Fault diagnosis method and system for avionics system of aircraft

Abstract

The invention discloses an aircraft avionics system fault diagnosis method and system, relates to the technical field of avionics system fault detection, and is used for solving the technical problems that in the prior art, the extraction accuracy of avionics system fault features is low, and complex fault modes cannot be effectively recognized. The invention discloses an aircraft avionics system fault diagnosis method. The aircraft avionics system fault diagnosis method comprises the following steps: improving a Transform model to obtain a multi-scale frequency attention Transform model; the multi-scale frequency attention Transform model is integrated with a multi-scale frequency attention mechanism in a Transform encoder framework of the multi-scale frequency attention Transform model; in the training stage, aircraft avionics system fault information vectors are obtained to form a training sample set; the method comprises the following steps of: training a multi-scale frequency attention Transform model to obtain a trained multi-scale frequency attention Transform model; in the application stage, aircraft avionics system information is input into the trained multi-scale frequency attention Transform model, and fault information of an aircraft avionics system is output.

Description

Technical Field

[0001] The present invention relates to the technical field of avionics system fault detection, and more specifically, to a fault diagnosis method and system for an aircraft avionics system. Background Art

[0002] In industrial production and applications, the fault diagnosis of aviation machinery and equipment is an important link to ensure the reliable operation of equipment and safe production. With the rapid development of industrial automation and intelligence, traditional fault diagnosis methods have been difficult to meet the requirements of modern industry for high precision, high efficiency, and high robustness.

[0003] In recent years, with the rise of deep learning technology, data-driven fault diagnosis methods have gradually become a research hotspot. Deep learning models can automatically learn feature representations from a large amount of data, thereby realizing the recognition and classification of fault patterns. However, existing deep learning models still have some limitations in processing mechanical vibration signals. For example, although the Convolutional Neural Network (CNN) performs well in local feature extraction, due to the fixed size of its convolutional kernels, it is difficult to capture long-term dependencies in long sequence data. The Recurrent Neural Network (RNN) and its variants can handle time series data, but they often have difficulty effectively identifying and locating fault signals in the face of complex non-linear changes. In addition, existing deep learning models have poor robustness in a noisy environment, and noise may interfere with the model's learning and recognition of fault features, thereby reducing the accuracy of diagnosis.

[0004] To overcome the above limitations, some researchers have begun to explore methods that combine signal processing technology with deep learning technology. For example, the Fast Fourier Transform (FFT) is used to transform vibration signals from the time domain to the frequency domain to better extract fault features. However, existing methods still have deficiencies in frequency domain feature encoding and noise suppression. Specifically, although traditional fast Fourier transform methods can transform signals to the frequency domain, they are difficult to effectively handle noise interference, resulting in inaccurate extraction of fault features. In addition, there are also deficiencies in multi-scale feature fusion and global information interaction, which limits the model's ability to recognize complex fault patterns. It can be seen that the existing technology has low accuracy in extracting fault features and cannot effectively identify complex fault patterns. Summary of the Invention

[0005] The present invention aims to provide an aircraft avionics system fault diagnosis method and system to address the technical problems of the prior art, which suffer from low accuracy in extracting avionics system fault features and an inability to effectively identify complex fault modes. In view of this, the present invention achieves this goal through the following solutions.

[0006] In a first aspect, the present invention provides an aircraft avionics system fault diagnosis method, wherein a Transformer model is improved to obtain a multi-scale frequency attention Transformer model; the multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism into its Transformer encoder architecture; the multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and implements data representation; The aircraft avionics system fault diagnosis method includes: In the training phase, the aircraft avionics system fault information vector is obtained to form a training sample set; The multi-scale frequency attention Transformer model is trained using the training sample set to obtain a trained multi-scale frequency attention Transformer model; in the multi-scale frequency attention Transformer model, the Adam optimizer is used for model training to accelerate model convergence and improve optimization effect; In the application phase, the aircraft avionics system information is input into the trained multi-scale frequency attention Transformer model, and the fault information of the aircraft avionics system is output.

