Aviation compressor fault prediction method and system

Through the periodic multi-dimensional decoupling prediction framework and parallel feature extraction strategy, combined with CNN and Transformer, the traditional fault diagnosis method has solved the accuracy and real-time shortcomings, and achieved more efficient aviation compressor fault prediction, which is suitable for a variety of purposes Model aviation compressors reduce maintenance costs and improve aircraft reliability.

CN119004226AActive Publication Date: 2024-11-22NORTHWESTERN POLYTECHNICAL UNIV

Patent Information

Application Number
CN202411488442.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-24
Publication Date
2024-11-22
Estimated Expiration
2044-10-24

AI Technical Summary

Technical Problem

Traditional aeronautical compressor fault diagnosis methods rely on manual feature extraction and have low accuracy. The existing 1DCNN and RNN variants have insufficient processing of multi-scale fault frequency characteristics, resulting in insufficient accuracy and real-time failure prediction.

Method used

Using the periodic multidimensional decoupled prediction framework (PDMDF), periodic features are extracted through Fourier transform, multidimensional data is converted into two-dimensional tensors, and parallel multi-modal feature extraction strategy of convolutional neural network (CNN) and Transformer , High-order feature fusion is used to optimize the accuracy and reliability of fault prediction.

Benefits of technology

It improves the accuracy and real-time prediction of aviation compressor failures, provides a more intelligent maintenance solution, suitable for a variety of models of aviation compressors, reduces maintenance costs, and improves the reliability and availability of the aircraft.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004226A_ABST
    Figure CN119004226A_ABST
Patent Text Reader

Abstract

The invention discloses an aviation compressor fault prediction method and system, and belongs to the field of aviation compressor fault prediction. The method comprises the following steps: constructing multi-period multi-sensor data; periodic feature extraction and data decoupling: Fourier transform is performed on the multi-period multi-sensor two-dimensional sample data, and periodic features are extracted; decoupling the multi-period multi-sensor two-dimensional sample into short-term and long-term multi-period sequences based on periodic characteristics, and reconstructing the decoupled multi-period sequences into a two-dimensional tensor; multi-mode feature extraction and fusion: carrying out parallel feature extraction by using a two-dimensional convolutional neural network CNN and a Transform structure, and carrying out high-order fusion of multi-mode features by using a gating cycle unit GRU; and carrying out two-dimensional convolutional neural network optimization and fault prediction. According to the method, through comprehensive decoupling and efficient fusion of periodic features, the accuracy and real-time performance of aviation compressor fault prediction are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of aviation compressor fault prediction, and in particular relates to an aviation compressor fault prediction method and system. Background Art

[0002] With the continuous development of information technology, the future all-airspace, all-weather, multi-element mission mode has put forward higher performance requirements for aviation equipment. At present, the system structure of aircraft is becoming more and more complex, and the functional integration is getting higher and higher, which makes the failure mode of aircraft more diverse in harsh working environments, and the impact of fault propagation is also more complicated. The traditional passive maintenance guarantee mode based on "after-the-fact maintenance" and "planned maintenance" can no longer meet the modern demand for aircraft integrity. Prognostics and health management (PHM) based on condition-based maintenance is a key supporting technology for improving aircraft safety, reliability and economy. Its advantages include high degree of automation of fault diagnosis, good accuracy and timeliness, and the ability to predict faults, reduce maintenance costs, and improve the real-time reliability of equipment. In order to solve the shortcomings of traditional methods, the aviation compressor fault prediction method came into being. By using advanced data acquisition technology and machine learning algorithms, by analyzing compressor operation data and sensor information, fault signs can be found and predicted in advance, thereby realizing early identification and prevention of compressor failures.

[0003] At present, traditional fault diagnosis methods rely on expert experience for manual feature extraction, and the detection accuracy is low. In recent years, deep learning has been widely used, and it can extract deep feature information, which is very suitable for the field of fault diagnosis. Among many deep learning methods, convolutional neural network (CNN) has strong feature extraction ability and generalization due to its weight sharing and sparse connection characteristics, which can greatly improve the accuracy of fault diagnosis and has become an important solution for fault diagnosis in engineering. However, traditional CNN converts one-dimensional signals into image form as input, and the powerful feature extraction ability of CNN has not been fully utilized. There is still a certain amount of manual intervention. Therefore, many scholars have begun to study the end-to-end 1DCNN (one-dimensional convolutional neural network) fault diagnosis method, directly inputting the original fault vibration signal into 1DCNN to complete the recognition and classification. For example, the aircraft engine fault data is collected through multiple sensors, and then the data of multiple sensors is fused and input into 1DCNN to achieve a high-precision classification. However, these 1DCNN methods applied to aero-engine fault diagnosis do not take into account the multi-scale characteristics of aero-engine fault frequency features, and their ability to extract the time domain features of the original signals of aero-engine faults and the multi-scale features of the fault frequency domain is still insufficient. Therefore, the feature extraction capability of aero-engine faults needs to be strengthened.

[0004] In addition to CNN, commonly used methods include recurrent neural network (RNN) and its variants, long short-term memory neural network (LSTM), gated recurrent unit (GRU), etc. RNN is mainly used to process time series data with interdependencies. Due to the advantages of its network structure, it can retain the state information of the hidden layer. However, due to the limitations of its structure, RNN is not effective in processing long-term sequences. As variants of RNN, LSTM and GRU were proposed to solve the problem of long-term sequence prediction. The three gates of LSTM are used to control the flow of information, namely the input gate, forget gate and output gate. Compared with LSTM, GRU has a simpler structure and only contains update gates and reset gates. For example, the publicly available prediction model using a multi-layer GRU network structure has improved the prediction accuracy, but it also has problems such as long training time and low efficiency. Therefore, it is necessary to consider starting from the internal structure of GRU, improving its network structure, and improving the learning efficiency and performance of the network. Summary of the invention

