Remaining useful life prediction method based on deep hybrid multi-task model

Abstract

This invention provides a method for predicting the remaining useful life (RUL) of operating equipment based on a deep hybrid multi-task model. The method includes: constructing an unlabeled dataset and a labeled dataset; generating training and validation data sets for the model using a sliding window method; constructing a first temporal model (DDTM) considering future sensor signal prediction for the unlabeled dataset; constructing a second temporal model (DTTM) considering future sensor signal prediction for the labeled dataset; constructing a deep hybrid multi-task model (DHMTM) considering time dependencies; training the final DHMTM model; using newly received sensor signals as input to the test set; and using the trained final DHMTM model to predict the RUL of operating equipment. This invention adds future sensor signals only during model training, trains two temporal models using unlabeled and labeled data respectively, and shares latent variables, thus capturing the temporal correlation of sensor signals to a greater extent and improving the accuracy of RUL prediction.

Description

Technical Field

[0001] This invention belongs to the field of fault prediction and lifetime estimation technology, and in particular, it is a method for predicting remaining useful life based on a deep hybrid multi-task model. Background Technology

[0002] Prognostic and Health Management (PHM) technology predicts failures through data monitoring and analysis, which is of great significance for developing reasonable maintenance and health management strategies and improving the reliability and safety of various equipment and products. As an important component of PHM, accurately estimating the remaining useful life (RUL) of equipment enables timely implementation of maintenance plans and reduces life-cycle costs. With advancements in sensing technology, deep learning (DL) methods have been successfully applied to RUL prediction. These methods possess strong feature extraction capabilities, require less manpower, and are more efficient.

[0003] To achieve accurate RUL (Recovery Limitation) prediction, valuable feature information can be extracted from degradation signals. This feature information characterizes the product's health status and is called Health Indicators (HIs). Constructing better HIs can improve RUL prediction capabilities. For example, a deep convolutional autoencoder based on a quadratic function can be introduced to learn HIs from raw bearing vibration signals; or the probabilistic relationship between HIs, sensor signals, RUL, and operating conditions can be considered; or trend spikes can be considered by applying convolutional neural networks to HI construction. These existing methods primarily utilize past and current sensor signals to learn current-time HIs. However, when constructing HIs, future sensor signals available in the training dataset must also be considered, as these future sensor signals also depend on the product's current health status. Furthermore, the effectiveness of HI learning is also affected by the prediction length of future sensor signals. Too short a prediction length limits the ability of HIs to characterize current and future degradation behavior, while too long a prediction length hinders the learning of better HIs due to accumulated prediction uncertainty caused by uncertain influencing factors. The size of the training dataset is another important factor affecting the learning effect of HIs. In practical engineering, failures can have serious consequences, and acquiring large amounts of training labeled data for RUL prediction is often costly, and in some cases where failures could lead to catastrophic consequences, it is even difficult to obtain labeled data. This limits the application of supervised deep learning methods. Therefore, in semi-supervised learning, it is of great significance to learn and construct better HIs by using both labeled and unlabeled data simultaneously. Among various semi-supervised learning methods, deep semi-supervised learning SSL generation methods utilizing generative adversarial network regression models, by considering the history of failures and interruptions, can be used not only for prediction tasks during testing but also for data generation tasks. However, there is often a mismatch between the distribution obtained by the model and the distribution of the real data, also known as model misspecification. Semi-supervised hybrid methods can overcome the problem of model misspecification. These methods combine generative methods with discriminative models. Using generative models enriches the flexibility of hybrid methods, while combining discriminative models allows hybrid methods to explicitly model conditional probabilities, thereby utilizing data more effectively. By using a deep hybrid model (DHM) that shares latent variables, both the generative and discriminative models can be trained simultaneously, improving the performance of each model and thus achieving higher accuracy in predicting RUL. However, time-series sensor signals are usually used as inputs to the RUL prediction model, and DHM cannot capture the temporal correlation between these multivariate time series, which greatly limits the model's ability to predict RUL.Therefore, to address this issue, it is urgent and necessary to seek a remaining useful life prediction method based on a deep hybrid multi-task model by adding an additional future sensor signal prediction task to improve feature extraction capabilities and obtain higher RUL accuracy. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a Remaining Usage (RUL) prediction method based on a deep hybrid multi-task model. The method includes constructing unlabeled and labeled datasets, generating training and validation data sets for the model using a sliding window method, constructing a first temporal model (DDTM) considering future sensor signal prediction for the unlabeled dataset, constructing a second temporal model (DTTM) considering future sensor signal prediction for the labeled dataset, constructing a deep hybrid multi-task model (DHMTM) considering time dependencies, training the final DHMTM model, using newly received sensor signals as input to the test set, and using the trained final DHMTM model to predict the RUL of operating equipment. This invention adds future sensor signals only during model training, trains two temporal models using unlabeled and labeled data respectively while sharing latent variables, thus capturing the temporal correlation of sensor signals to a greater extent and improving the accuracy of RUL prediction.