[0007] Compared with the prior art, the aircraft avionics system fault diagnosis method of the present invention utilizes an improved Transformer model, namely, the above-mentioned Multi-Scale Frequency Attention Transformer (MSFA-Trans) model. The Multi-Scale Frequency Attention Transformer model integrates a multi-scale frequency attention mechanism in its Transformer encoder architecture; the multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and realizes data representation; after training the Multi-Scale Frequency Attention Transformer model, the aircraft avionics system information is input, and the fault information of the aircraft avionics system can be output. In the above technical solution of the present invention, the multi-scale frequency attention mechanism is based on the Fast Fourier Transform (FFT) to model the correlation between signal frequencies and introduces an attention mechanism in the frequency domain, allowing the model to focus on valuable frequency information. At the same time, a learnable residual coefficient can be introduced to adjust the attention distribution, further improving the generalization ability and robustness of the model. This enables the present invention to achieve robust mechanical fault diagnosis in noisy environments and provides an efficient, accurate, and robust solution for the health monitoring of mechanical equipment in industrial production. The above technical solution of the present invention solves the technical problems of the existing technology in extracting avionics system fault features with low accuracy and inability to effectively identify complex fault modes.

[0008] Furthermore, in the aircraft avionics system fault diagnosis method of the present invention, the multi-scale frequency attention mechanism performs calculations in the frequency domain, including: Converting an input time domain signal into a frequency domain signal, wherein the frequency domain signal contains amplitude and phase information; The encoder-decoder architecture is used to compress the channel representation vector into a dimension of The hidden layer vector of ; where, represents the number of input feature vectors, Indicates the length of the input feature vector; Decoding the hidden layer vector back to its original dimension to obtain an output channel weight vector; wherein the encoding operation is performed by the first convolutional layer and the decoding operation is performed by the second convolutional layer; the first convolutional layer provides a nonlinear transformation function in combination with the ReLU function; the second convolutional layer uses the Sigmoid function to map the value to a range between 0 and 1; Introducing a learnable soft threshold parameter and a constant of 0 into the Fourier space-based deep convolutional network to improve the model's attention to important information; The activation function tanh is used to map the eigenvalues to any range of values, and the frequency domain signal is restored to the time domain signal through the inverse Fourier transform.

[0009] Furthermore, in the aircraft avionics system fault diagnosis method of the present invention, the multi-scale frequency attention mechanism performs calculations in the frequency domain, further comprising: Residual coefficients are introduced to adjust the distribution of attention; residual connections enable the network to easily learn identity mappings through direct paths across layers; Introducing learnable parameters for residual coefficients enables the model to effectively utilize residual connections, learn the optimal coefficients, and improve the model's expressive power.

[0010] Furthermore, in the aircraft avionics system fault diagnosis method of the present invention, the multi-scale frequency attention Transformer model includes a time domain dynamic convolution layer; The time domain dynamic convolution layer predicts the convolution filter according to the characteristics of each sample; The time-domain dynamic convolution layer has multiple convolution kernels; Among them, multiple convolution kernels share the same kernel size and input / output dimensions; multiple convolution kernels are aggregated by using attention weights, first using global average pooling to compress global spatial information, and using two fully connected layers and softmax function to generate normalized attention weights of multiple convolution kernels for the input vibration signal.

[0011] Furthermore, in the aircraft avionics system fault diagnosis method of the present invention, the aggregation process of the multiple convolution kernels is described as follows: ; ; ; in, y represents the extracted features, represents the activation function, is the weight after aggregation, represents the aggregated bias that shares the same attention weight, represents the input features, Indicates k The weight of the convolution kernel, Indicates k The bias of the convolution kernel, Indicates k The attention score of the convolution kernel, , .

[0012] Furthermore, in the aircraft avionics system fault diagnosis method of the present invention, after multiple convolution kernels are aggregated, convolution kernels of multiple different scales are used to filter out signal components at different levels from complex signals, and the multi-variable signal components of the input data are encoded and coupled with global and local correlations; for multiple different convolution kernels, denoted as , then there is: ; wherein, represents the signal output by applying the i th convolution kernel, represents convolution, y represents the original input signal, represents the i th convolution kernel, represents the n rd output value, represents the k th original input signal, represents the n - k th value in the convolution kernel.

[0013] Furthermore, in the aircraft avionics system fault diagnosis method of the present invention, the multi-scale frequency attention Transformer model includes a global frequency encoding layer; The global frequency encoding layer transforms the input signal from the time domain to the frequency domain, captures long-range dependencies through a weight sharing mechanism in the frequency domain dependence relationship, and deconstructs the signal into amplitude and phase information of different frequencies, enabling the multi-scale frequency attention Transformer model to understand the data from a broad perspective.