[0005] Technical issues to be solved: In order to avoid the shortcomings of the prior art, the present invention provides an aviation compressor fault prediction method and system, which is an aviation compressor fault prediction method based on the periodic multi-dimensional decoupling prediction framework (PDMDF); the method constructs a multi-dimensional sample reflecting the time characteristics by integrating the data of multi-period and multi-sensors, thereby improving the accuracy of feature extraction; Fourier transform is used to extract periodic features, and multi-period decoupling technology is used to convert multi-dimensional data into a two-dimensional tensor, which is convenient for capturing key features related to aviation compressor faults; combined with the parallel multi-mode feature extraction strategy of convolutional neural network (CNN) and Transformer, key information in data can be mined from different time scales, and the accuracy and reliability of fault prediction are optimized through a high-order feature fusion module, and the Mish activation function is introduced to improve the learning process of the network. The present invention not only improves the accuracy and real-time performance of aviation compressor fault prediction through comprehensive decoupling and efficient fusion of periodic features, but also provides a more intelligent solution for aircraft engine maintenance.

[0006] The technical solution of the present invention is: a method for predicting failure of an aviation compressor, the specific steps of which are as follows: Multi-period multi-sensor data construction: The sensors are used to collect the time-series multi-dimensional data of the normal operation, abnormal state and fault occurrence of the compressor, and the collected data is pre-processed. Then, multi-period multi-sensor two-dimensional samples are constructed by stacking the data of multiple time periods. Periodic feature extraction and data decoupling: Perform Fourier transform on multi-period multi-sensor two-dimensional sample data to extract periodic features; decouple multi-period multi-sensor two-dimensional samples into short-term and long-term multi-period sequences based on periodic features, and reconstruct the decoupled multi-period sequences into two-dimensional tensors; Multi-modal feature extraction and fusion: Use two-dimensional convolutional neural network (CNN) and Transformer structure for parallel feature extraction, and then use gated recurrent unit (GRU) for high-level fusion of multi-modal features; 2D convolutional neural network optimization and fault prediction.

[0007] A further technical solution of the present invention is: the process of preprocessing the collected data includes: Apply low-pass filters or median filters to remove high-frequency noise from the collected data, and detect and remove abnormal data points through statistical methods; For missing data in the collected data, linear interpolation or spline interpolation is used to fill in the missing data to ensure the continuity of the data; The processed data are normalized to a uniform scale range by Min-Max standardization or Z-score standardization, which scales the data to a specified range or converts them to zero mean and unit standard deviation, respectively. A further technical solution of the present invention is: the method for constructing a multi-period multi-sensor two-dimensional sample is: Based on Fourier transform, the periodic features are extracted and the time period length is set, that is, a single two-dimensional sample is composed of data within the corresponding time period; The preprocessed data set is divided into multiple two-dimensional samples according to the set period length; All time periods are stacked in time sequence to obtain multiple two-dimensional samples arranged in time sequence, that is, multi-period multi-sensor two-dimensional samples are constructed; the multi-period multi-sensor two-dimensional samples contain local change characteristics of the time series and global characteristics within the time period.

[0008] A further technical solution of the present invention is: the method of extracting periodic features by Fourier transform is: For discrete time series X(t) Perform Fourier transform to convert the time domain signal to the frequency domain to analyze the frequency components of the signal. The formula is as follows:

[0009] in, F(w) Indicates frequency w The Fourier transform result is the spectrum of the signal; X(t) Represents the signal value at time t of the time series; N represents the total number of time points or samples; w represents the frequency;j represents an imaginary unit; The complex form of the exponential function is used to perform Fourier transform and convert the signal value in the time domain X(t) Mapping to frequency domain; Analyze the spectrum, mark frequency components with significant amplitudes, and calculate the corresponding period lengths:

[0010] in, T Represents the cycle, w Indicates frequency.

[0011] A further technical solution of the present invention is: the steps of multi-period multi-sensor two-dimensional sample decoupling and two-dimensional tensor reconstruction are: According to the extracted periodic features, the multi-period multi-sensor two-dimensional samples are decoupled into short-term sequences and long-term sequences, as shown below:

[0012] in, It is a short-term series, containing the fast-changing part within the cycle, and is suitable for capturing local short-term changes; It is a long-term series, including slowly changing parts across multiple periods, and is suitable for capturing global long-term trends; X represents the signal value in the frequency domain, mod is a mathematical operation symbol; Reconstruct the decoupled short-term and long-term sequences into a two-dimensional tensor. , the short-term series and long-term series Combined to form a two-dimensional tensor , indicating the changes within the period; each element in the two-dimensional tensor represents a specific time change of the data within a given period, and retains the time change characteristics of multi-period multi-sensor two-dimensional samples.

[0013] A further technical solution of the present invention is: the mapping method of the periodic features is: mapping each period in the two-dimensional tensor to a selected periodic feature space, retaining the changes within and between periods; during the feature mapping process, using the previously extracted periodic features as a reference, mapping the time dimension of each two-dimensional tensor to a frequency domain feature, ensuring that information on multi-periodic changes is retained. A further technical solution of the present invention is: the steps of multi-mode feature extraction and fusion are: Use 2D convolutional neural network (CNN) to extract short-term features: Input 2D samples of multi-period and multi-sensor into 2D convolutional neural network; Use multiple convolution kernels for convolution operation, and convolution kernels of different sizes are used to capture short-term change features of different scales; Introduce Activation function, which performs nonlinear activation processing on the convolutional feature map; performs pooling operation on the extracted short-term features to further reduce the data dimension and retain key features; Use the Transformer structure to extract long-term features: Input the data sequence processed by the two-dimensional convolutional neural network into the Transformer module; Use the self-attention mechanism to extract long-term features by calculating the dependency between each time step and other time steps; Use the multi-head attention mechanism in the Transformer structure to extract long-term dependencies from different angles, thereby enhancing the comprehensiveness of features; Linearly project the long-term features so that they have the same dimension as the short-term features, facilitating subsequent fusion operations; Linear projection of short-term features and long-term features: linearly project the extracted short-term features and long-term features respectively; unify the projected features into the same feature dimension so that the short-term features and long-term features can be fused in the same space; Multimodal feature fusion uses the gated recurrent unit (GRU): short-term features and long-term features after linear projection are input into the gated recurrent unit; the gated recurrent unit fuses features of different time scales through its update gate and reset gate mechanism, and gradually refines a comprehensive feature representation; at each time step, the gated recurrent unit updates the current feature according to the input feature and the state of the previous time step, and finally outputs the integrated feature sequence; the fused features are further processed to obtain the final feature representation for fault prediction.

