Avionics equipment multi-scale fault feature extraction and sorting method

Abstract

The application discloses an avionics equipment multi-scale fault feature extraction and sorting method, and relates to the technical field of avionics equipment.The method specifically comprises the following steps: obtaining avionics equipment operation data, and performing normalization processing on the data; performing convolution operation on the normalized data by using convolution kernels of different scales, obtaining a plurality of time feature representations of different scales, and performing feature fusion; adopting a multi-path attention mechanism to encode the features after the fusion processing in each dimension, and calculating a weight vector, wherein the weight vector is composed of a mask tensor mapped to a global high-dimensional feature on the original input feature, and then the global high-dimensional feature is taken as the input of a subsequent task.The application extracts more representative fault features by using a multi-scale deep learning model, effectively reduces the dependence on expert knowledge, improves the accuracy and efficiency of data analysis, and provides better support for performance optimization and fault diagnosis of avionics equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of avionics equipment, and in particular to a method for extracting and sorting multi-scale fault features of avionics equipment. Background Art

[0002] Avionics equipment is critical for safe and efficient flight, encompassing flight control, navigation, communications, radar, meteorological, and emergency response systems. These systems provide a wealth of data, such as flight speed, altitude, course direction, and weather information, to help pilots and ground crew monitor and adjust aircraft status in real time, ensuring flight safety and accuracy. Traditional methods for extracting fault features from avionics equipment include time-domain feature extraction, frequency-domain feature extraction, and wavelet transform features. However, with the increasing integration and system complexity of avionics equipment, avionics equipment monitoring data exhibits high dimensionality and strong coupling, leading to low accuracy and implementation difficulties for these traditional methods. In recent years, with the advancement of artificial intelligence algorithms, methods based on deep learning networks have proven effective in processing high-dimensional, complex data and have gained widespread acceptance in various fields. Therefore, a multi-scale fault feature extraction and ranking method for avionics equipment is proposed. Summary of the Invention

[0003] The purpose of the present invention is to propose a multi-scale fault feature extraction and ranking method for avionics equipment, which uses a multi-scale network and a multi-way attention mechanism to extract key features in the data. This method can more accurately describe and explain the high-dimensional and complex features in avionics equipment data, improve the efficiency and accuracy of data processing, and provide better support for performance optimization and fault diagnosis of avionics equipment.

[0004] The specific technical solutions are as follows:

[0005] The present invention is a method for extracting and ranking multi-scale fault features of avionics equipment, comprising the following steps:

[0006] Obtain avionics equipment operation data and normalize the data;

[0007] Use convolution kernels of different scales to perform convolution operations on the normalized data to obtain temporal feature representations of multiple scales and perform feature fusion;

[0008] A multi-way attention mechanism is used to encode the fused features in each dimension and calculate their weight vectors. The weight vectors are composed into a mask tensor that maps the original input features to a global high-dimensional feature, which is then used as the input for subsequent tasks.

[0009] Furthermore, the data normalization step generally includes the following steps:

[0010] Determine the data normalization method and select the corresponding normalization method based on the distribution and characteristics of different data. Normalization methods include minimum-maximum normalization, z-score normalization, or decimal scaling normalization;

[0011] Calculate normalization parameters. For min-max normalization and decimal scaling normalization, you need to calculate the minimum and maximum values ​​to determine the normalization range. For z-score normalization, you need to calculate the mean and standard deviation.

[0012] Use the calculated normalization parameters to normalize the data to be within the specified range.

[0013] Furthermore, the min-max normalization usually scales the data to the range of [0, 1], while the fractional scaling normalization scales the data to the range of [-1, 1];

[0014] Among them, the minimum-maximum normalization converts the value of each feature into a value between 0 and 1. The specific calculation formula is:

[0015]

[0016] Among them, x min and x max are the minimum and maximum values ​​in the feature, x is any value in the feature, x norm is the normalized feature data.

[0017] Furthermore, the convolution operation is performed on the normalized data using convolution kernels of different scales to obtain multiple temporal feature representations of different scales. Specifically:

[0018] The convolution operation uses dilated causal convolution, which combines dilated convolution and causal convolution into one.

[0019] Dilated convolution introduces a dilation factor into the convolution kernel. The dilation factor is a positive integer that represents the number of intervals in the convolution kernel.

[0020] Causal convolution means that the convolution kernel can only access the previous data in the sequence, but cannot access the subsequent data in the sequence.

[0021] Furthermore, the feature fusion is to fuse the features extracted at different scales in the channel direction, specifically: to cascade or superimpose the features extracted at different scales.

