Method and system for detecting abnormality of aero-engine fuel control system based on multivariate time series regression model
By extracting long-term temporal features of the aero-engine fuel control system using a multivariate temporal regression model and variational autoencoder, the problems of reliance on data annotation and long processing time in existing technologies are solved, and efficient and accurate anomaly detection is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTHWESTERN POLYTECHNICAL UNIV
- Filing Date
- 2024-06-20
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies for detecting anomalies in aero-engine fuel control systems rely on data annotation, which is labor-intensive, has poor generalization ability, cannot utilize temporal features, has low accuracy, and is too time-consuming.
An anomaly detection method based on a multivariate time-series regression model is adopted. By acquiring data from multiple monitoring units of the aero-engine fuel control system, the maximum information coefficient is calculated to screen features, long-term time-series features are extracted using a variational autoencoder (VAE), and anomaly detection is performed by training a multivariate time-series regression model using the least squares method.
It achieves anomaly detection without relying on data annotation, can discover previously undiscovered anomalies, has strong generalization ability, short processing time and high accuracy.
Smart Images
Figure CN118709104B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to a method and system for anomaly detection in aero-engine fuel regulation system based on a multivariate time-series regression model. Background Technology
[0002] With the development of big data technologies, massive amounts of time-series data are generated every moment in human society, such as in the financial stock market, factory production, and equipment monitoring. Time-series data reflects the development status of observed phenomena and describes the process of change over time. By analyzing time-series data, we can understand the characteristics, patterns, and trends of these changes, thereby helping us to uncover more information and provide support for decision-making.
[0003] Anomaly detection in time series data is a key research area in data mining. In a stable process of change, irregular fluctuations in observed time series data can be considered anomalous events, usually caused by sudden changes in the data itself. The purpose of time series data anomaly detection is to transform detected anomalies into crucial information for specific applications. In many practical engineering applications, anomaly detection in time series data has significant practical value, enabling direct early warning of faults.
[0004] Since the 20th century, the invention of the airplane has greatly advanced human civilization. As the "heart" of an aircraft, the aero-engine has always been a key research focus for scholars worldwide. As the primary power source of an aircraft, a serious malfunction in the aero-engine during flight often leads to catastrophic consequences, resulting in the destruction of the aircraft and loss of life. However, due to its complex structure, high integration, and operation in extreme environments such as high temperature, high speed, high load, and strong vibration, long-term operation inevitably encounters abnormal conditions, especially in the fuel control system. These abnormal conditions are reflected in the monitored time-series data. Therefore, to ensure flight safety and stability, it is necessary to accurately detect any abnormalities in the aero-engine fuel control system during ground maintenance and promptly alert personnel for maintenance and support.
[0005] In existing technologies, several techniques can be used for anomaly detection in aero-engine fuel control system data. Those skilled in the art have proposed time-series anomaly pattern detection algorithms based on local density. This technique segments the data, calculates local density, and detects anomalies by measuring the similarity between the data and pre-labeled fault information. However, this technique relies heavily on data labeling, is labor-intensive, and cannot detect previously undiscovered anomalies, exhibiting poor generalization. Existing technologies disclose anomaly detection algorithms based on improved support vector machine (SVM) models. This technique combines semi-supervised dimensionality reduction methods and particle swarm optimization (PSO) algorithms with manually determined parameters in a semi-supervised dimensionality reduction model and a trained SVM to detect anomalies. This technique can only detect anomalies at isolated time points in the aero-engine fuel control system data, failing to utilize the temporal characteristics of the data. Furthermore, the accuracy of anomaly detection using this technique is low, making it difficult to meet practical needs. Existing technologies also disclose aero-engine data anomaly detection method based on ensemble learning and neural networks. This technique extracts data features through stacked denoising autoencoders and then uses convolutional neural networks and backpropagation (BP) neural networks for anomaly detection. This technology is too time-consuming to detect anomalies in the fuel control system of aircraft engines, and cannot quickly obtain anomaly detection results. Summary of the Invention
[0006] To address the aforementioned technical problems, this invention primarily targets the shortcomings of existing technologies, such as over-reliance on data annotation, poor generalization, inability to utilize the temporal characteristics of data, low accuracy, and excessive time consumption. This invention provides a method and system for anomaly detection in aero-engine fuel control systems based on a multivariate temporal regression model. This method does not rely on data annotation and can detect previously undiscovered anomalies, exhibiting strong generalization capabilities. Simulation experiments further demonstrate that this invention has the advantages of short processing time and high accuracy in anomaly detection of aero-engine fuel control systems.
