Remaining lifetime prediction method based on dual-attention multi-feature fusion network

Abstract

This invention proposes a method for predicting remaining useful life (RUL) based on a dual-attention multi-feature fusion network, comprising: S1, constructing and filtering a dataset from collected multi-sensor data of the device, and normalizing the filtered data; S2, constructing a multi-feature fusion network model based on a dual-attention mechanism based on training set samples; S3, inputting test samples into the trained dual-attention multi-feature fusion network model to achieve RUL prediction of the device; this invention integrates handcrafted features to construct a multi-feature fusion framework, achieving the final RUL prediction, thus improving the accuracy of RUL prediction for the device.

Description

Technical Field

[0001] This invention belongs to the technical field of equipment failure prediction and health management, and particularly relates to a method for predicting remaining useful life based on a dual-attention multi-feature fusion network. Background Technology

[0002] With the rapid development of technologies such as the Industrial Internet, the Internet of Things, and 5G, equipment is becoming larger, faster, more integrated, and more intelligent. These devices rapidly accumulate vast amounts of data during operation, providing new opportunities for advancements in equipment condition prediction and health management technologies. As a core research area within the PHM (Prognostics and Health Management) framework, remaining useful life prediction plays a crucial role in complex equipment (such as aircraft engines and high-speed trains) and their components.

[0003] Existing methods for predicting remaining useful life (JUV) can be broadly categorized into physical model-based methods and data-driven methods. Physical model-based methods predict JUV by modeling the physical behavior, wear mechanisms, and failure modes of equipment. These methods rely on the physical principles and operating conditions of the equipment, typically simulating its operating state and failure processes using mathematical or simulation models. Commonly used physical model-based JUV prediction methods include those based on the Gamma process and the Wiener process. However, with the rapid development of sensor technology and big data, the complexity and interconnectivity of mechanical systems are constantly increasing, gradually revealing the limitations of physical model-based JUV prediction methods. These methods face challenges when processing complex, multi-dimensional data, primarily because they rely too heavily on accurate modeling of the equipment's physical behavior and failure modes. Therefore, data-driven methods have gradually become a research hotspot. Data-driven methods extract features from equipment operating data and utilize machine learning algorithms to establish a mapping relationship between equipment degradation processes and JUV. In contrast, data-driven methods primarily rely on extracting features from equipment operation data and using machine learning and other methods to build predictive models. Data-driven methods for predicting remaining useful life (RUL) include artificial neural networks, extreme learning machines (ELM), support vector machines (SVM), support vector regression (SVR) models, radial basis function neural networks (RBF), and random forests (RF). These methods have all achieved remarkable results in RUL prediction applications. Currently, deep learning, as an end-to-end learning method, has received widespread attention in the field of RUL prediction. Unlike traditional machine learning methods, it eliminates the need for manual feature selection, automatically learning representative features from raw sensory data, making it more convenient and effective in application.

[0004] Currently, deep learning, as an end-to-end learning method, has received widespread attention in the field of Remaining Useful Life (RUL) prediction. Unlike traditional machine learning methods, it eliminates the need for manual feature selection, automatically learning representative features from raw sensory data, making it more convenient and effective in application. Among these deep learning-based methods, Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNNs), as two core deep learning models, have been widely used in RUL prediction.

[0005] Traditional RUL prediction methods based on LSTM and CNN still have some limitations: CNN networks have excellent feature extraction capabilities, but they do not weight different features, resulting in indiscriminate extraction of feature importance. LSTM-based RUL prediction methods are neural networks that analyze previous data sequentially, which often leads to long-term dependency problems. That is, they only perform regression or classification on features learned at the last time step, ignoring the contributions of features extracted at other times. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a remaining lifetime prediction method based on a dual-attention multi-feature fusion network, which combines CNN-CAM and LSTM-SAM multi-feature fusion models to extract features from the spatiotemporal dimension, comprehensively capture feature categories, and achieve the goal of accurately predicting the remaining lifetime.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for predicting remaining lifetime based on a dual-attention multi-feature fusion network, characterized by comprising the following steps:

[0008] Step 1: Collect data from multiple sensors of the device to build a dataset, merge and filter the collected data, and then normalize the filtered data.

