An engine life prediction method based on correlation analysis and multi-task learning

Abstract

The application discloses an engine life prediction method based on correlation analysis and multi-task learning, obtains data detected by a sensor during engine operation and processes the data; generates an initial health index HI; divides the initial health index into a normal stage and a degradation stage by using a mutation point; restores a random process and a probability density distribution of the initial health index degradation stage by using a SINN model after training; generates a contrast learning health index and a global health index, and then uses a cascade NTM model to predict the remaining service life of the engine based on the global health index. The application uses SINN to statistically analyze degradation characteristics, obtains a failure mode according to a probability density function and a KL divergence, and realizes a good trade-off between prediction performance and running time cost by using the cascade NTM model. The parameter migration based on a one-to-many model realizes automatic algorithms for failure analysis and life prediction, the NSGA-II fine-tuning algorithm adopted balances time and performance, and has practical application value.

Description

Technical Field

[0001] This invention relates to the field of aero-engine life prediction technology, and in particular to an engine life prediction method based on correlation analysis and multi-task learning. Background Technology

[0002] Aircraft engine components operate under relatively high-intensity environments for extended periods, leading to a continuous and cumulative degradation in their overall performance, causing them to deviate from their normal operating conditions. Predictive and health management technologies are crucial for analyzing engine condition monitoring data, identifying early signs of engine degradation, predicting degradation trends, and estimating remaining service life. A deeper understanding of engine health is needed to ultimately prevent serious failures, which is of great significance for ensuring flight safety.

[0003] Current mainstream prediction methods are mainly divided into two categories: physical models and data-driven methods. Physical models typically require establishing accurate physical laws and combining detailed failure modes and historical data for prediction. Due to the complexity and variability of engine failure modes and the dynamic nature of their operation, establishing accurate and reliable physical models is extremely difficult. Data-driven methods explore data patterns and extract potential information affecting aircraft engine performance degradation; these mainly include statistical methods and deep learning methods. Data-driven methods based on deep learning algorithms can learn better fitting functions on limited input samples and neural nodes, thereby improving prediction accuracy and efficiency. In the field of time series prediction, NTM achieves longer-term memory and higher prediction accuracy by adding external memory. Although NTM has sufficient memory space, it lacks effective representation learning capabilities to improve prediction performance.

[0004] To enhance representation capabilities, a contrastive learning method is proposed that compares unsupervised positive and negative samples. Its aim is to extract meaningful features by emphasizing the differences between instances within the same sequence while minimizing the differences between sequences. However, existing deep learning methods, primarily based on neural networks, still lack statistical methods for analyzing uncertainty degradation features. Statistical analysis methods offer excellent interpretability for degradation behaviors (such as probability density functions). Stochastic process models (SINNs) can recover the trajectories and statistical features of multiple stochastic processes, such as probability density functions (PDFs). However, in practical applications, it is necessary to comprehensively utilize the advantages of both deep learning methods and statistical analysis methods, encompassing two distinct tasks.

[0005] Multi-task learning, proposed by simultaneously processing multiple related tasks, can leverage common knowledge across tasks to improve generalization performance for each task. However, despite the popularity of multi-task learning methods, there are currently no multi-task learning methods specifically for RUL prediction and statistical fault analysis of aero-engines.

[0006] Therefore, a comprehensive multi-task learning method that combines the advantages of deep learning and fault feature analysis is needed. On the one hand, it effectively solves the problem of stable multi-task learning for rule-based prediction; on the other hand, it provides an effective RUL prediction method with uncertainty degradation characteristics. Furthermore, there is currently no multi-task learning method that simultaneously considers fault pattern recognition and RUL prediction. Therefore, it prevents the simultaneous resolution of key correlation analysis problems. Summary of the Invention

[0007] Purpose of the invention: To address the above problems, the purpose of this invention is to provide an engine life prediction method based on correlation analysis and multi-task learning.

[0008] Technical solution: The present invention provides an engine life prediction method based on correlation analysis and multi-task learning, comprising the following steps:

[0009] Step 1: Acquire the data detected by the sensors during engine operation and normalize the data to be dimensionless;

[0010] Step 2: Use the self-attention mechanism to perform multi-parameter fusion on the processed data to generate the initial health index HI;

[0011] Step 3: Construct a change point detection model based on NTM, train the change point detection model, use the trained change point detection model to find abrupt change points, and use the abrupt change points to divide the initial health index into normal stage and degenerative stage.

[0012] Step 4: Construct and train a SINN model. Use the trained SINN model to reconstruct the stochastic process and probability density distribution of the initial health index degradation stage, and identify different failure modes based on the differences in probability density distribution.

[0013] Step 5: Construct and train a contrastive learning model. Use the trained contrastive learning model to obtain a high-dimensional representation of the initial health index degradation stage and generate a contrastive learning health index.

[0014] Step 6: Concatenate the features of the stochastic process in the degradation stage with the contrastive learning health index to generate a global health index, and then use the cascaded NTM model to predict the remaining service life of the engine based on the global health index.

[0015] Step 7: Using one-to-many parameter transfer and NSGA-II fine-tuning algorithm, transfer the parameters of the trained change point detection model to subsequent tasks, use NSGA-II to optimize and fine-tune some parameters, use the dataset for instance validation, and perform ablation analysis.

[0016] Furthermore, the calculation expression for the normalized dimensionless processing in step 1 is as follows:

[0017] (1);

[0018] in, X represents the normalized data from sensor j. j (t) represents the raw data detected by sensor j, min(X) j ) and max(X j ) represent the minimum and maximum values ​​of the data detected by sensor j, respectively.

[0019] Furthermore, the process of generating the initial health index HI in step 2 includes:

[0020] The processed data is used as input data. Where n is the number of data points and d is the dimension of the data, for each input data point Calculate the corresponding query vector q i Key vector k j Sum vector v j The calculation expressions are as follows:

[0021] (2);

[0022] (3);

[0023] (4);

[0024] in, These are three trainable weight matrices in the self-attention mechanism, used to generate the query vector, key vector, and value vector, respectively.

[0025] Calculate each query vector With all key vectors within the window The similarity scores are normalized to obtain the self-attention weights, expressed as:

[0026] (5);

[0027] in, Self-attention weights for each sensor, This represents the dot product operation. The similarity score reflects the strength of the relationship between the query vector and the key vector. Normalization makes the sum of the weights equal to 1.

[0028] For each value vector The weighted summation based on the self-attention weights yields the fused one-dimensional data, expressed as:

[0029] (6);

[0030] Among them, HI i This is the initial health index.