[0005] This invention provides a method for predicting remaining useful life based on a deep hybrid multi-task model, which includes the following steps:

[0006] S1. Construct unlabeled and labeled datasets: Starting from N... U and N L Collect multi-sensor detection signals from similar devices to construct an unlabeled dataset D. U and label dataset D L :

[0007]

[0008] Among them, X n This indicates that data from the first to the Tth samples were collected from the nth device. n Sensor signals at any given time x t This represents the sensor signal at time t. Let represent the p-th dimension signal value of the sensor at time t, where P represents the total dimension of the sensor signal; Represents the real number field; Y n X represents n Corresponding remaining useful life (RUL) y t x represents t The corresponding remaining useful life (RUL) to be predicted;

[0009] S2. Training data set D for the model generated using the sliding window method. train and validation data group D valid ;

[0010] S3, For unlabeled dataset D U Construct a first-time model DDTM that considers predictions of future sensor signals;

[0011] S4, for the labeled dataset D L Construct a second time model DTTM that considers predictions of future sensor signals;

[0012] S41. Derive the second joint probability distribution p on the second time model DTTM. θ (X t ,Xpre t ,y t The second variational lower bound of evidence ELBO DTTM :

[0013]

[0014] Among them, X t The sensor signals representing past and present moments serve as the model's input samples; Xpre t Z represents the future sensor signal to be predicted; t Indicates from X t The corresponding health indicator HIs extracted from q Φ (Z t |X t ) represents a given X t Derivation of Z t inference network; p θ (X t ,Xpre t ,y t |Z t ) represents a given Z t Time X t Xpre t and y t The joint probability distribution of p; θ (Z t ) represents Z t The prior distribution of KL(q); Φ (Z t |X t ,Xpre t )||p θ (Z t )) represents q Φ (Z t |X t ,Xpret ) and p θ (Z t KL divergence values ​​between )

[0015] S42, to maximize the second variational evidence lower bound ELBO of the second time model DTTM. DTTM As the training objective, the generated component p is used. θ (X t Z t Reconstruct the input sensor signal Using the first discriminant component p θ (Xpre t |X t Z t Predicting future sensor prediction signals Using the second discriminant component p θ (y t |X t Z t Predict the remaining useful life to be predicted

[0016] S5. Construct a deep hybrid multitasking model (DHMTM) that considers time dependence, and train it to obtain the final DHMTM model.

[0017] S51. In each training round, the training data group D train As input to the deep hybrid multi-task model DHMTM, the training data set D is computed. train The loss function is determined, and the parameters of the deep hybrid multitasking model (DHMTM) are updated using the stochastic gradient descent algorithm.

[0018] S52, Based on the updated Deep Hybrid Multitask Model (DHMTM) and validation dataset D valid Calculate the validation data set D valid The verification loss;

[0019] S53. Repeat steps S51 and S52 until the maximum set number of rounds is reached, and use the deep hybrid multi-task model DHMTM with the minimum verification loss as the final DHMTM model.

[0020] S6. Transfer the newly received sensor signal D T As input to the test set, the final DHMTM model, trained using the Deep Hybrid Multitask Model (DHMTM), is used to predict the Remaining Usage (RUL) of the operating equipment.

[0021] Furthermore, step S3 specifically includes the following steps:

[0022] S31. Derive the first joint probability distribution p of the first-time model DDTM.θ (X t ,Xpre t The first variational evidence lower bound of the data likelihood is ELBO. DDTM :

[0023]

[0024] Where, q Φ (Z t |X t ,Xpre t ) represents a given X t and Xpre t Z t The posterior distribution of p; θ (X t ,Xpre t Z t ) represents X t Xpre t and Z t The joint probability distribution of p; θ (Z t |X t ,Xpre t ) represents a given X t and Xpre t Z t The prior distribution of p; θ (X t ,Xpre t |Z t ) represents a given Z t Time X t and Xpre t The joint probability distribution of ;

[0025] Considering the existence of the constraint q Φ (Z t |X t ,Xpre t ) = p θ (Z t |X t )=q Φ (Z t |X t ), where p θ (Z t |X t ) represents a given X t Z t Given the prior distribution, we have:

[0026]

[0027] S32, to maximize the first variational evidence lower bound ELBO of the first time model DDTM. DDTM As the training objective, the generated component p is used. θ (X t Z t Reconstruct the input sensor signal Using the first discriminant component p θ (Xpre t |X t Z t To predict future sensor prediction signals

[0028] Furthermore, the loss function of the Deep Hybrid Multitask Model (DHMTM) in step S5 is:

[0029] L DHMTM =αL DDTM +(1-α)L DTTM (10)

[0030] Among them, L DDTM L represents the first loss function; DTTM Let represent the second loss function; α represents the weight coefficient, and 0≤α≤1.