[0007] The first objective of this invention is to provide a method for anomaly detection in aero-engine fuel control systems based on a multivariate time-series regression model, comprising: Acquire dimensional data corresponding to multiple monitoring units during the operation of the aero-engine fuel control system; The maximum information coefficient between each monitoring unit is obtained based on data from multiple dimensions; Based on the maximum information coefficient between each monitoring unit, multiple monitoring units with a maximum information coefficient greater than the threshold are selected. The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as the training set to train the VAE network. After the training is completed, the encoder in the VAE network is obtained as the long-term feature extractor of the overall multivariate time-series regression model. Input the corresponding dimensional data from multiple times before the current time point into the long time series feature extractor of the overall multivariate time series regression model to obtain long time series features; Based on the dimensional data and long-term time-series features corresponding to multiple monitoring units with a maximum information coefficient greater than the threshold, a multivariate time-series regression model is constructed. The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as input to the multivariate time series regression model. The multivariate time series regression model is trained by the least squares method to obtain the trained multivariate time series regression model. The data to be detected is input into the trained multivariate time-series regression model to obtain the regression prediction results of the key feature dimensions of the fuel conditioning system at each time step. Anomalies in the data to be detected are determined based on the regression prediction results of the key feature dimensions of the fuel conditioning system at each time point.
[0008] Preferably, the abnormality of the data to be detected is judged, including: The regression prediction results of the key feature dimensions of the fuel conditioning system at each time point are subtracted from the actual values in the data to be tested to obtain the deviation at each time point. Time points where the deviation is greater than 10% of the actual value are marked as anomalies.
[0009] Preferably, the loss function used during the training of the VAE network includes mean squared error and KL divergence between the sampling distribution and the normal distribution.
[0010] Preferably, the KL divergence between the sampling distribution and the normal distribution is calculated using the following formula:
[0011] In the formula, Dimensional data The mean obtained by sampling after passing through the VAE network; The variance is obtained by sampling the dimensional data after it has passed through a VAE network. This represents the KL divergence between the sampled distribution and the normal distribution.
[0012] Preferably, the dimensional data is data monitored by the monitoring unit during aircraft flight and recorded by the onboard data recorder.
[0013] Preferably, the multivariate time-series regression model has the following formula:
[0014] In the formula, The dimension data of the monitoring unit to be detected at time T; These are the parameters for linear regression; The number of input data; To filter out the dimensional data of multiple monitoring units at time T whose maximum information coefficient is greater than a threshold; It is a 10-dimensional long-time series feature; and These are second-order difference data, specifically the data at time T minus the data at time T-1 and the data at time T minus the data at time T-2.
[0015] Preferably, during the training process of the VAE network, the learning rate is set to 0.01, the batch size is set to 10, and the number of iterations is set to 30.
[0016] The second objective of this invention is to provide an anomaly detection system for aero-engine fuel control system based on a multivariate time-series regression model, comprising: The data acquisition module is used to acquire dimensional data corresponding to multiple monitoring units during the operation of the aero-engine fuel control system; The data processing module is used to obtain the maximum information coefficient between each monitoring unit based on data from multiple dimensions; Based on the maximum information coefficient between each monitoring unit, multiple monitoring units with a maximum information coefficient greater than the threshold are selected. The network model training module is used to train the VAE network by using the dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold as the training set. After the training is completed, the encoder in the VAE network is obtained as the long-term feature extractor of the overall multivariate time-series regression model. Input the corresponding dimensional data from multiple times before the current time point into the long time series feature extractor of the overall multivariate time series regression model to obtain long time series features; Based on the dimensional data and long-term time-series features corresponding to multiple monitoring units with a maximum information coefficient greater than the threshold, a multivariate time-series regression model is constructed. The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as input to the multivariate time series regression model. The multivariate time series regression model is trained by the least squares method to obtain the trained multivariate time series regression model. The anomaly detection module is used to input the data to be detected into the trained multivariate time-series regression model to obtain the regression prediction results of the key feature dimensions of the fuel conditioning system at each time step. Anomalies in the data to be detected are determined based on the regression prediction results of the key feature dimensions of the fuel conditioning system at each time point.