[0031] Furthermore, the implementation process of step 3 includes:

[0032] Using LSTM as the controller of NTM to construct a change point detection model, the forward pass expression is as follows:

[0033] (7);

[0034] (8);

[0035] (9);

[0036] (10);

[0037] (11);

[0038] (12);

[0039] in, This represents the output of the forget gate at time t. This represents the output of the input gate at time t. This represents the output of the output gate at time t. This represents the state of the candidate cell at time t. x represents the state of the update unit at time t. t h represents the input at time t. t-1 W represents the hidden state value at time t-1; f W i W o and W a These represent the forget gate, input gate, output gate, and h during the feature extraction process, respectively. t-1 Weighting coefficients; V f V i V o and V a These represent the forget gate, input gate, output gate, and x during the feature extraction process, respectively. t Weighting coefficients; bf b i b o and b a represents the bias values ​​in the forget gate, input gate, output gate, and feature extraction process, respectively; tanh represents the tangent hyperbolic function, and σ represents the activation function Sigmoid; The Hadamard product is the element-wise multiplication symbol.

[0040] The input to the change point detection model is the fused initial health indicator HI, denoted as... , This represents the input sequence from time tm to time t. This indicates whether an engine unit whose health status and remaining service life are being monitored is in a normal operating phase or a degradation phase. A point before a change is marked as 0 (normal phase), and otherwise as 1 (degradation phase). This change-point detection model... Prediction based on multi-sensor data A nonlinear autoregressive function is used to capture the complex relationship between the initial health parameter HI and the points of change. The expression is as follows:

[0041] (13);

[0042] in, It is the parameter set of NTM, which contains the parameters in formulas (7) to (10). ;

[0043] In the task of change point detection, NTM is a classifier. The optimal parameter set of the change point detection model is trained by minimizing the binary cross-entropy, and the loss function is:

[0044] (14);

[0045] Where T is the number of time points in the training dataset;

[0046] Through training, the output of NTM The first moment when the threshold is exceeded is defined as the mutation point. As shown in formula (15):

[0047] (15);

[0048] Where threshold is the threshold for cross-validation to find the best classification; This indicates the moment when a search needs to be performed one by one. As a condition, that is, to guarantee Within the time frame t under consideration;

[0049] Based on mutation points The determination of the initial health index HI time series is divided into two stages. Before the mutation point, the engine unit has not degraded or the degradation is negligible, which is the normal stage. After the mutation point, the engine unit performance deteriorates and the remaining life gradually decreases, which is the degradation stage.

[0050] Furthermore, the implementation process of step 4 includes:

[0051] A SINN model is constructed using a two-layer LSTM. This SINN model outputs the trajectory and statistical characteristics of a random process based on the input random sequence, including the probability density distribution PDF and the autocorrelation function ACF.

[0052] The SINN model is trained using the loss functions of ACF and PDF. The inputs are the initial health index during the degradation phase, the corresponding target PDF, and the target ACF. The trained SINN model outputs both the stochastic process and the probability density distribution PDF. The expression for the ACF loss function is as follows:

[0053] (16);

[0054] The loss function for PDF is:

[0055] (17);

[0056] Among them, ACF T (τ) is the autocorrelation function of the objective stochastic process, τ is the lag number, and ACF O (τ) is the autocorrelation function output by the SINN model; Let be the probability density function of the target stochastic process, calculated using Gaussian kernel density estimation, where h is the width of the kernel function. The probability density function output by the SINN model;

[0057] Based on the PDFs of engine health indices corresponding to different failure modes output by the SINN model, the KL divergence between the various PDFs is calculated using the following expression:

[0058] (18);

[0059] Fault modes are identified based on the magnitude of KL divergence, as shown in formula (19):

[0060] (19);

[0061] Where p(x) and q(x) are probability density distributions, and α is the KL divergence threshold for the same fault type. As a fault mode, when the KL divergence between an unknown degraded PDF and a PDF of a known fault mode is less than a threshold α, they are determined to belong to the same fault mode; otherwise, they belong to different fault modes. Accurate fault identification is achieved through multiple comparisons.

[0062] Furthermore, the implementation process of step 5 includes:

[0063] First, data augmentation techniques are used to transform the initial health parameter HI during the degradation stage. These techniques include jittering, scaling, time shifting, and neighborhood segmentation, with the corresponding formulas as follows:

[0064] (20);

[0065] (twenty one);

[0066] (twenty two);

[0067] (twenty three);

[0068] Where Jitter is the jitter operation, t i Let A be the timestamp of the i-th sampling point, D be the ideal sampling interval, and N be the number of sampling points; Scaling is the scaling operation, and A is the scaling factor; Shifting is the time-shift operation. This refers to the time offset; Neighborhood Segmentation is the neighborhood segmentation operation. and Let f(x(t)) be the start and end times of the neighborhood, and f(x(t)) be a function that segments the signal outside the neighborhood.

[0069] The sample pairs generated after data transformation are denoted as positive samples. Then, an LSTM-based contrastive learning model is constructed. The positive samples are input into the model to extract high-dimensional feature representations. Negative samples are generated by pairing the feature representations of the positive samples with semantically unrelated feature representations from other samples in the dataset. The positive and negative pairs are then presented to a contrastive loss function. The InfoNCE loss function guides the contrastive learning model to learn feature representations, making positive pairs closer in the embedding space and pushing negative pairs further apart. The expression for the InfoNCE loss function is:

[0070] (twenty four);

[0071] Where, sim(A, P) i ) are anchor sample A and positive sample Similarity between them; sim(A, N) j ) are anchor sample A and negative sample The similarity between them, the anchor samples are automatically generated during the training process, It is a temperature parameter that needs to be preset to control the smoothness of the probability distribution;

[0072] By training a contrastive learning model, the high-dimensional features corresponding to the health parameters after separating positive and negative samples are output. This high-dimensional feature matrix is ​​denoted as the contrastive learning health index (CHI), and its expression is:

[0073] (25);

[0074] in, This represents the data at the t-th time point of the p-th dimension feature.

[0075] Furthermore, the implementation process of step 6 includes:

[0076] First, the stochastic process features output by the SINN model are concatenated with the CHI feature. The concatenated index is designated as the Global Health Index (OHI), and is represented as follows:

[0077] (26);

[0078] A prediction model is constructed, which is a cascaded NTM model composed of multiple identical NTMs connected in succession. Each NTM uses an LSTM as its controller, and the prediction model is trained accordingly. The read-write memory process of the LSTM is represented as follows:

[0079] (27);

[0080] (28);

[0081] (29);

[0082] (30);

[0083] (31);

[0084] (32)

[0085] Among them, w t w (i) represents the weight of the write header at time t, e t Let α be the erasure vector. t Both the added vector, erased vector, and added vector are generated by the controller; w t r (i) represents the weight of the read head at time t, and the vector r returned by the read head. tDefined as a weighted sum of rows in memory; w t c (i) is the content addressing weight of the i-th address, β t It represents the bond strength, and K is the similarity function. t It is a key vector, M t (i) is the i-th row of the storage matrix; It is the shift weight of the i-th address, w t (i) is the sharpening weight for the i-th address, w t g (j) is the interpolation weight of the j-th address, s t (ij) is the shift weight of the (ij)th element, γ t It is the sharpening coefficient. Based on the weight vector of the previous time step, an interpolation gate scalar output by the controller, a shift weight vector, and a sharpening factor scalar, a new weight vector is calculated, which represents the result obtained after interpolating, shifting, and sharpening the weight vector of the previous time step.

