A method for aircraft engine fault diagnosis based on multimodal sensors

Abstract

The present invention discloses an aircraft engine fault diagnosis method based on a multimodal sensor, comprising obtaining multimodal data of an aircraft engine; constructing a fault diagnosis model based on an interactive perception graph attention network, extracting spatial features of adjacent time steps of each signal from the multimodal data of the aircraft engine to construct a fully connected heterogeneous graph, and performing graph node screening on the fully connected heterogeneous graph and extracting spatial interaction relationship features between graph nodes to perform model training; and performing aircraft engine fault diagnosis using the trained fault diagnosis model based on the interactive perception graph attention network. The present invention improves the accuracy and robustness of aircraft engine fault diagnosis by converting the multimodal signal of an aircraft engine into a fully connected heterogeneous graph, and using an end-to-end neural network to directly screen sensors and mine the spatial interaction relationship between them.

Description

Technical Field

[0001] The invention relates to aircraft engine fault diagnosis, and in particular to an aircraft engine fault diagnosis method based on a multi-modal sensor. Background Art

[0002] Fault diagnosis technology is widely used in modern aviation systems, such as flight control systems and engine health management systems. As the core of the aerospace power system, the engine is often in extreme working conditions such as high temperature, high pressure, and strong vibration. This means that a small fault in any part of the engine may be transmitted to the entire system, causing huge losses or even threats to the economy and personnel safety. Therefore, it is crucial to accurately and timely detect operational faults in aerospace engines. This technology has important practical value and economic significance in reality. Existing technologies can collect data through different sensor networks under experimental conditions that simulate actual environments, such as rotation, low-pressure to high-pressure cycle replacement, and high temperature. Algorithm models such as whale algorithms and genetic algorithms are used to extract, fuse, aggregate, and optimize data, and models based on different fault abnormal conditions are established for fault diagnosis.

[0003] The operation anomaly detection methods for aerospace engines can be divided into three categories: model-driven, signal processing-based, and data-driven. The model-driven method requires the establishment of a dynamic model of the engine, but due to the complex structure and working conditions of the aerospace engine, this method is difficult to model and has low accuracy. The signal processing-based method mainly performs statistical analysis on the signals generated during the engine operation process, without relying on prior knowledge such as engine structure, but relies on expert experience and cannot extract deep features of the signal for diagnosis. Based on artificial intelligence, especially the rapid development of deep learning in recent years, the anomaly detection method can get rid of artificial experience and build a deep network to adaptively learn the inherent deep nonlinear features of the data. This method has been widely used in the field of mechanical fault diagnosis. Traditional fault diagnosis methods mainly focus on analyzing a single signal. However, a single signal only conveys part of the information related to the complex mechanical fault state and cannot fully monitor the degradation state of the complex machinery. Therefore, using multiple sensors to collect aerospace engine data has become a solution. Multiple signals provide more information resources, but how to fuse multi-sensor data to improve the accuracy of fault diagnosis has become a technical challenge. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an aircraft engine fault diagnosis method based on a multimodal sensor, which can diagnose aircraft engine faults based on multimodal signals and improve the accuracy of fault diagnosis.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] A method for diagnosing aircraft engine faults based on a multimodal sensor comprises the following steps:

[0007] S1. Acquire multimodal data of aircraft engines;

[0008] S2. Construct a fault diagnosis model based on interactive perception graph attention network, extract the spatial features of adjacent time steps of each signal from the multimodal data of aircraft engines to construct a fully connected heterogeneous graph, and perform graph node screening on the fully connected heterogeneous graph and extract the spatial interaction relationship features between graph nodes for model training;

[0009] S3. Use the trained fault diagnosis model based on interactive perceptual graph attention network to perform aircraft engine fault diagnosis.

[0010] Furthermore, step S1 specifically includes:

[0011] Multi-modal sensors are used to collect multi-modal data of aircraft engines at the fault detection locations of aircraft engines;

[0012] The fault detection parts of the aircraft engine include a fan, a low-pressure compressor, a high-pressure compressor, a combustion chamber or a burner, a high-pressure turbine and a low-pressure turbine.