[0014] A further technical solution of the present invention is: the steps of two-dimensional convolutional neural network optimization and fault prediction are: A two-dimensional convolutional neural network is constructed. The convolution kernels are set to different sizes of 3x3 and 5x5 to capture local features of different scales. The number of convolutional layers is designed to be 3 to extract features of different depths. Activation function; Use supervised learning to train a 2D convolutional neural network; Based on the experimental results of the validation set, the hyperparameters of the two-dimensional convolutional neural network are tuned to improve the performance of the two-dimensional convolutional neural network.

[0015] A further technical solution of the present invention is: the two-dimensional convolutional neural network introduces a cross entropy loss function and a domain adversarial loss, and uses the cross entropy loss function in the classification task to measure the difference between the predicted result and the actual category; the two-dimensional convolutional neural network is optimized by the domain adversarial loss to maintain consistency between different domains.

[0016] An aviation compressor fault prediction system includes a multi-period multi-sensor data construction module, a periodic feature extraction and data decoupling module, a multi-mode feature extraction and fusion module, and a two-dimensional convolutional neural network optimization and fault prediction module; The multi-period multi-sensor data construction module is used to collect the time-series multi-dimensional data of the normal operation, abnormal state and fault occurrence of the compressor, pre-process the collected data, and then construct a multi-period multi-sensor two-dimensional sample by stacking the data of multiple time periods; Through the periodic feature extraction and data decoupling module, the multi-period multi-sensor two-dimensional sample data is subjected to Fourier transform to extract periodic features; based on the periodic features, the multi-period multi-sensor two-dimensional samples are decoupled into short-term and long-term multi-period sequences, and the decoupled multi-period sequences are reconstructed into a two-dimensional tensor; The short-term and long-term features are extracted through the multi-modal feature extraction and fusion module, and the gated recurrent unit GRU is used for high-level fusion of multi-modal features. Compressor fault prediction is performed through the two-dimensional convolutional neural network optimization and fault prediction module, and the hyperparameters of the two-dimensional convolutional neural network are tuned to further improve the model performance.

[0017] Beneficial Effects The beneficial effects of the present invention are as follows: The method of the present invention combines the periodic feature extraction and multi-period decoupling technology of Fourier transform, as well as the parallel feature extraction strategy of convolutional neural network (CNN) and Transformer structure, effectively solving the shortcomings of traditional methods in terms of fault prediction accuracy and real-time performance. The specific advantages are as follows: 1. Introduce a periodic multi-dimensional decoupling framework to improve the accuracy of fault prediction.

[0018] The periodic multidimensional decoupling framework (PDMDF) is applied to aviation compressor fault prediction. The periodic features in multidimensional sensor data are identified through Fourier transform, and the original data are decomposed into multi-periodic short-term and long-term sequences and constructed as a two-dimensional tensor, which improves the model's ability to recognize and process periodic patterns.

[0019] 2. Use two-dimensional tensor reconstruction to improve the richness of feature extraction.

[0020] A method is proposed to reconstruct multi-period decoupled sequences into a two-dimensional tensor. The two-dimensional tensor retains more periodic information, enabling the model to simultaneously capture the temporal changes within and between periods, providing richer and more diverse feature representations.

[0021] 3. Parallel convolutional neural network and Transformer structure for multi-modal feature extraction.

[0022] Parallel convolutional neural networks (CNN) and Transformer structures are introduced to process short-term and long-term changes respectively, extract key features from different time scales, and improve the comprehensiveness of feature extraction.

[0023] 4. Integration of multi-sensor data: Comprehensive use of multi-dimensional sensor data, through periodic feature extraction and decoupling technology, provides comprehensive compressor operating status information, improving the comprehensiveness and reliability of fault prediction.

[0024] 5. Applicability and scalability: This method has good universality and scalability. It is applicable to various types of aviation compressors, provides advanced technical support for the intelligent maintenance of aircraft engines, and improves the accuracy and real-time performance of fault prediction.

[0025] The prediction method of the present invention is not only expected to achieve better performance in the field of aviation compressor fault diagnosis, but also can improve the accuracy and reliability of fault prediction, which will help airlines and maintenance teams take timely measures to reduce the impact of faults on flight safety and normal operation, while reducing maintenance costs and improving aircraft reliability and availability. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the overall framework of the model of the embodiment of the present invention; Figure 2 It is the overall architecture model based on the PDMDF method in the embodiment of the present invention; Figure 3 It is the periodic feature extraction and data decoupling in the embodiment of the present invention; Figure 4 This is the fault prediction application process of the embodiment of the present invention; Figure 5 The figure is a flow chart of the overall research scheme of the embodiment of the present invention. DETAILED DESCRIPTION

[0027] The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, but should not be construed as limiting the present invention.

[0028] Figure 2 The overall architecture model based on the periodic multi-dimensional decoupled prediction framework (PDMDF) of the present invention is presented. The model consists of multiple modules, including a data collection and preprocessing module, a periodic feature extraction and data decoupling module, a parallel multi-mode feature extraction and fusion module, and a final fault prediction module. These modules are connected in an orderly manner through data flow, and their specific functions include the collection and processing of multi-dimensional sensor data, the extraction and decomposition of periodic features, the multi-modal parallel processing of features, and the final generation of fault prediction results.

