Infrared optical equipment refrigeration life prediction method based on deep learning

Abstract

The invention discloses an infrared optical equipment refrigeration life prediction method based on deep learning, and aims to solve the problem that a traditional method is difficult to represent a complex nonlinear degradation track of a refrigeration system under multi-factor coupling. According to the method, multi-source sensor data such as vibration, cold finger temperature, input current and compression cavity outer wall temperature are collected, after screening and feature extraction, forward and reverse time sequence correlation is captured through a multi-channel bidirectional long-short-term memory network (Bi-LSTM), degeneration key time nodes are focused in combination with a time sequence attention mechanism, and a degeneration algorithm is established. And adaptively fusing multi-source information weights by using a channel attention mechanism, and finally outputting a residual life prediction result through a full connection layer. Experiments show that the method effectively breaks through the limitation of traditional evaluation, improves the accuracy, stability and reliability of prediction, and provides key technical support for intelligent operation and maintenance and full-life-cycle management of infrared detection equipment.

Description

Technical Field

[0001] This invention relates to the field of lifespan prediction for cooled infrared detection equipment, and more specifically to a prediction technology for cooled systems based on deep learning and multi-source sensor data fusion. Background Technology

[0002] As a core component of high-precision detection systems, the performance of infrared detection equipment directly determines the system's detection accuracy and operational efficiency. Cooled infrared detection equipment utilizes deep cooling technology, typically employing Stirling refrigerators or liquid helium cooling to cool the infrared detector to its operating temperature, thereby significantly suppressing dark current and thermal noise, and greatly improving the system's thermal sensitivity and detection range. This characteristic enables efficient target detection and tracking even in complex electromagnetic environments and low-light conditions, making it a key component in modern high-end optoelectronic systems. However, similar to all high-precision opto-mechatronic systems, infrared detection equipment faces severe challenges to lifespan and reliability during long-term operation. The cooling system, as a critical component affecting the overall lifespan, is prone to degradation under long-term storage and cyclic operation conditions, including mechanical wear, material fatigue, decreased refrigerant purity, or micro-leakage, leading to a gradual decline in performance. A typical external manifestation of this degradation process is the increase in "cooling time," i.e., the time required for the system to reach its rated operating temperature from startup gradually increases. The extended cooling time not only affects the system's response speed but also directly reflects the degradation of its internal health. Without intervention, the cooling capacity will continue to decline, key performance parameters of the detector will deteriorate, and ultimately the system will fail. Therefore, monitoring and analyzing the degradation process of the cooling time of cooled infrared detection equipment and establishing an accurate prediction method for its remaining service life can ensure the continuous availability of the system through accurate identification of fault precursors, providing technical support for the intelligent operation and maintenance and full life cycle management of opto-mechatronic systems.

[0003] For a long time, life assessment of such high-value equipment has mainly relied on physical failure model-based analysis and accelerated life testing. While traditional methods laid the foundation for early reliability engineering, their limitations have become increasingly apparent when dealing with complex systems involving the coupling of multiple physical fields (optics, mechanics, electronics, and heat) within detection equipment. Accelerated life testing requires consuming a large number of samples under harsh conditions, resulting in high costs and long cycles, making it difficult to adapt to the rapidly iterative needs of equipment evaluation. Furthermore, theoretical models are often based on simplifications and assumptions about actual systems, making it difficult to accurately describe microscopic failure mechanisms such as microscopic wear of piston rings in refrigerators and trace amounts of refrigerant contamination. Therefore, traditional methods struggle to achieve accurate "health diagnosis" and "life prediction" for specific detection equipment in service.

[0004] In recent years, the development of deep learning technology has provided a new technical path for solving the above problems. As a data-driven method, deep learning can "automatically learn complex patterns" from historical operating data without relying on precise physical models. In this research context, a mapping model from data to lifetime can be constructed by analyzing the correlation between a large amount of refrigeration process data and lifetime results. The advantages of deep learning are mainly reflected in two aspects: it has powerful automatic feature extraction capabilities and excellent sequence modeling capabilities. Recurrent neural networks (such as long short-term memory networks) can effectively capture the temporal dependencies formed by the data from each startup, thereby understanding the overall trend and dynamic evolution of system performance degradation. Combining local feature extraction with long-term trend modeling lays the technical foundation for accurate prediction of the remaining lifetime of refrigeration systems.

