Remaining life lightweight prediction method and system based on feature decoupling and sparse optimization

Abstract

The invention provides a residual life lightweight prediction method and system based on feature decoupling and sparse optimization, and the method comprises the steps: efficiently extracting equipment degradation features through a double-branch architecture: capturing a time sequence degradation law through a GRU network, and generating a deep feature h1 in combination with an attention mechanism, one branch directly extracts a key feature h2 from an original sensor signal through a sparse attention mechanism, the other branch directly extracts a key feature h2 from the original sensor signal through a sparse attention mechanism, then two feature vectors are fused and processed through a lightweight decoder, and finally a residual life prediction value is output. The balance between precision and calculation efficiency is realized through feature decoupling and sparse optimization, and the method is particularly suitable for deployment in an edge calculation environment with limited resources.

Description

Technical Field

[0001] This invention relates to the technical field of engine remaining service life prediction, specifically to a lightweight prediction method and system for remaining service life based on feature decoupling and sparse optimization. Background Technology

[0002] Complex electromechanical equipment, as key devices with precise structures and multifunctional integration, typically maintains a fault-free state for extended periods during actual operation. Remaining useful life prediction is a core aspect of equipment health management, crucial for achieving predictive maintenance, preventing safety accidents, and improving operational economy.

[0003] The rapid development of sensing technology, signal processing, and artificial intelligence has driven the widespread application of data-driven methods in the field of remaining useful life prediction. In particular, deep learning-based methods, through end-to-end feature learning mechanisms, can effectively and automatically extract deep features from raw data, overcoming the limitations of traditional methods that rely on manual feature engineering. Commonly used deep learning models include fully connected neural networks, recurrent neural networks, convolutional neural networks, and attention mechanisms.

[0004] Attention mechanisms, through adaptive weight allocation strategies, can automatically focus on key degradation features and capture long-sequence dependencies, significantly improving the performance of prediction models in complex industrial environments with multiple sensors and varying operating conditions. However, existing attention-based methods still have significant shortcomings: on the one hand, models that rely entirely on attention mechanisms lack a chain-like memory structure, making it difficult to effectively capture subtle, progressive degradation features during equipment operation; on the other hand, existing models are often too complex, consuming large amounts of computational resources, and failing to meet the real-time and lightweight requirements of industrial settings.

[0005] To overcome these limitations, there is an urgent need to develop a new lightweight model that can maintain prediction accuracy while having high computational performance. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a lightweight prediction method and system for remaining lifetime based on feature decoupling and sparse optimization, which effectively solves the problems of insufficient capture of progressive degradation features and high computational complexity in existing prediction methods by combining temporal feature extraction and attention weight allocation.

[0007] The embodiments of this application are implemented as follows: This application provides a lightweight prediction method for remaining lifetime based on feature decoupling and sparse optimization, characterized by the following steps: Step S1: Preprocess the raw sensor data, remove constant sensor signals, and then normalize and smooth the data distribution. Step S2: Input the preprocessed time series data into the encoder neural network to capture the progressive degradation features h1 during the operation of the device; Step S3: The sensor signal is processed using a dual attention mechanism to obtain attention weights, and the sensor signal is weighted to generate a second feature vector h2; Step S4: The two feature vectors h1 and h2 are fused along the feature dimension to form an aggregated feature. The decoder maps the fused feature to a decoding representation that reflects the device degradation state. Step S5: The RMSProp optimization algorithm is used for model training, and a dynamic learning rate decay strategy is used to improve convergence. The reliability of the evaluation is ensured by using independently partitioned training and validation sets, and the performance is evaluated by two dimensions: RMSE and comprehensive score, thereby optimizing the model's generalization ability. Step S6: Input the decoded features into a single-layer fully connected neural network for regression analysis. Use a weighted average method to combine the prediction results of each time step, and finally output the predicted value of the remaining service life of the device.

[0008] In some optional implementations, step S1 uses Z-SCORE normalization to normalize the data distribution, ensuring that the data has zero mean and unit variance. The calculation method is as follows:

[0009] in, Represents the normalized eigenvalues. σ represents the original feature value, μ represents the mean of the feature, and σ represents the standard deviation of the feature.

