Gas turbine residual life prediction method and system based on improved TTAO algorithm

Abstract

The invention discloses a gas turbine residual life prediction method and system based on an improved TTAO algorithm, and relates to the technical field of gas turbine residual life prediction, and the method comprises the steps: evaluating a sensor data contribution value through a random forest algorithm, screening key features, building an improved Autoformer model fusing a parallel multi-scale convolution kernel and an autocorrelation mechanism, and carrying out the prediction of the residual life of a gas turbine. A weighted loss function of JS divergence and a mean square error is constructed, an improved TTAO algorithm based on an adaptive disturbance balance strategy and a nonlinear tangential flight strategy is adopted to adjust model hyper-parameters, a prediction error is minimized through a back propagation algorithm, and finally high-precision residual life prediction is achieved. According to the method, redundant noise is reduced through feature screening, a dynamic convolution kernel is utilized to adapt to nonlinear degradation features, prediction precision and distribution consistency are balanced through a mixed loss function, convergence is accelerated by adopting a self-adaptive optimization algorithm, and the health management level of the gas turbine under complex working conditions is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of gas turbine remaining life prediction technology, and specifically to a gas turbine remaining life prediction method and system based on an improved TTAO algorithm. Background Technology

[0002] Gas turbines are a crucial component for powering thermal power plants, and predicting their remaining service life (UPS) is essential for their stable and reliable operation. The degradation process of gas turbines is complex, and the high dimensionality of sensor monitoring data presents a significant challenge to accurate UPS prediction. With the rapid development of deep learning technology in UPS prediction, utilizing reliable deep learning models to capture characteristic information during the gas turbine degradation process and predict its UPS offers new insights for gas turbine maintenance, support, and health management.

[0003] During long-term operation, gas turbines experience gradual performance degradation due to factors such as corrosion, fatigue, and wear, which can ultimately lead to aircraft malfunctions and accidents. In predicting remaining service life, deep learning technology utilizes the gas turbine's operating status information to automatically extract and regress features, revealing the patterns of performance degradation. While existing deep learning technologies have achieved some success in remaining service life prediction, they still have limitations. Convolutional Neural Networks (CNNs) excel at extracting local degradation features, but their ability to capture long-term temporal dependencies is insufficient; Recurrent Neural Networks (RNNs) and their variants (LSTM, GRU) can process temporal data, but they are weak in extracting subtle local features; Transformer models capture global dependencies through self-attention mechanisms, but they lack the ability to explicitly model the inherent structure of temporal data, making their efficiency unsatisfactory for equipment degradation data with significant periodicity.

[0004] In summary, current deep learning technologies lack an effective mechanism for automatic identification and information aggregation of periodic patterns in time-series data processing. Furthermore, hyperparameter selection relies on empirical operations and lacks systematic optimization methods, thus limiting the accuracy and generalization ability of gas turbine remaining life prediction. Therefore, there is an urgent need to propose a gas turbine remaining life prediction method to address these issues. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to address the problems of insufficient feature information extraction, difficulty in selecting hyperparameters, and poor generalization ability in existing remaining lifetime prediction methods.

[0006] This invention solves the above-mentioned technical problems through the following technical means: a gas turbine remaining life prediction method based on an improved TTAO algorithm, comprising: S1. Collect time-series monitoring data of the gas turbine degradation process, evaluate the contribution value of sensor data of each component through random forest algorithm, select the top N sensor data with high contribution value as input features, and randomly divide the selected feature dataset into training set and test set in a ratio of 8:2. S2. Input the training set into the improved Autoformer model, and use the parallel architecture of the ICBAM module and the autocorrelation mechanism to extract the temporal dependency. The ICBAM module contains a multi-scale convolutional kernel structure, which dynamically adjusts the size of the convolutional kernel to adapt to the nonlinear feature changes of the monitoring data. S3. Construct a weighted combination of JS divergence and root mean square error as the loss function, and use an improved TTAO algorithm based on adaptive perturbation balance strategy and nonlinear tangential flight strategy to adjust the model hyperparameters. Minimize the prediction error through backpropagation algorithm. S4. The feature vectors are mapped to the remaining service life prediction values ​​through a fully connected layer. The test set data is input into the trained model, and the remaining service life prediction results of the gas turbine are output.

[0007] Furthermore, the ICBAM module specifically comprises: The input data is processed in both the channel and spatial dimensions. The channel attention module performs global average pooling and global max pooling to obtain features. and Multi-scale convolution operations are performed on the two features using convolution kernels of three different sizes. The operational expression is as follows:

[0008]

[0009] in, and These represent the features obtained after global average pooling and global max pooling in the channel attention module, respectively. and They represent and Features obtained through multi-scale convolution operations This represents summation of elements. express convolution kernel, express convolution kernel, express The convolution kernel; The operational expression for channel attention is as follows:

[0010] in, For input variables, The channel attention feature weight vector of the input variables. It is the Sigmoid activation function. It is a multilayer perceptron neural network; The operational expression for ICBAM is as follows:

[0011] in, The numerical values ​​representing the computational values ​​of the convolutional attention mechanism. This indicates that the elements are multiplied and then added together one by one. This represents the spatial attention feature weight vector of the input vector.