[0086] In the cascaded NTM model, each NTM is considered a submodule. Starting from the second submodule, the prediction error of the previous submodule is continuously passed as output to the next submodule. That is, all subsequent submodules predict the output error of the previous submodule, thereby improving prediction accuracy by continuously reducing the prediction error. When the prediction error of the cascaded NTM model is less than the set prediction accuracy or the number of submodules reaches a set value, the cascaded NTM model will automatically terminate the process of expanding the submodules and output the error prediction value of the last submodule. Then, the OHI is input into the cascaded NTM model for engine remaining life prediction. The predicted remaining life is the sum of the outputs of all submodules, expressed as:

[0087] (33);

[0088] Among them, Y predi For the predicted output values ​​of each submodule;

[0089] The prediction model uses a quantile loss function to guide the upper and lower bounds of the output prediction interval; the quantile loss function is a piecewise formula, expressed as:

[0090] (34);

[0091] Where y is the target variable, and f(x) is the predicted value. It is a quantile; when When = 0.5, this loss function is the same as the loss function used for point prediction; when When the value is 0.95, the cascaded NTM model will try to make the predicted value greater than the true value to minimize the loss, which is the upper limit of the interval; when When the value is 0.05, the cascaded NTM model will try to make the predicted value less than the true value, which is the lower limit of the interval.

[0092] Furthermore, the implementation process of step 7 includes:

[0093] Let's denote the NTM-based change point detection model as the first task model. First, train the first task model using GridsearchCV on a pre-defined parameter space. The search is performed, and the parameter space is represented as follows:

[0094] (35);

[0095] Where, p k (k=1,2,..,N) represents a specific combination of hyperparameters, where N represents the number of combinations in the parameter grid;

[0096] For each hyperparameter combination p k Perform i cross-validations, k=1,2,...,N, and calculate the average score of the i cross-validations using the following expression:

[0097] (36);

[0098] Among them, CV j For the j-th cross-validation, score(p) k ,CV j ) represents the parameter combination p k The performance score in the j-th cross-validation;

[0099] Iterate through N hyperparameter combinations and determine the optimal combination:

[0100] (37);

[0101] The argmax operator returns avg_score(p k The hyperparameter combination p corresponding to the maximum value of the function k That is, the optimal hyperparameter combination p best The change point detection model is trained using this optimal combination of hyperparameters;

[0102] Then, the parameters of the change point detection model from Task 1 are transferred to the subsequent SINN model, contrastive learning model, and cascaded NTM model, and NSGA-II fine-tuning is performed: optimization objectives are defined for each model, including minimum MSE and shortest training time; Pareto optimization is performed, the crowding distance of the solution within each Pareto front is calculated, and a set of candidate solutions balancing time and performance is given.

[0103] Finally, a set of solutions is randomly selected to train the models for subsequent tasks and output the fine-tuned models. The engine remaining life prediction model based on fault analysis and multi-task learning is used to predict the engine remaining life, obtain the prediction structure, and then output the prediction curves and errors in sequence.

[0104] Beneficial effects: Compared with the prior art, the significant advantages of this invention are:

[0105] This invention utilizes SINN to perform statistical analysis on degradation features, and obtains failure modes based on probability density functions and KL divergence. The cascaded NTM model achieves a good trade-off between prediction performance and runtime cost. Based on parameter transfer of a one-to-many model, this invention automates the computation of failure analysis and lifetime prediction. The NSGA-II fine-tuning algorithm used balances time and performance, thus enabling high accuracy and interpretability of remaining lifetime prediction with less time and manpower, and has practical application value. Attached Figure Description

[0106] Figure 1 The flowchart shows an engine life prediction method based on correlation analysis and multi-task learning.

[0107] Figure 2 Flowchart for training a cascaded NTM model;

[0108] Figure 3 A flowchart of a multi-task learning algorithm;

[0109] Figure 4 A rendering of the Health Indicator (HI) based on data fusion;

[0110] Figure 5 The results of the change point test for the 81st engine;

[0111] Figure 6 The results of dividing the normal and degradation phases of the 81st engine;

[0112] Figure 7 The probability density distribution of the 81st engine during the normal and degradation phases;

[0113] Figure 8 Visualize the dimensionality reduction results of the comparative learning for the 81st engine;

[0114] Figure 9 The point prediction and interval prediction results for the 81st engine;

[0115] Figure 10 The results show the relative errors of the 81st engine at different stages of its lifespan.

[0116] Figure 11 Box plots for each stage of the predicted lifespan of the 81st engine;

[0117] Figure 12 The results of the ablation experiment for the 81st engine are shown in the comparison. Detailed Implementation

[0118] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments.

[0119] The engine life prediction method based on correlation analysis and multi-task learning described in this embodiment is illustrated in the flowchart below. Figure 1 As shown, the method includes the following steps:

[0120] Step 1: Acquire the data detected by the sensors during engine operation and normalize the data to be dimensionless.

[0121] Specifically, the calculation expression for the normalization dimensionless processing in step 1 is as follows:

[0122] (1);

[0123] in, X represents the normalized data from sensor j. j (t) represents the raw data detected by sensor j, min(X) j ) and max(X j ) represent the minimum and maximum values ​​of the data detected by sensor j, respectively.

[0124] Step 2: Use the self-attention mechanism to perform multi-parameter fusion on the processed data to generate the initial health index HI.

[0125] Specifically, the process of generating the initial health index HI in step 2 includes:

[0126] The processed data is used as input data. Where n is the number of data points and d is the dimension of the data, for each input data point Calculate the corresponding query vector q i Key vector k j Sum vector v j The calculation expressions are as follows:

[0127] (2);

[0128] (3);

[0129] (4);

[0130] in, These are three trainable weight matrices in the self-attention mechanism, used to generate the query vector, key vector, and value vector, respectively.

[0131] Calculate each query vector With all key vectors within the window The similarity scores are normalized to obtain the self-attention weights, expressed as:

[0132] (5);

[0133] in, Self-attention weights for each sensor, This represents the dot product operation. The similarity score reflects the strength of the relationship between the query vector and the key vector. Normalization makes the sum of the weights equal to 1.

[0134] For each value vector The weighted summation based on the self-attention weights yields the fused one-dimensional data, expressed as:

[0135] (6);

[0136] Among them, HI i This is the initial health index.

[0137] Step 3: Construct a change point detection model based on NTM, train the change point detection model, use the trained change point detection model to find abrupt change points, and use the abrupt change points to divide the initial health index into normal stage and degenerative stage.

