Turboshaft engine residual life prediction method based on long short-term memory network of multi-statistical decoupling

Abstract

This invention provides a method for predicting the remaining life of turbofan engines based on a long short-term memory network with multi-statistical decoupling, relating to the field of equipment life prediction technology based on neural networks. The method includes: collecting monitoring data from multi-dimensional sensors during the operation of industrial equipment; filtering and standardizing the monitoring data to form an original sample set; preprocessing the original sample set to obtain time-series samples; and predicting the remaining life using a remaining life prediction model. The advantage of this invention is that it can achieve high-precision remaining life prediction in complex, noisy industrial data environments.

Description

Technical Field

[0001] This invention relates to the field of equipment life prediction technology based on neural networks, and more specifically, to a method for predicting the remaining life of turbofan engines based on a long short-term memory network with multiple statistical decoupling. Background Technology

[0002] With the continuous increase in the complexity of industrial systems and the widespread application of intelligent manufacturing and condition monitoring technologies, large-scale, multi-dimensional sensor data generated during equipment operation is constantly accumulating. Against this backdrop, remaining useful life prediction, as a key component of fault prediction and health management systems, is of great significance for identifying potential faults in advance, developing reasonable maintenance plans, and avoiding sudden failures.

[0003] Traditional approaches to remaining useful life (UPS) prediction primarily focus on two categories: model-based and data-driven methods. Model-based methods typically rely on physical mechanism modeling or historical degradation patterns, constructing mathematical models to describe the system's degradation process. While offering some interpretability, these methods are heavily reliant on prior knowledge and struggle to accurately characterize the nonlinear degradation behavior of equipment in complex industrial scenarios. Furthermore, model building and parameter identification are costly. Data-driven methods, on the other hand, predict UPS by directly mining latent features and evolutionary patterns from sensor data. They offer greater flexibility and adaptability, but the traditional machine learning methods they rely on often face limitations in prediction accuracy and computational efficiency when feature engineering is complex or the data scale is large. To address the shortcomings of traditional approaches, deep learning methods have been widely applied in UPS prediction in recent years. In particular, Long Short-Term Memory (LSTM) networks have demonstrated excellent performance in time series modeling, effectively capturing the temporal dependencies in the equipment degradation process. However, existing LSTM-related methods still have certain limitations in practical industrial applications. On the one hand, industrial sensor data is inevitably subject to noise interference, which propagates layer by layer in the LSTM chain structure and can easily interfere with the extraction of effective feature information. On the other hand, the long-term dependency information contained in the hidden state in the LSTM network is not fully utilized, and when the effects of multidimensional sensor data and features at different degradation stages differ significantly, a single attention mechanism may introduce additional interference factors while strengthening key information, thereby affecting prediction accuracy.

[0004] Therefore, it is necessary to propose a method that can achieve high-precision remaining life prediction in complex and noisy industrial data environments. By enhancing the utilization of deep temporal features of LSTM, effectively integrating original sensor information, and suppressing noise interference, the model's ability to model equipment degradation processes and its predictive reliability can be improved, providing reliable technical support for intelligent operation and maintenance and health management of industrial equipment. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the remaining life of turbofan engines based on long short-term memory networks with multi-statistical decoupling, which can achieve high-precision remaining life prediction in complex and noisy industrial data environments.

[0006] This invention is achieved through the following technical solution: A method for predicting the remaining life of turbofan engines based on long short-term memory networks with multi-statistical decoupling includes the following steps: Collect monitoring data from multi-dimensional sensors during the operation of industrial equipment, screen the effectiveness of the monitoring data from the multi-dimensional sensors and standardize the data to form an original sample set; Data preprocessing is performed on the original sample set to obtain time series samples; Remaining lifetime prediction is performed using a remaining lifetime prediction model, including: Using a long short-term memory network as an encoder, temporal features are extracted based on time series samples, and deep temporal features are extracted based on the temporal features. Based on time series samples, initial features and global trends are extracted through self-attention mechanism and multi-statistic pooling aggregator, respectively. Then, a static gating mechanism is used to fuse the initial features and global trends to obtain fused features. Based on deep temporal features and fused features, the decoder outputs the prediction results of the remaining lifespan of the device.

[0007] Preferably, the method for effectiveness screening is as follows: Remove monitoring data corresponding to sensor dimensions whose values ​​remain unchanged or whose changes are less than a preset threshold throughout their entire lifecycle; The standardization process is as follows: The remaining monitoring data were processed using the Z-score normalization method.