[0012] Furthermore, the improved Autoformer model is specifically as follows:

[0013]

[0014] in, Represents the variable to be decomposed. Indicates the trend term. Indicates seasonal items, This represents the fill function. This represents the average pooling function; The formula for calculating the autocorrelation coefficient is:

[0015] in, Represents the autocorrelation coefficient. Indicates the time lag. Indicates Fourier transform, This indicates the conjugate operation. Indicates the inverse Fourier transform; The formula for calculating the autocorrelation mechanism is:

[0016] in, The numerical value representing the autocorrelation mechanism. Indicates the first A lag time, This indicates the number of autocorrelation coefficients selected. Represents the computation of vectors , The autocorrelation coefficient, Representing vectors of First-order lag vector, vector , , Depend on Obtained through linear transformation; The expression for the attention mechanism in parallel architecture is: .

[0017] Furthermore, the calculation formula for the adaptive perturbation balance strategy is as follows:

[0018] in, Indicates the current iteration number. Indicates the th population after the adaptive perturbation balance policy update. Individual, Indicates the first The first generation of the optimal population Individual, Indicates the first The first randomly selected individual from the optimal population. Indicates the first The second randomly selected individual from the optimal population. Represents the adaptive inertia factor. Indicates the adaptive perturbation factor; and The mathematical expression is:

[0019]

[0020] in, Represents a random number between [0,1], with weighting coefficients. and , , This indicates the maximum number of iterations.

[0021] Furthermore, the calculation formula for the nonlinear tangential flight strategy is as follows:

[0022] in, Dimensions representing the optimization process This indicates the generation of random numbers within a specified range; The mathematical expression for the local aggregation stage of the nonlinear tangential flight strategy optimization is as follows:

[0023] in, This represents the first [unit] in the population after the nonlinear tangential flight strategy update. Individual, Indicates the first The first generation of the optimal population Individual, Indicates the first The first generation of superior population Individual.

[0024] Furthermore, N is 14, and the sensor data includes: total outlet temperature of the low-pressure compressor, total outlet temperature of the high-pressure compressor, total outlet temperature of the low-pressure turbine, total outlet pressure of the high-pressure compressor, static pressure at the outlet of the high-pressure compressor, ratio of fuel flow rate to total outlet pressure of the high-pressure compressor, bypass ratio, bleed air enthalpy, cooling bleed air flow rate of the high-pressure turbine, cooling bleed air flow rate of the low-pressure turbine, physical rotational speed of the blades, corrected rotational speed of the blades, physical rotational speed of the core shaft, and corrected rotational speed of the core shaft.

[0025] Furthermore, the formula for calculating the loss function is as follows:

[0026] in, and For weighting coefficient control, and , Indicates the number of samples. This represents the distribution of the model output values. The distribution of actual measured values, Indicates the first The true value of each sample Indicates the first The predicted value for each sample, express and The JS divergence between them.

[0027] This invention also provides a gas turbine remaining life prediction system based on an improved TTAO algorithm, comprising: The data acquisition module is used to collect time-series monitoring data of the gas turbine degradation process. It evaluates the contribution value of sensor data of each component through random forest algorithm, selects the top N sensor data with high contribution values ​​as input features, and randomly divides the selected feature dataset into training set and test set in a ratio of 8:2. An improved Autoformer model core module is used to input the training set into the improved Autoformer model. The parallel architecture of the ICBAM module and the autocorrelation mechanism are used to extract temporal dependencies. The ICBAM module contains a multi-scale convolutional kernel structure, which dynamically adjusts the convolutional kernel size to adapt to the nonlinear feature changes of the monitoring data. The loss function construction module is used to construct a loss function with a weighted combination of JS divergence and root mean square error. It uses an improved TTAO algorithm based on an adaptive perturbation balance strategy and a nonlinear tangential flight strategy to adjust the model hyperparameters and minimizes the prediction error through the backpropagation algorithm. The prediction output module is used to map feature vectors to remaining service life prediction values ​​through a fully connected layer, input test set data into the trained model, and output the remaining service life prediction results of the gas turbine.

[0028] The present invention also provides a processing device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the above-described method steps by calling the program instructions.

[0029] The present invention also provides a computer-readable storage medium storing computer instructions that cause the computer to perform the above-described method steps.