[0138] Specifically, the implementation process of step 3 includes:

[0139] Using LSTM as the controller of NTM to construct a change point detection model, the forward pass expression is as follows:

[0140] (7);

[0141] (8);

[0142] (9);

[0143] (10);

[0144] (11);

[0145] (12);

[0146] in, This represents the output of the forget gate at time t. This represents the output of the input gate at time t. This represents the output of the output gate at time t. This represents the state of the candidate cell at time t. x represents the state of the update unit at time t. t h represents the input at time t. t-1 W represents the hidden state value at time t-1; f W i W o and W a These represent the forget gate, input gate, output gate, and h during the feature extraction process, respectively. t-1 Weighting coefficients; V f V i V o and V a These represent the forget gate, input gate, output gate, and x during the feature extraction process, respectively. t Weighting coefficients; b f b i b o and b a represents the bias values ​​in the forget gate, input gate, output gate, and feature extraction process, respectively; tanh represents the tangent hyperbolic function, and σ represents the activation function Sigmoid; The Hadamard product is the element-wise multiplication symbol.

[0147] The input to the change point detection model is the fused initial health indicator HI, denoted as... , This represents the input sequence from time tm to time t. This indicates whether an engine unit whose health status and remaining service life are being monitored is in a normal operating phase or a degradation phase. A point before a change is marked as 0 (normal phase), and otherwise as 1 (degradation phase). This change-point detection model... Prediction based on multi-sensor data A nonlinear autoregressive function is used to capture the complex relationship between the initial health parameter HI and the points of change. The expression is as follows:

[0148] (13);

[0149] in, It is the parameter set of NTM, which contains the parameters in formulas (7) to (10). ;

[0150] In the task of change point detection, NTM is a classifier. The optimal parameter set of the change point detection model is trained by minimizing the binary cross-entropy, and the loss function is:

[0151] (14);

[0152] Where T is the number of time points in the training dataset;

[0153] Through training, the output of NTM The first moment when the threshold is exceeded is defined as the mutation point. As shown in formula (15):

[0154] (15);

[0155] Where threshold is the threshold for cross-validation to find the best classification; This indicates the moment when a search needs to be performed one by one. As a condition, that is, to guarantee Within the time frame t under consideration;

[0156] Based on mutation points The determination of the initial health index HI time series is divided into two stages. Before the mutation point, the engine unit has not degraded or the degradation is negligible, which is the normal stage. After the mutation point, the engine unit performance deteriorates and the remaining life gradually decreases, which is the degradation stage.

[0157] Step 4: Construct and train a SINN model. Use the trained SINN model to reconstruct the stochastic process and probability density distribution of the initial health index degradation stage, and identify different failure modes based on the differences in probability density distribution.

[0158] Specifically, the implementation process of step 4 includes:

[0159] A SINN model is constructed using a two-layer LSTM. This SINN model outputs the trajectory and statistical characteristics of a random process based on the input random sequence, including the probability density distribution PDF and the autocorrelation function ACF.

[0160] The SINN model is trained using the loss functions of ACF and PDF. The inputs are the initial health index during the degradation phase, the corresponding target PDF, and the target ACF. The trained SINN model outputs both the stochastic process and the probability density distribution PDF. The expression for the ACF loss function is as follows:

[0161] (16);

[0162] The loss function for PDF is:

[0163] (17);

[0164] Among them, ACF T (τ) is the autocorrelation function of the objective stochastic process, τ is the lag number, and ACF O (τ) is the autocorrelation function output by the SINN model; Let be the probability density function of the target stochastic process, calculated using Gaussian kernel density estimation, where h is the width of the kernel function. The probability density function output by the SINN model;

[0165] Based on the PDFs of engine health indices corresponding to different failure modes output by the SINN model, the KL divergence between the various PDFs is calculated using the following expression:

[0166] (18);

[0167] Fault modes are identified based on the magnitude of KL divergence, as shown in formula (19):

[0168] (19);

[0169] Where p(x) and q(x) are probability density distributions, and α is the KL divergence threshold for the same fault type. As a fault mode, when the KL divergence between an unknown degraded PDF and a PDF of a known fault mode is less than a threshold α, they are determined to belong to the same fault mode; otherwise, they belong to different fault modes. Accurate fault identification is achieved through multiple comparisons.

[0170] Step 5: Construct and train a contrastive learning model. Use the trained contrastive learning model to obtain a high-dimensional representation of the initial health index degradation stage and generate a contrastive learning health index.

[0171] Specifically, the implementation process of step 5 includes:

[0172] First, data augmentation techniques are used to transform the initial health parameter HI during the degradation stage. These techniques include jittering, scaling, time shifting, and neighborhood segmentation, with the corresponding formulas as follows:

[0173] (20);

[0174] (twenty one);

[0175] (twenty two);

[0176] (twenty three);

[0177] Where Jitter is the jitter operation, t i Let be the timestamp of the i-th sampling point, D be the ideal sampling interval, and N be the number of sampling points. Random noise is added to the time series data in the time domain to increase the diversity and robustness of the data. Scaling is the scaling operation, where A is the scaling factor, which can simulate different amplitude changes in the data. Shifting is the time shift operation. The time offset can simulate different time delays in data; Neighborhood Segmentation is a neighborhood segmentation operation. and Let f(x(t)) be the start and end times of the neighborhood, and f(x(t)) be a function that segments the signal outside the neighborhood to simulate different local changes in the data.

[0178] The sample pairs generated after data transformation are denoted as positive samples. Then, an LSTM-based contrastive learning model is constructed. The positive samples are input into the model to extract high-dimensional feature representations. Negative samples are generated by pairing the feature representations of the positive samples with semantically unrelated feature representations from other samples in the dataset. The positive and negative pairs are then presented to a contrastive loss function. The InfoNCE loss function guides the contrastive learning model to learn feature representations, making positive pairs closer in the embedding space and pushing negative pairs further apart. The expression for the InfoNCE loss function is:

[0179] (twenty four)

[0180] Where, sim(A, P) i ) are anchor sample A and positive sample Similarity between them; sim(A, N) j ) are anchor sample A and negative sample The similarity between them, the anchor samples are automatically generated during the training process, It is a temperature parameter that needs to be preset to control the smoothness of the probability distribution;

[0181] By training a contrastive learning model, the high-dimensional features corresponding to the health parameters after separating positive and negative samples are output. This high-dimensional feature matrix is ​​denoted as the contrastive learning health index (CHI), and its expression is:

[0182] (25);

[0183] in, This represents the data at the t-th time point of the p-th dimension feature.

[0184] Step 6: Concatenate the features of the stochastic process in the degradation stage with the contrastive learning health index to generate a global health index, and then use the cascaded NTM model to predict the remaining service life of the engine based on the global health index.