[0029] In order to more clearly illustrate the embodiments, technical solutions and advantages of the present invention, the following is a detailed and complete description of the aviation compressor fault prediction method based on the periodic multi-dimensional decoupled prediction framework (PDMDF) of the present invention in conjunction with the accompanying drawings. The present invention improves the accuracy and timeliness of fault prediction through periodic feature analysis and multi-mode feature fusion.

[0030] Direct modeling of time series faces complex pattern problems. These changes can be of various types, including short-term fluctuations, declines, and rises, which often overlap with each other. Therefore, the time series is decomposed into short-term changes and long-term changes for modeling. Due to the diversity of short-term variations (such as fluctuations, rises, and declines), the "frequency slicing" technique is used to divide the lookback window into multiple subsequences to correspond to different periods. For long-term variations, researchers use "period patching" to extract changes in the corresponding time periods in all periods.

[0031] 1. Multi-period multi-sensor data construction Acquire multi-dimensional sensor data of aviation compressors under normal and abnormal conditions, including but not limited to temperature, pressure ratio, flow rate, vibration spectrum, etc. By installing sensors, multi-dimensional data of the compressor is collected in real time, and the normal operation and fault occurrence process are covered in the data collection cycle; pre-process the multi-sensor data to ensure data quality; perform noise filtering to remove noise and outliers in the data; perform data completion to reasonably fill in missing values; and perform standardization to normalize the data to a unified scale range, such as [-1, 1]. Multi-period multi-sensor two-dimensional samples are constructed by stacking data from multiple time periods. Assuming there are N sensors and M periods of data, construct multi-sensor data with K periods for each two-dimensional sample, such as the first sample contains data from the 1st to Kth periods, and the second sample contains data from the 2nd to K+1th periods. And so on, a total of M-K+1 two-dimensional samples are generated.

[0032] In order to realize the aviation compressor fault prediction method based on the periodic multi-dimensional decoupled prediction framework (PDMDF), it is necessary to first construct a multi-dimensional sample containing multi-period information. This process includes data collection, data preprocessing and multi-period two-dimensional sample construction.

[0033] During the data collection phase, it is necessary to install various types of sensors at key locations of the aviation compressor to fully acquire operating data. These sensors include temperature sensors, pressure sensors, flow sensors, and vibration sensors, which are installed at blades, bearings, casings, inlets, and outlets to monitor relevant parameters in real time. Data collection should cover the entire process of normal operation, abnormal conditions, and failures of the compressor to ensure the comprehensiveness of the data. The sampling frequency of sensor data should be set to an appropriate value (such as 1,000 times per second) to ensure that rapidly changing dynamic features are captured. In addition, each data point is marked with a timestamp and the data is time-aligned to ensure the time synchronization of data from different sensors.

[0034] After data collection is completed, preprocessing is required to improve data quality. First, a low-pass filter or a median filter is applied to remove high-frequency noise in the data, and abnormal data points are detected and removed through statistical methods. For missing data, linear interpolation or spline interpolation can be used to fill in the gaps to ensure data continuity. Next, the data is normalized to a uniform scale range (such as [-1, 1]), which can be achieved through Min-Max normalization or Z-score normalization, which scales the data to a specified range or converts it to zero mean and unit standard deviation, respectively.

[0035] After completing the preprocessing, we enter the multi-period two-dimensional sample construction stage. First, based on the periodic features extracted by Fourier transform, determine the appropriate period length. For example, assuming that the data period is 10 seconds, each two-dimensional sample consists of data within 10 seconds. Then, the entire data set is divided into multiple samples according to the determined period length. For example, suppose we have 1 hour of data, the sampling frequency is 1 time per second, and a total of 3600 seconds of data are collected. If the period length is 10 seconds, each sample contains 10 seconds of data, that is, 10 rows, and 3500 samples will eventually be formed.

[0036] To construct these samples, the data from the 1st to the 10th second is used as the first sample, the data from the 2nd to the 11th second is used as the second sample, and so on. The final sample retains the change information within multiple periods. This multi-period two-dimensional sample not only contains the local change characteristics of the time series, but also retains the global characteristics within the period. The constructed samples will be stored in a dataset or database. The file format can be selected as CSV, HDF5 or SQL database to ensure the accessibility and scalability of the data.

[0037] The multi-period multi-sensor data constructed in this way provides a high-quality input basis for subsequent periodic feature extraction and data decoupling, ensuring the accuracy and reliability of the prediction model when dealing with aviation compressor failures.

[0038] 2. Periodic feature extraction and data decoupling The main periodic features in multidimensional data are extracted by Fourier transform, key frequencies and period lengths are identified, and the multidimensional sensor data is Fourier transformed and converted into frequency domain representation. The spectrum is analyzed to identify the main periods and frequencies. Multidimensional data is decoupled, and the data is decoupled into short-term and long-term multi-periodic sequences based on periodic features and reconstructed into two-dimensional tensors. Based on the identified periodic frequencies, the data is decomposed into short-term and long-term multi-periodic sequences. The decoupled sequences are reconstructed into two-dimensional tensors, each of which reflects the changes within a cycle.

[0039] 2.1 Fourier transform to extract periodic features In the present invention, the periodic feature extraction of multi-dimensional sensor data of an aviation compressor is carried out, and the formula of periodic feature mapping is derived. The discrete time series X(t) is subjected to Fourier transform to convert the time domain signal into the frequency domain in order to analyze the frequency component of the signal. The specific formula is as follows:

[0040] in, F(w) Indicates frequency w The Fourier transform result is the spectrum of the signal. X(t) Representing time series t The signal value at the time; N Indicates the total number of time points or samples; w Indicates frequency; j represents an imaginary unit; The complex form of the exponential function is used to perform Fourier transform and convert the signal value in the time domain X(t) Mapping to frequency domain; This formula is based on the Fourier transform result. Fourier transform is used to extract periodic features to identify the frequency components in the data. The periodic features are extracted by converting the time domain signal into the frequency domain signal. Combined with spectrum analysis, key periodic features are identified and mapped to optimize the input representation of the fault prediction model.