[0030] The advantages of this invention are: This invention quantifies sensor contributions using the random forest algorithm, filters key features, effectively reduces data dimensionality and noise interference, and improves model training efficiency and prediction robustness. The parallel ICBAM module, combined with multi-scale convolutional kernels, dynamically adapts to nonlinear degradation features, enhancing the ability to capture long-term temporal dependencies and solving the problem of low computational efficiency in traditional Transformers. A weighted combination of JS divergence and mean squared error balances prediction accuracy and distribution consistency. An improved TTAO algorithm based on an adaptive perturbation balancing strategy and a nonlinear tangential flight strategy is used to optimize and adjust model hyperparameters. Backpropagation minimizes prediction errors, accelerating model convergence and improving the model's generalization performance under complex conditions. Fully connected layers map features to remaining lifetime predictions, providing interpretable and highly reliable decision-making basis for preventative maintenance. Attached Figure Description

[0031] Figure 1 This is a flowchart of the gas turbine remaining life prediction method based on the improved TTAO algorithm in Embodiment 1 of the present invention; Figure 2 This is a structural diagram of the improved channel attention module of ICBAM according to Embodiment 1 of the present invention; Figure 3 This is a structural diagram of the improved Autoformer according to Embodiment 1 of the present invention; Figure 4 This is a graph showing the relationship between the test function and the number of iterations in Embodiment 1 of the present invention; Figure 5 This is a sensor data contribution diagram from Embodiment 1 of the present invention, wherein, Figure 5a is a contribution map of different sensor data in the FD001 dataset; Figure 5 b is a contribution diagram of different sensor data in the FD003 dataset; Figure 6 This is a graph showing the predicted remaining life of a single gas turbine in the FD001 dataset of Embodiment 1 of the present invention; Figure 7 This is a graph showing the predicted remaining life of a single gas turbine in the FD003 dataset of Embodiment 1 of the present invention; Figure 8 The remaining life prediction results of the full gas turbines from the FD001 and FD003 datasets in Embodiment 1 of this invention; wherein, Figure 8 Figure a shows the prediction results for all gas turbines in the FD001 dataset. Figure 8 b is a graph showing the prediction results for all gas turbines in the FD003 dataset. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Example 1 like Figure 1 As shown, the gas turbine remaining life prediction method based on the improved TTAO algorithm includes: S1. Collect time-series monitoring data of the gas turbine degradation process, evaluate the contribution value of sensor data of each component through random forest algorithm, select the top N sensor data with high contribution value as input features, and randomly divide the selected feature dataset into training set and test set in a ratio of 8:2.

[0034] S2. Input the training set into the improved Autoformer model, and use the parallel architecture of the ICBAM module and the autocorrelation mechanism to extract the temporal dependencies. The ICBAM module contains a multi-scale convolutional kernel structure, which dynamically adjusts the size of the convolutional kernel to adapt to the nonlinear feature changes of the monitoring data.

[0035] Specifically, the improved Autoformer model, such as Figure 3 As shown.

[0036] This invention makes targeted improvements to the basic CBAM by adding convolutional kernels of different sizes after the pooling layer of the channel attention module in the traditional CBAM. The improved channel attention module can more comprehensively capture detailed and overall feature information in the data, thereby improving the accuracy and robustness of features and adapting to the complexity of multi-source sensor monitoring data from gas turbines. The structure of the improved channel attention module is as follows: Figure 2 As shown.

[0037] ICBAM (Improved Convolutional Block Attention Module) processes the input data in both the channel and spatial dimensions. The channel attention module performs global average pooling and global max pooling to obtain features. and Multi-scale convolution operations are performed on the two features using convolution kernels of three different sizes. The operational expression is as follows:

[0038]

[0039] in, and These represent the features obtained after global average pooling and global max pooling in the channel attention module, respectively. and They represent and Features obtained through multi-scale convolution operations This represents summation of elements. express convolution kernel, express convolution kernel, express The convolution kernel.

[0040] The convolutional part uses a 1×1 convolutional kernel to capture more local feature information, a 1×3 convolutional kernel to balance local and global feature information, and a 1×5 convolutional kernel to capture more global feature information. Different sizes of convolutional kernels can obtain features of different receptive fields, making the extracted feature information more complete. This multi-scale feature extraction enables the model to adapt to the complex and variable features in the gas turbine multi-source monitor data.

[0041] The operational expression for channel attention is as follows:

[0042] in, For input variables, The channel attention feature weight vector of the input variables. It is the Sigmoid activation function. It is a multilayer perceptron neural network.

[0043] The operational expression for ICBAM is as follows:

[0044] in, The numerical values ​​representing the computational values ​​of the convolutional attention mechanism. This indicates that the elements are multiplied and then added together one by one. This represents the spatial attention feature weight vector of the input vector.

[0045] Autoformer is a time-series forecasting model based on an improved Transformer architecture. Its core lies in the introduction of an embedded sequence decomposition mechanism and a periodic autocorrelation-based attention mechanism. This model replaces the traditional self-attention mechanism with an autocorrelation mechanism, significantly improving computational efficiency while more accurately capturing long-term dependencies by uncovering periodic patterns in time series. Structurally, Autoformer embeds sequence decomposition as a basic building block within the encoder-decoder framework. This allows the model to alternately perform decomposition and optimization during the prediction process, progressively separating trend and seasonal terms from the latent representation. This progressive extraction of time-series components effectively enhances the modeling ability of deep data patterns.