[0013] Furthermore, the multimodal data of the aircraft engine specifically include: total temperature at the fan inlet, total temperature at the low-pressure compressor outlet, total temperature at the high-pressure compressor outlet, total temperature at the low-pressure turbine outlet, pressure at the fan inlet, total pressure of the bypass duct, total pressure at the high-pressure compressor outlet, physical fan speed, physical core speed, engine pressure ratio, static pressure at the high-pressure compressor outlet, ratio of fuel flow to total pressure of the bypass duct, corrected fan speed, corrected core speed, bypass ratio, burner fuel-air ratio, exhaust enthalpy, required fan speed, required corrected fan speed, coolant flow rate and high-pressure turbine coolant flow rate.

[0014] Furthermore, step S1 also includes the following steps: sequentially performing data cleaning, data normalization, sliding window sampling and feature fusion on the acquired multimodal data of the aircraft engine.

[0015] Furthermore, sliding window sampling of the multimodal data of the aircraft engine specifically includes the following steps:

[0016] A1. Determine the window size and step size according to the sampling data length of the multi-modal data of the aircraft engine;

[0017] A2. Start sliding the window from the beginning of the sampled data sequence, and take out subsequences in sequence according to the window size and step size until the end of the sequence is reached;

[0018] A3. Generate corresponding sample data based on the subsequence taken out from the sliding window.

[0019] Furthermore, the fault diagnosis model based on the interactive perception graph attention network specifically includes: a feature extraction layer, a graph encoding layer and a graph classification layer connected in sequence;

[0020] The feature extraction layer is used to respectively extract the spatial features of adjacent time steps of each signal in the multimodal data using a gated convolutional layer to construct a fully connected heterogeneous graph;

[0021] The graph encoding layer is used to adaptively learn heterogeneous graph edge weights using a graph attention mechanism, and to filter heterogeneous graph nodes using degree centrality;

[0022] The graph classification layer is used to predict category probabilities based on the filtered graph connections.

[0023] Furthermore, the feature extraction layer specifically includes: a plurality of parallel gated convolutional layers; the gated convolutional layer includes a convolutional layer, a gated mechanism layer, a linear mapping layer and a position encoding layer connected in sequence;

[0024] The convolution layer is used to extract local feature information of the signal time series corresponding to the window size at the time step;

[0025] The gating mechanism layer is used to use a reset gate to re-read the original input information after the convolution layer performs the convolution operation, and use an update gate to use local context information to control the amplitude of the extracted features;

[0026] The linear mapping layer is used to adapt the dimension of the extracted local feature information to the dimension of the encoder layer;

[0027] The encoder layer is used to add position encoding to the local features after linear mapping and construct a fully connected heterogeneous graph.

[0028] Furthermore, the graph encoding layer specifically includes: a graph structure interaction perception layer and a multi-level cascaded graph attention layer and a graph pooling layer, and each level of the graph pooling layer is connected to a readout layer;

[0029] The graph structure interaction perception layer is used to calculate the relationship scores between graph nodes according to the input fully connected heterogeneous graph, and to filter the edges between graph nodes according to the relationship scores between graph nodes;

[0030] The graph attention layer is used to calculate the attention coefficient of each graph node based on the original features of the graph node and the features of its neighborhood, and update the node features in combination with the attention mechanism;

[0031] The graph pooling layer is used to filter graph nodes using Topk pooling;

[0032] The readout layer uses a cascade of average or cumulative pooling functions for output, and performs feature fusion on the output features of each readout layer.

[0033] Furthermore, the use of Topk pooling to screen graph nodes is specifically as follows:

[0034]

[0035] Among them, X ′ represents the node feature matrix after the update; X represents the node feature matrix before the update; ⊙ represents the matrix element multiplication; A ′ represents the updated adjacency matrix; A i,i represents the adjacency matrix formed after extracting the corresponding i-th row and i-th column; i represents the extraction index; z represents the importance matrix; top-k(·) represents the k-rank mechanism, k represents the pooling ratio; d(v i ) represents node v i The degree centrality of A i,j Represents node v i The adjacency matrix of .

[0036] Furthermore, the graph classification layer specifically includes: a first fully connected layer, a second fully connected layer and a softmax classification layer which are connected in sequence.