[0185] Specifically, the implementation process of step 6 includes:

[0186] First, the stochastic process features output by the SINN model are concatenated with the CHI feature. The concatenated index is designated as the Global Health Index (OHI), and is represented as follows:

[0187] (26);

[0188] A prediction model is constructed, which is a cascaded NTM model composed of multiple identical NTMs connected in succession. Each NTM uses an LSTM as its controller, and the prediction model is trained accordingly. The read-write memory process of the LSTM is represented as follows:

[0189] (27);

[0190] (28);

[0191] (29);

[0192] (30);

[0193] (31);

[0194] (32);

[0195] Among them, w t w (i) represents the weight of the write header at time t, e t Let α be the erasure vector. t Both the added vector, erased vector, and added vector are generated by the controller; w t r (i) represents the weight of the read head at time t, and the vector r returned by the read head. t Defined as a weighted sum of rows in memory; w t c (i) is the content addressing weight of the i-th address, β t K is the bond strength, and K is the similarity function (such as cosine similarity). t It is a key vector, M t (i) is the i-th row of the storage matrix; It is the shift weight of the i-th address, w t (i) is the sharpening weight for the i-th address, w t g (j) is the interpolation weight of the j-th address, s t (ij) is the shift weight of the (ij)th element, γ t It is the sharpening coefficient. Based on the weight vector of the previous time step, an interpolation gate scalar output by the controller, a shift weight vector, and a sharpening factor scalar, a new weight vector is calculated, which represents the result obtained after interpolating, shifting, and sharpening the weight vector of the previous time step.

[0196] like Figure 2 As shown, in the cascaded NTM model, each NTM is considered a submodule. Starting from the second submodule, the prediction error of the previous submodule is continuously passed as output to the next submodule. That is, all subsequent submodules predict the output error of the previous submodule, thereby improving prediction accuracy by continuously reducing the prediction error. When the prediction error of the cascaded NTM model is less than the set prediction accuracy or the number of submodules reaches the set value (in this embodiment, the maximum number of submodules is set to 10), the cascaded NTM model will automatically terminate the process of expanding the submodule cascading and output the error prediction value of the last submodule. Then, the OHI is input into the cascaded NTM model for engine remaining life prediction. The predicted value of the remaining life is the sum of the outputs of each submodule, expressed as:

[0197] (33);

[0198] Among them, Y predi For the predicted output values ​​of each submodule;

[0199] The prediction model uses a quantile loss function to guide the upper and lower bounds of the output prediction interval; the quantile loss function is a piecewise formula, expressed as:

[0200] (34);

[0201] Where y is the target variable, and f(x) is the predicted value. It is a quantile; when When = 0.5, this loss function is the same as the loss function used for point prediction; when When the value is 0.95, the cascaded NTM model will try to make the predicted value greater than the true value to minimize the loss, which is the upper limit of the interval; when When the value is 0.05, the cascaded NTM model will try to make the predicted value less than the true value, which is the lower limit of the interval.

[0202] Step 7: Using one-to-many parameter transfer and NSGA-II fine-tuning algorithm, transfer the parameters of the trained change point detection model to subsequent tasks, use NSGA-II to optimize and fine-tune some parameters, use the dataset for instance validation, and perform ablation analysis.

[0203] like Figure 3 As shown, specifically, the implementation process of step 7 includes:

[0204] Let's denote the NTM-based change point detection model as the first task model. First, train the first task model using GridsearchCV on a pre-defined parameter space. The search is performed, and the parameter space is represented as follows:

[0205] (35);

[0206] Where, p k This represents a specific combination of hyperparameters, k=1,2,..,N, where N represents the number of combinations in the parameter grid;

[0207] For each hyperparameter combination p k Perform i cross-validations and calculate the average score of the i cross-validations, expressed as:

[0208] (36);

[0209] Among them, CV j For the j-th cross-validation, score(p) k ,CV j ) represents the parameter combination p k The performance score in the j-th cross-validation;

[0210] Iterate through N hyperparameter combinations and determine the optimal combination:

[0211] (37);

[0212] The argmax operator returns avg_score(p k The hyperparameter combination p corresponding to the maximum value of the function k That is, p best (Optimal hyperparameter combination), and train the change point detection model with this optimal hyperparameter combination;

[0213] Then, the parameters of the change point detection model from Task 1 are transferred to the subsequent SINN model, contrastive learning model, and cascaded NTM model, and NSGA-II fine-tuning is performed: optimization objectives are defined for each model, including minimum MSE and shortest training time; Pareto optimization is performed, the crowding distance of the solution within each Pareto front is calculated, and a set of candidate solutions balancing time and performance is given.

[0214] Finally, a set of solutions is randomly selected to train the models for subsequent tasks and output the fine-tuned models. The engine remaining life prediction model based on fault analysis and multi-task learning is used to predict the engine remaining life, obtain the prediction structure, and then output the prediction curves and errors in sequence.

[0215] The following examples further illustrate the implementation process of the engine life prediction method based on correlation analysis and multi-task learning according to the present invention.

[0216] The C-MAPSS dataset is a publicly available dataset from NASA. This dataset simulates the degradation process of turbofan engines under different operating modes. Each degradation data point in the C-MAPSS dataset is 26-dimensional: the first dimension is the engine number, the second dimension is the time series, the third to fifth dimensions are the operating settings, and the sixth to twenty-sixth dimensions are sensor data. The dataset comprises four sub-datasets: FD001, FD002, FD003, and FD004, each containing data from 100 engines. This embodiment selected the 81st engine from FD001 for testing. The experimental results of this embodiment significantly outperformed other existing methods, further demonstrating the practicality of this invention.

[0217] The engine life prediction method based on correlation analysis and multi-task learning described in this invention includes the following steps:

[0218] (1) Data preprocessing

[0219] The data from each sensor were normalized according to Formula 1. Seven sensors (T2, P2, P15, epr, Bfa, Nf_dmd, and NRf_dmd) that remained unchanged were discarded, leaving 14 sensors as the fusion parameters. The self-attention weights of these 14 sensors, calculated according to Formulas 2-6, are shown in Table 1. The initial health index HI after weighted fusion is as follows: Figure 4 As shown.

[0220] Table 1. Sensor Self-Attention Weight Values ​​for Engine 81

[0221]

[0222] (2) Change point detection

[0223] The NTM-based change point detection model was trained, and the test output of the 81st engine is as follows: Figure 5 As shown, the horizontal line represents a threshold of 0.6. The first time an output value exceeds 0.6 is considered a mutation point. Using the mutation point as the stage division point, the performance data of the 81st engine is divided into a normal stage and a degradation stage, as shown below. Figure 6 As shown.

[0224] Compared to when the predicted value first exceeded 0.6, the change point for the 81st engine occurred at the 55th flight cycle, before 30% of its lifespan. Before the change point, the health value declined slowly, while after the change point, the health value declined rapidly. After the change point, the 81st engine's operational cycles from degradation to failure were 184.