[0046] The Autoformer model employs a stacked encoder-decoder architecture, where each component integrates three core modules: autocorrelation, sequence decomposition, and forward feedback. This model analyzes complex patterns in time series data through an embedded sequence decomposition mechanism. It extracts the trend component representing long-term changes based on a moving average algorithm and removes this component from the original sequence to obtain a seasonal component reflecting cyclical fluctuations. This structural design enables the model to effectively handle the unpredictability of future data. The calculation process is as follows:

[0047]

[0048] in, Represents the variable to be decomposed. Indicates the trend term. Indicates seasonal items, This represents the fill function. This represents the average pooling function.

[0049] The autocorrelation mechanism enhances the model's ability to model periodic patterns by identifying periodic dependencies in a sequence and fusing similar subsequences with the same time delay. The formula for calculating the autocorrelation coefficient is:

[0050] in, Represents the autocorrelation coefficient. Indicates the time lag. Indicates Fourier transform, This indicates the conjugate operation. This represents the inverse Fourier transform.

[0051] The formula for calculating the autocorrelation mechanism is:

[0052] in, The numerical value representing the autocorrelation mechanism. Indicates the first A lag time, This indicates the number of autocorrelation coefficients selected. Represents the computation of vectors , The autocorrelation coefficient, Representing vectors of First-order lag vector, vector , , Depend on It is obtained through linear transformation.

[0053] The improved Autoformer model designed in this invention features a parallel architecture for its multi-scale convolutional kernel-based improved convolutional attention mechanism and autocorrelation mechanism. This architecture enables the model to fully exploit the spatial features and temporal dependencies in the monitoring data, thereby comprehensively capturing the degradation characteristics of various gas turbine components. The attention mechanism operation expression of this parallel architecture is as follows: .

[0054] S3. Construct a loss function using a weighted combination of JS divergence and root mean square error. Use an improved TTAO algorithm based on an adaptive perturbation balancing strategy and a nonlinear tangential flight strategy to optimize the model hyperparameters and minimize the prediction error through backpropagation.

[0055] Specifically, (1) loss function When comparing the output values ​​of the model proposed in this invention with the actual values, a corresponding loss function needs to be established. In data regression tasks, the mean squared error (MSE) is generally used as the loss function for model training. However, since there are interactions between the state variables of the gas turbine, using only MSE as the loss function cannot fully account for the coupling relationship between the variables.

[0056] To improve the prediction accuracy and robustness of the proposed model, this invention designs a model loss function that combines Jensen-Shannon (JS) divergence with MSE. JS divergence is a variant of Kullback-Leibler (KL) divergence (i.e., relative entropy), which overcomes the inherent limitations of KL divergence through symmetrization and boundedness. JS divergence typically ranges from 0 to 1, and the formula for calculating the total loss function is as follows:

[0057] Among them, the weighting coefficient Controlling the contribution of the MSE component to the total loss, weighting coefficients Controlling the contribution of the JS divergence component to the total loss, and , Indicates the number of samples. This represents the distribution of the model output values. The distribution of actual measured values, Indicates the first The true value of each sample Indicates the first The predicted value for each sample, express and The JS divergence between them.

[0058] A single MSE loss function can easily lead to overfitting of the model to noise and specific features in the data. The model loss function of this invention can constrain the model to learn the distribution characteristics of the data comprehensively, thereby enhancing the model's generalization ability. MSE is more sensitive to outliers, while JS divergence measures at the distribution level and is less affected by prediction bias at individual points. The model loss function of this invention enables more stable model training and is more robust to noise in the data.

[0059] (2) Improved TTAO algorithm based on adaptive perturbation balance strategy and nonlinear tangential flight strategy Intelligent optimization algorithms possess certain global exploration and local exploitation capabilities, making them suitable for solving complex optimization problems. They are currently widely applied in the field of parameter optimization by scholars both domestically and internationally. The Triangular Topology Aggregation Optimizer (TTAO) algorithm is a meta-intelligent optimization algorithm derived from similar triangles in plane geometry, exhibiting excellent optimization performance.

[0060] The TTAO algorithm is an intelligent optimization algorithm that combines mathematical properties with stochastic search theory, inspired by the similar triangle structure in plane geometry. This algorithm constructs multiple similar triangular topological units with inherent structural symmetry and achieves the optimization objective by integrating high-quality vertex information of the triangles. In the TTAO algorithm, the two key steps of global aggregation and local aggregation correspond to the development and exploration phases of the intelligent optimization algorithm, respectively. Global aggregation generates new nodes through positive information exchange between different triangular topological units, while local aggregation constructs new topological units at optimal positions based on the best vertices of each unit. Although the TTAO algorithm performs well in optimization, it suffers from problems such as being prone to getting trapped in local optima and having low convergence accuracy during the development and exploration phases. Therefore, this invention proposes an improved triangular topology aggregation optimization (ITTAO) algorithm, which optimizes the TTAO algorithm using an adaptive perturbation balancing strategy and a nonlinear tangential flight strategy.