[0037] Compared with the prior art, the aircraft engine fault diagnosis method based on multimodal sensor provided by the present invention has the following advantages:

[0038] Beneficial effects:

[0039] The present invention improves the accuracy and robustness of aircraft engine fault diagnosis by converting the multimodal signals of aircraft engines into a fully connected heterogeneous graph and using an end-to-end neural network to directly screen sensors and mine the spatial interaction relationship between them. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic flow chart of an aircraft engine fault diagnosis method based on a multimodal sensor provided in an embodiment of the present invention;

[0041] Figure 2 This is a layout diagram of an aircraft engine multi-modal sensor in an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the structure of a fault diagnosis model based on an interactive perception graph attention network in an embodiment of the present invention;

[0043] Figure 4 This is a normal state monitoring diagram of an aircraft engine in an embodiment of the present invention;

[0044] Figure 5 This is a fault status monitoring diagram of an aircraft engine in an embodiment of the present invention;

[0045] Figure 6 It is a schematic diagram of the structure of the aircraft engine sensor after screening in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0047] In actual anomaly detection, the diversity of engine failure modes leads to difficulties in fault simulation, which hinders detection methods based on data-driven methods such as signal processing and deep learning. In addition, the health status of aerospace engines is monitored by multiple sensors, but traditional methods usually analyze each data source independently. At the same time, it is difficult to screen the effectiveness of multi-sensing when prior knowledge is insufficient, and it is also impossible to effectively mine the spatial interaction relationship between sensors. It is difficult to comprehensively evaluate the true health status of the engine, and the accuracy of anomaly detection is not high. Therefore, the present invention provides a fault diagnosis method for aerospace engines based on a multimodal interactive perception graph attention network, which converts the remaining life estimation problem of multi-sensor monitoring equipment into a time series graph prediction problem. The core idea of ​​the method is to convert multiple signals into a fully connected heterogeneous graph, and use an end-to-end neural network to directly screen sensors and mine the spatial interaction relationship between them. This method can realize sensor screening without prior knowledge, and mine the spatial interaction relationship between sensors, so as to delete the nodes that have no effect or even a negative effect on fault diagnosis to realize sensor screening. Through the above work, the spatial interaction between multiple sensors is fully utilized, and the sparsification process is also completed, thereby improving the accuracy and robustness of the network in fault diagnosis.

[0048] like Figure 1 As shown, the aircraft engine fault diagnosis method based on a multimodal sensor provided by an embodiment of the present invention includes the following steps S1 to S3:

[0049] S1. Acquire multimodal data of aircraft engines;

[0050] In an optional embodiment of the present invention, step S1 specifically includes:

[0051] The multi-modal sensor is used to collect the multi-modal data of the aircraft engine for the fault detection parts of the aircraft engine; the fault detection parts of the aircraft engine include the fan, low-pressure compressor, high-pressure compressor, combustion chamber or burner, high-pressure turbine and low-pressure turbine. Specifically, Figure 2 As shown, in this embodiment, multi-modal sensors are arranged at the six main components of the aircraft engine, namely, the fan, the low-pressure compressor, the high-pressure compressor, the combustion chamber or the burner, the high-pressure turbine and the low-pressure turbine, to obtain corresponding multi-modal signals, and respectively collect 21 signals including the total temperature at the fan inlet, the total temperature at the low-pressure compressor outlet, the total temperature at the high-pressure compressor outlet, the total temperature at the low-pressure turbine outlet, the pressure at the fan inlet, the total pressure of the bypass duct, the total pressure at the high-pressure compressor outlet, the physical fan speed, the physical core speed, the engine pressure ratio, the static pressure at the high-pressure compressor outlet, the ratio of the fuel flow rate to the total pressure of the bypass duct, the corrected fan speed, the corrected core speed, the bypass ratio, the burner fuel-air ratio, the exhaust enthalpy value, the required fan speed, the required corrected fan speed, the coolant flow rate, the high-pressure turbine coolant flow rate, etc. Because the ratio of the engine speed to the aircraft movement speed is different at different wind speeds. According to the target wind speed and the actual wind speed, the physical fan speed and the physical core speed are corrected to obtain the corrected fan speed and the corrected core speed. Required fan speed, required corrected fan speed - the fan speed required to achieve the target thrust and the corrected fan speed. Corrected fan (core) speed = physical fan (core) speed * (target wind speed / actual wind speed).

[0052] In an optional embodiment of the present invention, step S1 further includes the following steps:

[0053] The acquired multimodal data of aircraft engines are sequentially subjected to data cleaning, data normalization, sliding window sampling and feature fusion.

