CN-122020115-A - Deep learning-based infrared optical equipment refrigeration life prediction method

CN122020115ACN 122020115 ACN122020115 ACN 122020115ACN-122020115-A

Abstract

The invention discloses a deep learning-based infrared optical equipment refrigeration life prediction method, and aims to solve the problem that a traditional method is difficult to characterize a complex nonlinear degradation track under multi-factor coupling of a refrigeration system. The method collects multi-source sensor data such as vibration, cold finger temperature, input current and compression cavity outer wall temperature, captures forward and reverse time sequence association through a multi-channel two-way long and short term memory network (Bi-LSTM), focuses a degradation key time node by combining a time sequence attention mechanism, adaptively fuses multi-source information weights by utilizing a channel attention mechanism, and finally outputs a residual life prediction result through a full-connection layer. Experiments show that the method effectively breaks through the traditional evaluation limitation, improves the prediction accuracy, stability and reliability, and provides key technical support for intelligent operation and maintenance and full life cycle management of infrared detection equipment.

Inventors

ZHAO XIN
FANG HONGZHENG
LI RUI
XIONG YI
YU JIAHAO
HUANG YULONG
CHEN FEI

Assignees

北京航天测控技术有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (6)

1. A method for predicting the refrigerating life of infrared optical equipment based on deep learning is characterized by comprising the following steps: 1) Collecting multi-source sensor monitoring data of an infrared optical equipment refrigerating system, and respectively carrying out deep feature extraction and sensitive feature screening on the monitoring data of each sensor channel to obtain a sensitive feature sequence corresponding to each channel; 2) Respectively inputting the sensitive characteristic sequences into corresponding two-way long-short-term memory network units for time sequence modeling, dynamically weighting different time step hidden states output by the two-way long-short-term memory network units through a time sequence attention mechanism, and fusing historical time sequence information to obtain advanced characteristic characterization of each channel; 3) The method comprises the steps of representing advanced features of all channels into a channel attention mechanism, adaptively distributing weights of all channels through the channel attention mechanism, highlighting key channel contributions, realizing adaptive integration of multi-source features, and outputting comprehensive feature vectors fused with multi-source information; 4) And inputting the comprehensive feature vector into a fully-connected neural network to obtain a residual life prediction result of the refrigerating system of the infrared optical equipment.
2. The method according to claim 1, characterized in that: The bidirectional long-short-term memory network unit adopts a bidirectional circulating structure, and input sequences are processed in the forward direction and the reverse direction respectively through two independent hidden layers, and bidirectional output is spliced to obtain hidden states so as to capture context information before and after.
3. The method according to claim 1, characterized in that: the time step weight output by the time sequence attention mechanism The method is characterized by comprising the following steps: Wherein, the Is a hidden state of the two-way long-short-period memory network unit after being spliced at the t moment; Is that A representation obtained after passing through a single-layer perceptron and a tanh () activation function; And Is the weight matrix and the bias matrix of the single-layer perceptron; is a one-dimensional feature vector which is introduced and initialized randomly, uses the more meaningful information in each moment to be measured in an adaptive way, and finally outputs the weight of the importance degree of the hidden state at each moment 。
4. The method according to claim 1, characterized in that: The working process of the channel attention mechanism comprises the steps of taking output characteristics of each channel processed by the time sequence attention mechanism as input, respectively carrying out information aggregation through three parallel operations of global average pooling, global maximum pooling and full connection layers, generating three one-dimensional feature vectors v, m and n with the same dimensions as the number of sensor channels, respectively inputting the vectors v and m into a multi-layer sensor mining channel with a hidden layer for association, splicing the output of the vector v and the vector n, then carrying out activation function processing to obtain channel attention weight beta, and carrying out weighted summation on the output characteristics of each channel by utilizing the weight beta to generate a comprehensive feature vector fused with multi-source sensitive information.
5. The method according to claim 1, characterized in that: the number of the channels is at least 3.
6. The method according to any one of claims 1-5, wherein: the multisource sensor monitoring data comprise vibration signals, cold finger temperature signals, input current signals and compression cavity outer wall temperature signals; the sensitive characteristic sequences of the channels are corresponding to a spectrum kurtosis and margin index time sequence of a vibration signal, a cold finger temperature average time sequence, an input current average time sequence and a compression cavity outer wall temperature average time sequence.