[0017] The present invention has at least the following beneficial effects: This invention provides a method and system for anomaly detection in aero-engine fuel control systems based on a multivariate temporal regression model. This method automatically eliminates irrelevant or redundant dimensional data by performing feature selection on aero-engine fuel control system data, reducing data processing time. By training a variational autoencoder (VAE) neural network, an encoder mapping from high to low dimensions is obtained, enabling the extraction and utilization of long-term temporal features. A multivariate temporal regression model is established by combining the encoder in the VAE network to achieve regression prediction of aero-engine fuel control system data. By comparing the regression prediction results with actual data, abnormal data is filtered out based on the degree of deviation, ultimately achieving anomaly detection for aero-engine fuel control system data. This invention does not rely on data annotation and can detect previously undiscovered anomalies, exhibiting strong generalization ability. Simulation experiments further demonstrate that this invention has the advantages of short processing time and high accuracy in anomaly detection of aero-engine fuel control systems. Attached Figure Description
[0018] Figure 1 The present invention provides a flowchart of an anomaly detection method for aero-engine fuel control system based on a multivariate time-series regression model. Detailed Implementation
[0019] In order to illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description is provided in conjunction with the embodiments.
[0020] See Figure 1 As shown, a method for anomaly detection in aero-engine fuel control system based on a multivariate time-series regression model includes: Step 1: Acquisition of data from the aero-engine fuel control system Acquire dimensional data corresponding to multiple monitoring units during the operation of the aero-engine fuel control system; The dimensional data refers to the data monitored by the monitoring unit during aircraft flight and recorded by the onboard data recorder.
[0021] In this embodiment, during aircraft operation, the engine fuel control system is monitored by multiple monitoring units installed on the aircraft, such as fuel needle position and combustion chamber temperature. The generated data is recorded and saved by the onboard data recorder. After the aircraft completes its flight mission, the onboard data recorder stores the engine fuel control system data from the flight process. On the ground, the dimensional data corresponding to each monitoring unit recorded by the data recorder is read and saved to a computer in ".CSV" file format. One file represents the fuel control system data for one flight, each line in the file represents the data at a certain moment during the flight, and each column corresponds to a different moment. The obtained data file is used as the object for subsequent anomaly detection processing.
[0022] Step 2, Feature selection of data The maximum information coefficient between each monitoring unit is obtained based on data from multiple dimensions; Based on the maximum information coefficient between each monitoring unit, multiple monitoring units with a maximum information coefficient greater than the threshold are selected. In this embodiment, the maximum information coefficient (MIC) between each monitoring unit is obtained based on its corresponding dimensional data, wherein the MIC is defined as: (1) In equation (1), , This refers to the dimensional data corresponding to the two monitoring units currently being calculated, such as the position of the fuel needle and the combustion chamber temperature. For data length, The number of squares set for the algorithm is a hyperparameter used to count the number of data points within each square. Calculated. and Different grid division methods , The grid distribution of directions must satisfy the following conditions. . yes and The mutual information is calculated using the following formula: (2) In equation (2), For data length, For the first Count the number of data points in all cells. For the first The number of data points in all cells of a row. For the first Line 1 The number of data points in the cells of a column.
[0023] Therefore, calculate all that satisfy In the given situation, the grid partitioning methods are compared, and the method that yields the largest mutual information value is determined. and The MIC value.