[0010] In some optional implementations, step S1 employs exponential smoothing to smooth the data, as follows: The exponentially smoothed predicted value y at time t t :

[0011]

[0012]

[0013] in, This represents the sensor observation at time ti, where i ranges from 0 to t, representing historical data at different time points. This represents the weighting coefficient, used to weight sensor values ​​at different time points. Weighting is applied; γ is the attenuation factor, γ∈[0, 1]; s is the smoothing coefficient, s>1.

[0014] In some optional implementations, step S2 employs an encoder GRU to extract temporal degradation features, as detailed below: First, the reset gate state is calculated. The calculation process for the hidden unit activation value at time step t is as follows:

[0015] Where σ represents the logical Sigmoid function, W r and U r These represent resetting the door. Input the weight matrix and the cyclic weight matrix. This indicates the hidden state at the previous moment. This represents the input vector at time step t; New memory cells are accessed via a reset gate. It is generated together with the tanh activation layer, and its computation function is... Represented as:

[0016] The GRU model simplifies the LSTM structure by removing the memory and forget gates and introducing an update gate z. t As an alternative, update gate z t The calculation formula is:

[0017] in, It is the hidden state weight matrix; It is the current state weight matrix; The hidden state h at the current moment t By updating gate z t Regarding the state h at the previous time step t-1 and the current candidate state h̃ t The weighted fusion is performed to obtain the result, and its calculation formula is as follows:

[0018] Input vector x = (x1, ..., x2) t , ..., x L ), where x t ∈R n Let x represent an n-dimensional vector, t∈[1,L] represent the time step, n represent the number of sensors, and L represent the sliding window length; the encoder GRU takes vector x as input and outputs the hidden state hm=(h m，1 h m，2 , ..., h m，L ):

[0019] Where m∈[1,M], m represents the number of layers in the encoder GRU. These represent the learnable parameters of the encoder; By assigning different weights to the features at each time step of the encoder output using a standard attention mechanism, the influence of key time points is highlighted, and the first feature vector h1 is generated: Use the hidden state of all time steps in the last layer = As a key-value pair, the two dimensions of this two-dimensional vector correspond to the time step L and the number of hidden units k, respectively. Output vector for the last time step of all layers = Averaging is performed on the two dimensions of this two-dimensional vector, which correspond to the number of GRU layers m and the number of hidden units k, respectively. The average is obtained by averaging along the first dimension. As a query vector:

[0020] Calculated using the ADDITIVE attention scoring function and The similarity is expressed as follows:

[0021] Where b is and The sum of each element; The attention score is obtained using the following formula. :

[0022] in, Represents the weight vector; Attention weights are calculated using the softmax function. As shown in the following formula:

[0023] Through weighted summation and final hidden state Obtain lightweight feature vectors As shown in the following formula:

[0024] Where M represents the total number of layers in the neural network; t represents the t-th time step of the input sequence; This represents the additive attention weight at time step t.

[0025] In some optional implementations, step S3 includes the following specific details: The original degenerate features are processed using a sparse attention mechanism, with different weights assigned to each time step. This significantly reduces computational complexity while preserving key information, generating a second feature vector h2. In the sparse attention layer, the input vector is Wherein, the input vector at time step t Here, n represents the number of sensors, and t represents the time step. Two identical vectors are concatenated along the sensor dimension to calculate attention scores for different time steps. :

[0026]

[0027] in, This indicates the output projection vector; Represents the linear transformation weight matrix; Attention Score The attention weights are converted using the softmax function, as shown below:

[0028] Weighting the input information to obtain :

[0029] in, This represents the sparse attention weights at time step t.

[0030] In some alternative implementations, the aggregation feature described in step S4 is as follows:

[0031] in, , , ; Input h into the decoder GRU network to obtain the decoding features. :

[0032] in, 'e' represents the number of hidden units in the decoder's GRU. This represents the GRU parameters of the decoder.