[0041] Analyze the spectrum, mark frequency components with significant amplitudes, and calculate the corresponding period lengths:

[0042] in, T Represents the cycle, w Indicates frequency.

[0043] 2.2 Multi-period multi-sensor 2D sample decoupling and 2D tensor reconstruction According to the extracted periodic features, the data decomposition formula is derived to decouple the multi-period multi-sensor two-dimensional samples into short-term series and long-term series:

[0044] in, It is a short-term series, containing the fast-changing part within the cycle, and is suitable for capturing local short-term changes; It is a long-term series, including slowly changing parts across multiple periods, and is suitable for capturing global long-term trends; X represents the signal value in the frequency domain, mod is a mathematical operation symbol; Reconstruct the decoupled short-term and long-term sequences into a two-dimensional tensor. , the short-term series and long-term series Combined to form a two-dimensional tensor , indicating the changes within the period; each element in the two-dimensional tensor represents a specific time change of the data within a given period, and retains the time change characteristics of multi-period multi-sensor two-dimensional samples.

[0045] 2.3 Mapping of periodic features: Periodic feature mapping is based on the periodic features obtained by Fourier transform, which are mapped to different time dimensions of the two-dimensional tensor to reflect the differences between different periods. Therefore, the relationship between the two is: Fourier transform provides periodic features, and periodic feature mapping maps time dimension data to a new feature space based on these features to ensure that multi-periodic changes are retained and reflected.

[0046] Each cycle in the two-dimensional tensor is mapped to a specific periodic feature space, retaining the changes within and between cycles. The specific periodic feature space should be based on the significant frequency components after Fourier transform. Specifically, by analyzing the Fourier transform results, the frequency components with the largest amplitude or those that play a key role in fault prediction can be selected; through spectrum analysis, these key periodic components can be screened out as a specific periodic feature space.

[0047] Feature mapping: Using the previously extracted periodic features as a reference, the time dimension of each 2D tensor is mapped into frequency domain features, ensuring that information about multi-periodic variations is preserved.

[0048] Through the above process, the multi-dimensional sensor data is effectively decoupled into short-term and long-term multi-periodic sequences and constructed into a two-dimensional tensor. The data processed in this way can fully reflect the characteristics of periodic changes and provide rich basic information for subsequent feature extraction and fault prediction.

[0049] Periodic feature extraction and data decoupling are used to separate short-term and long-term feature data and reconstruct them into two-dimensional tensors. This part provides the basis for subsequent multi-modal feature extraction and fusion. In other words, through the two-dimensional tensor after data decoupling, short-term features (through CNN) and long-term features (through Transformer) can be extracted separately, and then high-order fusion is performed through the gated recurrent unit (GRU). The periodic features and decoupled data extracted in step 2 provide input for multi-modal feature extraction and fusion. Short-term features are extracted from the decoupled short-term sequences, and long-term features are extracted from the long-term sequences. The two parts of the features are processed in parallel through different network structures (CNN and Transformer), and then combined and fused with GRU.

[0050] 3. Multimodal feature extraction and fusion Convolutional neural network (CNN) and Transformer structure are used for parallel feature extraction. Short-term features are extracted using two-dimensional convolutional network (CNN): convolution kernels of different sizes are used to mine local time-varying features in the data. Long-term feature extraction: long-term features are extracted using the Transformer structure, and the long-term dependency features of the data are captured through the self-attention mechanism; multi-modal feature fusion: the gated recurrent unit (GRU) is used to replace the traditional fully connected layer for high-order fusion of multi-modal features. Linearly project the short-term and long-term features extracted by CNN and Transformer. Feature fusion is performed through GRU to integrate multi-modal features at different levels and time scales; the specific process is as follows: Short-term feature extraction uses a two-dimensional convolutional neural network (CNN): Convolution kernels of different sizes are designed to extract short-term features. By using convolution kernels of different sizes (e.g., 3x3, 5x5), local variation features of data can be extracted at different temporal and spatial scales, thereby capturing rapid changes within a cycle.

[0051] Implementation steps: The two-dimensional samples of multi-period and multi-sensor are input into the two-dimensional convolutional neural network.

[0052] Use multiple convolution kernels to perform convolution operations. Convolution kernels of different sizes are used to capture short-term variation features of different scales.

[0053] Perform nonlinear activation processing on the convolutional feature map (introducing activation function) to improve the nonlinear expression ability.

[0054] The extracted short-term features are pooled to further reduce the data dimension and retain the key features.

[0055] Long-term feature extraction uses the Transformer structure: By introducing the self-attention mechanism in the Transformer structure, the inventors designed a module that captures the long-term dependencies of data. The Transformer structure can effectively process key features in long time series and extract long-term changes by calculating the dependencies between different time steps.

[0056] Implementation steps: The data sequence processed by the convolutional network is input into the Transformer module.

[0057] Using the self-attention mechanism, long-term features are extracted by calculating the dependencies between each time step and other time steps. This process can effectively capture long-term dependencies in multi-period data.

[0058] The multi-head attention mechanism in the Transformer structure can extract long-term dependencies from different angles, enhancing the comprehensiveness of features.

[0059] The long-term features are linearly projected to make them have the same dimension as the short-term features, which is convenient for subsequent fusion operations.

[0060] Linear projection of short-term and long-term features: After extracting short-term and long-term features, the two features need to be mapped to the same dimensional space. The inventors designed a linear projection layer to unify the features extracted by the convolutional neural network and the Transformer to facilitate subsequent feature fusion.

[0061] Linear projection is performed on the short-term features extracted from CNN and the long-term features extracted from Transformer respectively.

[0062] The projected features are unified into the same feature dimension so that short-term and long-term features can be fused in the same space.

[0063] Multimodal feature fusion using Gated Recurrent Unit (GRU): The design uses the gated recurrent unit (GRU) to perform high-order fusion of multi-modal features at different levels and time scales. Compared with traditional RNN, GRU has a simpler structure and more efficient feature fusion capability, and can better handle feature associations in time series.