[0061] In the global aggregation phase of the TTAO algorithm, the generation of new individuals depends on a linear combination between the current best individual and randomly selected individuals. This approach is prone to the loss of key information when dealing with complex optimization problems. Especially when individual differences are significant, new individuals are unlikely to effectively inherit the superior characteristics of their parents, thus limiting the algorithm's global exploration capability and increasing the risk of the algorithm getting trapped in local optima.

[0062] The adaptive perturbation balancing strategy aims to enhance the algorithm's exploration capability and convergence performance during the global search phase. This strategy introduces an adaptive inertia factor and an adaptive perturbation factor during the global aggregation phase. The adaptive inertia factor dynamically adjusts the weight ratio between the current optimal individual and random individuals when generating new individuals, while the adaptive perturbation factor endows the algorithm with stronger exploration capabilities in the early iterations by introducing controllable perturbations into the search direction to expand the search range. As iterations progress, the perturbation factor gradually decreases, allowing the algorithm to focus more on fine-grained searches of local regions in later stages, thus achieving an effective balance between exploration and development. The mathematical expression of the adaptive perturbation balancing strategy is as follows:

[0063] in, Indicates the current iteration number. Indicates the th population after the adaptive perturbation balance policy update. Individual, Indicates the first The first generation of the optimal population Individual, Indicates the first The first randomly selected individual from the optimal population. Indicates the first The second randomly selected individual from the optimal population. Represents the adaptive inertia factor. This represents the adaptive perturbation factor. The value gradually decreases from large to small, which makes it more capable of exploration in the early iteration stage and more capable of utilization in the later iteration stage, thus achieving a better balance between the exploration and utilization stages.

[0064] and The mathematical expression is as follows:

[0065]

[0066] in, Represents a random number between [0,1]. Indicates the maximum number of iterations, weight coefficients and , The value will gradually decrease, and it can adjust. The rate of change.

[0067] The adaptive perturbation balancing strategy constructs a composite update formula containing linear crossover and dynamic perturbation during the global aggregation stage. This allows new individuals to inherit the information of the best individual while also introducing necessary randomness and diversity through the perturbation term. This design improves the dispersion of the population during the search process, effectively avoids premature convergence caused by information homogenization, and makes the algorithm more robust in complex function optimization.

[0068] To address the premature convergence problem during the local aggregation phase, the nonlinear tangential flight strategy utilizes the nonlinear properties of the tangent function to simulate the stochastic transition behavior of individuals. The mathematical expression of the nonlinear tangential flight operator is as follows:

[0069] in, Dimensions representing the optimization process This indicates that a random number within a specified range will be generated.

[0070] The mathematical expression for the local aggregation stage of the nonlinear tangential flight strategy optimization is as follows:

[0071] in, This represents the first [unit] in the population after the nonlinear tangential flight strategy update. Individual, Indicates the first The first generation of the optimal population Individual, Indicates the first The first generation of superior population Individual.

[0072] The nonlinear tangential flight strategy enables individuals to make random perturbations with large and small steps near the local optimum, thereby promoting the non-uniform expansion of the local aggregation region. This strategy not only enhances the algorithm's global search capability but also effectively reduces the probability of premature convergence in the optimization problem.

[0073] The ITTAO algorithm is compared with the TTAO algorithm, Sparrow Search Algorithm (SSA), and Grey Wolf Optimizer (GWO). Single-peaked and multi-peaked test functions are used as examples for testing. The population size is set to 30, and the maximum number of iterations is 1000. The above functions are tested respectively, and the results are as follows. Figure 4 As shown in the test plot of the unimodal function, the convergence curve of ITTAO exhibits a trend of high accuracy and fast convergence speed. Among the four optimization algorithms, ITTAO requires the fewest iterations at the same level of accuracy. This indicates that the adaptive perturbation balancing strategy enhances the algorithm's exploration ability and convergence performance in the global search phase, enabling it to find excellent fitness values ​​earlier. The nonlinear tangential flight strategy reduces the possibility of the algorithm repeatedly getting trapped in local optima, accelerating the convergence speed. The ITTAO algorithm demonstrates superiority in both convergence speed and optimization accuracy. Therefore, applying the ITTAO algorithm to hyperparameter optimization of the model can achieve high convergence accuracy and fast convergence speed.

[0074] ITTAO was applied to the hyperparameter tuning of the proposed model. Hyperparameters included the learning rate, dropout rate, batch size, and the number of heads in the multi-head autocorrelation mechanism. The fitness function was the diagnostic error rate. After data was input into the model, training and test sets were generated. During model training, ITTAO was continuously applied to optimize the model's hyperparameters until the optimal values ​​were obtained, completing the training. The remaining service life of the gas turbine was then predicted using the test set.

[0075] S4. The feature vectors are mapped to the remaining service life prediction values ​​through a fully connected layer. The test set data is input into the trained model, and the remaining service life prediction results of the gas turbine are output.