[0005] In summary, traditional methods for assessing the lifespan of infrared detection equipment cooling systems rely primarily on physical failure model analysis and accelerated life testing. However, their inherent limitations make them ill-suited to the precise operation and maintenance needs of modern high-tech equipment. On one hand, accelerated life testing requires a large number of samples, incurs high costs, and is time-consuming, failing to keep pace with the rapid iteration of equipment assessments. On the other hand, physical models are often based on simplifications and assumptions about actual systems, making it difficult to accurately characterize complex failure mechanisms such as piston ring micro-wear and refrigerant contamination under the coupling of multiple physical fields (optical, mechanical, electrical, and thermal) within the cooling system. This results in the inability to achieve precise individual-level health diagnosis and remaining lifespan prediction for equipment in service. Although the development of deep learning technology has provided a new technical path for data-driven lifespan prediction, it still requires targeted innovation in key areas such as modeling complex degradation trajectories of cooling systems and effectively fusing multi-source sensor information.

[0006] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art. Summary of the Invention

[0007] In view of this, this invention proposes a deep learning-based method for predicting the cooling lifespan of infrared optical devices. The aim is to predict the lifespan of infrared detection device cooling systems under conditions of complex physical mechanisms and variable operating environments, thereby solving the degradation problem of current infrared detection device cooling systems. The specific technical solution is as follows: 1. A method for predicting the cooling lifetime of infrared optical devices based on deep learning, comprising the following steps: 1) Collect multi-source sensor monitoring data of the infrared optical equipment cooling system, perform deep feature extraction and sensitive feature screening on the monitoring data of each sensor channel, and obtain the sensitive feature sequence corresponding to each channel; 2) Input the sensitive feature sequences into the corresponding bidirectional long short-term memory network units for temporal modeling. Dynamically weight the hidden states at different time steps output by the bidirectional long short-term memory network units through a temporal attention mechanism, and fuse historical temporal information to obtain high-level feature representations for each channel. 3) Input the high-level feature representations of all channels into the channel attention mechanism, and through the channel attention mechanism, adaptively allocate the weights of each channel and highlight the contributions of key channels to achieve adaptive integration of multi-source features and output a comprehensive feature vector that integrates multi-source information; 4) Input the comprehensive feature vector into a fully connected neural network to obtain the remaining lifetime prediction result of the infrared optical device cooling system.

[0008] Furthermore, the bidirectional long short-term memory network unit adopts a bidirectional recurrent structure, which processes the input sequence in the forward and reverse directions through two independent hidden layers, and splices the bidirectional outputs to obtain the hidden state, so as to capture forward and backward context information.

[0009] Furthermore, the time step weights output by the temporal attention mechanism The details are as follows:

[0010]

[0011] in, This is the hidden state of the bidirectional long short-term memory network units after being spliced ​​together at time t; yes The representation obtained after passing through a single-layer perceptron and the tanh() activation function; and These are the weight matrix and bias matrix of a single-layer perceptron; It introduces randomly initialized one-dimensional feature vectors to adaptively measure more meaningful information at each time step, and finally outputs the weights of the importance of the hidden states at each time step. .

[0012] Furthermore, the working process of the channel attention mechanism is as follows: taking the output representation of each channel after processing by the temporal attention mechanism as input, information is aggregated through three parallel operations: global average pooling, global max pooling, and fully connected layers, to generate three one-dimensional feature vectors v, m, and n with the same dimensions as the number of sensor channels; vectors v and m are respectively input into a multilayer perceptron with hidden layers to mine channel associations, and their outputs are concatenated with vector n and processed by an activation function to obtain the channel attention weight β; the output representations of each channel are weighted and summed using the weight β to generate a comprehensive feature vector that integrates multi-source sensitive information.

[0013] Furthermore, the number of channels is at least three.

[0014] Furthermore, the multi-source sensor monitoring data includes vibration signals, cold finger temperature signals, input current signals, and compression cavity outer wall temperature signals; the sensitive feature sequences of each channel correspond to: the spectral kurtosis and margin index time series of the vibration signal, the average cold finger temperature time series, the average input current time series, and the average compression cavity outer wall temperature time series.