[0009] Step 2: Based on the training set samples, construct a multi-feature fusion network model based on the dual attention mechanism. First, combine CNN and CAM to initially extract features in the spatial dimension; then, combine LSTM and SAM to extract features in the temporal dimension, thus completing the integration of feature information in the spatial and temporal dimensions.

[0010] Step 3: Input the test samples into the trained dual-attention multi-feature fusion network model to predict the remaining lifespan of the device.

[0011] According to the above scheme, the data to be removed in the sensor data screening in step one are sensor data that show stable performance in the degradation experiment, do not show degradation characteristics, and cannot provide obvious degradation information.

[0012] According to the above scheme, the data normalization process in step one includes the following:

[0013] The data is normalized using the Min-Max method, and the specific formula is as follows:

[0014]

[0015] In the formula: x i This represents the value at the current time point; and These refer to the maximum and minimum values ​​of the sensor at all points in time; x i ′ represents the data obtained after normalization calculation.

[0016] According to the above scheme, the construction of the CNN-CAM module combining CNN and CAM in step two includes the following:

[0017] The original time series data is preprocessed to form the input data, and the characteristics of the input data can be expressed as...

[0018] As shown:

[0019] X = [x1, x2, x3, ..., x k ],whereX∈R T×K

[0020] In the formula: x k This represents the Kth channel in the feature map X; while T and K represent the time step and the amount of sensor data, respectively.

[0021] Convolution is calculated as follows:

[0022]

[0023] In the formula: X i-1 W represents the input of the i-th convolutional layer; i This represents the i-th convolutional kernel; and b i These represent the activation function and the bias term of the function, respectively;

[0024] The CAM process is used to process the data output from the convolutional layer. After compressing the feature map X output by the convolutional layer, the spatial information contained in the data is integrated through a global average pooling layer and a global max pooling layer. Given that the spatial dimension of the feature map is h×w×c, it becomes 1×1×c after the compression operation. After the compression operation, a corresponding channel descriptor P is generated for each channel of the feature map X. K The formula is as follows:

[0025]

[0026] In the formula: x k (i,j) represents the element in the i-th row and j-th column of the feature map X;

[0027] P K Perform the stimulus operation to generate the extraction descriptor L, represented as:

[0028] L=E(P,FC)=tanh(FC2tanh(FC1P))

[0029] In the formula: FC1 represents a dimension reduction layer; FC2 represents a dimension increase layer; tanh is the activation function;

[0030] To avoid the central symmetry problem, the tanh function sets the output weights to the range [-1, 1], and then uses a weighted processing method to recalibrate the original features in the channel dimension. Finally, in order to integrate the data from the global average pooling layer and the global max pooling layer, the element-wise summation method is used to generate the output feature map. The softmax function is used to normalize the weights, and the normalized weights are distributed to the features of each channel through scaling operations. The final output of this module is then directly input into the next network layer.

[0031] According to the above scheme, the construction of the LSTM-SAM module combining LSTM and SAM in step two includes the following:

[0032] Before data flows into the LSTM network, it first passes through a forget gate. Its main function is to selectively forget the data transmitted from the previous time step. The update formula is as follows:

[0033] f t =σ(W f [h t-1 ,x t ]+b f )

[0034] In the formula: x t It is the input of time step t; h t-1 It is the hidden state at time step t-1; W f Represents the weight matrix; bf The term represents the bias term; σ represents the sigmoid function.

[0035] After being filtered by the forget gate, the data is passed to the input gate. The LSTM network input signal is processed by the input gate to determine the information that needs to be updated and the content to be updated, as shown below:

[0036] i t =σ(W i [h t-1 ,x t ]+b i )

[0037]

[0038] In the formula: W i W C Represents the weight matrix; b i b C Indicates the bias term; Indicates x t and h t-1 State variables that integrate state information; C t Indicates the current state;

[0039] The output gate calculates the cell state to be output. The input signal passes through the output gate to determine the useful information for output and immediately updates the current LSTM neural network cell state. The update calculation formula is as follows:

[0040] o t =σ(W o [h t-1 ,x t ]+b o )