[0014] Furthermore, in the aircraft avionics system fault diagnosis method of the present invention, the deconstructing the signal into amplitude and phase information of different frequencies includes: Using the global frequency encoding layer, for the input data, C represents the number of input feature vectors, L represents the length of the input feature vector. Using the one-dimensional fast Fourier transform to convert the time domain signal into a frequency domain signal, then there is: ; wherein, represents the frequency domain signal after Fourier transform, represents that the dimension of the output frequency domain signal is , represents the number of input feature vectors, represents the length of the input feature vector, represents the result of the one-dimensional fast Fourier transform, which is a set of complex variables containing amplitude and phase information and is extracted from the signal through the following formula: ; ; Among them, represents the amplitude information of the frequency-domain signal, represents the real part of the frequency-domain signal, represents the imaginary part of the frequency-domain signal, represents the phase information of the frequency-domain signal; In the global frequency encoding layer, a learnable filter is introduced for frequency-domain convolution. For two different sets of learnable weights, they are respectively denoted as and , and the two sets of learnable weights are regarded as learnable frequency filters for different hidden dimensions. The filter size is k , and it keeps the same dimension as the input data. Then there is: ; Among them, represents the filtered feature, represents the learnable filter, represents the frequency-domain signal, represents the dot product operation, represents that the dimension of the output signal is, , represents the number of input feature vectors, represents the length of the input feature vector.

[0015] Furthermore, in the aircraft avionics system fault diagnosis method of the present invention, the element-wise multiplication between the transformed data and the filter is expressed as: ; ; Among them, represents the amplitude signal, represents the learnable amplitude weight, represents the phase signal, represents the learnable phase weight; Multiply the obtained amplitude and phase information after transformation by the learnable amplitude and phase weights, globally adjust the frequency of the signal according to the task requirements, extract different frequency feature information of the time series data by enhancing the frequencies of interest and attenuating the irrelevant frequencies, and then reconvert the reconstructed complex signal back to the time domain through the inverse Fourier transform.

[0016] Based on the above technical solutions, compared with the prior art, in the aircraft avionics system fault diagnosis method of the present invention, by introducing the above specific time-domain dynamic convolution layer and global frequency encoding layer (Global Frequency Encoding Layer, GFE-Layer) into the above multi-scale frequency attention Transformer model, the accurate extraction of the mechanical vibration signal characteristics of the aircraft avionics system can be realized; further, the above time-domain dynamic convolution layer can adaptively adjust the convolution kernel weights to capture the time dynamic characteristics of the input signal, while the global frequency encoding layer can transform the signal from the time domain to the frequency domain through the fast Fourier transform, extract frequency characteristics, and enhance the feature extraction ability through global information interaction; based on the above technical solutions, that is, the above special feature extraction method, the present invention can achieve high-precision fault diagnosis under different signal-to-noise ratio conditions; further, the present invention can enhance the adaptability and generalization ability of the model to different input signals by introducing learnable thresholds and residual coefficients; during the training process, the model can automatically learn the optimal thresholds and residual coefficients, so as to maintain stable performance under different noise levels and fault modes. In addition, the above multi-scale frequency attention mechanism can dynamically adjust the degree of attention to different frequency components, enabling the model to better adapt to different input signal characteristics, and the present invention can exhibit excellent generalization ability on multiple data sets.

[0017] In a second aspect, the present invention provides an aircraft avionics system fault diagnosis system, including: An aircraft avionics system fault diagnosis system, which improves the Transformer model to obtain a multi-scale frequency attention Transformer model. The multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism in the Transformer encoder architecture. The multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and realizes data representation; the aircraft avionics system fault diagnosis system includes: A training module, configured to: obtain a training sample set composed of aircraft avionics system fault information vectors; train the multi-scale frequency attention Transformer model with the training sample set to obtain a trained multi-scale frequency attention Transformer model; during the training process of the multi-scale frequency attention Transformer model, the Adam optimizer is used for model training to accelerate model convergence and improve the optimization effect; An application module, configured to: input the aircraft avionics system information into the trained multi-scale frequency attention Transformer model and output the fault information of the aircraft avionics system.