[0022] Furthermore, the multi-way attention mechanism is used to encode the fused features in each dimension. Specifically:

[0023] For input features X=[x1,x2,...,xc ]∈R C×H×W , use one-dimensional pooling in its vertical and horizontal directions to get the feature encoding of the c-th channel at height h as The feature encoding of the cth channel at width w is Among them, h∈00H represents the height; w∈000 represents the width.

[0024] Furthermore, the weight vectors are calculated for the feature encoding in the vertical and horizontal directions. Specifically:

[0025] The feature codes in the vertical and horizontal directions are input into two parallel fully connected-nonlinear activation-fully connected networks respectively, and then a sigmoid function is used to obtain the final vertical attention weight vector g. h ∈R C×H×1 and the horizontal attention weight vector g w ∈R C×1×W .

[0026] Furthermore, the weight vector is formed into a mask tensor that maps the original input feature to a global high-dimensional feature, which is then used as the input for subsequent tasks, specifically:

[0027] Vertical attention weight vector g h ∈R C×H×1 and the horizontal attention weight vector g w ∈R C×1×W Outer product, get a mask tensor M∈R C×H×W , the shape of the mask tensor is consistent with the shape of the input feature X, the mask tensor M∈R C×H×W Each attention coefficient in represents the importance of the element at the corresponding position in the input feature X, and the input feature X is combined with the mask tensor M∈R using the Hadamard product. C×H×W The final global high-dimensional features are obtained by multiplication and serve as input data for performance optimization or fault diagnosis of avionics equipment.

[0028] The application provides an avionics equipment multi-scale fault feature extraction and sorting method adopting a multi-scale time convolution network and a multi-path attention mechanism fusion. The data generated by avionics equipment usually has the characteristics of high dimension and strong coupling, and the traditional data processing method cannot effectively extract useful features therein. The method uses a multi-scale network and a multi-path attention mechanism to extract key features in the data. The multi-scale network can extract features of different granularities at different scales, and the multi-path attention mechanism can weight different features during feature extraction, thereby improving the importance and interpretability of the features. Through this method, the high-dimensional complex features in the avionics equipment data can be more accurately described and explained, the efficiency and accuracy of data processing are improved, and better support is provided for performance optimization and fault diagnosis of avionics equipment.

[0029] The beneficial effects of the application are as follows:

[0030] The avionics equipment multi-scale fault feature extraction and sorting method provided by the application can extract high-dimensional features of avionics equipment data, which also brings extensive economic benefits to the aviation industry. By extracting and analyzing the features of the equipment data, the reliability and safety of the equipment can be effectively improved, and the occurrence of faults and accidents can be reduced, thereby avoiding unnecessary personnel casualties and property losses. Feature extraction also helps to optimize the maintenance plan of the equipment, reduce the cost and time of maintenance, and improve the efficiency and quality. In addition, by analyzing the features, information related to flight control and navigation can be extracted, providing more accurate and reliable flight control and navigation data for pilots, thereby reducing the incidence of flight accidents and improving the safety of flight. Finally, high-dimensional feature extraction plays a positive role in promoting the development and innovation of the avionics equipment industry and the development and progress of the entire aviation industry. Therefore, high-dimensional feature extraction has become an important direction for the application of avionics equipment data, and the economic benefits brought by it are significant. BRIEF DESCRIPTION OF DRAWINGS

[0031] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiments will be briefly introduced as follows. It should be understood that the following drawings only show some embodiments of the application, and therefore should not be regarded as a limitation on the scope. Other related drawings can also be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0032] Figure 1 is a method flowchart of the application;

[0033] Figure 2 is a multi-scale convolution structure diagram of the application;

[0034] Figure 3This is a flowchart of the multi-way attention mechanism processing of the present invention. DETAILED DESCRIPTION

[0035] In order to make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present invention and are not intended to limit the present invention. That is, the embodiments described herein are only some embodiments of the present invention, not all embodiments. Generally, the components of the embodiments of the present invention described and illustrated in the drawings herein may be arranged and designed in various different configurations.

[0036] It should be noted that the terms "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, product or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, product or apparatus.

[0037] Example 1

[0038] The present invention is a method for extracting and sorting multi-scale fault features of avionics equipment. The specific implementation steps are as follows: Figure 1 As shown, the details are as follows:

[0039] Step 1: Obtain avionics equipment operating data and determine a data normalization method. The corresponding normalization method is selected based on the distribution and characteristics of different data. The normalization methods include minimum-maximum normalization, z-score normalization, or decimal scaling normalization.

[0040] Step 2: Calculate the normalization parameters. For min-max normalization and decimal scaling normalization, you need to calculate the minimum and maximum values ​​to determine the normalization range. For z-score normalization, you need to calculate the mean and standard deviation.