[0015] Beneficial effects 1. This invention uses a bidirectional long short-term memory network to construct the core of time series modeling. It processes the input sequence in the forward and reverse directions through two independent hidden layers, and simultaneously captures the forward and reverse time series correlation of "history-current-future" in the monitoring data. It breaks through the limitation of traditional unidirectional LSTM that can only use historical information. It effectively solves the technical problem that traditional methods are difficult to characterize complex nonlinear degradation trajectories under the combined action of multiple factors such as multi-physics coupling, material aging and mechanical wear. It achieves accurate description of the degradation process of refrigeration system. 2. This invention introduces a temporal attention mechanism, which is unique in that it can dynamically evaluate the importance of the hidden states of all time steps of the bidirectional LSTM output and adaptively focus on key time nodes that are strongly correlated with system degradation (such as the moment of performance change and the stage of accelerated degradation). This overcomes the defects of traditional recurrent neural networks, such as the "information bottleneck" and the easy loss of key degradation details in the early stage. As a result, the method performs well in the early identification and evolution trend prediction of the performance degradation of the refrigeration system, and significantly improves the stability and reliability of the prediction results. 3. This invention innovatively adopts a channel attention mechanism, using three parallel and independent information aggregation methods—global average pooling, global max pooling, and fully connected layers—to mine the response characteristics of multi-source sensor data from different dimensions. Then, a multilayer perceptron is used to mine channel correlations and adaptively allocate contribution weights for each channel. This not only strengthens the response characteristics of channels sensitive to degradation but also suppresses redundant interference. It solves the problems of traditional multi-source information fusion lacking specificity and relying on manual weight setting, achieving a leap from statistical lifetime assessment to accurate individual prediction. Simultaneously, it overcomes the limitations of traditional physical models, such as high cost, long cycle time, and reliance on simplified assumptions. This provides key technical support for intelligent health management of high-value optoelectronic equipment, improving system readiness and operational efficiency, and optimizing the entire lifecycle cost. Attached Figure Description

[0016] Figure 1 LSTM cell structure; Figure 2 A schematic diagram of a bidirectional LSTM structure; Figure 3 A schematic diagram of the temporal attention mechanism; Figure 4Schematic diagram of the channel attention mechanism structure; Figure 5 A schematic diagram of the prediction model of this invention; Figure 6 Flowchart for predicting the lifespan of a refrigeration system; Figure 7 Data on cooling time of infrared detection equipment before aging; Figure 8 Data on cooling time after aging of infrared detection equipment; Figure 9 Example Prediction Model; Figure 10 , 1 Results of the lifespan prediction method for the No. 1 refrigerator based on the deep integration of long short-term memory network and dual attention mechanism; Figure 11 , 2 Results of the lifespan prediction method for refrigerator No. 1 based on the deep integration of long short-term memory network and dual attention mechanism. Detailed Implementation

[0017] The following will provide a detailed explanation from two aspects: principle analysis and specific implementation.

[0018] The lifespan prediction method based on long short-term memory networks combined with dual attention mechanisms mainly consists of three parts: long short-term memory (LSTM) networks, temporal attention mechanisms, and channel attention mechanisms.

[0019] (a) Long Short-Term Memory Network As a variant of the traditional recurrent neural network, LSTM improves upon the traditional RNN structure by introducing a gating mechanism to control the flow of information within neurons, thus mitigating the problems of gradient vanishing and gradient exploding. It can process variable-length time data and overcome the limitation of traditional RNNs, which can only handle short time series.

[0020] The specific structure of the LSTM unit is as follows: Figure 1 As shown. Among them, This represents the input information of the network unit at time t; , , and These represent the hidden state and memory unit at the corresponding time, respectively.

[0021] LSTM cells are further divided into the following 5 main substructures: (1) Temporary memory unit

[0022] (1.1) in, , It is the weight matrix in the cell. Let be the activation function. From the above equation, we can obtain... The value of is determined by the input at time t. and Hidden state of time Determined comprehensively.

[0023] (2) Input gate

[0024] Input gate It is the first gating structure of the LSTM unit, which determines how the input information of the neuron at the current time step is retained in the memory unit at the current time step. The degree of, that is, through and Control temporary memory unit The degree of retention is as follows: (1.2) in, It is the sigmoid activation function.