[0008] Preferably, the method for performing data preprocessing is as follows: The monitoring data in the original sample set were smoothed using the Savitzky-Golay filtering method. The time series sample is constructed by segmenting the filtered monitoring data using a sliding window method.

[0009] Preferably, the method for extracting the temporal features based on the time series samples is to perform temporal modeling on the time series samples using a long short-term memory network as an encoder:

[0010]

[0011]

[0012]

[0013]

[0014]

[0015] in, The time series sample at time step t, The hidden state at time step t-1, It is the Sigmoid activation function. Let be the output vector of the forget gate at time step t. Let be the output vector of the input gate at time step t. For candidate cell states at time step t, The cell state at time step t. The cell state at time step t-1. The output vector of the output gate at time step t. , , and These are all weight matrices for Long Short-Term Memory networks. , , and These are all bias matrices for Long Short-Term Memory (LSTM) networks. The hyperbolic tangent activation function is given, and the temporal feature is... The t-th element This represents the hidden state at time step t. , The total number of time steps for the time series samples.

[0016] Preferably, the method for extracting the deep temporal features based on the temporal features is as follows: feature extraction is performed using a deep temporal feature deep extractor, which includes a feature decoupling aggregator and a global average pooling layer. ; ; ; ; ; in, For the first The monitoring data from the dimensional sensor is output by the feature decoupling aggregator. The output sequence of the feature decoupling aggregator is the monitoring data from all sensors. For the first The monitoring data of the dimensional sensor is the j-th element of the time-series feature. The corresponding value in the text, For the dimensions of the sensor, and These are the weights and biases of the feature decoupling aggregator updated during training. For the first The output sequence of the sensor's monitoring data in the global average pooling layer. The output of the global average pooling layer is the monitoring data from all sensors. This refers to the deep temporal features.

[0017] Preferably, the method for extracting the initial features is as follows: ; ; ; in, For the time series sample, The hyperbolic tangent activation function is used. and These are the first layer weight matrix and the second layer weight matrix, respectively. As a score vector for the importance of time steps, Let be a weight vector and let the t-th element be... , Represents time step The weight, For the initial feature, for function, For time step The time series samples, The total number of time steps for the time series samples. This represents a request for transposition.

[0018] Preferably, the method for extracting the global trend is to perform global trend modeling on the time series samples using a multi-statistic pooling aggregator: ; ; ; ; ; in, , , and These are the minimum value parameter, the maximum value parameter, the standard deviation parameter, and the mean value parameter, respectively. Represents the time step in the time series sample The time series samples The first in Monitoring data from dimensional sensors, min and max are functions for finding the minimum and maximum values, respectively. It is the first time step in all time steps of the time series sample. The average value of the monitoring data from the dimensional sensor. For the global trend, and These are the weights and biases of the multi-statistic pooling aggregator, respectively.

[0019] Preferably, the method for feature fusion using a static gating mechanism to integrate initial features and global trends is as follows: ; in, For the fusion feature, It is the Sigmoid activation function. These are the model parameters for the static gating mechanism.

[0020] Preferably, the method for predicting the remaining lifespan of the device based on deep temporal features and fused features through the decoder outputs the following: ; in, It is the Sigmoid activation function. For the deep temporal features, For the fusion feature, and The weights and biases of the decoder, The sub-network consists of a single-layer long short-term memory network and a linear layer, and the single-layer long short-term memory network shares model parameters with the long short-term memory network of the encoder. This is a prediction of the remaining service life of the equipment.

[0021] Preferably, the model parameters of the remaining lifespan prediction model are optimized through supervised learning.

[0022] The technical solution of the present invention has at least the following advantages and beneficial effects: This invention fully mines the effective information in the temporal hidden state and enhances the modeling ability of long-term dependencies by coordinating the design of the encoder's long short-term memory network and the deep temporal feature extractor. This invention effectively balances local key information with the overall degradation trend by using a dual-path initial feature representation of a self-attention mechanism and a multi-statistic pooling aggregator, thereby reducing the risk of interference information introduced by the attention mechanism. This invention combines filtering and multi-stage feature fusion strategies to significantly improve the model's prediction robustness in noisy environments; This invention is rationally designed and applicable to predicting the remaining service life of complex industrial equipment, providing reliable technical support for predictive maintenance and health management systems. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling, as provided in Embodiment 1 of the present invention. Detailed Implementation

[0024] 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 with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0025] Example This embodiment provides a method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multiple statistical decoupling. (See also...) Figure 1 This includes the following steps: Step S1: Collect monitoring data from multi-dimensional sensors during the operation of industrial equipment. Perform validity screening and standardization on the monitoring data to form an original sample set. Validity screening aims to reduce redundant information, while standardization aims to unify the dimensions and scales of different sensor variables, reducing the impact of feature distribution differences on model training.