[0018] Compared with the prior art, the beneficial effects of the aircraft avionics system fault diagnosis system of the present invention are the same as those of the aircraft avionics system fault diagnosis method described in the above technical solution, and will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings described herein are used to provide a further understanding of the present invention, and constitute a part of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention. In the drawings: Figure 1 is a schematic flowchart of the aircraft avionics system fault diagnosis method of the present invention; Figure 2 is a schematic architecture diagram of the multi-scale frequency attention Transformer model of the present invention; Figure 3 is a schematic architecture diagram of the time-domain dynamic convolution layer in the present invention; Figure 4 is a schematic architecture diagram of the multi-scale frequency attention mechanism module; Figure 5 is a schematic topology diagram of a 24-pulse ATRU; Figure 6 is a schematic diagram of an experimental ATRU model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer, the present invention will be further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0021] It should be noted that when an element is referred to as being "fixed to" or "disposed on" another element, it can be directly on the other element or indirectly on the other element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element.

[0022] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present invention, "a plurality" means two or more unless otherwise specifically defined. The meaning of "several" is one or more unless otherwise specifically defined.

[0023] Some researchers have explored methods that combine signal processing with deep learning techniques. Specifically, the Fast Fourier Transform (FFT) is used to convert vibration signals from the time domain to the frequency domain, thereby better extracting fault features. However, existing methods still have shortcomings in frequency domain feature encoding and noise suppression. For example, while the traditional Fast Fourier Transform method can convert signals to the frequency domain, it has difficulty effectively processing noise interference, resulting in inaccurate fault feature extraction. Furthermore, there are deficiencies in multi-scale feature fusion and global information interaction, which limits the model's ability to identify complex fault modes. Therefore, existing technologies have technical problems such as low fault feature extraction accuracy and an inability to effectively identify complex fault modes.

[0024] See also Figure 1 In order to solve the above technical problems, the present invention provides an aircraft avionics system fault diagnosis method, in which a Transformer model is improved to obtain a multi-scale frequency attention Transformer model; the multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism into its Transformer encoder architecture; the multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and realizes data representation; The method for diagnosing a fault in an avionics system of an aircraft includes: S100, obtaining a training sample set; obtaining an aircraft avionics system fault information vector to form a training sample set; S200, training a model using a training sample set; training the multi-scale frequency attention Transformer model using the training sample set to obtain a trained multi-scale frequency attention Transformer model; using an Adam optimizer to train the multi-scale frequency attention Transformer model to accelerate model convergence and improve optimization effect; S300, using the trained model to perform fault diagnosis; inputting the aircraft avionics system information into the trained multi-scale frequency attention Transformer model, and outputting the fault information of the aircraft avionics system.

[0025] When the above technical solution is adopted, the aircraft avionics system fault diagnosis method of the present invention utilizes an improved Transformer model, namely the above-mentioned multi-scale frequency attention Transformer model. The multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism into its Transformer encoder architecture. The multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and implements data representation. After the multi-scale frequency attention Transformer model is trained, aircraft avionics system information is input and aircraft avionics system fault information is output. In the above technical solution of the present invention, the multi-scale frequency attention mechanism is based on the Fast Fourier Transform (FFT) to model the correlation between signal frequencies and introduces an attention mechanism in the frequency domain, allowing the model to focus on valuable frequency information. At the same time, a learnable residual coefficient can be introduced to adjust the attention distribution, further improving the generalization ability and robustness of the model. This enables the present invention to achieve robust mechanical fault diagnosis in noisy environments, providing an efficient, accurate, and robust solution for mechanical equipment health monitoring in industrial production. The above technical solution of the present invention solves the technical problems of the prior art in that the accuracy of extracting avionics system fault features is low and complex fault modes cannot be effectively identified.

[0026] In order to better understand the present invention, the content of the present invention is further explained below in conjunction with specific examples, but the content of the present invention is not limited to the following examples. Example 1

[0027] This embodiment provides an aircraft avionics system fault diagnosis method, including: Step 1: Improve the Transformer model to obtain a multi-scale frequency attention Transformer model; the multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism into its Transformer encoder architecture; the multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and realizes data representation; Step 2: Obtain aircraft avionics system fault information vectors to form a training sample set; Step 3: Using the training sample set to train the multi-scale frequency attention Transformer model to obtain a trained multi-scale frequency attention Transformer model; in the multi-scale frequency attention Transformer model, the Adam optimizer is used for model training to accelerate model convergence and improve optimization effect; Step 4, application stage, input the avionics system information of the aircraft into the trained multi-scale frequency attention Transformer model, and output the fault information of the avionics system of the aircraft. Embodiment 2