[0225] (3) Failure Mode Analysis

[0226] The SINN model was trained using loss functions 14 and 15, and the probability density distributions of the 81st engine during the normal and degradation phases were tested and output as follows: Figure 7 As shown in Table 2, the basic information of the probability density distribution is as follows. It can be seen that, due to the continuous decrease in engine health value, the mean PDF in the normal stage is higher than that in the degradation stage. The skewness values ​​are all greater than 0, indicating that most values ​​are located to the left of the mean. The kurtosis values ​​are all very small, indicating that most values ​​are dispersed on both sides of the mean. Compared with the high peak and thin tail of the normal stage, the degradation stage exhibits the characteristics of a low peak and thick tail.

[0227] Table 2 Probability density distribution of the 81st engine

[0228]

[0229] To determine the threshold α for Equation 17, the KL divergences for the 81st engine with a single fault (HPC), the 61st engine with a double fault (HPC, fan), the 5th engine with a triple fault (HPC efficiency, HPC flow, LPT flow), and the 10th engine with a quadruple fault (LPC efficiency, LPC flow, HPC efficiency, HPC flow) were calculated to be 0.026, 0.033, 0.012, and 0.019, respectively. The KL divergences were calculated using multiple engines, with α set to 0.05.

[0230] (4) Comparative learning analysis

[0231] The initial health index HI is augmented using formulas 18 to 21 to obtain positive and negative sample pairs for contrastive learning. These sample pairs are then presented to the contrastive loss function (formula 22). An LSTM-based autoencoder separates the positive and negative sample pairs. To visualize the representation results of contrastive learning, t-SNE is used to reduce the high-dimensional representation obtained from the contrastive learning of the 81st engine's degradation stage to two dimensions, as shown below. Figure 8 As shown.

[0232] By minimizing the contrastive loss function, the distance between positive and negative samples is maximized. Figure 8 It can be seen that the positive and negative samples are separated to a certain extent, indicating that contrastive learning is effective. Further quantitative calculations, cosine similarity is calculated for similar samples and variance is calculated for dissimilar samples. The cosine similarity of the 81st engine is 1, which means that the representations of similar samples are highly similar because the autoencoder can accurately reconstruct the input samples; the variance is 0.53, which indicates that dissimilar samples are separated to a certain extent, but there is still overlap or similarity.

[0233] (5) Remaining life prediction analysis

[0234] Based on the Global Health Index (OHI) generated in steps 4 and 5, a cascaded NTM is trained using steps 6 and 7, and the point prediction and interval prediction results for the 81st engine are output, such as... Figure 9 As shown in the figure, the model described in this invention effectively captures both the long-term and short-term trends of the degradation data. With the accumulation of operating cycles, the point prediction results show a clear downward trend, proving the model's effectiveness in predicting degradation data. Table 3 presents the evaluation metrics for the prediction results of the 81st engine.

[0235] Table 3 Predictive Evaluation Indicators for the 81st Engine

[0236]

[0237] Table 3 shows that the RMSE of the 81st engine was 0.07, achieving high prediction accuracy, and the PICP was 0.894, indicating that the prediction interval largely covered the original value. From a longitudinal perspective, the upper and lower limits of the interval are close to the maximum and minimum values ​​of the original value fluctuations. PINAW reflects the scale of the prediction interval relative to the original value, indicating high clarity of the prediction interval. From a lateral perspective, with the accumulation of operating cycles, the prediction interval gradually narrows, and the uncertainty gradually decreases, reflecting the good prediction performance of the model of this invention. With the increase of operating cycles, the NTM's external memory stores more and more historical information, thereby achieving better performance.

[0238] To demonstrate the superiority of the model in this invention, models such as RNN, LSTM, and NTM were selected for comparison. Figure 10The comparison of relative errors at 50%, 60%, 70%, 80%, and 90% of the lifespan of the 81st engine is presented. It can be seen that the RNN has the worst prediction accuracy, while the model proposed in this invention has the best. As the lifespan progresses, the prediction performance of the RNN improves only slightly, while the prediction performance of the model proposed in this invention improves significantly. The prediction error is minimized when the lifespan reaches 90%, demonstrating the advantage of this method in predicting long-sequence data. For the 81st engine, in the last 90% of its lifespan, the RE of the model proposed in this invention is 3.86% lower than that of the NTM, 5.41% lower than that of the LSTM, and 7.81% lower than that of the RNN.

[0239] Figure 11 The data distribution of the predicted lifespan of the 81st engine at various stages is shown. It can be seen that the median continuously decreases over time, indicating a gradual decline in engine performance. The thickness of the box represents the interquartile range (IQR), and the two short horizontal lines above and below the box represent the upper and lower limits. As the lifespan increases, the box thickens or the distance between the upper and lower horizontal lines increases, indicating a gradually increasing rate of decline in the predicted value. The data declines most rapidly after reaching 90% of the lifespan, ultimately leading to complete failure, which provides a valid explanation for engine degradation.

[0240] (6) Ablation research

[0241] Ablation studies are conducted by removing or altering a portion of our model while keeping other conditions constant, helping us understand the effects of different components. The results of the ablation studies are shown in Table 4. Models A, B, C, and D all used traditional methods to manually tune parameters and output the best results, while the model of this invention uses a one-to-many parameter transfer and fine-tuning algorithm. The results show that Model A yielded the worst results, while the model of this invention yielded the best results, consistent with the expected outcome.

[0242] Table 4. Results of the ablation test for the 81st engine.

[0243]

[0244] (a) The difference between Model B and Model A lies in the input features. OHI is better at predicting RUL than CHI. The RMSE of the 81st engine improved by 4.2%. Compared to CHI, OHI adds statistical features of stochastic processes and increases the diversity of data patterns, which helps the model to understand the engine degradation process more comprehensively. In addition, the randomness of the data provided by the stochastic process improves the generalization performance.

[0245] (b) The difference between Model D and Model B lies in whether a cascaded algorithm is used. Using a cascaded NTM to predict RUL is better than using a single NTM. The RMSE of the 81st engine improved by 3.1%, and the cascaded NTM effectively improved prediction performance compared to the NTM. By setting the error threshold and upper limit of the submodules, an early stopping strategy can be implemented, which can effectively save running time and avoid overfitting.

[0246] (c) The difference between the model of this invention and model D lies in the parameter transfer of the one-to-many model and the NSGA-II fine-tuning algorithm. Using parameter transfer to train the model is better than traditional parameter tuning methods. The RMSE of the 81st engine improved by 10.2%. Compared with tuning the parameters of each task individually, the multi-task learning algorithm of this invention improves prediction performance by training the first task model NTM and transferring its parameters to subsequent tasks, and then using NSGA-II for optimization and fine-tuning.

[0247] Figure 12 The results show a comparison of the prediction results of the proposed model, Model B, and Model D on the 81st engine. The results indicate that, compared to other ablation models, the proposed model exhibits smaller fluctuations in prediction results, better stability, and is closer to the original values. Stable RUL prediction results can better reflect the engine's degradation status.