[0024] The correlation threshold (MIC) values between the key parameter dimension of the combustion control system and all other dimensions are determined sequentially using the above method. Based on experience, a correlation threshold of 0.5 is set, and the dimension data corresponding to the monitoring units with MIC values greater than the threshold are selected and subsequently input into the VAE network model and the multivariate time series regression model.
[0025] Step 3, train the VAE encoder The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as the training set to train the VAE network. After the training is completed, the encoder in the VAE network is obtained as the long-term feature extractor of the overall multivariate time-series regression model. During the training of the VAE network, the loss functions used include mean squared error and KL divergence between the sampling distribution and the normal distribution.
[0026] During the training of the VAE network, the learning rate was set to 0.01, the batch size was set to 10, and the number of iterations was set to 30.
[0027] In this embodiment, a VAE network with a structure of 100*50*10*50*100 is built on the TensorFlow framework in a Python 3.7 environment on the computer, and the loss function is set as follows: (3) In equation (3), The mean squared error before and after passing through the VAE network is calculated using the following formula: (4) In equation (4), For data length, The input data to the VAE network is in vector form. This refers to the data output by the VAE network.
[0028] In equation (3), The KL divergence between the sampling distribution and the normal distribution is calculated using the following formula: (5) In equation (5), Dimensional data The mean obtained by sampling after passing through the VAE network; The variance is obtained by sampling the dimensional data after it has passed through a VAE network. This represents the KL divergence between the sampled distribution and the normal distribution.
[0029] In this embodiment, the optimizer in the neural network training is set to "rmsprop", the learning rate is set to 0.01, the batch size is set to 10, and the number of iterations is set to 30. The feature dimension vectors obtained in the previous step are input into the VAE network for training. After training, the encoder part of each VAE network is extracted as the long-term feature extractor of the overall multivariate temporal regression model.
[0030] Step 4: Train the multivariate time series regression model Input the corresponding dimensional data from multiple times before the current time point into the long time series feature extractor of the overall multivariate time series regression model to obtain long time series features; Based on the dimensional data and long-term time-series features corresponding to multiple monitoring units with a maximum information coefficient greater than the threshold, a multivariate time-series regression model is constructed. The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as input to the multivariate time series regression model. The multivariate time series regression model is trained by the least squares method to obtain the trained multivariate time series regression model. In this embodiment, flight data is input into a VAE encoder to extract long-term temporal features for each time point. The extraction method involves inputting the corresponding data from the previous 100 time points into each trained VAE encoder, outputting a 10-dimensional feature vector, which represents the long-term temporal features. Then, a multivariate time-series regression model is constructed. The formula for the multivariate time-series regression model is as follows: (6) In the formula, The dimension data of the monitoring unit to be detected at time T; These are the parameters for linear regression; The number of input data; To filter out the dimensional data of multiple monitoring units at time T whose maximum information coefficient is greater than a threshold; It is a 10-dimensional long-time series feature; and These are second-order difference data, specifically the data at time T minus the data at time T-1 and the data at time T minus the data at time T-2. Following the above input format, input the data collected during previous flights from the existing equipment of the aero-engine fuel control system. Since the probability of engine anomalies is low, the previously collected data contains far more normal data than abnormal data; therefore, no filtering is performed during training, ensuring that the model proposed in this patent predicts data under normal conditions. Finally, the model parameters in equation (6) are determined through least squares training.
[0031] Step 5: Anomaly detection in the engine fuel control system data The data to be detected is input into the trained multivariate time-series regression model to obtain the regression prediction results of the key feature dimensions of the fuel conditioning system at each time step. Anomalies in the data to be detected are determined based on the regression prediction results of the key feature dimensions of the fuel conditioning system at each time point.
[0032] Judgment of anomalies in the data to be tested includes: The regression prediction results of the key feature dimensions of the fuel conditioning system at each time point are subtracted from the actual values in the data to be tested to obtain the deviation at each time point. Time points where the deviation is greater than 10% of the actual value are marked as anomalies.