[0033] In some optional implementations, the two-dimensional performance evaluation described in step S5 includes the following: The root mean square error (RMSE) is used as one of the evaluation indicators, and its calculation formula is as follows:

[0034] Where N represents the number of data samples, and These are the model's predicted remaining useful life and the actual remaining useful life, respectively. The smaller the RMSE value, the better the model's predictive ability. The RMSE metric cannot distinguish between early and late predictions. Therefore, the Score metric is introduced. By applying a larger penalty factor to prediction lags, the applicability of the model in safety-critical scenarios can be reasonably evaluated. The Score metric is defined as follows: .

[0035] In some optional implementations, the formula for calculating the final output remaining useful life value RUL in step S6 is as follows:

[0036] in, This represents the prediction result for the m-th engine. This represents the weight of the i-th point.

[0037] In some alternative implementations, the weight values ​​are... .

[0038] A lightweight remaining lifetime prediction system based on feature decoupling and sparse optimization, characterized in that it includes: The data processing module removes constant sensor signals and performs normalization and smoothing processing on the data distribution. The encoder module, GRU, captures the progressive degradation characteristics during device operation and outputs a hidden state sequence containing rich temporal information; The attention mechanism module deeply mines the effective feature information in the GRU network and compresses it into lightweight feature vectors to improve the model's expressive power and reduce computational complexity. The model training and optimization module employs the RMSProp optimization algorithm for model training, combined with a dynamic learning rate decay strategy to improve convergence. Independently partitioned training and validation sets ensure evaluation reliability, and performance is evaluated using both RMSE and comprehensive score metrics to optimize model generalization ability. The RUL prediction output module inputs the decoded features into a single-layer fully connected neural network for regression analysis. It then uses a weighted average method to synthesize the prediction results at each time step and finally outputs the predicted remaining lifespan of the device.

[0039] The beneficial effects of this application are as follows: This application provides a lightweight remaining lifetime prediction method and system based on feature decoupling and sparse optimization. By innovatively integrating a dual attention mechanism and a gated recurrent unit (GRU), it achieves high-precision remaining lifetime prediction while significantly reducing parameter scale and computational complexity. The encoder GRU effectively compensates for the shortcomings of the attention mechanism in temporal feature modeling, and the decoder GRU further compresses computational overhead through a single-time-step prediction structure. Simultaneously, exponential smoothing technology is introduced to suppress environmental noise in sensor signals, enhancing feature discriminability while improving the model's generalization ability and prediction stability. This method provides a highly efficient and accurate remaining lifetime prediction solution for resource-constrained edge computing environments. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a structural diagram of the prediction model in an embodiment of this application; Figure 2 This is a diagram of the GRU network structure according to an embodiment of this application; Figure 3 This is a diagram of the timing feature layer structure in an embodiment of this application. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0043] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0044] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0045] The features and performance of this application will be further described in detail below with reference to the embodiments.

[0046] The lightweight remaining lifetime prediction method based on feature decoupling and sparse optimization of this invention constructs a novel lightweight prediction model by fusing a dual attention mechanism and a gated recurrent unit (GRU) with exponential smoothing technique (see [link]). Figure 1 The model employs a dual GRU encoding / decoding structure, which maintains the global modeling advantages of the attention mechanism while possessing the chain memory capability of GRU. At the same time, its lightweight design ensures efficient deployment in edge computing environments.

[0047] Follow these steps to implement the procedure: Specifically, the steps include the following: Step 1: Data Preprocessing. The raw sensor data is preprocessed. First, constant sensor signals are removed. Then, Z-SCORE normalization is used to normalize the data distribution, ensuring zero mean and unit variance. Based on this, exponential smoothing is applied to remove high-frequency noise and redundant features, retaining key information reflecting equipment degradation trends.

[0048] Step 2, Temporal Feature Encoding: The preprocessed temporal data is input into the encoder GRU network, which captures the progressive degradation features during device operation through its chain memory structure. The encoder GRU outputs a hidden state sequence containing rich temporal information.

[0049] Step 3, Attention Weight Allocation: A dual attention mechanism is used to process the encoder output. First, different weights are assigned to the features at each time step of the encoder output using a standard attention mechanism to highlight the influence of key time points, generating the first feature vector h1. Simultaneously, a sparse attention mechanism is used to process the original degenerate features, assigning different weights to each time step, significantly reducing computational complexity while preserving key information, generating the second feature vector h2.