[0025] (3) Gate of Oblivion

[0026] Forgotten Gate It is the second gate structure of the LSTM unit, according to and control Time-based memory units The degree of retention at the current moment. The calculation process is as follows: (1.3) 4) Memory unit

[0027] memory unit This allows for the long-term retention of information. The input gate is calculated... With the Gate of Oblivion After obtaining the value, use and Temporary memory units and By weighting and merging the memory units at different times, a memory unit can be obtained. .

[0028] (1.4) 5) Output gate

[0029] Output gate It is used to calculate the output value of network units. The calculation process is shown in the following formula: (1.5) Depend on and To determine the memory unit The degree to which information is retained. The specific calculation process for the final output value is as follows: (1.6) In complex equipment condition monitoring, the monitoring signals exhibit strong time dependence, meaning that the system state at a given moment is not only related to past history but also interdependent with future short-term evolution trends. While standard LSTM units can effectively utilize historical information, their unidirectional structure cannot capture this forward-backward temporal correlation. Therefore, this invention introduces a bidirectional recurrent structure based on LSTM to construct a prediction network. This structure processes the input sequence from both the forward and backward directions through two independent hidden layers, thereby simultaneously capturing information from both past and future contexts, significantly improving the completeness of state representation and prediction accuracy. The network structure is as follows: Figure 2 As shown.

[0030] bidirectional LSTM output The concatenation of the outputs from two hidden layers is shown in the following formula: (1.7) (II) Attention Mechanism As research has deepened, attention mechanisms have gradually become one of the core components of neural networks, and have been successfully applied in various fields such as image recognition, natural language processing, speech recognition, fault diagnosis, and remaining life prediction. Its core idea originates from the human visual attention mechanism, which guides the model to dynamically assign appropriate weights to different parts of the input information, thereby highlighting key features, suppressing redundancy and noise interference, and ultimately improving the model's expressive and generalization abilities.

[0031] 1) Temporal attention mechanism Although the hidden states of a recurrent neural network aggregate global information about the sequence in the final moments, directly using them as the network output would create an information bottleneck, leading to the loss of crucial details from earlier parts of the sequence. To overcome this limitation, this invention introduces a temporal attention mechanism. This mechanism dynamically evaluates the importance of the hidden states at all time steps to the current task, thereby adaptively focusing on the temporal features most relevant to the system degradation process, achieving more comprehensive information utilization. The temporal attention mechanism is as follows: Figure 3 As shown.

[0032] Weights in temporal attention mechanisms The calculation is as follows: (1.8) (1.9) in, It is the hidden state of the bidirectional LSTM after splicing at time t; yes The representation obtained after passing through a single-layer perceptron and the tanh() activation function; and These are the weight matrix and bias matrix of a single-layer perceptron; It introduces randomly initialized one-dimensional feature vectors to adaptively measure more meaningful information at each time step, and finally outputs the weights of the importance of the hidden states at each time step. .

[0033] 2) Channel attention mechanism To effectively fuse the output representations of multi-channel networks and achieve adaptive integration of multi-source sensor information, this invention introduces a channel attention mechanism at the network endpoint. The structure of this module is as follows: Figure 4 As shown.

[0034] The process of the channel attention mechanism is as follows: Global average pooling, global max pooling, and fully connected layers are used to aggregate the output features of each channel, generating three one-dimensional feature vectors equal to the number of channels, denoted as follows: v、m and n These three vectors characterize the response properties of each channel from different perspectives, and their calculation process is shown in the following formula: (1.10) (1.11) (1.12) In the formula, i The i-th element of the one-dimensional feature for each channel. I It is the size represented by the output of each channel. j Number of sensor channels , It is the first j Bias and weights of the fully connected channel layer It is the first j Output representation of the channel.

[0035] Will v and m The inputs are fed into a multilayer perceptron with one hidden layer to obtain the hidden relationships between different channels, and the degradation information contained in the output representation of each channel is evaluated. The output results are then compared with... n The components are concatenated and then processed by an activation function. Obtain the channel attention weights ,for: (1.13) in, , , and It is the weight matrix of two multilayer perceptrons.

[0036] Using channel attention weights The output representations of each channel are weighted and summed to obtain the fused representation. Z ,Right now: (1.14) This invention introduces a channel attention mechanism, which can effectively enhance feature responses sensitive to device degradation processes while suppressing interference from irrelevant information. After processing using this mechanism, a high-level feature vector fusing multi-source information is generated.