[0064] Implementation steps: The short-term and long-term features after linear projection are input into the GRU unit.

[0065] GRU integrates features of different time scales through its update gate and reset gate mechanism, and gradually refines a comprehensive feature representation.

[0066] At each time step, GRU updates the current features based on the input features and the state of the previous time step, and finally outputs the integrated feature sequence.

[0067] The fused features are further processed to obtain the final feature representation for fault prediction.

[0068] Network optimization and prediction: Construct a two-dimensional convolutional neural network; Convolution kernel design: The convolution kernel size can be set to 3x3 and 5x5 convolution kernels of different sizes to capture local features of different scales. Layer design: Three convolution layers are designed to extract features of different depths. Activation function: Introduce The activation function replaces the traditional The activation function aims to improve the nonlinear expression ability and learning efficiency of the network.

[0069] Use supervised learning to train a 2D convolutional neural network; Based on the experimental results of the validation set, the hyperparameters of the two-dimensional convolutional neural network are tuned to further improve the performance of the two-dimensional convolutional neural network.

[0070] By introducing Activation function optimizes the learning efficiency and performance of the network. In the process of feature extraction and fusion, application Activation function improves gradient flow and enhances the nonlinear expression ability of the network; uses the fused features for fault prediction and integrates all feature information to generate prediction results, outputs the fault prediction results, and compares them with the actual values ​​to verify the effectiveness of the model.

[0071] Model training method: Multi-dimensional sensor data collection 1) Sensor deployment: Temperature sensors, pressure sensors, flow sensors and vibration sensors are deployed at key parts of the aircraft compressor (such as blades, bearings, shafts, etc.) to ensure real-time collection of multi-dimensional data.

[0072] 2) Data collection frequency: Set an appropriate sampling frequency (such as 1000 times per second) to capture high-frequency dynamic changes. The time range of collected data should cover the entire process of normal operation, abnormal fluctuations and faults.

[0073] 3) Data synchronization and marking: The consistency of sensor data is ensured through timestamp and synchronization mechanism. All data points must be accurately marked with the collection time to ensure the accuracy of data analysis.

[0074] Data preprocessing Noise Filtering: 1) Bandpass filter: Use a bandpass filter to remove high-frequency and low-frequency noise from the data. The specific settings are: 2) Highpass cutoff frequency: Set to the lowest frequency in the signal spectrum, usually slightly above zero.

[0075] 3) Low-pass cutoff frequency: Set above the frequency of most interest in the signal, but below any noise interference frequencies.

[0076] 4) Resonance Signal Decomposition (RSSD): Perform resonance signal decomposition on the data to divide the signal into resonant components and non-resonant components, and retain the resonant components to enhance the fault characteristics.

[0077] Missing value filling: Linear interpolation: For missing values ​​in time series, linear interpolation is used to fill them to ensure the continuity of the data.

[0078] Data Standardization: Min-Max Normalization: Normalize all data to the range of [-1, 1] to eliminate the impact of dimension.

[0079] Time Alignment: Data alignment and resampling: Align the sensor data by timestamp to ensure that all data points are analyzed on the same time scale. Resample the data if necessary to unify the time step of the data.

[0080] Specific training steps Supervised learning and model optimization: 1. Model initialization and settings: Network structure setup: Initialize parallel convolutional neural network (CNN) and Transformer modules for extracting short-term and long-term features, respectively. Set the convolution kernel size (e.g., 3x3, 5x5) and number of layers (e.g., 3 convolutional layers), as well as the number of attention heads and layers in the Transformer (e.g., 4 attention heads, 2 layers of Transformer encoder).

[0081] 2. Loss function and evaluation indicators: Cross-entropy loss function: The cross-entropy loss function is used in classification tasks to measure the difference between the predicted results and the actual categories: Domain adversarial loss: In order to enhance domain adaptability, domain adversarial training is introduced into the model. Setting the domain discriminator , and optimize the model through domain adversarial loss to maintain consistency across different domains (e.g., different equipment and working conditions): 3. Training strategy and optimization algorithm: 5-fold cross validation: Divide the dataset into 5 subsets, use 4 of them for training each time, and use the 5th subset for validation. Repeat this cycle 5 times, updating the model parameters each time. Calculate the average of the 5 validation results to get the overall performance of the model.

[0082] Adam optimizer: The Adam optimizer is used for parameter optimization, and the initial learning rate is set to $0.001$ and automatically adjusted according to the model performance. The specific parameters of the optimizer are set to $\beta_1 = 0.9, \beta_2 = 0.999$, and $\epsilon = 1e-8$ is used to prevent division by zero errors.

[0083] Learning rate adjustment: During training, use a learning rate scheduler (such as ) Dynamically adjust the learning rate based on the validation set loss. When the validation set loss stops decreasing, the learning rate is reduced proportionally.

[0084] Model training and iteration Training iteration: In each training iteration, batches of data are fed into the model, predictions are calculated through forward propagation, and loss values ​​are calculated through the loss function. The model parameters are updated using the backpropagation algorithm to minimize the loss function.

[0085] Domain adversarial training: In each iteration, the classifier loss is optimized alternately to ensure that the model performs well in both fault diagnosis and domain adaptation tasks.

[0086] Model Validation 1. Construction and use of validation set: Validation set division: When dividing the training set, retain 20% of the data as an independent validation set for real-time evaluation of model performance. The validation set should cover a variety of fault conditions to test the comprehensiveness and reliability of the model.

[0087] Evaluation indicators: The model performance was evaluated using the following metrics: 1) Confusion Matrix: Used to evaluate the accuracy in classification tasks, showing the number of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN).

[0088] 2) ROC curve and AUC value: The ROC curve shows the sensitivity (True Positive Rate, TPR) and specificity (False Positive Rate, FPR) of the model at different thresholds. The higher the AUC value, the better the model performance.

[0089] 3) Accuracy and Recall: Accuracy measures the overall correct classification ability of the model, and recall measures the model's ability to identify the positive class.