[0041] Among them, min-max normalization usually scales the data to the range of [0,1], while fractional scaling normalization scales the data to the range of [-1,1];

[0042] Among them, the minimum-maximum normalization converts the value of each feature into a value between 0 and 1. The specific calculation formula is:

[0043]

[0044] Among them, x min and x max are the minimum and maximum values ​​in the feature, x is any value in the feature, x norm is the normalized feature data.

[0045] Step 3: Use the calculated normalization parameters to normalize the data so that it is within the specified range. Normalizing the data can improve the stability and performance of the deep learning model.

[0046] Step 4, multi-scale convolution: Use convolution kernels of different scales to perform convolution operations on the normalized data to obtain temporal feature representations of multiple scales. Specifically:

[0047] The convolution operation uses dilated causal convolution, which combines dilated convolution and causal convolution into one.

[0048] Dilated convolution introduces a dilation factor into the convolution kernel. The dilation factor is a positive integer that represents the number of intervals in the convolution kernel. For example, if the dilation factor is 2, each element in the convolution kernel is separated by 2 elements. The purpose of introducing the dilation factor is to expand the receptive field of the convolution operation without increasing the size of the convolution kernel, thereby effectively processing long sequence data. Causal convolution means that the convolution kernel can only access the previous data in the sequence, but cannot access the data after the sequence. The dilated causal convolution structure is as follows: Figure 2 shown.

[0049] At the same time, dilated causal convolution is also paired with a residual structure. The residual structure introduces a skip connection, which refers to a direct connection introduced into the network, adding the input directly to the network's output. In this way, the network can learn the residual mapping, that is, the difference between the output and the input, rather than directly learning the input-to-output mapping, avoiding the vanishing gradient problem.

[0050] Step 5, multi-scale merging: Multi-scale merging performs feature fusion on the temporal feature representation after the convolution operation, that is, the features extracted at different scales are fused in the channel direction. Specifically, the features extracted at different scales are cascaded or superimposed to obtain a new feature representation.

[0051] Steps 6, 7, and 8 use a multi-way attention mechanism to process features. Figure 3 shown.

[0052] Step 6, feature encoding: Use a multi-way attention mechanism to encode the fused features in each dimension. Specifically:

[0053] For input features X=[x1,x2,...,x c ]∈R C×H×W , use one-dimensional pooling in its vertical and horizontal directions to get the feature encoding of the c-th channel at height h as The feature encoding of the cth channel at width w is Among them, h∈00H represents the height; w∈000 represents the width.

[0054] Step 7, feature coding correction: Calculate the weight vector for the feature coding in the vertical and horizontal directions. Specifically:

[0055] The feature codes in the vertical and horizontal directions are input into two parallel fully connected-nonlinear activation-fully connected networks respectively, and then a sigmoid function is used to obtain the final vertical attention weight vector g. h ∈R C×H×1 and the horizontal attention weight vector g w ∈R C×1×W .

[0056] Step 8, global feature generation: The weight vector is formed into a mask tensor that maps the original input feature to a global high-dimensional feature, which is then used as the input for subsequent tasks, specifically:

[0057] Vertical attention weight vector g h ∈R C×H×1 and the horizontal attention weight vector g w ∈R C×1×W Outer product, get a mask tensor M∈R C×H×W , the shape of the mask tensor is consistent with the shape of the input feature X, the mask tensor M∈R C×H×W Each attention coefficient in represents the importance of the element at the corresponding position in the input feature X, and the input feature X is combined with the mask tensor M∈R using the Hadamard product. C×H×W The final global high-dimensional features are obtained by multiplication and serve as input data for performance optimization or fault diagnosis of avionics equipment.

[0058] The advantage of the multi-way attention mechanism is that it can automatically learn which features in the input feature sequence are important, and can select different input features to pay attention to at different times, which enables the model to better capture the relationship in the input feature sequence, thereby improving the accuracy and performance of the model.

[0059] Compared to TCN (Temporal Convolutional Network), currently popular LSTM (Long Short-Term Memory) and RNN (Recurrent Neural Network) deep learning networks are susceptible to vanishing or exploding gradients due to their use of recurrent structures. These models perform poorly when processing long sequences and require more parameters and computing resources to process the same amount of data. Furthermore, due to the limitations of their recurrent structures, they may not be able to capture longer-term dependencies. In contrast, the TCN used in this technology is a new sequence modeling method based on convolutional neural networks. It can effectively capture long-term dependencies by stacking convolutional layers and can process sequences of arbitrary length. It also has higher parallelization capabilities and a smaller number of parameters.