[0037] The specific process of the prediction method based on the deep integration of long short-term memory networks and dual attention mechanisms is as follows: (1) Deep feature extraction and sensitive feature screening are performed on the monitoring data of each sensor channel; (2) The sensitive features obtained from each channel are input into the corresponding long short-term memory network units, and the hidden states at different time steps are dynamically weighted using the temporal attention mechanism to effectively integrate historical temporal information; (3) Adaptively integrate high-level feature representations from multiple source sensors through channel attention mechanism to highlight the contribution of key channels; (4) Input the fused integrated feature vector into a fully connected neural network to achieve accurate prediction of remaining lifespan.

[0038] The method and process are as follows Figure 5 As shown.

[0039] The following detailed descriptions are exemplary and intended to provide further illustration of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0040] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0041] Where there is no conflict, the embodiments and features described herein can be combined with each other.

[0042] To achieve reliable condition monitoring and accurate lifespan prediction of the cooling system of infrared detection equipment, acquiring key monitoring data that effectively characterizes its health status is a primary prerequisite. This embodiment uses a Stirling refrigerator degradation dataset as the basis for analysis. This dataset systematically records multi-source sensor signals from multiple refrigerators during the degradation process, including information such as vibration, cold finger temperature, input current, and compression chamber outer wall temperature. Among these, cold finger temperature, input current, and compression chamber outer wall temperature are typical slowly varying signals, with their amplitudes remaining relatively stable within a single sampling period, and can be considered characteristic parameters characterizing the system degradation process.

[0043] The experimental procedure for life prediction of refrigeration systems is as follows: Figure 6 As shown, it specifically includes the following steps.

[0044] (1) Data Acquisition: Taking the cooling time data of an infrared detection device as an example, the following explanation is provided. During normal use or bench testing of the infrared detection device, time-series data is collected in real time for each cooling start-up; such as... Figure 7 , 8 The data show the cooling time of the infrared detection equipment before and after aging. It can be observed that, for the same temperature drop from room temperature to the operating temperature of -188℃, the cooling time required after aging is significantly longer. When the cooling time exceeds the specified requirements after aging to a certain extent, it can be considered a failure of the cooling component of the detection equipment.

[0045] (2) Data filtering and cleaning: that is, removing data with format errors, abnormal amplitudes or uncollected data, and standardizing the data; (3) Feature extraction: Feature extraction and screening are performed on the monitoring signals of different channels. Among them, the average value of the compression cavity outer wall temperature, input current and cold finger temperature signals, which are slowly changing signals, is taken as their feature index at each sampling time. Finally, the feature index corresponding to each channel signal is obtained, and the feature set is constructed. The cooling start point and the end point of reaching the stable working temperature are automatically identified, and the cooling time, cooling rate, steady-state temperature difference and other indicators are calculated, and the cooling performance feature vector is constructed. (4) After completing the deep feature extraction and sensitive feature screening of the monitoring data of each sensor channel, a life prediction model of the cooling system of the infrared detection equipment was constructed. The model consists of 4 channels, and the feature sets input to each channel are: the time series of spectral kurtosis and margin index of vibration signal, the time series of average input current, the time series of average outer wall temperature, and the time series of average cold finger temperature; (5) Input the selected vibration signal kurtosis and margin index time series, input current mean time series, outer wall temperature mean time series, and cold finger temperature mean time series into the corresponding LSTM units for time series modeling. The LSTM outputs the hidden state sequence of all time steps, as shown in Equation (1.7). This sequence is a deep encoding representation of the input features, containing long-term dependencies and time pattern information; (6) Inputting the hidden state sequence into the time attention mechanism will assign different weights to the hidden states at different time steps in the sequence, as shown in Equation (1.9), thereby highlighting the more critical moments or stages for lifetime prediction and ultimately generating a weighted feature representation. (7) The feature vectors of the four channels, weighted by temporal attention, are concatenated and used as the input to the channel attention mechanism. The channel attention mechanism analyzes the features of each channel and assigns corresponding weights according to their importance. The weight calculation and weighted fusion process is shown in Equation (1.13). The output is a fused feature vector highlighting the key channel information; (8) The output feature vectors of the four independent channels after processing by their respective "temporal attention mechanisms" are concatenated and used as the input of the channel attention mechanism. The channel attention mechanism analyzes the features of these four channels and assigns corresponding weights according to their importance as shown in Equation (1.13), and then outputs a weighted fused feature vector as shown in Equation (1.13). This output vector fuses the information of all channels, but highlights the channels that contribute more. (9) Input the fused integrated feature vector into a fully connected neural network to achieve accurate prediction of remaining lifespan.