[0050] Step 4, Feature Fusion and Decoding: The two feature vectors h1 and h2 generated by the attention mechanism are fused along the feature dimension to form a comprehensive feature representation. This fused feature is input into the optimized decoder GRU, which employs a single-time-step computation structure, significantly reducing computational resource consumption. The decoder maps the fused feature into a high-level representation reflecting the device's degradation state.

[0051] Step 5, Model Training and Optimization: The model training employs the RMSProp optimization algorithm, combined with a dynamic learning rate decay strategy to improve convergence. Independently partitioned training and validation sets ensure evaluation reliability, and performance is evaluated using both RMSE and a comprehensive score to optimize the model's generalization ability.

[0052] Step 6, RUL Prediction Output: Input the decoded features into a single-layer fully connected neural network for regression analysis. Combine the prediction results of each time step using the weighted average method to finally output the predicted value of the remaining service life of the device.

[0053] Explanation of each layer of the lightweight prediction model: 1. Encoder layer: Before applying the attention mechanism, a gated recurrent unit (GRU) is first employed (see...). Figure 2 GRU extracts temporal degradation features. Through the special design of its update and reset gates, GRU effectively alleviates the vanishing and exploding gradient problems in recurrent neural networks. Compared to Long Short-Term Memory (LSTM) networks, GRU has a simpler network structure and has become a widely used variant in recurrent neural networks.

[0054] First, the reset gate state is calculated. The calculation process for the hidden unit activation value at time step t is as follows: (1) Where σ represents the logical Sigmoid function, W r and U r These represent resetting the door. Input the weight matrix and the cyclic weight matrix. This indicates the previously hidden state; This represents the input vector at time step t.

[0055] New memory cells are accessed via a reset gate. It is generated together with the tanh activation layer, and its computation function is... Represented as: (2) The GRU model simplifies the LSTM structure by removing the memory and forget gates and introducing an update gate z. t As an alternative, update gate z t The calculation formula is: (3) in, It is the hidden state weight matrix; It is the current state weight matrix.

[0056] The hidden state h at the current moment t By updating gate z t Regarding the state h at the previous time step t-1 and the current candidate state h̃ t The weighted fusion is performed to obtain the result, and its calculation formula is as follows: (4) Input vector x = (x1, ..., x2) t , ..., x L ), where x t ∈R n Let x represent an n-dimensional vector, t∈[1,L] represent the time step, n represent the number of sensors, and L represent the sliding window length; the encoder GRU takes vector x as input and outputs the hidden state hm=(h m，1 h m，2 , ..., h m，L ): (5) Where m∈[1,M], m represents the number of layers in the encoder GRU. These represent the learnable parameters of the encoder.

[0057] 2. Temporal Feature Layer This invention designs a novel attention mechanism for deep mining of effective feature information in GRU networks and compressing it into lightweight feature vectors to improve model expressiveness and reduce computational complexity. This attention mechanism consists of three core components: a query vector, a key vector, and a value vector. Unlike existing self-attention mechanisms where these three components typically originate from the same vector, this invention innovatively decouples the query vector from the key-value pairs, specifically including: Key-value pair construction: using the hidden states of all time steps in the last layer. = As key-value pairs, the two dimensions of this two-dimensional vector correspond to the number of time steps (L) and the number of hidden units (k), respectively.

[0058] Query vector generation: Output vector for the last time step of all layers. = The average is then applied (as shown in Equation (6)). The two dimensions of this two-dimensional vector correspond to the number of GRU layers (M) and the number of hidden units (k), respectively. h is obtained by averaging along the first dimension (layer dimension). mean ∈ R k As a query vector.

[0059] like Figure 3 As shown, select h mean Using query vectors is based on the inherent characteristics of the GRU chain structure, namely that the hidden states at the end of the sequence can effectively aggregate key degradation information in the entire temporal context. mean The expression is shown in formula (6). h is calculated using the ADDITIVE attention scoring function. M with h mean The similarity is shown in equation (7), where b is the similarity of h. Mwith h mean The element-wise summation. Obtain the attention score s. enc Then, the attention weights are calculated using the softmax function. As shown in equation (9). With h mean Encoded features with higher similarity will receive greater weight, indicating their greater importance. Ultimately, through... The weighted sum yields a lightweight feature vector. As shown in equation (10).

