method, a state and residual life prediction device and a medium. Background The working efficiency of the large-caliber equipment is limited by multiple factors, and in an extreme working environment, under the combined action of a high-temperature high-pressure working condition and a high impact load, the system has obvious squat load and mechanical vibration, so that the fatigue damage accumulation rate of key parts exceeds the expected rate, and finally the problems of precision index deviation and the like are caused. In the related technology, the current equipment guarantee system generally adopts a maintenance strategy based on historical emission frequency and wearing part replacement period, and the mode cannot effectively evolve dynamic damage in a complex environment. When working condition variables such as high-strength countermeasure, cross-region rapid transition and the like are met, the traditional mode is difficult to effectively realize equipment fault detection and service life prediction. Disclosure of Invention Embodiments described herein provide a model training method, state and remaining life prediction method, apparatus, and medium that overcome the above-described problems. According to the first aspect, the invention provides a device state and residual life prediction model training method, which is applied to a large language model, wherein the large language model comprises a first branch, a second branch, a third branch and a fourth branch, the first branch is used for extracting spatial features and frequency domain features of a device, and the second branch is used for capturing a forward hidden state and a backward hidden state in the running process of the device through bidirectional time sequence modeling; The method comprises the following steps: Acquiring multi-sensor time sequence data of sample equipment; extracting characteristics of multi-sensor time sequence data of the sample equipment through the first branch to obtain space-frequency fusion characteristics; the second branch is used for capturing the bidirectional state of the space-frequency fusion characteristic to obtain a bidirectional hidden state sequence; historical state association is carried out on the bidirectional hidden state sequence through the third branch, so that an attention weight set for representing different states is obtained; Distributing attention weights to all states in the bidirectional hidden state sequence based on the attention weight set so as to update the weight of the bidirectional hidden state sequence; performing characteristic compression and nonlinear transformation on the updated bidirectional hidden state sequence through the fourth branch, and predicting the belonging state and the residual service life of the sample equipment; And training the large language model by adopting the preset loss function based on the belonging state and the residual service life of the sample equipment to obtain an equipment state and residual service life prediction model. In a second aspect, according to the present disclosure, there is provided a device status and remaining life prediction method, including: Acquiring multi-sensor time sequence data of target equipment; inputting the multi-sensor time sequence data of the target equipment into an equipment state and residual service life prediction model to obtain the belonging state and residual service life of the target equipment; Wherein the device state and remaining life prediction model are generated by training the method of the first aspect. According to the third aspect, the device state and residual life prediction model training device is applied to a large language model, wherein the large language model comprises a first branch, a second branch, a third branch and a fourth branch, the first branch is used for extracting spatial features and frequency domain features of a device, and the second branch is used for capturing a forward hidden state and a backward hidden state in the running process of the device through bidirectional time sequence modeling; The device comprises: The first acquisition module is used for acquiring multi-sensor time sequence data of the sample equipment; the extraction module is used for extracting characteristics of multi-sensor time sequence data of the sample equipment through the first branch to obtain space-frequency fusion characteristics; the capturing module is used for capturing the bidirectional state of the space-frequency fusion characteristic through the second branch to obtain a bidirectional hidden state sequence; The association module is used for carrying out historical state association on the bidirectional hidden state sequence through the third bran