[0076] This invention predicts the remaining life of a gas turbine by analyzing the operating parameters of its various components. The experimental data is the C-MAPSS full-life simulation dataset. The sensors used to collect data in this dataset are mainly distributed across key components such as the combustion chamber, low-pressure rotor, high-pressure rotor, low-pressure compressor, high-pressure compressor, low-pressure turbine, high-pressure turbine, and nozzles. It monitors 21 parameters, including temperature, pressure, and speed, throughout the gas turbine's operating cycle. The dataset includes four sets of full-life cycle operating data for the gas turbine under different operating conditions and failure modes. The C-MAPSS dataset records the monitoring characteristic parameter values ​​of the 21 sensors during gas turbine operation. Here, subsets FD001 and FD003 from the C-MAPSS dataset are selected to verify the effectiveness of the proposed model.

[0077] To reduce redundancy in the input data of the transmission model, the contribution of monitoring data from each component is evaluated using the random forest algorithm, such as... Figure 5 As shown, data from 7 sensors with relatively low contribution rates were removed, and data from 14 sensors with higher correlation to the remaining service life of the gas turbine were selected as input. These included: total outlet temperature of the low-pressure compressor, total outlet temperature of the high-pressure compressor, total outlet temperature of the low-pressure turbine, total outlet pressure of the high-pressure compressor, static pressure of the high-pressure compressor outlet, ratio of fuel flow rate to total outlet pressure of the high-pressure compressor, bypass ratio, bleed air enthalpy, cooling bleed air flow rate of the high-pressure turbine, cooling bleed air flow rate of the low-pressure turbine, blade physical speed, blade corrected speed, core shaft physical speed, and core shaft corrected speed. This improved the efficiency of service life prediction. After removing data with no obvious degradation trend, the data underwent max-min normalization to scale the sensor data to the range [0,1]. The normalization formula is as follows:

[0078] in, and The first Minimum and maximum values ​​of sensor data, For the first The first sensor Observational data for each sequence, For the first The first sensor The data after normalization of the sequence.

[0079] Choose an appropriate sliding window to process the normalized data. The normalized data was processed into rectangular data using an appropriate sliding time window to facilitate processing by the deep learning model. Experiments showed that a sliding window size of 30 ensured the samples contained sufficient time dimensions while avoiding data redundancy. The experiment employed a piecewise linear degradation model, dividing the remaining lifespan of the gas turbine into stable operation and linear degradation stages.

[0080] The proposed model yields predictions for the remaining life of the gas turbine. Figure 6 The remaining life prediction results for two randomly selected gas turbines from the FD001 dataset are presented. Figure 7 The remaining life prediction results for two gas turbines randomly selected from the FD003 test set are presented. Figure 8 The remaining service life prediction results for 100 gas turbines in the FD001 and FD003 datasets are presented. Figure 6 As shown in Figures 7-7, the model's remaining service life predictions exhibit stability in the initial stages of operation. However, when the remaining service life of the gas turbine begins to decline, the model's predictions show some bias. This is mainly because the characteristic information in the early stages of degradation is not obvious enough. As the gas turbine's service life decreases, the degradation characteristic information in the time-series data becomes increasingly significant, enabling the model to output more accurate prediction results. Figure 8 As can be seen, the deviation between the predicted value and the actual value is small, and the model has high accuracy in predicting the remaining life of the gas turbine.

[0081] Example 2 Based on Example 1, Example 2 also provides a gas turbine remaining life prediction system based on the improved TTAO algorithm, including: The data acquisition module is used to collect time-series monitoring data of the gas turbine degradation process. It evaluates the contribution value of sensor data of each component through random forest algorithm, selects the top N sensor data with high contribution values ​​as input features, and randomly divides the selected feature dataset into training set and test set in a ratio of 8:2.

[0082] Specifically, N is 14, and the sensor data includes: total outlet temperature of the low-pressure compressor, total outlet temperature of the high-pressure compressor, total outlet temperature of the low-pressure turbine, total outlet pressure of the high-pressure compressor, static pressure at the outlet of the high-pressure compressor, ratio of fuel flow rate to total outlet pressure of the high-pressure compressor, bypass ratio, bleed air enthalpy, cooling bleed air flow rate of the high-pressure turbine, cooling bleed air flow rate of the low-pressure turbine, physical blade speed, corrected blade speed, physical core shaft speed, and corrected core shaft speed.

[0083] An improved Autoformer model core module is used to input the training set into the improved Autoformer model. The parallel architecture of the ICBAM module and the autocorrelation mechanism are used to extract temporal dependencies. The ICBAM module contains a multi-scale convolutional kernel structure, which dynamically adjusts the convolutional kernel size to adapt to the nonlinear feature changes of the monitoring data.

[0084] Specifically, the ICBAM module is as follows: The input data is processed in both the channel and spatial dimensions. The channel attention module performs global average pooling and global max pooling to obtain features. and Multi-scale convolution operations are performed on the two features using convolution kernels of three different sizes. The operational expression is as follows:

[0085]

[0086] in, and These represent the features obtained after global average pooling and global max pooling in the channel attention module, respectively. and They represent and Features obtained through multi-scale convolution operations This represents summation of elements. express convolution kernel, express convolution kernel, express The convolution kernel; The operational expression for channel attention is as follows:

[0087] in, For input variables, The channel attention feature weight vector of the input variables. It is the Sigmoid activation function. It is a multilayer perceptron neural network.