[0041] h t =o t *tanh(C t )

[0042] In the formula: W o Represents the weight matrix; b o Indicates the bias term;

[0043] The learned features output by the LSTM network can be represented as X = {X1, X2, ..., X...} n} T , where T represents the transpose operation, and x is the input to SAM. i ={x1,x2,…,x m} T , where m represents the time step of the data, and the specific formula is as follows:

[0044] u i=softmax(W T x i +b)

[0045]

[0046] In the formula: u i W represents attention weights; T b are the weight matrix and bias term in the attention network, and a i It represents the magnitude of the influence of the i-th feature value on the target sequence data, and all a i The sum is 1;

[0047] The feature sequences are weighted to obtain the output sequence, which is represented as follows:

[0048]

[0049] The beneficial effects of this invention are: it provides a remaining useful life prediction method based on a dual-attention multi-feature fusion network. The method employs a dual-attention mechanism framework, where CNN-CAM assigns different weights to channels in the spatial dimension to extract important features; while LSTM-SAM further determines the importance of features in the temporal dimension, applying higher weights to more important features and time steps. Simultaneously, this invention integrates handcrafted features to construct a multi-feature fusion framework, achieving the final RUL prediction and improving the accuracy of device remaining useful life prediction. Attached Figure Description

[0050] Figure 1 This is a flowchart of an embodiment of the present invention.

[0051] Figure 2 This is a diagram of a multi-feature fusion network for a dual-attention mechanism according to an embodiment of the present invention.

[0052] Figure 3a , Figure 3b , Figure 3c , Figure 3d The image shows the fitting plot of the prediction results and the true RUL of four subset datasets FD001, FD002, FD003, and FD004 in one embodiment of the present invention. Detailed Implementation

[0053] To better understand the present invention, the invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] To compensate for the shortcomings of CNNs in feature weighting, this invention introduces the Channel Attention Mechanism (CAM). First proposed by Hu et al., the Channel Attention Mechanism adaptively allocates channel weights, allowing the network to focus on more important features. Combining CNN and CAM not only enhances the model's feature extraction capabilities but also overcomes the shortcomings of traditional CNNs in feature focus, thereby improving the accuracy of remaining lifetime prediction.

[0055] Meanwhile, the Spatial Attention Mechanism (SAM) is considered an effective method for addressing long-term dependencies. First proposed by Bahdanau, SAM provides direct correlations between different states in a sequence within a predictive model. As the name suggests, SAM assigns different attention weights to different parts of the sequence, allowing the model to focus more on important information during training. SAM effectively overcomes the long-term dependency problem of LSTM networks, and combining it with LSTM (LSTM-SAM) can better extract features in the temporal dimension.

[0056] like Figure 1 , Figure 2 As shown, this invention proposes a method for predicting remaining lifetime based on a dual-attention multi-feature fusion network, comprising the following steps:

[0057] Step 1: Collect data from multiple sensors of the device to build a dataset, merge and filter the collected data, and then normalize the filtered data.

[0058] 1. Among the various sensors, the data in the degradation experiment showed relatively stable performance, meaning that these data did not exhibit degradation characteristics and could not provide significant degradation information. Therefore, the data from these sensors were removed, and the data from the remaining sensors were used to train the model proposed in this invention.

[0059] 2. The data has a high level of noise, so it is necessary to normalize the data to eliminate this effect.

[0060] The data is normalized using the Min-Max method, and the specific formula is as follows:

[0061]

[0062] In the formula: x i This represents the value at the current time point; and These refer to the maximum and minimum values ​​of the sensor at all points in time; x i ′ represents the data obtained after normalization calculation.

[0063] Step two: Based on the training set samples, a multi-feature fusion network model based on the dual attention mechanism is constructed. First, CNN and CAM are combined to initially extract features in the spatial dimension. Then, LSTM and SAM are combined to extract features in the temporal dimension, completing the integration of feature information in the spatial and temporal dimensions and enhancing the feature extraction capability. Finally, the trained neural network model is used for RUL prediction.

[0064] Given the excellent feature extraction capabilities of CNNs, this invention proposes a CNN-CAM module that integrates CNNs and CAMs.