[0054] The sliding window sampling of the multimodal data of the aircraft engine specifically includes the following steps:

[0055] A1. Determine the window size and step size according to the sampling data length of the multi-modal data of the aircraft engine;

[0056] A2. Start sliding the window from the beginning of the sampled data sequence, and take out subsequences in sequence according to the window size and step size until the end of the sequence is reached;

[0057] A3. Generate corresponding sample data based on the subsequence taken out from the sliding window.

[0058] Specifically, this embodiment uses the 21 multimodal signals collected above as source data and performs data preprocessing, specifically, sequentially performing data cleaning, data normalization, sliding window sampling and feature fusion on the acquired multimodal data of the aircraft engine.

[0059] Data cleaning specifically involves cleaning the data of each mode, removing noise, removing some values ​​that have no meaningful changes, etc., to ensure data quality.

[0060] Data normalization is specifically to use maximum and minimum value normalization for processing, so that different modal data pairs have the same importance in the network, and the maximum and minimum values ​​in the signal data of the modality are selected.

[0061] The normalized data is sampled using sliding windows to obtain enough samples for training and testing. The specific steps are as follows: first, the window size and step size are determined according to the length of the sampled data. The window size determines the length of each subsequence, and the step size specifies the distance between adjacent subsequences. The sliding window starts from the starting position of the original sequence, and the subsequences are taken out in sequence according to the window size and step size until the end of the sequence is reached.

[0062] This embodiment divides the processed data according to the failure mode of the aircraft engine when it is collected, and randomly selects part of the data from the normal sampling of the aircraft engine in normal operation and various failure modes as the training set and the test set, and according to the actual situation that the normal operation of the aircraft engine accounts for a relatively large proportion, the proportion of the data of the normal operation of the aircraft engine in the training and test sets is adjusted to 60%.

[0063] S2. Construct a fault diagnosis model based on interactive perception graph attention network, extract the spatial features of adjacent time steps of each signal from the multimodal data of aircraft engines to construct a fully connected heterogeneous graph, and perform graph node screening on the fully connected heterogeneous graph and extract the spatial interaction relationship features between graph nodes for model training;

[0064] In an optional embodiment of the present invention, Figure 3 As shown, the fault diagnosis model based on the interactive perception graph attention network constructed in this embodiment specifically includes: a feature extraction layer, a graph encoding layer and a graph classification layer connected in sequence.

[0065] The feature extraction layer is used to respectively extract the spatial features of adjacent time steps of each signal in the multimodal data using a gated convolutional layer to construct a fully connected heterogeneous graph;

[0066] The graph encoding layer is used to adaptively learn heterogeneous graph edge weights using a graph attention mechanism, and to filter heterogeneous graph nodes using degree centrality;

[0067] The graph classification layer is used to predict category probabilities based on the filtered graph connections.

[0068] The feature extraction layer of this embodiment specifically includes: a plurality of parallel gated convolutional layers; the gated convolutional layer includes a convolutional layer, a gated mechanism layer, a linear mapping layer and a position encoding layer connected in sequence.

[0069] The convolution layer is used to extract local feature information of the signal time series corresponding to the window size at the time step;

[0070] The gating mechanism layer is used to use a reset gate to re-read the original input information after the convolution layer performs the convolution operation, and use an update gate to use local context information to control the amplitude of the extracted features;

[0071] The linear mapping layer is used to adapt the dimension of the extracted local feature information to the dimension of the encoder layer;

[0072] The encoder layer is used to add position encoding to the local features after linear mapping and construct a fully connected heterogeneous graph.

[0073] Specifically, this embodiment takes into account that the signals collected by a single sensor may have strong dependencies at adjacent time steps, and therefore proposes a gated convolution layer consisting of convolution operations, gating mechanisms, linear mappings, and position encoding; the calculation content of this layer is cross-graph operations between the same nodes.

[0074] The convolution layer extracts local information of the window size corresponding to the time series at the time step; the gating mechanism is used in the feature extraction layer to act on the original sensor data and the corresponding output from the convolution operation. It consists of two control schemes, namely the reset gate and the update gate. Among them, the reset gate is used to re-read the original input information after the convolution operation, and the update gate is used to use the local context information to control the amplitude of the feature; the linear mapping layer is used in the feature extraction layer to adapt the dimension of the local feature to the dimension of the encoder layer to obtain the output of the linear mapping; the position encoding is used in the feature extraction layer to allow the model to effectively learn through relative position, and the position encoding uses cosine and sine functions of different frequencies. Finally, the final output of the feature extraction layer is represented as v s ; After completing the local feature extraction, this layer is output to the graph encoding layer.