[0088] The operational expression for ICBAM is as follows:

[0089] in, The numerical values ​​representing the computational values ​​of the convolutional attention mechanism. This indicates that the elements are multiplied and then added together one by one. This represents the spatial attention feature weight vector of the input vector.

[0090] The improved Autoformer model is as follows:

[0091]

[0092] in, Represents the variable to be decomposed. Indicates the trend term. Indicates seasonal items, This represents the fill function. This represents the average pooling function; The formula for calculating the autocorrelation coefficient is:

[0093] in, Represents the autocorrelation coefficient. Indicates the time lag. Indicates Fourier transform, This indicates the conjugate operation. This represents the inverse Fourier transform.

[0094] The formula for calculating the autocorrelation mechanism is:

[0095] in, The numerical value representing the autocorrelation mechanism. Indicates the first A lag time, This indicates the number of autocorrelation coefficients selected. Represents the computation of vectors , The autocorrelation coefficient, Representing vectors of First-order lag vector, vector , , Depend on It is obtained through linear transformation.

[0096] The expression for the attention mechanism in parallel architecture is: .

[0097] The loss function construction module is used to construct a loss function that uses a weighted combination of JS divergence and root mean square error. It employs an improved TTAO algorithm based on an adaptive perturbation balancing strategy and a nonlinear tangential flight strategy to optimize and adjust the model hyperparameters, and minimizes the prediction error through a backpropagation algorithm.

[0098] Specifically, the formula for calculating the loss function is as follows:

[0099] in, and For weighting coefficient control, and , Indicates the number of samples. This represents the distribution of the model output values. The distribution of actual measured values, Indicates the first The true value of each sample Indicates the first The predicted value for each sample, express and The JS divergence between them.

[0100] The calculation formula for the adaptive perturbation balancing strategy is as follows:

[0101] in, Indicates the current iteration number. Indicates the th population after the adaptive perturbation balance policy update. Individual, Indicates the first The first generation of the optimal population Individual, Indicates the first The first randomly selected individual from the optimal population. Indicates the first The second randomly selected individual from the optimal population. Represents the adaptive inertia factor. This represents the adaptive perturbation factor.

[0102] and The mathematical expression is:

[0103]

[0104] in, Represents a random number between [0,1]. Indicates the maximum number of iterations, weight coefficients and , .

[0105] The calculation formula for the nonlinear tangential flight strategy is as follows:

[0106] in, Dimensions representing the optimization process This indicates that a random number within a specified range will be generated.

[0107] The mathematical expression for the local aggregation stage of the nonlinear tangential flight strategy optimization is as follows:

[0108] in, This represents the first [unit] in the population after the nonlinear tangential flight strategy update. Individual, Indicates the first The first generation of the optimal population Individual, Indicates the first The first generation of superior population Individual.

[0109] The prediction output module is used to map feature vectors to remaining service life prediction values ​​through a fully connected layer, input test set data into the trained model, and output the remaining service life prediction results of the gas turbine.

[0110] Example 3 Based on Embodiment 1, Embodiment 3 of the present invention also provides a processing device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the method steps of Embodiment 1 by calling the program instructions.

[0111] Example 4 Based on Embodiment 1, Embodiment 4 of the present invention also provides a computer-readable storage medium storing computer instructions that cause the computer to perform the steps of the method described in Embodiment 1.

[0112] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the remaining life of a gas turbine based on an improved TTAO algorithm, characterized in that, include: S1. Collect time-series monitoring data of the gas turbine degradation process, evaluate the contribution value of sensor data of each component through random forest algorithm, select the top N sensor data with high contribution value as input features, and randomly divide the selected feature dataset into training set and test set in a ratio of 8:

2. S2. Input the training set into the improved Autoformer model, and use the parallel architecture of the ICBAM module and the autocorrelation mechanism to extract the temporal dependency. The ICBAM module contains a multi-scale convolutional kernel structure, which dynamically adjusts the size of the convolutional kernel to adapt to the nonlinear feature changes of the monitoring data. S3. Construct a weighted combination of JS divergence and root mean square error as the loss function, and use an improved TTAO algorithm based on adaptive perturbation balance strategy and nonlinear tangential flight strategy to adjust the model hyperparameters. Minimize the prediction error through backpropagation algorithm. S4. The feature vectors are mapped to the remaining service life prediction values ​​through a fully connected layer. The test set data is input into the trained model, and the remaining service life prediction results of the gas turbine are output.