[0031] Preferably, in step S3, the first loss function L of the first temporal model DDTM is... DDTM The calculation specifically includes the following steps:

[0032] S311. The mean squared error (MSE) is used to measure the prediction error, and the reparameterization technique Z is applied. t =μ t +σ t ·ε, for the first variational evidence lower bound ELBO in step S31 DDTM The KL divergence term in the equation is defined as:

[0033]

[0034] Where, μ t ,σ t They represent q respectively Φ (Z t |X t The output means and variance; ε represents the factor and has ε ~ N(0,1), N(0,1) means that it follows a normal distribution with a mean of 0 and a variance of 1; This indicates that the expression follows a mean of μ. t The variance is σ t The normal distribution;

[0035] S312, based on the first variational evidence lower bound ELBODDTM The first loss function L is obtained during the training process of the first time model DDTM. DDTM :

[0036]

[0037] Preferably, in step S4, the second loss function L of the second time model DTTM DTTM for:

[0038]

[0039] Preferably, step S2 specifically includes the following steps:

[0040] S21. Based on the size W of the sliding window and the distance s the window moves, define the variables for generating samples:

[0041]

[0042] Where, x t-W+1 The sensor signal at time t-W+1; z t x represents t The corresponding health indicators HIs; z t-W+1 x represents t-W+1 The corresponding health indicators (HIs);

[0043] S22. For each input sample of the model, the sliding window method is used to generate the future sensor signal Xpre to be predicted. t :

[0044] Xpre t ={x t+1 ,...,x t+L} (3)

[0045] Where L represents the length of the future sensor signal to be predicted; x t+1 The sensor signal at time t+1; x t+L The sensor signal at time t+L;

[0046] S23, will use the unlabeled dataset D U and label dataset D L The generated samples are divided into training data groups D train and validation data group D valid .

[0047] Preferably, the deep hybrid multitasking model DHMTM interpolates between the second temporal model DTTM (α=0) based on supervised learning and the first temporal model DDTM (α=1) based on unsupervised learning.

[0048] Preferably, the device includes a turbofan engine, a lithium-ion phosphate battery, or a graphite battery.

[0049] Compared with the prior art, the technical effects of the present invention are as follows:

[0050] 1. The present invention designs a method for predicting remaining useful life based on a deep hybrid multi-task model. During the model training process, an additional future sensor signal prediction task is added. Combined with the future sensor signal prediction method, the model can learn better HIs and improve the prediction accuracy of RUL. The future sensor signal is only used during the model training process, while during the model testing process, only the sensor signals at the past and current times are used to predict the RUL of the test sample, and the future sensor signal is not required.

[0051] 2. This invention presents a residual useful life (RUL) prediction method based on a deep hybrid multi-task model. For the RUL prediction task, it proposes a semi-supervised learning (SSL) method based on multi-task learning. This method can efficiently utilize unlabeled and labeled data. By training two temporal models, DDTM and DTTM, using unlabeled and labeled data respectively, the two temporal models can share latent variables, enabling multi-task co-learning based on semi-supervised learning (SSL). In addition, this method can capture the temporal correlation of sensor signals to a large extent, further improving the prediction accuracy of RUL. Attached Figure Description

[0052] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0053] Figure 1 This is a flowchart of the remaining useful life prediction method based on a deep hybrid multi-task model of the present invention;

[0054] Figure 2 This is a diagram illustrating the process of generating the first and second sets of training samples using the sliding window method of the present invention.

[0055] Figure 3 This is a block diagram of the first time-based model DDTM of the present invention;

[0056] Figure 4 This is a block diagram of the second time model DTTM of the present invention;

[0057] Figure 5 This is a block diagram of the Deep Hybrid Multitasking Model (DHMTM) of this invention;

[0058] Figure 6 This is the training process of the Deep Hybrid Multitask Model (DHMTM) of the present invention;

[0059] Figure 7 This is a computational flowchart of the Deep Hybrid Multitask Model (DHMTM) of this invention;

[0060] Figure 8 The prediction accuracy of the deep hybrid multitasking model DHMTM trained with different L values ​​in the present invention is shown in scenarios S1 and S2, respectively. Detailed Implementation

[0061] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0062] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0063] Figure 1 The present invention illustrates a method for predicting remaining useful life based on a deep hybrid multi-task model, the method comprising the following steps:

[0064] S1. Construct unlabeled and labeled datasets: For turbofan engines, lithium-ion phosphate batteries, or graphite batteries, respectively, from N... U and N L Collect multi-sensor detection signals from similar devices to construct an unlabeled dataset D. U and label dataset D L :

[0065]

[0066] Among them, X n This indicates that data from the first to the Tth samples were collected from the nth device. n Sensor signals at any given time x t This represents the sensor signal at time t. Let represent the p-th dimension signal value of the sensor at time t, where P represents the total dimension of the sensor signal; Represents the real number field; Y n X represents n Corresponding remaining useful life (RUL) y t x represents t The corresponding remaining useful life (RUL) to be predicted.

[0067] S2. Training data set D for the model generated using the sliding window method. trainand validation data group D valid .

[0068] To better capture the temporal dependencies between multiple time-series sensor signals, this invention employs a Long Short-Term Memory (LSTM) network. To enable the use of LSTM to model dynamic sensor signal inputs, a sliding window method is used to generate samples.