[0075] In this embodiment, the data obtained by sliding window sampling of each signal in 21 multimodal signals is taken as a node, and the attributes of the node are determined; in order to construct a fully connected heterogeneous graph, each node therein has an edge with the remaining nodes, and an adjacency matrix with a diagonal of 0 and the rest of the diagonal being 1 is formed, and a fully connected heterogeneous graph is constructed. In this graph, the feature sets of nodes and edges are represented respectively.

[0076] This embodiment proposes a gated convolution method for the feature extraction layer to better extract the features of signals of adjacent time steps. This method solves the deficiency of GAT that cannot realize cross-graph feature extraction, realizes feature extraction across time steps on the time series graph, and retains the spatial interaction between sensors. This method can further improve the accuracy of predicting the remaining life of the equipment. The gated convolution method of this embodiment is based on a gating mechanism, which can better extract the features of signals of adjacent time steps. Compared with the traditional graph convolution method, adding gated convolution can realize cross-graph feature extraction and retain the relationship of temporal interaction between sensors, thereby better extracting the spatiotemporal characteristics of the signal, further improving the accuracy of predicting the remaining life of the equipment.

[0077] The graph encoding layer of this embodiment specifically includes: a graph structure interaction perception layer and a multi-level cascaded graph attention layer and a graph pooling layer, and each level of the graph pooling layer is connected to a readout layer;

[0078] The graph structure interaction perception layer is used to calculate the relationship scores between graph nodes according to the input fully connected heterogeneous graph, and to filter the edges between graph nodes according to the relationship scores between graph nodes;

[0079] The graph attention layer is used to calculate the attention coefficient of each graph node based on the original features of the graph node and the features of its neighborhood, and update the node features in combination with the attention mechanism;

[0080] The graph pooling layer is used to filter graph nodes using Topk pooling;

[0081] The readout layer uses a cascade of average or cumulative pooling functions for output, and performs feature fusion on the output features of each readout layer.

[0082] Specifically, in this embodiment, the signal passing through the feature extraction layer is first used as a node, and multiple pairs of interactions between nodes constitute an implicit sensor network. According to domain knowledge, the implicit sensor network can be converted into an explicit graph structure, and the signal is input into the graph structure interaction perception layer according to the fully connected heterogeneous graph constructed previously.

[0083] In this embodiment, the relationship score r between nodes is calculated at the graph structure interaction perception layer. ij , to represent the node v i and v j The interaction relationship between them is calculated, and the invalid connection is removed by using the relationship score; this embodiment defines a shared single-layer feedforward neural network, which consists of a weight vector Parameterized, this weight vector is applied to each node. Then, we use the attention mechanism to calculate the weight vector according to the input feature matrix Compute Node v i and vj The p-th edge relationship score between:

[0084]

[0085] Where CAT{·} represents the concatenation operation. is the attention weight vector.

[0086] According to the output result relationship score A high relationship score means that there is a strong connection between the nodes. According to the set threshold, the edges between these nodes should be retained. A relatively low value of the relationship score reflects that the interaction between the two nodes is weak, and the edges between these nodes should be deleted.

[0087] In order to ensure that the relationship scores between nodes can more accurately reflect the strength of the direct connection between the nodes, this embodiment uses a normalization function to normalize the attention coefficients of all nodes, and deletes the edges with relationship scores less than or equal to the threshold, thereby achieving graph sparsification and improving robustness.

[0088] After the graph structure interaction perception layer, the graph has been sparse, and we input the sparse graph into the graph attention layer.

[0089] In the graph attention layer, this embodiment first inputs the original features of node v and the features of its neighborhood into the softmax layer and calculates the attention coefficient of each node through the corresponding formula, and obtains the updated features of the node by performing the inner product of the generated attention coefficient and the features of its corresponding adjacent nodes. This embodiment also uses multi-head attention to update the updated features of the node again, specifically obtaining the concatenation or average value of the node features under different attentions.