[0065] First, the original time series data is preprocessed to form the input data, and the characteristics of the input data can be represented as follows:

[0066] X = [x1, x2, x3, ..., x k ],whereX∈R T×K

[0067] In the formula: x k This represents the Kth channel in the feature map X; while T and K represent the time step and the amount of sensor data, respectively.

[0068] Convolution is calculated as follows:

[0069]

[0070] In the formula: X i-1 W represents the input of the i-th convolutional layer; i This represents the i-th convolutional kernel; and b i These represent the activation function and the bias term of the function, respectively;

[0071] The CAM process is used to process the data output from the convolutional layer. After compressing the feature map X output by the convolutional layer, the spatial information contained in the data is integrated through a global average pooling layer and a global max pooling layer. Given that the spatial dimension of the feature map is h×w×c, it becomes 1×1×c after the compression operation. After the compression operation, a corresponding channel descriptor P is generated for each channel of the feature map X. K The formula is as follows:

[0072]

[0073] In the formula: x k (i,j) represents the element in the i-th row and j-th column of the feature map X;

[0074] P K Perform the stimulus operation to generate the extraction descriptor L, represented as:

[0075] L=E(P,FC)=tanh(FC2tanh(FC1P))

[0076] In the formula: FC1 represents a dimension reduction layer; FC2 represents a dimension increase layer; tanh is the activation function.

[0077] The two fully connected layers in the activation layer reduce model complexity through dimensionality reduction, which greatly simplifies the computation process and reduces the number of parameters. To avoid the centrosymmetry problem, the tanh function sets the output weights to the range [-1, 1], and then a weighted processing method is used to recalibrate the original features in the channel dimension. Finally, in order to integrate the data from the global average pooling layer and the global max pooling layer, element-wise summation is used to generate the output feature map. The softmax function is used to normalize the weights, and the normalized weights are distributed to the features of each channel through scaling operations. The final output of this module is directly input into the next network layer.

[0078] The construction of the LSTM-SAM module, which combines LSTM and SAM, includes the following:

[0079] Before data flows into the LSTM network, it first passes through a forget gate. Its main function is to selectively forget the data transmitted from the previous time step. The update formula is as follows:

[0080] f t =σ(W f [h t-1 ,x t ]+b f )

[0081] In the formula: x t It is the input of time step t; h t-1 It is the hidden state at time step t-1; W f Represents the weight matrix; b f The term represents the bias term; σ represents the sigmoid function.

[0082] After being filtered by the forget gate, the data is passed to the input gate. The LSTM network input signal is processed by the input gate to determine the information that needs to be updated and the content to be updated, as shown below:

[0083] i t =σ(W i [h t-1 ,x t ]+b i )

[0084]

[0085] In the formula: W i W CRepresents the weight matrix; b i b C Indicates the bias term; Indicates x t and h t-1 State variables that integrate state information; C t Indicates the current state;

[0086] The output gate calculates the cell state to be output. The input signal passes through the output gate to determine the useful information for output and immediately updates the current LSTM neural network cell state. The update calculation formula is as follows:

[0087] o t =σ(W o [h t-1 ,x t ]+b o )

[0088] h t =o t *tanh(C t )

[0089] In the formula: W o Represents the weight matrix; b o Indicates the bias term;

[0090] The learned features output by the LSTM network can be represented as X = {X1, X2, ..., X...} n} T , where T represents the transpose operation, and x is the input to SAM. i ={x1,x2,…,x m} T , where m represents the time step of the data, and the specific formula is as follows:

[0091] u i =softmax(W T x i +b)

[0092]

[0093] In the formula: u i W represents attention weights; T b are the weight matrix and bias term in the attention network, and a i It represents the magnitude of the influence of the i-th feature value on the target sequence data, and all a i The sum is 1;

[0094] The feature sequences are weighted to obtain the output sequence, which is represented as follows:

[0095]

[0096] Step 3: Input the test samples into the trained dual-attention multi-feature fusion network model to predict the remaining lifespan of the device.