[0248] (7) Multi-task analysis

[0249] Table 5 compares the parameter transfer multi-task model algorithm of this invention with the empirical independent task model and the grid search-based parameter transfer algorithm. Evaluation metrics from each task are used: KL divergence for Task 2 to measure the accuracy of fault mode identification; uniformity for Task 3 to measure the performance of contrastive learning; and RMSE for Task 4 to evaluate the accuracy of lifetime prediction, as well as the training time for each task model. Among all tasks, the method of this invention performs best, followed by the grid search-based multi-task optimization algorithm, and worst is the empirically determined parameter independent task optimization algorithm. The grid search method is the most time-consuming, while the method of this invention is second only to the empirically determined parameter method. This is because the present invention expands the original data space through parameter transfer, improving model performance. The NSGA-II optimization algorithm used in the fine-tuning step effectively searches the parameter space and converges to the optimal solution, while grid search is computationally intensive and may miss the optimal parameter combination. The method of this invention balances time and performance, enabling the acquisition of a better-performing model in a shorter time.

[0250] Table 5 Multi-task analysis of the 81st engine

[0251]

[0252] This invention's cascaded NTM model can predict the remaining life of an engine. On one hand, it leverages the advantages of NTM in processing long-term data series and improves the prediction accuracy of the original NTM by combining statistical process analysis and comparative learning. Furthermore, the cascaded algorithm and multi-task optimization algorithm enhance the accuracy and efficiency of model training. On the other hand, this cascaded NTM model can also provide an explanation for degradation through the probability density function, enabling effective identification of different fault modes based on differences in probability density.

Claims

1. An engine life prediction method based on correlation analysis and multi-task learning, characterized in that, Includes the following steps: Step 1: Acquire the data detected by the sensors during engine operation and normalize the data to be dimensionless; Step 2: Use the self-attention mechanism to perform multi-parameter fusion on the processed data to generate the initial health index HI; Step 3: Construct a change point detection model based on NTM, train the change point detection model, use the trained change point detection model to find abrupt change points, and use the abrupt change points to divide the initial health index into normal stage and degenerative stage. Step 4: Construct and train a SINN (Statistics-Informed Neural Network) model. Use the trained SINN model to reconstruct the stochastic process and probability density distribution of the initial health index degradation stage, and identify different failure modes based on the differences in probability density distribution. Step 5: Construct and train a contrastive learning model. Use the trained contrastive learning model to obtain a high-dimensional representation of the initial health index degradation stage and generate a contrastive learning health index. Step 6: The stochastic process of the degradation stage and the contrastive learning health index are concatenated to generate a global health index. Then, the cascaded NTM model is used to predict the remaining service life of the engine based on the global health index. The cascaded NTM model is composed of multiple identical NTMs connected in stages, and each NTM uses an LSTM as the controller. Step 7: Using one-to-many parameter transfer and NSGA-II fine-tuning algorithm, transfer the parameters of the trained change point detection model to subsequent tasks, use NSGA-II to optimize and fine-tune some parameters, use the engine dataset for instance verification, and perform ablation analysis.

2. The engine life prediction method based on association analysis and multi-task learning according to claim 1, characterized in that, The calculation expression for normalization and dimensionless processing in step 1 is: ； in, X represents the normalized data from sensor j. j (t) represents the raw data detected by sensor j, min(X) j ) and max(X j ) represent the minimum and maximum values ​​of the data detected by sensor j, respectively.

3. The engine life prediction method based on association analysis and multi-task learning according to claim 1, characterized in that, Step 2, which generates the initial health index HI, includes: The processed data is used as input data. Where n is the number of data points and d is the dimension of the data, for each input data point Calculate the corresponding query vector q i Key vector k j Sum vector v j The calculation expressions are as follows: ； ； ； in, These are three trainable weight matrices in the self-attention mechanism, used to generate the query vector, key vector, and value vector, respectively. Calculate each query vector With all key vectors within the window The similarity scores are normalized to obtain the self-attention weights, expressed as: ； in, Self-attention weights for each sensor, This represents the dot product operation. The similarity score reflects the strength of the relationship between the query vector and the key vector. Normalization makes the sum of the weights equal to 1. For each value vector The weighted summation based on the self-attention weights yields the fused one-dimensional data, expressed as: ； Among them, HI i This is the initial health index.

4. The engine life prediction method based on association analysis and multi-task learning according to claim 3, characterized in that, Step 3 includes the following steps: Using LSTM as the controller of NTM to construct a change point detection model, the forward pass expression is as follows: ； ； ； ； ； ； in, This represents the output of the forget gate at time t. This represents the output of the input gate at time t. This represents the output of the output gate at time t. This represents the state of the candidate cell at time t. x represents the state of the update unit at time t. t h represents the input at time t. t-1 W represents the hidden state value at time t-1; f W i W o and W a These represent the forget gate, input gate, output gate, and h during the feature extraction process, respectively. t-1 Weighting coefficients; V f V i V o and V a These represent the forget gate, input gate, output gate, and x during the feature extraction process, respectively. t Weighting coefficients; b f b i b o and b a represents the bias values ​​in the forget gate, input gate, output gate, and feature extraction process, respectively; tanh represents the tangent hyperbolic function, and σ represents the activation function Sigmoid; The Hadamard product is the element-wise multiplication symbol. The input to the change point detection model is the fused initial health indicator HI, denoted as... , This represents the input sequence from time tm to time t; This indicates whether an engine unit whose health status and remaining service life are being monitored is in a normal operating phase or a degradation phase. A point before a change is marked as 0 (normal phase), and otherwise as 1 (degradation phase). This change-point detection model... Prediction based on multi-sensor data A nonlinear autoregressive function is used to capture the complex relationship between the initial health parameter HI and the points of change. The expression is as follows: ； in, It is the parameter set of NTM, containing formulas -formula Parameters in ; In the task of change point detection, NTM is a classifier. The optimal parameter set of the change point detection model is trained by minimizing the binary cross-entropy, and the loss function is: ； Where T is the number of time points in the training dataset; Through training, the output of NTM The first moment when the threshold is exceeded is defined as the mutation point. , as in the formula As shown: ； Where threshold is the threshold for cross-validation to find the best classification; This indicates the moment when a search needs to be performed one by one. As a condition, that is, to guarantee Within the time frame t under consideration; Based on mutation points The determination of the initial health index HI time series is divided into two stages. Before the mutation point, the engine unit has not degraded or the degradation is negligible, which is the normal stage. After the mutation point, the engine unit performance deteriorates and the remaining life gradually decreases, which is the degradation stage.