[0069] S21. Based on the size W of the sliding window and the distance s the window moves, define the variables for generating samples:

[0070]

[0071] Among them, X t The sensor signals representing past and present moments are the input samples for the model; Z t Indicates from X t The corresponding health indicator HIs extracted from it; x t-W+1 The sensor signal at time t-W+1; z t x represents t The corresponding health indicators HIs; z t-W+1 x represents t-W+1 The corresponding health indicators HIs.

[0072] S22. For each input sample of the model, the sliding window method is used to generate the future sensor signal Xpre to be predicted. t :

[0073] Xpre t ={x t+1 ,...,x t+L} (3)

[0074] Where L represents the length of the future sensor signal to be predicted; x t+1 The sensor signal at time t+1; x t+L This represents the sensor signal at time t+L.

[0075] S23, will use the unlabeled dataset D U and label dataset D L The generated samples are divided into training data groups D train and validation data group D valid .

[0076] The process of generating the first and second sets of training samples using the sliding window method is as follows: Figure 2 As shown, attention should be paid to X. t Xpre serves as the input sample during both model training and testing. t and y tUsed only during model training.

[0077] S3. To learn a better HIs, for the unlabeled dataset D U Constructing a first-time model DDTM that considers future sensor signal prediction, such as Figure 3 As shown, solid lines represent the generator network, dashed lines represent the inference network, and double-headed arrows represent p... θ (Z t |X t )=q Φ (Z t |X t The equivalent constraint relationship. The first-time model DDTM includes two learning tasks: reconstructing the input sensor signal and predicting the future sensor signal.

[0078] S31. Derive the first joint probability distribution p of the first-time model DDTM. θ (X t ,Xpre t The first variational evidence lower bound of the data likelihood is ELBO. DDTM :

[0079]

[0080] Where, q Φ (Z t |X t ,Xpre t ) represents a given X t and Xpre t Z t The posterior distribution of p; θ (X t ,Xpre t Z t ) represents X t Xpre t and Z t The joint probability distribution of p; θ (Z t |X t ,Xpre t ) represents a given X t and Xpre t Z t The prior distribution of p; θ (X t ,Xpre t |Z t ) represents a given Z t Time X t and Xpre t The joint probability distribution of KL(q) Φ (Z t |Xt ,Xpre t )||p θ (Z t )) represents q Φ (Z t |X t ,Xpre t ) and p θ (Z t KL divergence values ​​between ); p θ (Z t ) represents Z t The prior distribution of .

[0081] Considering the existence of the constraint q Φ (Z t |X t ,Xpre t ) = p θ (Z t |X t )=q Φ (Z t |X t ), where p θ (Z t |X t ) represents a given X t Z t The prior distribution of q Φ (Z t |X t ) represents a given X t Derivation of Z t The reasoning network then has:

[0082]

[0083] S32, ELBO of the first variational evidence lower bound of the first-time model DDTM DDTM As the training objective, the generated component p is used. θ (X t Z t Reconstruct the input sensor signal Using the first discriminant component p θ (Xpre t |X t Z t To predict future sensor prediction signals

[0084] The first loss function of the first-time model DDTM is L DDTM The calculation specifically includes the following steps:

[0085] S311. The mean squared error (MSE) is used to measure the prediction error, and the reparameterization technique Z is applied. t =μ t +σ t ·ε, for the lower bound of the first variational evidence ELBO in step S31 DDTM The KL divergence term in the equation is defined as:

[0086]

[0087] Where, μ t ,σ t They represent q respectively Φ (Z t |X t The output means and variance; ε represents the factor and has ε ~ N(0,1), N(0,1) means that it follows a normal distribution with a mean of 0 and a variance of 1; This indicates that the expression follows a mean of μ. t The variance is σ t It follows a normal distribution.

[0088] S312. According to the lower bound of the first variational evidence ELBO DDTM The first loss function L is obtained during the training process of the first-time model DDTM. DDTM :

[0089]

[0090] S4, for the labeled dataset D L Construct a second time model DTTM that considers predictions of future sensor signals, such as Figure 4 As shown, only the first-time model DDTM is supplemented with a remaining useful life (RUL) prediction task.

[0091] S41. Derive the second joint probability distribution p on the second time model DTTM. θ (X t ,Xpre t ,y t The second variational lower bound of evidence ELBO DTTM :

[0092]

[0093] Where, p θ (X t ,Xpre t ,y t |Z t ) represents a given Z t Time X t Xpre t and yt The joint probability distribution.

[0094] S42, ELBO with maximizing the second variational evidence lower bound of the second time model DTTM DTTM As the training objective, the generated component p is used. θ (X t Z t Reconstruct the input sensor signal Using the first discriminant component p θ (Xpre t |X t Z t Predicting future sensor prediction signals Using the second discriminant component p θ (y t |X t Z t Predict the remaining useful life to be predicted The second time model DTTM proposed in this invention can learn information that is more conducive to HIs extraction by jointly training these three learning tasks.