[0060] (6) (7) (8) (9) (10) in, Represents the weight vector; This represents the additive attention weight at time step t.

[0061] GRU is responsible for modeling temporal degradation features, while the attention mechanism can fully extract useful information from all hidden states in deep GRU networks. In contrast, traditional methods only focus on the hidden states of the last layer, resulting in the loss of effective information.

[0062] 3. Sparse attention layer This invention improves the traditional attention calculation method by adopting the Sparse Multi-Head Attention Mechanism (SMSA). Based on the idea of ​​sparse attention, this mechanism enhances the ability to model changes in the operating state of the device by selectively focusing on key feature segments in the time series data of the sensor, while significantly reducing the computational complexity of the model.

[0063] In traditional multi-head attention mechanisms, the similarity between all feature points needs to be calculated in the feature map after feature fusion. This global calculation method is prone to noise interference, leading to the dilution of key features and thus affecting prediction accuracy. In contrast, the sparse multi-head attention mechanism adopted in this invention implements a selective attention strategy: first, it calculates the similarity of all feature pairs between the query vector and the key vector; then, it retains only the Top K most relevant feature pairs for each feature point to generate a sparse attention mask; next, it performs softmax normalization on the Top K elements in each row, setting the remaining elements to zero; finally, it multiplies the sparse weight matrix with the value vector matrix to obtain the output. This method effectively avoids the problem in traditional attention mechanisms where the softmax function over-amplifies secondary similarities, weakening the weights of key features, ensuring that the model always focuses on the most discriminative feature segments, and significantly reducing computational and memory overhead while improving state modeling capabilities.

[0064] In the sparse attention layer, the input vector is ,in , where n represents the number of sensors and t represents the time step. As shown in formulas (11)-(12), two identical vectors are connected along the sensor dimension to calculate the attention scores for different time steps. Then, the vector passes through two fully connected layers in sequence. , Attention score The input information is converted into attention weights using the softmax function, as shown in formula (13). Finally, the input information is weighted to obtain h2, as shown in formula (14).

[0065] (11) (12) (13) (14) in This indicates the output projection vector; Represents the linear transformation weight matrix; This represents the sparse attention weights at time step t.

[0066] 4. Decoder layer Inspired by the Transformer encoder-decoder structure, this invention employs a second GRU network as the decoder to compute the RUL. Since the chain structure of the decoder GRU is decomposed into single-time-step processing, as shown in Equation (15), the two lightweight vectors h1 and h2 are concatenated to compute the aggregate feature h: (15) in, , , ; Input h into the decoder GRU network to obtain the decoding features. : (16) in, 'e' represents the number of hidden units in the decoder's GRU. This represents the GRU parameters of the decoder.

[0067] Example 1 In this embodiment, the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset released by NASA is used for model training and validation. This dataset simulates the state evolution of aero-engines under various operating conditions and failure modes through high-fidelity engine performance degradation simulation. The dataset contains multi-dimensional time-series data collected by a monitoring network composed of 21 sensors, covering various key performance parameters such as rotor speed, inlet and outlet temperatures, and pressure. It comprehensively reflects the full-cycle state changes of the engine from normal operation to performance degradation, providing a realistic and reliable training and validation foundation for the remaining service life prediction model.

[0068] As shown in Table 1, the C-MAPSS dataset contains four independent subsets, each containing multidimensional time-series signals collected by 21 sensors, and is strictly divided into training and test sets. In the training set, the operating data of each engine is recorded from its normal state until performance failure occurs, while providing the complete remaining service life (RUL) as a supervision label. The test set simulates the incomplete data acquisition characteristics of actual industrial scenarios, truncating the data at a certain point before engine failure, and only providing a partial operating sequence. The objective of this invention is to accurately predict the number of cycles an engine can continue to operate before failure, i.e., the remaining service life (RUL), based on the incomplete test sequence.