2. The gas turbine remaining life prediction method based on the improved TTAO algorithm according to claim 1, characterized in that, The ICBAM module is specifically: The input data is processed in both the channel and spatial dimensions. The channel attention module performs global average pooling and global max pooling to obtain features. and Multi-scale convolution operations are performed on the two features using convolution kernels of three different sizes. The operational expression is as follows: in, and These represent the features obtained after global average pooling and global max pooling in the channel attention module, respectively. and They represent and Features obtained through multi-scale convolution operations This represents summation of elements. express convolution kernel, express convolution kernel, express The convolution kernel; The operational expression for channel attention is as follows: in, For input variables, The channel attention feature weight vector of the input variables. It is the Sigmoid activation function. It is a multilayer perceptron neural network; The operational expression for ICBAM is as follows: in, The numerical values ​​representing the computational values ​​of the convolutional attention mechanism. This indicates that the elements are multiplied and then added together one by one. This represents the spatial attention feature weight vector of the input vector.

3. The gas turbine remaining life prediction method based on the improved TTAO algorithm according to claim 1, characterized in that, The improved Autoformer model is specifically as follows: in, Represents the variable to be decomposed. Indicates the trend term. Indicates seasonal items, This represents the fill function. This represents the average pooling function; The formula for calculating the autocorrelation coefficient is: in, Represents the autocorrelation coefficient. Indicates the time lag. Indicates Fourier transform, This indicates the conjugate operation. Indicates the inverse Fourier transform; The formula for calculating the autocorrelation mechanism is: in, The numerical value representing the autocorrelation mechanism. Indicates the first A lag time, This indicates the number of autocorrelation coefficients selected. Represents the computation of vectors , The autocorrelation coefficient, Representing vectors of First-order lag vector, vector , , Depend on Obtained through linear transformation; The expression for the attention mechanism in parallel architecture is: 。 4. The gas turbine remaining life prediction method based on the improved TTAO algorithm according to claim 1, characterized in that, The calculation formula for the adaptive perturbation balance strategy is as follows: in, Indicates the current iteration number. Indicates the th population after the adaptive perturbation balance policy update. Individual, Indicates the first The first generation of the optimal population Individual, Indicates the first The first randomly selected individual from the optimal population. Indicates the first The second randomly selected individual from the optimal population. Represents the adaptive inertia factor. Indicates the adaptive perturbation factor; and The mathematical expression is: in, Represents a random number between [0,1], with weighting coefficients. and , , This indicates the maximum number of iterations.

5. The gas turbine remaining life prediction method based on the improved TTAO algorithm according to claim 1, characterized in that, The calculation formula for the nonlinear tangential flight strategy is as follows: in, Dimensions representing the optimization process This indicates the generation of random numbers within a specified range; The mathematical expression for the local aggregation stage of the nonlinear tangential flight strategy optimization is as follows: in, This represents the first [unit] in the population after the nonlinear tangential flight strategy update. Individual, Indicates the first The first generation of the optimal population Individual, Indicates the first The first generation of superior population Individual.

6. The gas turbine remaining life prediction method based on the improved TTAO algorithm according to claim 1, characterized in that, N is 14, and the sensor data includes: total outlet temperature of low-pressure compressor, total outlet temperature of high-pressure compressor, total outlet temperature of low-pressure turbine, total outlet pressure of high-pressure compressor, static pressure of high-pressure compressor outlet, ratio of fuel flow rate to total outlet pressure of high-pressure compressor, bypass ratio, bleed air enthalpy, cooling bleed air flow rate of high-pressure turbine, cooling bleed air flow rate of low-pressure turbine, physical speed of blades, corrected speed of blades, physical speed of core shaft, and corrected speed of core shaft.

7. The gas turbine remaining life prediction method based on the improved TTAO algorithm according to claim 1, characterized in that, The formula for calculating the loss function is as follows: in, and For weighting coefficient control, and , Indicates the number of samples. This represents the distribution of the model output values. The distribution of actual measured values, Indicates the first The true value of each sample Indicates the first The predicted value for each sample, express and The JS divergence between them.

8. A gas turbine remaining life prediction system based on an improved TTAO algorithm, characterized in that, include: The data acquisition module is used to collect time-series monitoring data of the gas turbine degradation process. It evaluates the contribution value of sensor data of each component through random forest algorithm, selects the top N sensor data with high contribution values ​​as input features, and randomly divides the selected feature dataset into training set and test set in a ratio of 8:

2. An improved Autoformer model core module is used to input the training set into the improved Autoformer model. The parallel architecture of the ICBAM module and the autocorrelation mechanism are used to extract temporal dependencies. The ICBAM module contains a multi-scale convolutional kernel structure, which dynamically adjusts the convolutional kernel size to adapt to the nonlinear feature changes of the monitoring data. The loss function construction module is used to construct a loss function with a weighted combination of JS divergence and root mean square error. It uses an improved TTAO algorithm based on an adaptive perturbation balance strategy and a nonlinear tangential flight strategy to adjust the model hyperparameters and minimizes the prediction error through the backpropagation algorithm. The prediction output module is used to map feature vectors to remaining service life prediction values ​​through a fully connected layer, input test set data into the trained model, and output the remaining service life prediction results of the gas turbine.

9. A processing device, characterized in that, The method includes at least one processor and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the method as described in any one of claims 1 to 7 by invoking the program instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause the computer to perform the method as described in any one of claims 1 to 7.

Images