[0095] The second loss function L of the second time model DTTM DTTM for:

[0096]

[0097] S5. To efficiently utilize both unlabeled and labeled data to learn more valuable information, a Deep Hybrid Multitask Model (DHMTM) considering time dependencies is constructed and trained to obtain the final DHMTM model. The final DHMTM model is the trained Deep Hybrid Multitask Model (DHMTM) containing the optimal combination of weight parameters.

[0098] Inspired by Deep Hybrid Models (DHM), the Deep Hybrid Multitask Model (DHMTM) is well-suited for predicting Remaining Lifetime (RUL) in semi-supervised scenarios. DHMTM combines a first-time model (DDTM) for unlabeled data and a second-time model (DTTM) for labeled data by sharing latent variables. DHMTM works as follows: Figure 5 As shown, the solid line represents the first time-limited model DDTM, and the dashed line represents the second time-limited model DTTM.

[0099] The training process of the Deep Hybrid Multitask Model (DHMTM) is as follows: Figure 6 As shown.

[0100] S51. In each training round, the training data group D trainAs input to the deep hybrid multi-task model DHMTM, the training data set D is computed. train The loss function is determined, and the parameters of the deep hybrid multitasking model (DHMTM) are updated using the stochastic gradient descent algorithm.

[0101] S52, Based on the updated Deep Hybrid Multitask Model (DHMTM) and validation dataset D valid Calculate the validation data set D valid The verification loss.

[0102] S53. Repeat steps S51 and S52 until the maximum set number of rounds is reached, and use the Deep Mixed Multi-Task Model (DHMTM) with the minimum validation loss as the final DHMTM model. The Deep Mixed Multi-Task Model (DHMTM) with the minimum validation loss obtained in a given training period is the Deep Mixed Multi-Task Model (DHMTM).

[0103] The computational flowchart of the Deep Hybrid Multitask Model (DHMTM) is as follows: Figure 7 As shown, solid lines represent the feedforward network for unlabeled data, and dashed lines represent the feedforward network for labeled data. The networks used for future sensor signal prediction and input sensor signal reconstruction can process both unlabeled and labeled data simultaneously, while the network used for remaining lifetime (RUL) prediction only processes labeled data. During the training of the Deep Hybrid Multitask Model (DHMTM), the network parameters are updated using the stochastic gradient descent algorithm.

[0104] The loss function of the Deep Hybrid Multitask Model (DHMTM) is:

[0105] L DHMTM =αL DDTM +(1-α)L DTTM (10)

[0106] Where α represents the weighting coefficient, and 0≤α≤1.

[0107] The Deep Hybrid Multitask Model (DHMTM) interpolates between the supervised learning-based second-time model (DTTM, α=0) and the unsupervised learning-based first-time model (DDTM, α=1).

[0108] S6. Transfer the newly received sensor signal D T As input to the test set, the final DHMTM model, trained using the Deep Hybrid Multi-Task Model (DHMTM), is used to predict the Remaining Usage (RUL) of the operating equipment. The newly received sensor signal D... T This refers to sensor signals whose time series are located after the labeled and unlabeled datasets.

[0109] In one specific embodiment, two time-series datasets—the C-MAPSS dataset and the lithium battery dataset—were used to verify the effectiveness and superiority of the proposed deep hybrid multi-task model DHMTM in predicting remaining useful life (RUL).

[0110] The experiment compared two contrasting models with the Deep Hybrid Multitask Model (DHMTM): Model 1 is a regression model based on a Long Short-Term Memory (LSTM) network, trained using only labeled data; Model 2, inspired by the DHM model, is a hybrid model combining a Variational Auto-Encoder (VAE) and a Gaussian Stochastic Neural Network (GSNN), used to verify the benefits of considering future sensor signal prediction in HIs learning for Remaining Lifetime (RUL) prediction. Compared to the DHMTM model, Model 2 only reduces the task of predicting future sensor signals. Furthermore, the model's structural parameters will be optimized and adjusted based on different datasets.

[0111] This invention selects the root mean square error (RMSE) as the evaluation index for the accuracy of RUL prediction. The specific calculation method is as follows:

[0112]

[0113] in, and represents the actual and predicted remaining useful life (RUL), respectively; N represents the total sample size.

[0114] (1) C-MAPSS dataset

[0115] C-MAPSS is a tool for simulating real large commercial turbofan engines. It simulates an engine model with a thrust level of 90,000 pounds, capable of flying at altitudes from sea level to 40,000 feet, Mach numbers from 0 to 0.90, and sea-level temperatures from -60 to 103°F. The equipment includes transient aerodynamic models and engine controllers, allowing the engine to operate at various thrust levels across the entire range of flight conditions.

[0116] The C-MAPSS dataset describes the degradation process of aircraft engines and contains four subsets with different fault types and operating conditions. This invention selects the FD001 dataset, which has only one fault type and constant operating conditions. This dataset includes a training dataset (train_FD001) and a test dataset (test_FD001), both containing cyclic information and measurement data collected from three operating condition recording devices and 26 sensor devices. Measurement data collected from 14 sensor devices that change over time are selected as input sensor signals.