[0069] Table 1. C-MAPSS Dataset Information

[0070] The C-MAPSS dataset provides monitoring data from 21 sensors, but 7 of these sensors have constant output values ​​throughout the entire engine operating cycle, failing to reflect degradation trends, and were therefore removed. The final selected sensors include 14 sensors with index numbers 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, and 21. Because the selected sensor signals have different units and dimensions, and their numerical ranges vary significantly, directly using the raw data for modeling can easily lead to the neglect of smaller numerical features and make neural network training convergence difficult. To address this issue, this invention employs the Z-SCORE normalization method to unify the data processing. The calculation method is as follows: (17) in, Represents the normalized eigenvalues. σ represents the original feature value, μ represents the mean of the feature, and σ represents the standard deviation of the feature.

[0071] Before standardization, the raw sensor signals exhibit a non-uniform distribution and inconsistent numerical scales, interfering with the effective learning of degradation features by the model. After Z-SCORE standardization, the sensor data are converted into a standard distribution with zero mean and unit variance, eliminating dimensional differences, improving feature comparability, and facilitating more stable and efficient extraction of degradation-related features by subsequent neural network models.

[0072] Raw sensor signals often contain significant noise, which can cause models to learn features irrelevant to prediction. Since the influence of historical sensor values ​​on future values ​​diminishes with increasing time intervals, a more effective approach is to use a weighted average as the prediction. Exponential smoothing meets this requirement; it has a simple recursive form that reduces signal oscillations and filters out redundant noise while preserving the overall trend of the sensor signal.

[0073] As shown in equations (18)-(19), in exponential smoothing, data points closer to the current time are given greater weight, while the weight of data points further away gradually decreases. The smoothness is adjusted by the decay factor γ (γ∈ [0, 1]), and s is the smoothing coefficient, s>1. Equation (20) describes the relationship between γ and s.

[0074] (18) (19) (20) Among them, y t This represents the exponentially smoothed forecast value at time t; This represents the sensor observation at time ti, where i ranges from 0 to t, representing historical data at different time points. This represents the weighting coefficient, used to weight sensor values ​​at different time points. Weighting is applied.

[0075] This invention addresses the problem of noise and oscillations in directly estimated Remaining Useful Life (RUL) predictions, which lead to increased prediction errors. It proposes a weighted averaging method to smooth the prediction results. This method effectively suppresses fluctuations and improves the stability and reliability of the prediction results by calculating the weighted average of the current point and several consecutive data points preceding it. The final RUL value is calculated as follows: (twenty one) in, This represents the prediction result for the m-th engine. This represents the weight of the i-th point. In this embodiment, the weight value is... .

[0076] To objectively evaluate model performance, this invention uses root mean square error (RMSE) as one of the evaluation metrics, and its calculation formula is as follows: (twenty two) Where N represents the number of data samples, and These are the model's predicted remaining useful life and the actual remaining useful life, respectively. The smaller the RMSE value, the better the model's predictive ability.

[0077] The RMSE metric cannot distinguish between early and late predictions. Therefore, the Score metric is introduced. By applying a larger penalty factor to prediction lags, the applicability of the model in safety-critical scenarios can be reasonably evaluated. The Score metric is defined as follows: (twenty three) To verify the effectiveness of the proposed model, a systematic test was first conducted on the training data of four subsets, FD001 to FD004. The raw sensor data was first input into the encoder GRU network to extract temporal degradation features. Then, the features were weighted and fused using an additive attention mechanism to generate a weighted feature vector h1. Simultaneously, a sparse self-attention mechanism was used to assign weights to the original data, resulting in another weighted feature vector h2. h1 and h2 were then fused through a concatenation operation and input into the decoder for further feature mapping. Finally, the predicted remaining lifetime (RUL) was output through a fully connected layer.

[0078] During model training, a learning rate decay strategy was employed, with an initial learning rate of 0.002, which decayed to 10% of its original value every 10 training epochs. The RMSPROP algorithm was used for optimization, with a batch size of 128. The training and validation sets were divided at a ratio of 90% to 10%, ensuring the independence of model training and validation.

[0079] To avoid the model getting trapped in local optima, this invention employs a multi-random seed initialization strategy, performing 50 independent training iterations on each subset of the dataset. The model with the smallest root mean square error (RMSE) on the validation set is selected as the optimal model for performance evaluation on the test set, and its parameters are saved. The configuration of the model's hyperparameters is shown in Table 2.