[0060] Currently, most feature extraction techniques use single-scale models, which can only process input data at a fixed time scale. This makes single-scale models ill-suited to input data of varying time scales, potentially failing to capture details and variations in the input data. Furthermore, single-scale models may lose some temporal information when processing input data of varying time scales, thus impacting the model's predictive capabilities. This technology fuses multi-scale models with TCN to propose a multi-scale TCN structure that can process input data at different time scales and better capture the complexity and variation of the data, thereby improving the model's performance and generalization capabilities.

[0061] Models without an attention mechanism cannot effectively focus on the important parts of the input data, which may limit the model's ability to model the input data. These models may ignore some key information in the input data or be disturbed by some irrelevant information in the input data, thereby affecting the model's performance and generalization ability. In contrast, the multi-way attention mechanism designed in this patent can adaptively allocate attention based on the importance of the input data, thereby better focusing on the important parts of the input data and improving the model's performance and generalization ability. This attention mechanism can also help the model remember more distant information when processing long sequence data, thereby improving the model's performance on sequence data.

[0062] High-dimensional feature extraction from avionics equipment data has applications across multiple fields. First, feature extraction and analysis of data can be used to build fault prediction and diagnosis models, enabling early warning and rapid location of equipment failures, thereby improving equipment reliability and safety. Finally, feature extraction and analysis of data can identify features related to flight control and navigation, providing pilots with more accurate and reliable flight control and navigation information. Therefore, high-dimensional feature extraction is a crucial foundation for the application of avionics equipment data in multiple fields.

[0063] The present invention is not limited to the aforementioned specific embodiments, but extends to any new features or any new combination disclosed in this specification, as well as any new method or process steps or any new combination disclosed.

Claims

1. A method for extracting and ranking multi-scale fault features of avionics equipment, characterized in that: The following steps are involved: Obtain avionics equipment operation data and normalize the data; Use convolution kernels of different scales to perform convolution operations on the normalized data to obtain temporal feature representations of multiple scales and perform feature fusion; A multi-way attention mechanism is used to encode the fused features in each dimension and calculate their weight vectors. The weight vectors are combined into a mask tensor that maps the original input features to a global high-dimensional feature, which is then used as the input for subsequent tasks. The convolution operation is performed on the normalized data using convolution kernels of different scales to obtain multiple temporal feature representations of different scales. Specifically: The convolution operation uses dilated causal convolution, which combines dilated convolution and causal convolution into one. Dilated convolution introduces a dilation factor into the convolution kernel. The dilation factor is a positive integer that represents the number of intervals in the convolution kernel. Causal convolution means that the convolution kernel can only access the data before the sequence, but cannot access the data after the sequence; The feature fusion is to fuse the features extracted at different scales in the channel direction, specifically: cascading or superimposing the features extracted at different scales; The multi-way attention mechanism is used to encode the fused features in various dimensions. Specifically: For input features , use one-dimensional pooling in its vertical and horizontal directions to get the feature encoding of the c-th channel at height h as , the feature encoding of the cth channel at width w is , where h∈0~H represents height; w∈0~W represents width; Calculate the weight vector for feature encoding in the vertical and horizontal directions. Specifically: The feature codes in the vertical and horizontal directions are input into two parallel fully connected-nonlinear activation-fully connected networks respectively, and then a sigmoid function is used to obtain the final vertical attention weight vector. and the horizontal attention weight vector ; The weight vector is formed into a mask tensor that maps the original input features to a global high-dimensional feature, which is then used as the input for subsequent tasks, specifically: Vertical attention weight vector and the horizontal attention weight vector Outer product, get a mask tensor , the shape of the mask tensor and the input features The shape is consistent with the mask tensor Each attention coefficient in represents the input feature The importance of the corresponding position elements in the input feature is converted into With masked tensors The final global high-dimensional features are obtained by multiplication and serve as input data for performance optimization or fault diagnosis of avionics equipment.

2. The method for extracting and ranking multi-scale fault features of avionics equipment according to claim 1, characterized in that: The data normalization step generally includes the following steps: Determine the data normalization method and select the corresponding normalization method based on the distribution and characteristics of different data. Normalization methods include minimum-maximum normalization, z-score normalization, or decimal scaling normalization; Calculate normalization parameters. For min-max normalization and decimal scaling normalization, you need to calculate the minimum and maximum values ​​to determine the normalization range. For z-score normalization, you need to calculate the mean and standard deviation. Use the calculated normalization parameters to normalize the data to be within the specified range.

3. The method for extracting and ranking multi-scale fault features of avionics equipment according to claim 2, characterized in that: The min-max normalization usually scales the data to the range [0, 1], while the fractional scaling normalization scales the data to the range [-1, 1]; Among them, the minimum-maximum normalization converts the value of each feature into a value between 0 and 1. The specific calculation formula is: in, and are the minimum and maximum values ​​of the feature, is any value in the feature, is the normalized feature data.

Images