[0117] Machines and equipment begin to degrade and eventually fail after operating in a healthy state for a certain number of cycles. By setting a maximum remaining useful life (RUL) value, the cycle life can be divided into a healthy phase and a degradation phase. For the C-MAPSS dataset, a maximum remaining useful life (RUL) of 120 to 130 cycles is appropriate; therefore, the maximum RUL is set to 120 cycles. Next, the input sensor signals and their corresponding remaining useful life (RUL) are standardized using the Min-Max normalization method, and training samples are generated using the sliding window method.

[0118] The training set was set to train_FD001, and the test set was set to test_FD001. To ensure the sufficiency of the results, two semi-supervised scenarios were set up, denoted as S1 and S2. In scenario S1, 50 engine data points from the 100 engines in train_FD001 were randomly selected as the D dataset. L The data for another 50 engines was tagged and used as D. U And ensure D train and D valid The ratio of the number of engines in the S2 scenario is 4:1. In scenario S2, there is less labeled data compared to scenario S1; that is, data from 10 randomly selected engines are used as D. L The data for the other 90 engines were tagged and used as D. U .

[0119] After experimenting with different values, the hyperparameters were set as follows: batch size = 512, initial learning rate = 0.0003, sliding window size W = 30, and window sliding length s = 1. Furthermore, during the training of the Deep Hybrid Multitask Model (DHMTM), the interpolation parameter α varied from 0 to 1 in steps of 0.1, and the length L of the future sensor signal to be predicted varied from 5 to 30 in steps of 5. The structures of the DHMTM model and the two comparative models are listed in Tables 1-3, respectively.

[0120]

[0121] Table 1

[0122]

[0123] Table 2

[0124]

[0125] Table 3

[0126] Based on the above experimental setup, the deep hybrid multi-task model DHMTM with the minimum validation loss was obtained, using the test set D. T The Remaining Lifetime (RUL) prediction accuracy of the three models was evaluated. Table 4 shows the highest prediction accuracy of the Deep Hybrid Multitask Model (DHMTM) in scenarios S1 and S2. Specifically, in scenario S1, the optimal L and α values ​​for DHMTM are 10 and 0.8, respectively, while in scenario S2, they are 10 and 0.7, respectively.

[0127] Training data (number of labeled engines) All data (100) S1(50) S2(10) Model 1 (using only label data) 14.76 15.13 15.95 Model 2 - 15.00 15.78 DHMTM - 14.69 15.66

[0128] Table 4

[0129] By fixing α and changing L, and observing the predicted Remaining Useful Life (RUL) under different L values, we can analyze the impact of the value of L on the prediction accuracy. Figure 8 The prediction accuracy of the deep hybrid multitasking model DHMTM trained with different values ​​of L is shown in scenarios S1 (α = 0.8) and S2 (α = 0.7). It can be observed that different values ​​of L will produce different remaining lifetime (RUL) prediction results.

[0130] In addition, the remaining useful life (RUL) prediction accuracy of the two comparative models is also given in Table 4, where α is 0.4 and 0.2 when Model 2 has the highest prediction accuracy in S1 and S2 scenarios, respectively.

[0131] (2) Lithium battery dataset

[0132] The lithium battery dataset comes from an accelerated degradation experiment, which used 124 lithium-ion phosphate (LFP) / graphite batteries manufactured by A123 Systems (APR18650M1A), placed in a horizontal cylindrical fixture on a 48-channel Arbin LBT potentiostat, with the forced convection temperature set to 30°C.

[0133] These batteries, with a nominal capacity of 1.1 Ah and a nominal voltage of 3.3 V, were cycled under different fast-charging and identical discharge conditions. The dataset was divided into three batches, named according to either the "batch date" or the date the test began. We selected the batch from June 30, 2017, for our experimental study. Each batch contained seven degradation data points: discharge capacity, charge capacity, internal resistance, maximum temperature, average temperature, minimum temperature, and charging time. These degradation data were used as input sensor signals to the proposed model. Furthermore, battery cycle life was defined as the number of cycles required for the actual capacity to reach 80% of the nominal capacity. Similar to Case 1, samples were generated using Min-Max normalization and a sliding window method.

[0134] This invention considers degradation data from 39 batteries in a selected batch and randomly selects 4 batteries as D. T An additional 35 batteries were used as the training dataset. Similar to Case 1, two semi-supervised scenarios were set up, denoted as S3 and S4 respectively. In scenario S3, 10 batteries were randomly selected as D. L Another 25 batteries had their labels removed as D U In scenario S4, randomly select 5 batteries as D. L Another 30 batteries had their labels removed as D U Similar to Case 1, set D. train and D valid The ratio of batteries in the mixture is 4:1.

[0135] After experimenting with different values, the hyperparameter settings were as follows: batch size = 512, initial learning rate = 0.0003, sliding window size W = 30, window sliding length s = 1, interpolation parameter α varied from 0 to 1 in steps of 0.1, and the length of the future sensor signal to be predicted L varied from 5 to 50 in steps of 5. Using the lithium battery dataset, the structures of the deep hybrid multitasking model DHMTM and the two contrasting models are listed in Tables 5-7, respectively.