[0090] In this embodiment, the output result of the graph attention layer is input into the graph pooling layer, in which Topk pooling is adopted and the importance of nodes is changed to be measured by the degree of nodes. Finally, the top k nodes with the highest scores are retained to achieve node screening. The calculation formula is as follows:

[0091]

[0092] Among them, X ′ represents the node feature matrix after the update; X represents the node feature matrix before the update; ⊙ represents the matrix element multiplication; A ′ represents the updated adjacency matrix; A i,i represents the adjacency matrix formed after extracting the corresponding i-th row and i-th column; i represents the extraction index; z represents the importance matrix; top-k(·) represents the k-rank mechanism, k∈(0,1], represents the pooling ratio; d(v i ) represents node v iThe degree centrality of A i,j Represents node v i The adjacency matrix of i,j The number of elements.

[0093] In this embodiment, the result of Topk pooling output is input to the readout layer, where we average or accumulate the cascade of pooling functions for output, thereby obtaining the output of the entire graph encoding layer.

[0094] Since traditional remaining life estimation methods usually require the use of prior knowledge to perform statistical analysis of signals, this method has many limitations. For example, prior knowledge may not be accurate or complete, and it is impossible to fully consider the spatial interaction relationship between sensors. Therefore, the present invention adopts a graph attention mechanism to adaptively learn the weights of edges and uses degree centrality to screen nodes, which can better mine the spatial interaction relationship between sensors, thereby improving the accuracy of predicting the remaining life of the equipment.

[0095] By comparing the experimental results with prior knowledge, it can be confirmed that the present invention can effectively remove redundant sensors or invalid sensors, thereby verifying that this method is very effective in screening sensors and mining the spatial interaction relationship between sensors, and can help engineers better understand the performance and remaining life of the equipment.

[0096] The graph classification layer of this embodiment specifically includes: a first fully connected layer, a second fully connected layer, and a softmax classification layer connected in sequence.

[0097] Specifically, this embodiment inputs the output of the graph encoding layer to the graph classification layer, where it passes through two fully connected layers and a softmax classifier to obtain a fault diagnosis result, thereby achieving multiple fault diagnosis of aircraft engines.

[0098] S3. Use the trained fault diagnosis model based on interactive perceptual graph attention network to perform aircraft engine fault diagnosis.

[0099] The present invention constructs the spatial interaction relationship between nodes through the graph structure interaction perception layer in the graph coding layer, thereby completing the graph composition without prior knowledge, and uses Topk pooling to complete the node screening, thereby deleting the nodes that have no effect or even a negative effect on fault diagnosis, so as to achieve sensor screening. Through the above work, the spatial interaction between multiple sensors is fully utilized, and the sparse process is also completed, thereby improving the accuracy and robustness of the network in fault diagnosis.

[0100] The present invention has conducted experiments on aerospace engine data, using 21 sensors to collect multi-source data at different positions of the aerospace engine as a data set, including 12,000 one-dimensional signals from the normal operation of the engine, and 2,000 engine operation data under four fault modes. The time series is preprocessed to obtain 20,000 two-dimensional samples, 14,000 samples are selected as the training set, and the remaining 6,000 samples are used as the test set. Applying the detection method proposed in the present invention, the accuracy rates on the four types of faults and normal data sets are 98.86%, 96.09%, 94.12%, 96.83%, and 98.89%, respectively. It can be seen that compared with other advanced methods, the aerospace engine fault diagnosis system based on multimodal sensors of the present invention provides excellent prediction accuracy.

[0101] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0102] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0103] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0104] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0105] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. An aircraft engine fault diagnosis method based on a multimodal sensor, characterized in that: The following steps are involved: S1. Acquire multimodal data of aircraft engines; S2. Construct a fault diagnosis model based on interactive perception graph attention network, extract the spatial features of adjacent time steps of each signal from the multimodal data of aircraft engines to construct a fully connected heterogeneous graph, and perform graph node screening on the fully connected heterogeneous graph and extract the spatial interaction relationship features between graph nodes for model training; The fault diagnosis model based on the interactive perception graph attention network specifically includes: a feature extraction layer, a graph encoding layer and a graph classification layer connected in sequence; The feature extraction layer is used to respectively extract the spatial features of adjacent time steps of each signal in the multimodal data using a gated convolutional layer to construct a fully connected heterogeneous graph; The graph encoding layer is used to adaptively learn heterogeneous graph edge weights using a graph attention mechanism, and to filter heterogeneous graph nodes using degree centrality; The graph classification layer is used to predict category probabilities based on the filtered graph connections; The graph encoding layer specifically includes: a graph structure interaction perception layer and a multi-level cascaded graph attention layer and a graph pooling layer, and each level of the graph pooling layer is connected to a readout layer; The graph structure interaction perception layer is used to calculate the relationship scores between graph nodes according to the input fully connected heterogeneous graph, and to filter the edges between graph nodes according to the relationship scores between graph nodes; The graph attention layer is used to calculate the attention coefficient of each graph node based on the original features of the graph node and the features of its neighborhood, and update the node features in combination with the attention mechanism; The graph pooling layer is used to filter graph nodes using Topk pooling; The readout layer uses a cascade of average or cumulative pooling functions for output, and performs feature fusion on the output features of each readout layer; S3. Use the trained fault diagnosis model based on interactive perceptual graph attention network to perform aircraft engine fault diagnosis.