[0097] Example 1

[0098] An aerospace propulsion simulation system simulator was used to simulate 90,000 pounds of thrust. The engine was configured to operate at a sea level of 40,000 feet, Mach numbers from 0 to 0.90, and temperatures ranging from -60 to 103°F. Engine components included the fan, gas chamber, low-pressure turbine, low-pressure compressor, high-pressure compressor, high-pressure turbine, and nozzles. The C-MAPSS dataset comprises four sub-datasets: FD001, FD002, FD003, and FD004. Each sub-dataset is further divided into training and test datasets. Both training and test datasets consist of sensor data from a specific engine in each operating cycle, with the aim of predicting the true RUL (Real Limiting Rank) of each engine in the test dataset using given sensor measurements. The datasets are arranged in an Nx26 matrix, where each data sample includes 26 variables: engine unit number for each engine, degradation cycle for each turbofan engine, three turbofan engine operating settings parameters, and sampling data from 21 sensors on the turbofan engine in each operating cycle, where N corresponds to the number of sensor signals recorded for each engine. Table 1 provides detailed information about the dataset. The dataset also contains multivariate time data from 21 sensors, including 100 time series from operation to failure.

[0099] Table 1. C-MAPSS Dataset Information

[0100]

[0101]

[0102] To accurately verify the effectiveness and accuracy of the model, this invention proposes two evaluation metrics: Root Mean Square Error (RMSE) and a score function. Both metrics, the smaller the value, the better the predictive performance. This study will comprehensively utilize these two evaluation metrics to assess the model's predictive performance. The formula and definition of Root Mean Square Error are as follows:

[0103]

[0104] In the formula: N represents the total number of samples, and y i These represent the predicted RUL value and the actual RUL value, respectively.

[0105] The formula and definition of the scoring function are as follows:

[0106]

[0107] In the formula: N represents the total number of samples, and y i These represent the predicted RUL value and the actual RUL value, respectively.

[0108] To verify the contribution of each component of the model, this invention selected four methods for comparative experiments: no CNN-CAM (NLH), no LSTM-SAM (CNH), no handcrafted features (CLN), and the model used in this invention. The results are shown in Table 2.

[0109] Table 2 shows the RMSE results of the ablation experiments on four datasets.

[0110]

[0111] The results show that, under the same conditions of fusing handcrafted features, the model without channel attention mechanism achieves scores of 15.19, 18.11, 14.69, and 18.31 in each subset, while the model without LSTM-AM achieves scores of 14.55, 15.62, 12.99, and 16.49 in the subsets. This indicates that channel attention mechanism is more effective than the temporal attention mechanism emphasized by LSTM-AM in improving the accuracy of RUL prediction. The fusion of handcrafted features does indeed optimize the model performance to some extent, suggesting that some handcrafted features with domain knowledge may provide effective assistance in RUL prediction. Finally, the proposed method reduces the scores by 2.49, 3.64, 2.66, and 2.11 in the four subsets, representing relative reductions of 16.39%, 20.1%, 18.1%, and 11.52%, respectively, demonstrating that the fusion model used in this invention is accurate and effective.

[0112] like Figure 3a , Figure 3b , Figure 3c , Figure 3d As shown, a fitting study was conducted between the predicted results and the actual RUL values. The predicted RUL values ​​of the four datasets matched the actual RUL values ​​well. The fitting curves on FD001 and FD003 were quite close to the actual values, indicating that the model can effectively predict RUL. It can also be seen that the frequency of early predictions is higher than that of late predictions, which to some extent demonstrates the model's stability.

[0113] Tables 3 and 4 compare the RMSE and Score of some state-of-the-art algorithms with those of this experiment. In the tables, results in bold indicate the best results achieved by all methods. As can be seen from the tables, the architecture proposed in this invention outperforms some state-of-the-art algorithms on subsets FD002, FD003, and FD004, reducing RMSE by 0.77, 0.1, and 1.956 respectively, equivalent to reductions of 5.05%, 0.83%, and 10.77%.

[0114] Table 3. Results of the research method of this invention and other advanced methods on RMSE.

[0115]

[0116]

[0117] Table 4. Results of the research method of this invention and other advanced methods on the Score.