Description

Deep learning-based infrared optical equipment refrigeration life prediction method Technical Field The invention relates to the field of life prediction of refrigeration type infrared detection equipment, in particular to a refrigeration system prediction technology integrating deep learning and multi-source sensing data. Background The infrared detection equipment is used as a core component of the high-precision detection system, and the performance of the infrared detection equipment directly determines the detection precision and the working efficiency of the system. The refrigerating type infrared detection equipment cools the infrared detector to the working temperature by a Stirling refrigerator or liquid helium refrigeration mode through a deep refrigeration technology, so that dark current and thermal noise of the detector are obviously restrained, and the thermal sensitivity and detection distance of the system are greatly improved. The characteristic enables the device to still realize efficient target detection and tracking under the complex electromagnetic environment and low illumination condition, and the device becomes a key component in a modern high-end photoelectric system. However, similar to all high precision opto-electromechanical systems, infrared detection devices face significant challenges in lifetime and reliability during long-term operation. As a key link for influencing the service life of the whole machine, the refrigerating system is easy to generate degradation phenomena such as mechanical abrasion, material fatigue, refrigerant purity reduction or micro leakage and the like under long-term storage and cyclic working conditions, so that the performance of the refrigerating system is gradually attenuated. One typical appearance of this degradation process is characterized by an extension of the "cool down time", i.e., the time required for the system to reach the nominal operating temperature from start-up to the detector. The prolonged refrigeration time not only affects the response speed of the system, but also directly reflects the degradation of the internal health state. If no intervention is carried out, the refrigeration capacity is continuously reduced, the key performance parameters of the detector are deteriorated, and finally the system function failure is caused. Therefore, the degradation process of the refrigerating time of the refrigerating type infrared detection equipment is monitored and analyzed, an accurate prediction method of the residual service life of the refrigerating type infrared detection equipment is established, the continuous availability of the system can be ensured through the accurate identification of the fault precursors, and technical support is provided for intelligent operation and maintenance and full life cycle management of the optical-mechanical-electrical system. For a long time, life assessment for such high value devices has relied primarily on analysis based on physical failure models and accelerated life testing. Although the traditional method lays a foundation for early reliability engineering, the limitation of the traditional method is gradually highlighted when the traditional method is used for coping with a complex system of optical, mechanical, electric and thermal multi-physical field coupling inside the detection equipment. The accelerated life test consumes a large amount of samples under severe conditions, has high cost and long period, and is difficult to adapt to the equipment evaluation requirement of rapid iteration. In addition, the theoretical model is mostly based on simplification and assumption of an actual system, and microcosmic failure mechanisms such as microcosmic abrasion of a piston ring of a refrigerator, microcosmic pollution of a refrigerant and the like are difficult to accurately describe. Therefore, it is difficult to achieve accurate "health diagnosis" and "life prediction" for a particular in-service detection device with conventional methods. In recent years, the development of deep learning technology provides a brand new technical path for solving the above problems. As a data driving method, deep learning can automatically learn complex rules from historical operation data without relying on an accurate physical model. In this research context, a mapping model from data to lifetime can be constructed by analyzing correlations between a large number of refrigeration process data and lifetime results. The advantage of deep learning is mainly characterized in that the method has strong automatic feature extraction capability and excellent sequence modeling capability. The cyclic neural network (such as a long-term and short-term memory network) can effectively capture the time sequence dependency relationship formed by the starting data of the past time, so that the overall trend and the dynamic evolution rule of the system performance degradation are understood. The local feature extract