[0026] The following is an example of monitoring data from a multi-dimensional sensor: The data collected by the multi-dimensional sensor has 14 fields, with the following meanings: Total temperature at LPCoutlet (low-pressure compressor outlet total temperature, unit °R), Total temperature at HPC outlet (high-pressure compressor outlet total temperature, unit °R), Total temperature at LPT outlet (low-pressure turbine outlet total temperature, unit °R), Total pressure at HPC outlet (high-pressure compressor outlet total pressure, unit psia), Physical fan speed (fan physical speed, unit rpm), Physical core speed (core physical speed, unit rpm), Static pressure at HPC outlet (high-pressure compressor outlet static pressure, unit psia), Ratio of fuelflow to Ps30 (fuel flow to Ps30 ratio, unit pps / psi), Corrected fan speed (fan equivalent speed, unit rpm), Corrected core speed (core equivalent speed, unit rpm), Bypass Ratio (bypass ratio, unit --), Bleed Enthalpy (bleed air enthalpy, unit --), HPT Coolant bleed (high pressure turbine cooling bleed air, unit: lbm / s), LPT coolant bleed (low pressure turbine cooling bleed air, unit: lbm / s). Example of sensor data for one acquisition time: [641.82, 1589.70, 1400.60, 554.36, 2388.06, 9046.19, 47.47, 521.66, 2388.02, 8138.62, 8.4195, 392, 39.06, 23.4190].

[0027] First, the method for effectiveness screening is as follows: Remove monitoring data corresponding to sensor dimensions whose values ​​remain unchanged or whose changes are less than a preset threshold throughout their entire lifecycle.

[0028] Secondly, the method for standardizing the remaining data is as follows: The remaining monitoring data were processed using the Z-score normalization method. The formula for Z-score normalization is: ; in, For standardized data, For data before standardization, and These are the mean and standard deviation of the corresponding sensor data, respectively. The standardized data is used for subsequent processing steps.

[0029] Step S2: Perform data preprocessing on the original sample set to obtain time series samples. In this embodiment, the data preprocessing method is as follows: To reduce the interference of random noise in industrial sensor data on model training and preserve degradation trend features, the Savitzky-Golay filtering method is used to smooth the monitoring data in the original sample set, thereby reducing high-frequency noise components while preserving the signal degradation trend. The expression is as follows: ; ; ; in, for The final correction value, It is the fitting function Order coefficient, , The parameters that need to be specified for the fitting function. It means Surrounding values ​​relative to the window center Location index difference, This refers to solving for specified coefficients. The partial derivative is the index of the position of the coefficient. The number representing the specified coefficient. , This refers to half the width of the window. Refers to the time step The corresponding original data value.

[0030] To form a time-series sample set that can be directly input into the model, it is usually necessary to perform sliding sampling on the time-series data based on a preset window length and step size to construct a sample sequence of fixed length. Therefore, this embodiment needs to use a sliding window method to segment and sample the filtered monitoring data to construct the time-series sample with a total of m time steps. The j-th element is , The time series sample representing time step j includes monitoring data from an n-dimensional sensor. : .

[0031] Step S3: Predict remaining lifetime using the remaining lifetime prediction model, including: In feature extraction, on the one hand, since the Long Short-Term Memory Network can control the flow of information through the forget gate, input gate and output gate, the Long Short-Term Memory Network is used as the encoder to extract temporal features based on time series samples. The temporal features are the temporal dependencies in the device operation process. Then, deep temporal features are extracted based on the temporal features. Specifically, the method for extracting the temporal features based on the time series samples is as follows: A long short-term memory network is used as an encoder to perform temporal modeling on the time series samples.

[0032]

[0033]

[0034]

[0035]

[0036]

[0037] in, For time step The time series samples, The hidden state at time step t-1, It is the Sigmoid activation function. For time step The output vector of the forget gate, For time step The output vector of the input gate, For time step Candidate cell state, For time step The state of cells, The cell state at time step t-1. For time step The output vector of the output gate. , , and These are all weight matrices for Long Short-Term Memory networks. , , and These are all bias matrices for Long Short-Term Memory (LSTM) networks. The hyperbolic tangent activation function is given, and the temporal feature is... The first of them element Represents a time step The hidden state, , Let be the total number of time steps for the time series samples. Let the encoder function be... Its mapping relationship to the sliding window sequence is expressed as: .