[0033] In this embodiment, data from multiple normal flight sorties are prepared to obtain a trained VAE encoder and a multivariate temporal regression model. The flight data to be detected is input into the trained multivariate temporal regression model to obtain the regression prediction results of the dimensional data corresponding to the fuel control system monitoring unit at each time point. This result is subtracted from the actual value in the flight data to be detected to obtain the deviation at each time point. Then, the data at time points where the deviation is greater than 10% of the actual value are marked as abnormal. If there are abnormal time points in a flight, it is considered an abnormal flight; otherwise, if there are no abnormal time points, it is considered a normal flight. This achieves anomaly detection for the aero-engine fuel control system.
[0034] To illustrate the effectiveness of the detection method provided by this invention, the following simulation experiment is used for further explanation.
[0035] 1. Simulation conditions This invention is a simulation performed using Python software on a CPU with an Intel® i5-3470 3.2GHz CPU, 8GB of memory, and a Windows 7 operating system.
[0036] The experiment used the NASA-provided C-MAPSS dataset for aero-turbine engines. The dataset consists of four parts: FD001, FD002, FD003, and FD004. FD001 contains complete records of 100 turbine engines under one environmental and failure condition, from start to finish. The simulation dataset was constructed by extracting normal data segments and manually adding anomalous data points.
[0037] 2. Simulation Content Anomaly detection experiments were conducted on the dataset. To compare the effectiveness of this invention, multilayer perceptron (MLP), support vector machine (SVM), relevance vector regression (RVR), convolutional neural network (CNN), and convolutional gated recurrent neural network (CNNGRU) were selected as comparison methods. The accuracy (ACC) was calculated, and the comparison results are shown in Table 1 (higher values are better, and the optimal values are indicated in bold).
[0038] Table 1 Comparison of Accuracy Rates in Experimental Results
[0039] The time taken for anomaly detection was calculated, and the comparison results are shown in Table 2 (the smaller the index, the better; the optimal value is marked in bold).
[0040] Table 2 Comparison of Experimental Time Consumption
[0041] As shown in Table 1, the anomaly detection performance of this method is superior to other comparative methods. As shown in Table 2, the processing time of this method is shorter than other comparative methods. The effectiveness of this invention can be verified through the above simulation experiments.
[0042] This invention provides an anomaly detection system for aero-engine fuel control system based on a multivariate time-series regression model, comprising: The data acquisition module is used to acquire dimensional data corresponding to multiple monitoring units during the operation of the aero-engine fuel control system; The data processing module is used to obtain the maximum information coefficient between each monitoring unit based on data from multiple dimensions; Based on the maximum information coefficient between each monitoring unit, multiple monitoring units with a maximum information coefficient greater than the threshold are selected. The network model training module is used to train the VAE network by using the dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold as the training set. After the training is completed, the encoder in the VAE network is obtained as the long-term feature extractor of the overall multivariate time-series regression model. Input the corresponding dimensional data from multiple times before the current time point into the long time series feature extractor of the overall multivariate time series regression model to obtain long time series features; Based on the dimensional data and long-term time-series features corresponding to multiple monitoring units with a maximum information coefficient greater than the threshold, a multivariate time-series regression model is constructed. The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as input to the multivariate time series regression model. The multivariate time series regression model is trained by the least squares method to obtain the trained multivariate time series regression model. The anomaly detection module is used to input the data to be detected into the trained multivariate time-series regression model to obtain the regression prediction results of the key feature dimensions of the fuel conditioning system at each time step. Anomalies in the data to be detected are determined based on the regression prediction results of the key feature dimensions of the fuel conditioning system at each time point.