[0090] 2. Experimental results and model tuning: Experimental results: Experiments were conducted on the validation set, and the classification accuracy, AUC value, and other performance indicators of the model on different fault types were recorded. The experimental results show that the PDMDF method of the present invention has achieved a classification accuracy of more than 98% in various fault modes, especially when processing high-noise data and small sample data.

[0091] Model tuning: Based on the experimental results, we tuned the model's hyperparameters (such as adjusting the convolution kernel size, number of Transformer layers, learning rate, etc.) to further improve the model's performance. We continuously optimized the model through experiments on the validation set to ensure its efficiency and robustness in practical applications.

[0092] Experimental analysis During the experiment, NASA’s CMAPSS dataset was used to compare the PDMDF model with various existing fault prediction models (including SVM, LSTM, CNN, and The experimental results are mainly evaluated by two indicators: root mean square error (RMSE) and score function (Score).

[0093] Table 1 Experimental analysis and comparison

[0094] The RMSE and Score values ​​of the PDMDF model on the test set are 5.2 and 45.8 respectively, which are significantly better than other models. This shows that PDMDF has higher efficiency and accuracy in periodic feature extraction and feature fusion.

[0095] The SVM model performed poorly, with an RMSE of 8.3 and a Score of 72.5. Especially when dealing with complex time series data, its prediction accuracy was low.

[0096] The LSTM model performs better in processing time dependency and nonlinear relationships, but is still inferior to the PDMDF model in extracting periodic features, with an RMSE of 6.5 and a Score of 58.3.

[0097] The CNN model performs slightly better than the SVM, but due to its limited ability to extract periodic features, the RMSE is 7.0 and the Score is 63.7.

[0098] The model shows certain advantages through multi-scale feature processing, with an RMSE of 5.9 and a Score of 49.2, but it is still inferior to the PDMDF model in effectively integrating periodic features.

[0099] The present embodiment provides an aviation compressor fault prediction system, including a multi-period multi-sensor data construction module, a periodic feature extraction and data decoupling module, a multi-mode feature extraction and fusion module, and a two-dimensional convolutional neural network optimization and fault prediction module; the multi-period multi-sensor data construction module is used to collect time-series multi-dimensional data of normal operation, abnormal state and fault occurrence of the compressor, the collected data is preprocessed, and then a multi-period multi-sensor two-dimensional sample is constructed by stacking data of multiple time periods; the multi-period multi-sensor two-dimensional sample data is Fourier transformed by the periodic feature extraction and data decoupling module to extract periodic features; the multi-period multi-sensor two-dimensional sample data is decoupled into short-term and long-term multi-period sequences based on the periodic features, and the decoupled multi-period sequences are reconstructed into a two-dimensional tensor; the short-term features and long-term features are extracted by the multi-mode feature extraction and fusion module, and then the gated recurrent unit GRU is used to perform high-order fusion of multi-mode features; the compressor fault is predicted by the two-dimensional convolutional neural network optimization and fault prediction module, and the two-dimensional convolutional neural network is hyperparameter tuned to further improve the model performance.

[0100] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention without departing from the principles and intent of the present invention.

Claims

1. A method for predicting failure of an aviation compressor, characterized in that The specific steps are as follows: Multi-period multi-sensor data construction: The sensors are used to collect the time-series multi-dimensional data of the normal operation, abnormal state and fault occurrence of the compressor, and the collected data is pre-processed. Then, multi-period multi-sensor two-dimensional samples are constructed by stacking the data of multiple time periods. Periodic feature extraction and data decoupling: Fourier transform the multi-period multi-sensor two-dimensional sample data to extract periodic features; Based on periodic features, multi-period multi-sensor two-dimensional samples are decoupled into short-term and long-term multi-period sequences, and the decoupled multi-period sequences are reconstructed into a two-dimensional tensor. Multi-modal feature extraction and fusion: Use two-dimensional convolutional neural network (CNN) and Transformer structure for parallel feature extraction, and then use gated recurrent unit (GRU) for high-level fusion of multi-modal features; 2D convolutional neural network optimization and fault prediction.

2. The method for predicting aviation compressor failure according to claim 1, characterized in that: The process of preprocessing the collected data includes: Apply low-pass filters or median filters to remove high-frequency noise from the collected data, and detect and remove abnormal data points through statistical methods; For missing data in the collected data, linear interpolation or spline interpolation is used to fill in the missing data to ensure the continuity of the data; The processed data are normalized to a uniform scale range by Min-Max standardization or Z-score standardization, which scales the data to a specified range or converts them to zero mean and unit standard deviation, respectively.

3. The method for predicting aviation compressor failure according to claim 2, characterized in that: The method for constructing a multi-period multi-sensor two-dimensional sample is: Based on Fourier transform, the periodic features are extracted and the time period length is set, that is, a single two-dimensional sample is composed of data within the corresponding time period; The preprocessed data set is divided into multiple two-dimensional samples according to the set period length; All time periods are stacked in time sequence to obtain multiple two-dimensional samples arranged in time sequence, that is, multi-period multi-sensor two-dimensional samples are constructed; the multi-period multi-sensor two-dimensional samples contain local change characteristics of the time series and global characteristics within the time period.

4. The method for predicting aviation compressor failure according to claim 3, characterized in that: The method of extracting periodic features by Fourier transform is: For discrete time series X(t) Perform Fourier transform to convert the time domain signal to the frequency domain to analyze the frequency components of the signal. The formula is as follows: in, F(w) Indicates frequency w The Fourier transform result is the spectrum of the signal; X(t) Represents the signal value at time t of the time series; N represents the total number of time points or samples; w represents the frequency; j represents an imaginary unit; The complex form of the exponential function is used to perform Fourier transform and convert the signal value in the time domain X(t) Mapping to frequency domain; Analyze the spectrum, mark frequency components with significant amplitudes, and calculate the corresponding period lengths: in, T Represents the cycle, w Indicates frequency.