[0136]

[0137] Table 5

[0138]

[0139] Table 6

[0140]

[0141] Table 7

[0142] Next, we will use test set D. TThe highest prediction accuracy of the Deep Hybrid Multitask Model (DHMTM) in scenarios S3 and S4 is shown in Table 8. In scenario S3, the L and α of DHMTM are 15 and 0.4, respectively, while in scenario S4, the L and α are 20 and 0.8, respectively.

[0143] Training data (number of battery tags) All data (35) S3(10) S4(5) Model 1 (using only label data) 25.47 27.41 69.56 Model 2 - 26.87 53.89 DHMTM - 25.31 42.21

[0144] Table 8

[0145] Similarly, by fixing α and changing L, prediction results were obtained for different L values ​​in the lithium battery dataset, and the impact of L value on prediction accuracy was analyzed. Figure 7 The prediction accuracy of the deep hybrid multitasking model DHMTM trained with different L values ​​is shown in the S3 scenario (α = 0.4) and the S4 scenario (α = 0.8).

[0146] First, in the two cases mentioned above, the method proposed in this invention achieved higher prediction accuracy compared to the other two comparative models. This indirectly proves that the Deep Hybrid Multi-Task Model (DHMTM) can more effectively utilize unlabeled and labeled data, and learn better HIs by embedding future sensor signal prediction tasks. Furthermore, from... Figure 6 and Figure 7 It was observed that by adjusting the length L of the future sensor signal to be predicted, there exists an optimal value of L during the model training process for constructing HIs. This value varies depending on the specific problem and the input data.

[0147] Secondly, comparing the RMSE results, both the Deep Hybrid Multitask Model DHMTM and Model 2, based on semi-supervised learning SSL, have lower RMSE scores than Model 1, based on supervised learning. This demonstrates that with only a small amount of labeled data, the semi-supervised learning SSL model utilizing unlabeled data can effectively improve the prediction accuracy of RUL. Furthermore, in both cases, the RMSE value of the Deep Hybrid Multitask Model DHMTM is lower than that of Model 2, further proving that considering future sensor signal predictions can help learn better HIs and further improve the prediction accuracy of RUL.

[0148] Finally, it was observed that the minimum RMSE value of the Deep Hybrid Multitask Model (DHMTM) in scenarios S1 and S3 was similar to that of Model 1, which was trained using all labeled data. This indicates that in the case of semi-supervised learning SSL, the DHMTM can efficiently utilize unlabeled data and achieve prediction accuracy similar to that of supervised learning models using a large amount of labeled data. Since obtaining a large amount of labeled data is often costly and even difficult in some cases during Remaining Lifetime (RUL) prediction, the method proposed in this invention is more efficient in practical applications and significantly reduces prediction costs.

[0149] In summary, this invention proposes a deep hybrid multi-task model (DHMTM) that considers future sensor signal prediction based on semi-supervised learning (SSL). This model can be used to construct better HIs (Hypothetical Indicators), thereby improving the prediction accuracy of Remaining Life (RUL). Based on existing generative models, this invention proposes two temporal models: DDTM for unlabeled data and DTTM for labeled data. These DDTM and DTTM are combined by sharing latent variables to obtain the deep hybrid multi-task model (DHMTM), which is then used to complete the RUL prediction task in a semi-supervised scenario. By embedding an additional future sensor signal prediction task during model training, DHMTM achieves higher RUL prediction accuracy, indirectly demonstrating its superior ability in HIs learning. Furthermore, DHMTM can fully utilize unlabeled data, enabling accurate RUL prediction results even with a limited amount of labeled data. The effectiveness of the proposed method is verified using C-MAPSS and lithium battery datasets. Experimental results show that DHMTM can learn better HIs, thus achieving higher RUL prediction accuracy compared to comparative models.

[0150] This invention presents a residual lifetime (RUL) prediction method based on a deep hybrid multi-task model. During model training, an additional future sensor signal prediction task is added, combining this method with future sensor signal prediction to enable the model to learn better HIs (Latent Hierarchy Signs) and improve RUL prediction accuracy. Future sensor signals are only used during model training; during model testing, only past and current sensor signals are needed to predict the RUL of test samples, eliminating the need for future sensor signals. For the RUL prediction task, a semi-supervised learning (SSL) method based on multi-task learning is proposed. This method efficiently utilizes both unlabeled and labeled data, training two temporal models (DDTM and DTTM) using unlabeled and labeled data respectively. These two models share latent variables, enabling multi-task co-learning based on semi-supervised learning (SSL). Furthermore, this method can capture the temporal correlation of sensor signals to a greater extent, further improving the RUL prediction accuracy.