[0046] In the model validation experiment, refrigerators No. 1 and No. 2 in the dataset were selected as the test set, and the data of the remaining equipment were used as the training set. For the remaining life prediction task, feature extraction and screening were first performed, and then the extracted features were input into a four-channel bidirectional long short-term memory neural network. The main hyperparameter combinations of the model were determined through grid search, as shown in the table below.

[0047] Table 1-1 Computation time when data points are randomly arranged

[0048] After completing all hyperparameter settings, the predicted remaining lifespan of chillers 1 and 2 is as follows: Figure 10 , 11 As shown.

[0049] Analysis results show that, from the initial degradation point (i.e., when the performance parameters of the refrigeration unit first exceed the preset degradation threshold), the prediction method proposed in this invention can accurately capture the performance degradation pattern of the refrigeration unit, and its predicted trajectory matches the actual lifespan curve. This indicates that the method has excellent degradation trend learning ability and lifespan prediction accuracy, thus providing an effective technical basis for lifespan prediction and predictive maintenance of refrigeration systems.

Claims

1. A method for predicting the cooling lifetime of infrared optical devices based on deep learning, characterized in that: 1) Collect multi-source sensor monitoring data of the infrared optical equipment cooling system, perform deep feature extraction and sensitive feature screening on the monitoring data of each sensor channel, and obtain the sensitive feature sequence corresponding to each channel; 2) Input the sensitive feature sequences into the corresponding bidirectional long short-term memory network units for temporal modeling. Dynamically weight the hidden states at different time steps output by the bidirectional long short-term memory network units through a temporal attention mechanism, and fuse historical temporal information to obtain high-level feature representations for each channel. 3) Input the high-level feature representations of all channels into the channel attention mechanism, and through the channel attention mechanism, adaptively allocate the weights of each channel and highlight the contributions of key channels to achieve adaptive integration of multi-source features and output a comprehensive feature vector that integrates multi-source information; 4) Input the comprehensive feature vector into a fully connected neural network to obtain the remaining lifetime prediction result of the infrared optical device cooling system.

2. The method according to claim 1, characterized in that: The bidirectional long short-term memory network unit adopts a bidirectional recurrent structure. It processes the input sequence in the forward and backward directions through two independent hidden layers, and splices the bidirectional outputs to obtain the hidden state in order to capture forward and backward context information.

3. The method according to claim 1, characterized in that: The time step weights output by the temporal attention mechanism The details are as follows: in, This is the hidden state of the bidirectional long short-term memory network units after being spliced ​​together at time t; yes The representation obtained after passing through a single-layer perceptron and the tanh() activation function; and These are the weight matrix and bias matrix of a single-layer perceptron; It introduces randomly initialized one-dimensional feature vectors to adaptively measure more meaningful information at each time step, and finally outputs the weights of the importance of the hidden states at each time step. .

4. The method according to claim 1, characterized in that: The channel attention mechanism works as follows: The output representations of each channel after processing by the temporal attention mechanism are used as input. Information is aggregated through three parallel operations: global average pooling, global max pooling, and a fully connected layer, generating three one-dimensional feature vectors v, m, and n, each with the same dimensions as the number of sensor channels. Vectors v and m are then input into a multilayer perceptron with hidden layers to mine channel associations. Their outputs are concatenated with vector n and processed by an activation function to obtain channel attention weights β. The weights β are then used to weight and sum the output representations of each channel to generate a comprehensive feature vector that integrates multi-source sensitive information.

5. The method according to claim 1, characterized in that: There are at least three channels.

6. The method according to any one of claims 1-5, characterized in that: The multi-source sensor monitoring data includes vibration signals, cold finger temperature signals, input current signals, and compression chamber outer wall temperature signals. The sensitive feature sequences of each channel correspond to: the time series of spectral kurtosis and margin indices of the vibration signal, the time series of the average cold index temperature, the time series of the average input current, and the time series of the average temperature of the outer wall of the compression cavity.

Images