2. The method for diagnosing aircraft engine faults based on a multimodal sensor according to claim 1, characterized in that: Step S1 specifically includes: Multi-modal sensors are used to collect multi-modal data of aircraft engines at the fault detection locations of aircraft engines; The fault detection parts of the aircraft engine include a fan, a low-pressure compressor, a high-pressure compressor, a combustion chamber or a burner, a high-pressure turbine and a low-pressure turbine.

3. The aircraft engine fault diagnosis method based on multimodal sensor according to claim 1 or 2, characterized in that: The multimodal data of the aircraft engine specifically include: the total temperature at the fan inlet, the total temperature at the low-pressure compressor outlet, the total temperature at the high-pressure compressor outlet, the total temperature at the low-pressure turbine outlet, the pressure at the fan inlet, the total pressure of the bypass duct, the total pressure at the high-pressure compressor outlet, the physical fan speed, the physical core speed, the engine pressure ratio, the static pressure at the high-pressure compressor outlet, the ratio of the fuel flow to the total pressure of the bypass duct, the corrected fan speed, the corrected core speed, the bypass ratio, the burner fuel-air ratio, the exhaust enthalpy, the required fan speed, the required corrected fan speed, the coolant flow rate and the high-pressure turbine coolant flow rate.

4. The method for diagnosing aircraft engine faults based on a multimodal sensor according to claim 1, characterized in that: Step S1 also includes the following steps: sequentially performing data cleaning, data normalization, sliding window sampling and feature fusion on the acquired multimodal data of the aircraft engine.

5. The method for diagnosing aircraft engine faults based on a multimodal sensor according to claim 4, characterized in that: Sliding window sampling of multimodal data of aircraft engines specifically includes the following steps: A1. Determine the window size and step size according to the sampling data length of the multi-modal data of the aircraft engine; A2. Start sliding the window from the beginning of the sampled data sequence, and take out subsequences in sequence according to the window size and step size until the end of the sequence is reached; A3. Generate corresponding sample data based on the subsequence taken out from the sliding window.

6. The method for diagnosing aircraft engine faults based on a multimodal sensor according to claim 1, characterized in that: The feature extraction layer specifically includes: a plurality of parallel gated convolutional layers; the gated convolutional layer includes a convolutional layer, a gated mechanism layer, a linear mapping layer and a position encoding layer connected in sequence; The convolution layer is used to extract local feature information of the signal time series corresponding to the window size at the time step; The gating mechanism layer is used to use a reset gate to re-read the original input information after the convolution layer performs the convolution operation, and use an update gate to use local context information to control the amplitude of the extracted features; The linear mapping layer is used to adapt the dimension of the extracted local feature information to the dimension of the encoder layer; The encoder layer is used to add position encoding to the local features after linear mapping and construct a fully connected heterogeneous graph.

7. The method for diagnosing aircraft engine faults based on a multimodal sensor according to claim 1, characterized in that: The use of Topk pooling to screen graph nodes is specifically as follows: Among them, X ′ represents the node feature matrix after the update; X represents the node feature matrix before the update; ⊙ represents the matrix element multiplication; A ′ represents the updated adjacency matrix; A i,i represents the adjacency matrix formed after extracting the corresponding i-th row and i-th column; i represents the extraction index; z represents the importance matrix; top-k(·) represents the k-rank mechanism, k represents the pooling ratio; d(v i ) represents node v i The degree centrality of A i,j Represents node v i The adjacency matrix of .

8. The method for diagnosing aircraft engine faults based on a multimodal sensor according to claim 1, characterized in that: The graph classification layer specifically includes: a first fully connected layer, a second fully connected layer and a softmax classification layer which are connected in sequence.

Images