[0043] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method and system for anomaly detection in aero-engine fuel control system based on a multivariate time-series regression model, characterized in that, include: Acquire dimensional data corresponding to multiple monitoring units during the operation of the aero-engine fuel control system; The maximum information coefficient between each monitoring unit is obtained based on data from multiple dimensions; Based on the maximum information coefficient between each monitoring unit, multiple monitoring units with a maximum information coefficient greater than the threshold are selected. The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as the training set to train the VAE network. After the training is completed, the encoder in the VAE network is obtained as the long-term feature extractor of the overall multivariate time-series regression model. Input the corresponding dimensional data from multiple times before the current time point into the long time series feature extractor of the overall multivariate time series regression model to obtain long time series features; Based on the dimensional data and long-term time-series features corresponding to multiple monitoring units with a maximum information coefficient greater than the threshold, a multivariate time-series regression model is constructed. The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as input to the multivariate time series regression model. The multivariate time series regression model is trained by the least squares method to obtain the trained multivariate time series regression model. The data to be detected is input into the trained multivariate time-series regression model to obtain the regression prediction results of the key feature dimensions of the fuel conditioning system at each time step. Anomalies in the data to be detected are judged based on the regression prediction results of the key feature dimensions of the fuel conditioning system at each time point; The multivariate time-series regression model has the following formula: In the formula, The dimension data of the monitoring unit to be detected at time T; These are the parameters for linear regression; The number of input data; To filter out the dimensional data of multiple monitoring units at time T whose maximum information coefficient is greater than a threshold; It is a 10-dimensional long-time series feature; and These are second-order difference data, specifically the data at time T minus the data at time T-1 and the data at time T minus the data at time T-2.
2. The method for anomaly detection in aero-engine fuel control system based on a multivariate time-series regression model according to claim 1, characterized in that, Judgment of anomalies in the data to be tested includes: The regression prediction results of the key feature dimensions of the fuel conditioning system at each time point are subtracted from the actual values in the data to be tested to obtain the deviation at each time point. Time points where the deviation is greater than 10% of the actual value are marked as anomalies.
3. The method for anomaly detection in aero-engine fuel control system based on a multivariate time-series regression model according to claim 1, characterized in that, During the training of the VAE network, the loss functions used include mean squared error and KL divergence between the sampling distribution and the normal distribution.
4. The method for anomaly detection in aero-engine fuel control system based on a multivariate time-series regression model according to claim 3, characterized in that, The KL divergence between the sampling distribution and the normal distribution is calculated using the following formula: In the formula, Dimensional data The mean obtained by sampling after passing through the VAE network; The variance is obtained by sampling the dimensional data after it has passed through a VAE network. This represents the KL divergence between the sampled distribution and the normal distribution.
5. The method for anomaly detection in aero-engine fuel control system based on a multivariate time-series regression model according to claim 1, characterized in that, The dimensional data refers to the data monitored by the monitoring unit during aircraft flight and recorded by the onboard data recorder.
6. The method for anomaly detection in aero-engine fuel control system based on a multivariate time-series regression model according to claim 1, characterized in that, During the training of the VAE network, the learning rate was set to 0.01, the batch size was set to 10, and the number of iterations was set to 30.
7. A system for anomaly detection in an aero-engine fuel control system based on a multivariate time-series regression model as described in claim 1, characterized in that, include: The data acquisition module is used to acquire dimensional data corresponding to multiple monitoring units during the operation of the aero-engine fuel control system; The data processing module is used to obtain the maximum information coefficient between each monitoring unit based on data from multiple dimensions; Based on the maximum information coefficient between each monitoring unit, multiple monitoring units with a maximum information coefficient greater than the threshold are selected. The network model training module is used to train the VAE network by using the dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold as the training set. After the training is completed, the encoder in the VAE network is obtained as the long-term feature extractor of the overall multivariate time-series regression model. Input the corresponding dimensional data from multiple times before the current time point into the long time series feature extractor of the overall multivariate time series regression model to obtain long time series features; Based on the dimensional data and long-term time-series features corresponding to multiple monitoring units with a maximum information coefficient greater than the threshold, a multivariate time-series regression model is constructed. The dimensional data corresponding to multiple monitoring units with the maximum information coefficient greater than the threshold are used as input to the multivariate time series regression model. The multivariate time series regression model is trained by the least squares method to obtain the trained multivariate time series regression model. The anomaly detection module is used to input the data to be detected into the trained multivariate time-series regression model to obtain the regression prediction results of the key feature dimensions of the fuel conditioning system at each time step. Anomalies in the data to be detected are determined based on the regression prediction results of the key feature dimensions of the fuel conditioning system at each time point.