5. The engine life prediction method based on correlation analysis and multi-task learning according to claim 4, characterized in that, Step 4 includes the following steps: A SINN model is constructed using a two-layer LSTM. This SINN model outputs the trajectory and statistical characteristics of a random process based on the input random sequence, including the probability density distribution PDF and the autocorrelation function ACF. The SINN model is trained using the loss functions of ACF and PDF. The inputs are the initial health index during the degradation phase, the corresponding target PDF, and the target ACF. The trained SINN model outputs both the stochastic process and the probability density distribution PDF. The expression for the ACF loss function is as follows: ； The loss function for PDF is: ； Among them, ACF T (τ) is the autocorrelation function of the objective stochastic process, τ is the lag number, and ACF O (τ) is the autocorrelation function output by the SINN model; Let be the probability density function of the target stochastic process, calculated using Gaussian kernel density estimation, where h is the width of the kernel function. The probability density function output by the SINN model; Based on the PDFs of engine health indices corresponding to different failure modes output by the SINN model, the KL divergence between the various PDFs is calculated using the following expression: ； Fault modes are identified based on the magnitude of KL divergence, as shown in the formula. As shown: ； Where p(x) and q(x) are probability density distributions, and α is the KL divergence threshold for the same fault type. As a fault mode, when the KL divergence between an unknown degraded PDF and a PDF of a known fault mode is less than a threshold α, they are determined to belong to the same fault mode; otherwise, they belong to different fault modes. Accurate fault identification is achieved through multiple comparisons.

6. The engine life prediction method based on association analysis and multi-task learning according to claim 5, characterized in that, Step 5 includes the following steps: First, data augmentation techniques are used to transform the initial health parameter HI during the degradation stage. These techniques include jittering, scaling, time shifting, and neighborhood segmentation, with the corresponding formulas as follows: ； ； ； ； Where Jitter is the jitter operation, t i Let A be the timestamp of the i-th sampling point, D be the ideal sampling interval, and N be the number of sampling points; Scaling is the scaling operation, and A is the scaling factor; Shifting is the time-shift operation. This represents the time offset; NeighborhoodSegmentation is the neighborhood segmentation operation. and Let f(x(t)) be the start and end times of the neighborhood, and f(x(t)) be a function that segments the signal outside the neighborhood. The sample pairs generated after data transformation are denoted as positive samples. Then, an LSTM-based contrastive learning model is constructed. The positive samples are input into the model to extract high-dimensional feature representations. Negative samples are generated by pairing the feature representations of the positive samples with semantically unrelated feature representations from other samples in the dataset. The positive and negative pairs are then presented to a contrastive loss function. The InfoNCE loss function guides the contrastive learning model to learn feature representations, making positive pairs closer in the embedding space and pushing negative pairs further apart. The expression for the InfoNCE loss function is: ； Where, sim(A, P) i ) are anchor sample A and positive sample Similarity between them; sim(A, N) j ) are anchor sample A and negative sample The similarity between them, the anchor samples are automatically generated during the training process, It is a temperature parameter that needs to be preset to control the smoothness of the probability distribution; By training a contrastive learning model, the high-dimensional features corresponding to the health parameters after separating positive and negative samples are output. This high-dimensional feature matrix is ​​denoted as the contrastive learning health index (CHI), and its expression is: ； in, This represents the data at the t-th time point of the p-th dimension feature.

7. The engine life prediction method based on association analysis and multi-task learning according to claim 6, characterized in that, The implementation process of step 6 includes: First, the stochastic process features output by the SINN model are concatenated with the CHI feature. The concatenated index is designated as the Global Health Index (OHI), and is represented as follows: ； Construct a prediction model, which is a cascaded NTM model, where the read-write memory process of LSTM is represented as follows: ； ； ； ； ； ； Among them, w t w (i) represents the weight of the write header at time t, e t Let α be the erasure vector. t Both the added vector, erased vector, and added vector are generated by the controller; w t r (i) represents the weight of the read head at time t, and the vector r returned by the read head. t Defined as a weighted sum of rows in memory; w t c (i) is the content addressing weight of the i-th address, β t It represents the bond strength, and K is the similarity function. t It is a key vector, M t (i) is the i-th row of the storage matrix; It is the shift weight of the i-th address, w t (i) is the sharpening weight for the i-th address, w t g (j) is the interpolation weight of the j-th address, s t (ij) is the shift weight of the ij-th element, γ t It is the sharpening coefficient. Based on the weight vector of the previous time step, an interpolation gate scalar output by the controller, a shift weight vector, and a sharpening factor scalar, a new weight vector is calculated, which represents the result obtained after interpolating, shifting, and sharpening the weight vector of the previous time step. In the cascaded NTM model, each NTM is considered a submodule. Starting from the second submodule, the prediction error of the previous submodule is continuously passed as output to the next submodule. That is, all subsequent submodules predict the output error of the previous submodule, thereby improving prediction accuracy by continuously reducing the prediction error. When the prediction error of the cascaded NTM model is less than the set prediction accuracy or the number of submodules reaches a set value, the cascaded NTM model will automatically terminate the process of expanding the submodules and output the error prediction value of the last submodule. Then, the OHI is input into the cascaded NTM model for engine remaining life prediction. The predicted remaining life is the sum of the outputs of all submodules, expressed as: ； Among them, Y predi For the predicted output values ​​of each submodule; The prediction model uses a quantile loss function to guide the upper and lower bounds of the output prediction interval; the quantile loss function is a piecewise formula, expressed as: ； Where y is the target variable, and f(x) is the predicted value. It is a quantile; when When = 0.5, this loss function is the same as the loss function used for point prediction; when When the value is 0.95, the cascaded NTM model will try to make the predicted value greater than the true value to minimize the loss, which is the upper limit of the interval; when When the value is 0.05, the cascaded NTM model will try to make the predicted value less than the true value, which is the lower limit of the interval.

8. The engine life prediction method based on correlation analysis and multi-task learning according to claim 7, characterized in that, The implementation process of step 7 includes: Let's denote the NTM-based change point detection model as the first task model. First, train the first task model using GridsearchCV on a pre-defined parameter space. The search is performed, and the parameter space is represented as follows: ； in, Each represents a specific combination of hyperparameters, and N represents the number of combinations in the parameter grid. For each hyperparameter combination p k Perform i cross-validations, k=1,2,...,N, and calculate the average score of the i cross-validations using the following expression: ； Among them, CV j For the j-th cross-validation, score(p) k ,CV j ) represents the parameter combination p k The performance score in the j-th cross-validation; Iterate through N hyperparameter combinations and determine the optimal combination: ； The argmax operator returns avg_score(p k The hyperparameter combination p corresponding to the maximum value of the function k That is, the optimal hyperparameter combination p best The change point detection model is trained using this optimal combination of hyperparameters; Then, the parameters of the first task model are transferred to the subsequent SINN model, contrastive learning model, and cascaded NTM model and NSGA-II fine-tuning is performed: optimization objectives are defined for each model, including minimum MSE and shortest training time; Pareto optimization is performed, the crowding distance of the solution within each Pareto front is calculated, and a set of candidate solutions balancing time and performance is given. Finally, a set of solutions is randomly selected to train the models for subsequent tasks and output the fine-tuned models. The engine remaining life prediction model based on fault analysis and multi-task learning is used to predict the engine remaining life, obtain the prediction structure, and then output the prediction curves and errors in sequence.

Images