[0080] Table 2 Hyperparameters of the proposed model

[0081] This invention achieves state-of-the-art comprehensive prediction performance on all four subsets of the datasets FD001 to FD004. Experimental results (as shown in Tables 3 and 4) demonstrate that on the more complex and noisier FD002 and FD004 datasets, the RMSE and Score of this invention's model significantly outperform existing representative methods, exhibiting excellent accuracy and engineering applicability. On the FD001 and FD003 datasets, this invention also achieves prediction levels comparable to the known best methods, without exhibiting the performance fluctuations or significant degradation common in other models, demonstrating stable generalization ability and robustness. In summary, this invention achieves optimal or tied-optimal comprehensive performance on all four standard test sets, verifying its technical advantages of combining high accuracy and high stability under complex operating conditions.

[0082] Table 3 Comparison of RMSE with different methods

[0083] Table 4 Comparison of SCORE with different methods

Claims

1. A residual life lightweight prediction method based on feature decoupling and sparse optimization, characterized in that, Comprising the following steps: Step S1, preprocessing the original sensor data, eliminating constant sensor signals, and then normalizing and smoothing the data distribution; Step S2, inputting the preprocessed time series data into the encoder neural network to capture the progressive degradation characteristics h1 during the operation of the device; Step S3, using a double attention mechanism to process the sensor signals to obtain attention weights, and generating a second feature vector h2 by weighting the sensor signals; Step S4, fusing the two feature vectors h1 and h2 in the feature dimension to form an aggregated feature, and the decoder maps the fused feature to a decoding representation reflecting the degradation state of the device; Step S5, using the RMSProp optimization algorithm for model training, combining a dynamic learning rate decay strategy to improve convergence, ensuring evaluation reliability through independent division of training and validation sets, and evaluating performance in two dimensions with RMSE and comprehensive Score indicators to optimize model generalization ability; Step S6, inputting the decoded feature into a single-layer fully connected neural network for regression analysis, and finally outputting the predicted value of the remaining useful life of the device by weighted average method to integrate the prediction results of each time step.

2. The method for remaining useful life prognosis based on feature decoupling and sparse optimization according to claim 1, characterized in that, In step S1, the data distribution is normalized by Z-SCORE standardization method to ensure that the data has zero mean and unit variance, and the calculation method is as follows: wherein, represents the normalized feature value, represents the original feature value, μ represents the mean value of the feature, and σ represents the standard deviation of the feature.

3. The remaining useful life lightweight prediction method based on feature decoupling and sparse optimization according to claim 1 or 2, characterized in that, In step S1, the data is smoothed by exponential smoothing method, as follows: Exponential smoothing prediction value y at time t t : wherein, represents the sensor observation value at time t-i, i from 0 to t, represents the historical data at different time points; represents the weight coefficient, used for weighting the sensor values at different time points ; γ is the attenuation factor, γ ∈ [0, 1]; s is the smoothing coefficient, s >

1.

4. The method for remaining useful life prognosis based on feature decoupling and sparse optimization according to claim 3, characterized in that, In step S2, the encoder GRU is used to extract the time series degradation feature, as follows: First, calculate the reset gate state, and the hidden unit activation value at time step t is calculated as follows: where σ denotes a logistic sigmoid function, W r and U r denote the weights of the reset gate, respectively an input weight matrix and a recurrent weight matrix, denotes a hidden state of the previous time step, denotes an input vector at time step t; New memory cell through reset gate Co-generated with tanh activation layer, whose computational function is represented as: The GRU model simplifies the LSTM structure by removing the memory gate and the forget gate and introducing an update gate z t As an alternative, the update gate z t The calculation formula is: wherein, is a hidden state weight matrix; is a current state weight matrix; the hidden state h at the current time t by updating gate z t the previous time state h t-1 and the current candidate state h̃ t weighted fusion, and the calculation formula is: Input vector x = (x1, ..., x2) t , ..., x L ), where x t ∈R n Let x represent an n-dimensional vector, t∈[1,L] represent the time step, n represent the number of sensors, and L represent the sliding window length; the encoder GRU takes vector x as input and outputs the hidden state hm=(h m，1 h m，2 , ..., h m，L ): wherein m e [1, M], m represents the number of layers of the encoder GRU, represent the learnable parameters of the encoder. Through the standard attention mechanism, different weights are assigned to the features output by the encoder at each time step, highlighting the influence of key time points, and generating the first feature vector h1: The hidden state of the last layer at all time steps = As a key-value pair, the two dimensions of the two-dimensional vector correspond to the number of time steps L and the number of hidden units k, respectively; the last time step output vector of all layers = average processing, the two dimensions of the two-dimensional vector correspond to the GRU layer number m and the hidden unit number k respectively, and the average operation along the first dimension is obtained as the query vector: The ADDITIVE attention score function is calculated by with similarity, as follows: wherein b is with the element-wise addition of The attention score is obtained by the following formula : wherein denotes a weight vector; The attention weight is calculated via a softmax function As follows: weighted sum of the passing and final hidden states obtaining a light-weight feature vector as follows: wherein M represents the total number of neural network layers; t represents the tth time step of the input sequence; denotes the additive attention weight at time step t.