[0038] Based on the above results, in order to fully explore... The effective information in the hidden state needs to be extracted through a collaborative processing method of feature decoupling aggregation and global information convergence to enhance the extraction of effective information and suppress noise interference accumulated during layer-by-layer propagation. In this embodiment, feature extraction is performed using a deep temporal feature depth extractor, which includes a feature decoupling aggregator and a global average pooling layer. ; ; ; ; ; in, For the first The monitoring data from the dimensional sensor is output by the feature decoupling aggregator. The output sequence of the feature decoupling aggregator is the monitoring data from all sensors. For the first The monitoring data of the dimensional sensor is the j-th element of the time-series feature. The corresponding value in the text, For the dimensions of the sensor, and These are the weights and biases of the feature decoupling aggregator updated during training. For the first The output sequence of the sensor's monitoring data in the global average pooling layer. The output of the global average pooling layer is the monitoring data from all sensors. This refers to the deep temporal features.

[0039] On the other hand, when extracting features, based on time series samples, initial features and global trends are extracted respectively through self-attention mechanism and multi-statistic pooling aggregator. Then, a static gating mechanism is used to fuse the initial features and global trends to obtain fused features. In this step, self-attention mechanism is used to characterize the difference in importance of different time steps in the degradation process, while multi-statistic pooling aggregator characterizes the overall change trend of sensor data through various statistical indicators, supplementing the initial information expression from a global perspective. To supplement deep temporal features, a self-attention mechanism is introduced to model the initial features of the time series samples. In other words, the method for extracting these initial features is as follows: ; ; ; in, For the time series sample, The hyperbolic tangent activation function is used. and These are the first layer weight matrix and the second layer weight matrix, respectively. As a score vector for the importance of time steps, Let be a weight vector and where the first is a weight vector. Element is , Represents time step The weight, For the initial feature, for function, For time step The time series samples, The total number of time steps for the time series samples. This represents a request for transposition.

[0040] To avoid the attention mechanism from over-amplifying local noise, a multi-statistic pooling aggregator is used to model the global trend of the time series samples to extract the global trend: ; ; ; ; ; in, , , and These are the minimum value parameter, the maximum value parameter, the standard deviation parameter, and the mean value parameter, respectively. Represents the time step in the time series sample The time series samples The first in Monitoring data from dimensional sensors, min and max are functions for finding the minimum and maximum values, respectively. It is the first time step in all time steps of the time series sample. The average value of the monitoring data from the dimensional sensor. For the global trend, and These are the weights and biases of the multi-statistic pooling aggregator, respectively.

[0041] Based on this, a static gating mechanism is used to fuse initial features and global trends, balancing local key time step information with overall trend information. The method is as follows: ; in, For the fusion feature, It is the Sigmoid activation function. These are the model parameters for the static gating mechanism.

[0042] After obtaining the deep temporal features and fused features, the remaining lifespan of the device can be predicted by the decoder based on these features. The method is as follows: ; in, It is the Sigmoid activation function. For the deep temporal features, For the fusion feature, and The weights and biases of the decoder, The sub-network consists of a single-layer long short-term memory network and a linear layer, and the single-layer long short-term memory network shares model parameters with the long short-term memory network of the encoder. This is a prediction of the remaining service life of the equipment.

[0043] It is worth noting that the model parameters of the remaining life prediction model are optimized through supervised learning. After training, an optimal parameter set is obtained. During the prediction phase, newly acquired sensor data is processed using the same procedure and input into the model, outputting the corresponding prediction results for the remaining life of the equipment. Stabilization strategies can be used to ensure the reliability and robustness of the model convergence process. Model selection and system deployment are completed based on the evaluation results. During long-term operation, the model can be periodically updated based on new data to maintain prediction accuracy and stability.

[0044] During training, a health index or its remaining lifespan can be constructed based on the equipment's operating status as a monitoring label. The expression is as follows: ; in, This indicates the engine's health index. This indicates the total number of operating cycles of a certain engine. This indicates the number of dynamic operating cycles of a certain engine.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling, characterized in that, Includes the following steps: Collect monitoring data from multi-dimensional sensors during the operation of industrial equipment, screen the effectiveness of the monitoring data from the multi-dimensional sensors and standardize the data to form an original sample set; Data preprocessing is performed on the original sample set to obtain time series samples; Remaining lifetime prediction is performed using a remaining lifetime prediction model, including: Using a long short-term memory network as an encoder, temporal features are extracted based on time series samples, and deep temporal features are extracted based on the temporal features. Based on time series samples, initial features and global trends are extracted through self-attention mechanism and multi-statistic pooling aggregator, respectively. Then, a static gating mechanism is used to fuse the initial features and global trends to obtain fused features. Based on deep temporal features and fused features, the decoder outputs the prediction results of the remaining lifespan of the device.

2. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling as described in claim 1, characterized in that, The method for effectiveness screening is as follows: Remove monitoring data corresponding to sensor dimensions whose values ​​remain unchanged or whose changes are less than a preset threshold throughout their entire lifecycle; The standardization process is as follows: The remaining monitoring data were processed using the Z-score normalization method.

3. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling as described in claim 1, characterized in that, The method for performing data preprocessing is as follows: The monitoring data in the original sample set were smoothed using the Savitzky-Golay filtering method. The time series sample is constructed by segmenting the filtered monitoring data using a sliding window method.

4. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling as described in claim 1, characterized in that, The method for extracting the temporal features based on the time series samples is as follows: A long short-term memory network is used as an encoder to perform temporal modeling on the time series samples. in, For time step The time series samples, The hidden state at time step t-1, It is the Sigmoid activation function. For time step The output vector of the forget gate, Let be the output vector of the input gate at time step t. For time step Candidate cell state, For time step The state of cells, The cell state at time step t-1. For time step The output vector of the output gate. , , and These are all weight matrices for Long Short-Term Memory networks. , , and These are all bias matrices for Long Short-Term Memory (LSTM) networks. The hyperbolic tangent activation function is given, and the temporal feature is... The first of them element Represents a time step The hidden state, , The total number of time steps for the time series samples.

5. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling, as described in claim 4, is characterized in that... The method for extracting the deep temporal features based on the aforementioned temporal features is as follows: feature extraction is performed using a deep temporal feature deep extractor, which includes a feature decoupling aggregator and a global average pooling layer. ； ； ； ； ； in, For the first The monitoring data from the dimensional sensor is output by the feature decoupling aggregator. The output sequence of the feature decoupling aggregator is the monitoring data from all sensors. For the first The monitoring data of the dimensional sensor is the j-th element of the time-series feature. The corresponding value in the text, For the dimensions of the sensor, and These are the weights and biases of the feature decoupling aggregator updated during training. For the first The output sequence of the sensor's monitoring data in the global average pooling layer. The output of the global average pooling layer is the monitoring data from all sensors. This refers to the deep temporal features.

6. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling according to claim 1, characterized in that, The method for extracting the initial features is as follows: ； ； ； in, For the time series sample, The hyperbolic tangent activation function is used. and These are the first layer weight matrix and the second layer weight matrix, respectively. As a score vector for the importance of time steps, Let be a weight vector and where the first is a weight vector. Element is , Represents time step The weight, For the initial feature, for function, For time step The time series samples, The total number of time steps for the time series samples. This represents a request for transposition.

7. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling as described in claim 6, characterized in that, The method for extracting the global trend is to perform global trend modeling on the time series samples using the multi-statistic pooling aggregator: ； ； ； ； ； in, , , and These are the minimum value parameter, the maximum value parameter, the standard deviation parameter, and the mean value parameter, respectively. Represents the time step in the time series sample The time series samples The first in Monitoring data from dimensional sensors, min and max are functions for finding the minimum and maximum values, respectively. It is the first time step in all time steps of the time series sample. The average value of the monitoring data from the dimensional sensor. For the global trend, and These are the weights and biases of the multi-statistic pooling aggregator, respectively.

8. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling, as described in claim 7, is characterized in that... The method for feature fusion using a static gating mechanism to integrate initial features and global trends is as follows: ； in, For the fusion feature, It is the Sigmoid activation function. These are the model parameters for the static gating mechanism.

9. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling according to claim 1, characterized in that, The method for predicting the remaining lifespan of a device based on deep temporal features and fused features through a decoder output is as follows: ； in, It is the Sigmoid activation function. For the deep temporal features, For the fusion feature, and The weights and biases of the decoder, The sub-network consists of a single-layer long short-term memory network and a linear layer, and the single-layer long short-term memory network shares model parameters with the long short-term memory network of the encoder. This is a prediction of the remaining service life of the equipment.

10. The method for predicting the remaining life of a turbofan engine based on a long short-term memory network with multi-statistical decoupling according to claim 1, characterized in that, The model parameters of the remaining lifespan prediction model are optimized using supervised learning.

Images