[0118]

[0119] In summary, compared with some existing advanced algorithms, the model of this invention has achieved good results on both RMSE and SCORE evaluation metrics, and has been validated on four subset datasets, demonstrating the superiority and reliability of the model of this invention.

[0120] The above are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent changes made in accordance with the claims of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for predicting remaining lifetime based on a dual-attention multi-feature fusion network, characterized in that, Includes the following steps: Step 1: Collect data from multiple sensors of the device to build a dataset, merge and filter the collected data, and then normalize the filtered data. Step two: Based on the training set samples, construct a multi-feature fusion network model based on a dual attention mechanism. First, combine CNN and CAM to initially extract features in the spatial dimension, specifically including the following: The original time series data, after data preprocessing, forms the input data. The characteristics of the input data can be represented as follows: In the formula: In the feature map The Middle One channel; and and These represent the time step and the amount of sensor data, respectively. Convolution is calculated as follows: In the formula: Indicates the Input to a convolutional layer; Indicates the One convolutional kernel; and These represent the activation function and the bias term of the function, respectively; CAM is used to process the data output by the convolutional layer, and the feature maps output by the convolutional layer are analyzed. After the squeezing operation, the spatial information contained in the data is integrated through a global average pooling layer and a global max pooling layer, resulting in a spatial dimension of the given feature map. After being squeezed, it becomes After the extrusion operation, the feature map of each channel A corresponding channel descriptor will be generated. The formula is as follows: In the formula: In the feature map The Middle Line number Column elements; right Perform stimulus operations to generate extraction descriptors. , is represented as: In the formula: This represents a dimensionality reduction layer; This represents a layer with an increasing dimension. It is an activation function; To avoid the problem of central symmetry, The function sets the output weights to the range [-1, 1], and then uses a weighted processing method to recalibrate the original features in the channel dimension. Finally, in order to integrate the data from the two branches of the global average pooling layer and the global max pooling layer, the element-wise summation method is used to generate the output feature map. The softmax function is used to normalize the weights, and the normalized weights are distributed to the features of each channel through scaling operations. The final output of this module is directly input into the next network layer. Next, LSTM and SAM are combined to extract features in the time dimension, completing the integration of feature information in the spatial and temporal dimensions. Specifically, this includes the following: Before data flows into the LSTM network, it first passes through a forget gate, which selectively updates the data from the previous time step. The update formula for the forgotten data transmitted at any given moment is: In the formula: It is the time step Input; It is the time step The hidden state; Represents the weight matrix; Indicates the bias term; Represents the sigmoid function; After being filtered by the forget gate, the data is passed to the input gate. The LSTM network input signal is processed by the input gate to determine the information that needs to be updated and the content to be updated, as shown below: In the formula: , Represents the weight matrix; , Indicates the bias term; Indicates will and State variables that integrate state information; Indicates the current state; The output gate calculates the cell state to be output. The input signal passes through the output gate to determine the useful information for output and immediately updates the current LSTM neural network cell state. The update calculation formula is as follows: In the formula: Represents the weight matrix; Indicates the bias term; The learned features output by the LSTM network can be represented as ,in This represents the transpose operation, and is used as the input to SAM. , The time step of the data is represented by the following formula: In the formula: Indicates attention weight; and These are the weight matrix and bias terms in an attention network. It is the first The magnitude of the influence of each feature value on the target sequence data, all The sum is 1; The feature sequences are weighted to obtain the output sequence, which is represented as follows: ； Step 3: Input the test samples into the trained dual-attention multi-feature fusion network model. To enable prediction of the remaining service life of equipment.

2. The remaining lifetime prediction method based on a dual-attention multi-feature fusion network according to claim 1, characterized in that, The data removed in the sensor data screening described in step one are sensor data that showed stable performance in the degradation experiment, did not exhibit degradation characteristics, and could not provide obvious degradation information.

3. The remaining lifetime prediction method based on a dual-attention multi-feature fusion network according to claim 2, characterized in that, The data normalization process described in step one includes the following: The data is normalized using the Min-Max method, and the specific formula is as follows: In the formula: This represents the value at the current time point; and These refer to the current sensor at all points in time. The maximum and minimum values ​​within; This represents the data obtained after normalization calculation.

Images