5. The method for remaining useful life prognosis based on feature decoupling and sparse optimization according to claim 4, characterized in that, In step S3, the following specific contents are included: Through the sparse attention mechanism, the original degradation feature is processed, and different weights are assigned to each time step, which significantly reduces the computational complexity while retaining key information, and generates the second feature vector h2: In the sparse attention layer, the input vector is where the t-th time step input vector n denotes the number of sensors, t denotes the time step, and two identical vectors are concatenated in the sensor dimension to compute the attention scores for different time steps ​ wherein, represents an output projection vector; represents a linear transformation weight matrix; Attention score converted to attention weights by a softmax function as follows: weighting the input information yields : wherein, denotes the sparse attention weights at time step t.

6. The method for remaining useful life prognosis based on feature decoupling and sparse optimization according to claim 5, characterized in that, The aggregated feature in step S4 is as follows: wherein , , ; inputting h into a decoder GRU network to obtain decoding features : wherein, , e represents the number of hidden units in the decoder GRU, represent the decoder GRU parameters.

7. The method for remaining useful life prognosis based on feature decoupling and sparse optimization according to claim 6, characterized in that, The two-dimensional performance evaluation in step S5 includes the following contents: The root mean square error RMSE is used as one of the evaluation indicators, and its calculation formula is as follows: wherein N represents the number of data samples, and are the model predicted residual useful life value and the real residual useful life value, respectively, and the smaller the value of RMSE, the better the prediction ability of the model. The RMSE index cannot distinguish whether the prediction result is ahead of time or lagging behind, so the Score index is introduced, which applies a larger penalty factor to the prediction lag, and reasonably evaluates the practicality of the model in safety-critical scenarios. The definition of this index is as follows: 。 8. The method for remaining useful life prognosis based on feature decoupling and sparse optimization according to claim 7, characterized in that, The calculation formula of the final output remaining useful life value RUL in step S6 is as follows: wherein, represents the prediction result of the mth engine, represents the weight of the ith point.

9. The method for remaining useful life prognosis based on feature decoupling and sparse optimization according to claim 8, characterized in that, The weight values .

10. A remaining useful life lightweight prediction system based on feature decoupling and sparse optimization, characterized in that, It includes: Data processing module, eliminating constant sensor signals, and normalizing and smoothing the data distribution; Encoder module, GRU captures the progressive degradation characteristics during the operation of the device, and outputs a hidden state sequence containing rich time series information; Attention mechanism module, deeply excavates the effective feature information in the GRU network, and compresses it into a lightweight feature vector to improve the model expression ability and reduce the computational complexity; Model training and optimization module, model training adopts RMSProp optimization algorithm, combined with dynamic learning rate decay strategy to improve convergence; Through independent division of training set and validation set to ensure the reliability of evaluation, and evaluate the performance in two dimensions of RMSE and comprehensive score index to optimize the generalization ability of the model; RUL prediction output module, the decoded features are input into a single-layer fully connected neural network for regression analysis, and the prediction results of each time step are integrated by weighted average method to finally output the remaining useful life prediction value of the device.

Images