[0151] Finally, it should be noted that the above embodiments are for illustration only and not for limiting the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention without departing from the spirit and scope of the present invention. Any modifications or partial substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting the remaining useful life based on a deep hybrid multi-task model, characterized in that, It comprises the following steps: S1, constructing a no-label dataset and a label dataset: respectively collecting multi-sensor detection signals from and the same type of devices to construct a no-label dataset and a label dataset : （1）； wherein, represents the sensor signal at time t from the first to the Tth dimension of the n th device, n , represents the sensor signal at time t, , represents the sensor signal value at time t in the pth dimension, P represents the total dimension of the sensor signal; represents the real number field; represents the corresponding remaining useful life RUL, ; represents the corresponding remaining useful life RUL to be predicted;​ S2, generating a training data set for the model using a sliding window approach and a validation data set ; S3, for the unlabeled dataset constructing a first temporal model DDTM that takes into account the prediction of future sensor signals; S4, for the label dataset constructing a second temporal model DTTM that takes into account future sensor signal predictions; S41、deriving a second joint probability distribution on the second temporal model DTTM under a second variational evidence lower bound : （6）； wherein, denotes the sensor signals representing past and current time instants, i.e. the input samples for the model; denotes the future sensor signals to be predicted; denotes the corresponding health indicators HIs extracted from the ; denotes the given ; derivation of the inference network; denotes the joint probability distribution of the given ; , and ; denotes the prior distribution of the ; denotes the KL divergence value between the and ; S42, maximizing the second variational evidence lower bound of the second temporal model DTTM for training purposes, using a generating component reconstructing input sensor signals , using a first discriminative component predicting future sensor prediction signals , using a second discriminative component predicting a remaining useful life RUL to be predicted ; S5, constructing a deep hybrid multi-task model DHMTM considering time dependence, and training to obtain a final DHMTM model; S51. In each training round, a training data set As input to the deep hybrid multi-task model DHMTM, a loss function of the training data set is calculated and the parameters of the deep hybrid multi-task model DHMTM are updated by a stochastic gradient descent algorithm. S52, calculate a validation loss of the validation data set based on the updated deep hybrid multi-task model DHMTM and the validation data set ;​ S53, repeating the step S51 and the step S52 until a maximum set round is reached, and taking a deep hybrid multi-task model DHMTM with a minimum validation loss as the final DHMTM model; S6, newly received sensor signals As input for the test set, the final DHMTM model trained with the deep hybrid multi-task model DHMTM is used to complete the remaining useful life RUL prediction of the running device. 2.The method of claim 1, wherein, The step S3 specifically comprises the following steps: S31, deriving a first joint probability distribution of the first temporal model DDTM a first variational evidence lower bound on the data likelihood : （4）； wherein, denotes the posterior distribution of and denotes the joint probability distribution of and denotes the prior distribution of , and denotes the joint probability distribution of and denotes the prior distribution of and denotes the joint probability distribution of and denotes the joint probability distribution of and denotes the joint probability distribution of Taking into account the existence of set constraints where representing a given below a prior distribution of, then we have: （5）； S32, maximizing the first variational evidence lower bound of the first temporal model DDTM for training purposes, using a generative component reconstructing input sensor signals , using a first discriminative component to predict future sensor prediction signals . 3.The deep hybrid multi-task model based remaining useful life prediction method of claim 1, wherein, The loss function of the deep hybrid multi-task model DHMTM in the step S5 is: （10）； wherein, represents a first loss function; represents a second loss function; represents a weight coefficient, and . 4.The method of claim 2, wherein, the first loss function of the first temporal model DDTM in the step S3 comprises in particular the following steps: S311, the mean square error (MSE) is used to measure the prediction error, and a reparameterization trick is applied The KL divergence term in the first variational lower bound in step S31 is defined as: The KL divergence term in the first variational lower bound in step S31 is defined as: （7）； wherein respectively denote the mean and variance of the output; denote factors and have , denote a normal distribution with mean 0 and variance 1; denote a normal distribution with mean and variance . S312、according to the first variational evidence lower bound , obtaining the first loss function in the first temporal model DDTM training process : （8）。 5.The method of claim 3, wherein, the second loss function of the second temporal model DTTM in the step S4 is: （9）。 6.The deep hybrid multi-task model based remaining useful life prediction method of claim 1, wherein, The step S2 specifically comprises the following steps: S21, defining a variable of generating samples based on a size W of a sliding window and a distance s of window movement: （2）； wherein, represents the sensor signal at the time instant; represents the corresponding health indicator His; represents the corresponding health indicator His; S22, for each input sample of the model, a sliding window method is adopted to generate the future sensor signals to be predicted : （3）； wherein L denotes the length of the future sensor signal to be predicted; denotes the sensor signal at time instant t; denotes the sensor signal at time instant t; S23, the use of a no-label dataset and a labeled dataset The generated samples are divided into a training data set and a validation data set . 7.The deep hybrid multi-task model based remaining useful life prediction method of claim 1, wherein, The deep hybrid multi-task model DHMTM interpolates between the second time model DTTM based on supervised learning ( ) and the first time model DDTM based on unsupervised learning ( ). 8.The method of predicting the remaining useful life based on the deep hybrid multi-task model according to claim 1, wherein, The device includes a turbofan engine, a lithium ion phosphate battery, or a graphite battery.

Images