5. The method for predicting aviation compressor failure according to claim 4, characterized in that: The steps of multi-period multi-sensor two-dimensional sample decoupling and two-dimensional tensor reconstruction are: According to the extracted periodic features, the multi-period multi-sensor two-dimensional samples are decoupled into short-term sequences and long-term sequences, as shown below: in, It is a short-term series, containing the fast-changing part within the cycle, and is suitable for capturing local short-term changes; It is a long-term series, including slowly changing parts across multiple periods, and is suitable for capturing global long-term trends; X represents the signal value in the frequency domain, mod is a mathematical operation symbol; Reconstruct the decoupled short-term and long-term sequences into a two-dimensional tensor. , the short-term series and long-term series Combined to form a two-dimensional tensor , indicating the changes within the period; each element in the two-dimensional tensor represents a specific time change of the data within a given period, and retains the time change characteristics of multi-period multi-sensor two-dimensional samples.

6. The method for predicting aviation compressor failure according to claim 5, characterized in that: The mapping method of the periodic features is as follows: each period in the two-dimensional tensor is mapped to a selected periodic feature space, and the changes within and between periods are retained; during the feature mapping process, the previously extracted periodic features are used as a reference to map the time dimension of each two-dimensional tensor to a frequency domain feature, ensuring that the information of multi-periodic changes is retained.

7. The method for predicting aviation compressor failure according to claim 6, characterized in that: The steps of multi-mode feature extraction and fusion are: Use 2D convolutional neural network (CNN) to extract short-term features: Input 2D samples of multi-period and multi-sensor into 2D convolutional neural network; Use multiple convolution kernels for convolution operation, and convolution kernels of different sizes are used to capture short-term change features of different scales; Introduce Activation function, which performs nonlinear activation processing on the convolutional feature map; performs pooling operation on the extracted short-term features to further reduce the data dimension and retain key features; Use the Transformer structure to extract long-term features: Input the data sequence processed by the two-dimensional convolutional neural network into the Transformer module; Use the self-attention mechanism to extract long-term features by calculating the dependency between each time step and other time steps; Use the multi-head attention mechanism in the Transformer structure to extract long-term dependencies from different angles, thereby enhancing the comprehensiveness of features; Linearly project the long-term features so that they have the same dimension as the short-term features, facilitating subsequent fusion operations; Linear projection of short-term features and long-term features: linearly project the extracted short-term features and long-term features respectively; unify the projected features into the same feature dimension so that the short-term features and long-term features can be fused in the same space; Multimodal feature fusion uses the gated recurrent unit (GRU): short-term features and long-term features after linear projection are input into the gated recurrent unit; the gated recurrent unit fuses features of different time scales through its update gate and reset gate mechanism, and gradually refines a comprehensive feature representation; at each time step, the gated recurrent unit updates the current feature according to the input feature and the state of the previous time step, and finally outputs the integrated feature sequence; the fused features are further processed to obtain the final feature representation for fault prediction.

8. The method for predicting aviation compressor failure according to claim 7, characterized in that: The steps of two-dimensional convolutional neural network optimization and fault prediction are: A two-dimensional convolutional neural network is constructed. The convolution kernels are set to different sizes of 3x3 and 5x5 to capture local features of different scales. The number of convolutional layers is designed to be 3 to extract features of different depths. Activation function; Use supervised learning to train a two-dimensional convolutional neural network; Based on the experimental results of the validation set, the hyperparameters of the two-dimensional convolutional neural network are tuned to improve the performance of the two-dimensional convolutional neural network.

9. The method for predicting aviation compressor failure according to claim 8, characterized in that: The two-dimensional convolutional neural network introduces a cross entropy loss function and a domain adversarial loss. The cross entropy loss function is used in classification tasks to measure the difference between the predicted results and the actual categories. The two-dimensional convolutional neural network is optimized by the domain adversarial loss to maintain consistency between different domains.

10. An aviation compressor fault prediction system, used to implement the aviation compressor fault prediction method according to any one of claims 1 to 9, characterized in that: It includes multi-period multi-sensor data construction module, periodic feature extraction and data decoupling module, multi-mode feature extraction and fusion module, and two-dimensional convolutional neural network optimization and fault prediction module; The multi-period multi-sensor data construction module is used to collect the time-series multi-dimensional data of the normal operation, abnormal state and fault occurrence of the compressor, pre-process the collected data, and then construct a multi-period multi-sensor two-dimensional sample by stacking the data of multiple time periods; Through the periodic feature extraction and data decoupling module, the multi-period multi-sensor two-dimensional sample data is Fourier transformed to extract the periodic features; Based on periodic features, multi-period multi-sensor two-dimensional samples are decoupled into short-term and long-term multi-period sequences, and the decoupled multi-period sequences are reconstructed into a two-dimensional tensor. The short-term and long-term features are extracted through the multi-modal feature extraction and fusion module, and the gated recurrent unit GRU is used for high-level fusion of multi-modal features. Compressor fault prediction is performed through the two-dimensional convolutional neural network optimization and fault prediction module, and the hyperparameters of the two-dimensional convolutional neural network are tuned to further improve the model performance.

Citation Information

Patent Citations

  • Depth calculation model for aero-engine gas circuit fault diagnosis

    CN110321603A

  • Power load prediction method and device and medium

    CN118199054A

  • Complex equipment fault diagnosis system and diagnosis method based on multivariate data fusion

    CN118468165A

  • Fault prediction model training method, fault prediction method and related device

    CN118520287A

  • Pattern detection in time-series data

    US20190379589A1

Cited By

  • Power distribution network fault prediction method and system

    CN119644038A

  • Transform-based air compressor fault prediction method and system

    CN119885068A

  • Tap water loss detection method and device based on Internet of Things

    CN119939228A

  • Tap Water Loss Detection Method and Device Based on Internet of Things

    CN119939228B

  • Intelligent modularized prefabricated cabin type tunnel fire station terminal dynamic state based on edge calculation and multi-source perception

    CN120977063A