[0028] This embodiment provides a fault diagnosis method for an aircraft avionics system, including: S100, improve the Transformer model to obtain a multi-scale frequency attention Transformer model; the multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism in its Transformer encoder architecture; the multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and realizes data representation; wherein, the multi-scale frequency attention mechanism performs calculations in the frequency domain, including: Transform the input time-domain signal into a frequency-domain signal, and the frequency-domain signal includes amplitude and phase information; Adopt an encoder-decoder architecture to compress the channel representation vector into a hidden layer vector with a dimension of ; wherein, represents the number of input feature vectors, represents the length of the input feature vector; Decode the hidden layer vector back to the original dimension to obtain an output channel weight vector; wherein, the encoding operation is performed by the first convolutional layer, and the decoding operation is performed by the second convolutional layer; the first convolutional layer provides a non-linear transformation function in combination with the ReLU function; the second convolutional layer uses the Sigmoid function to map the value to the range between 0 and 1; Introduce a learnable soft threshold parameter and a constant of 0 in the deep convolutional network based on the Fourier space to improve the model's attention to important information; Use the activation function tanh to map the eigenvalue to any range value, and restore the frequency-domain signal to the time-domain signal through the inverse Fourier transform; Introduce a residual coefficient to adjust the distribution of attention; the residual connection enables the network to easily learn the identity mapping through a direct path across layers; Introduce a learnable parameter of the residual coefficient, so that the model can effectively utilize the residual connection, learn the optimal coefficient, and improve the model's expression ability; Furthermore, the multi-scale frequency attention Transformer model includes a time-domain dynamic convolution layer, which predicts convolution filters according to the features of each sample; the time-domain dynamic convolution layer has multiple convolution kernels; among them, the multiple convolution kernels share the same kernel size and input / output dimensions; the multiple convolution kernels are aggregated by using attention weights. First, global average pooling is used to compress global spatial information, and two fully connected layers and a softmax function are used to generate normalized attention weights for multiple convolution kernels of the input vibration signal; the aggregation process of the multiple convolution kernels is described as: ; ; ; Among them, y represents the extracted features, represents the activation function, is the aggregated weight, represents the aggregated bias sharing the same attention weight, represents the input features, represents the k th weight of the convolution kernel, represents the k th bias of the convolution kernel, represents the k th attention score of the convolution kernel, , ; Furthermore, after the multiple convolution kernels are aggregated, multiple convolution kernels with different scales are used to filter out signal components at different levels from the complex signal. The multivariate signal components of the input data are encoded and coupled with global and local correlations; for multiple different convolution kernels, denoted as , then there is: ; Among them, represents the signal output by applying the i th convolution kernel, represents convolution, y represents the original input signal, represents the i th convolution kernel, represents the n th output value, represents the k th original input signal, represents the n - k th value in the convolution kernel; Furthermore, the multi-scale frequency attention Transformer model includes a global frequency encoding layer; the global frequency encoding layer transforms the input signal from the time domain to the frequency domain, captures long-range dependencies through a weight sharing mechanism in the frequency domain dependency, and deconstructs the signal into amplitude and phase information of different frequencies, enabling the multi-scale frequency attention Transformer model to understand data from a wide perspective; Furthermore, the deconstructing the signal into amplitude and phase information of different frequencies includes: using the global frequency encoding layer, for the input data, C denotes the number of input feature vectors, L denotes the length of the input feature vector, and the time domain signal is converted to the frequency domain signal using the one-dimensional fast Fourier transform, then there is: ; where, denotes the frequency domain signal after Fourier transform, denotes that the dimension of the output frequency domain signal is , denotes the number of input feature vectors, denotes the length of the input feature vector, denotes the result of the one-dimensional fast Fourier transform, which is a set of complex variables containing amplitude and phase information and is extracted from the signal through the following formula: ; ; where, denotes the amplitude information of the frequency domain signal, denotes the real part of the frequency domain signal, denotes the imaginary part of the frequency domain signal, denotes the phase information of the frequency domain signal; In the global frequency encoding layer, a learnable filter is introduced for frequency domain convolution. For two different sets of learnable weights, they are respectively denoted as and , and the two sets of learnable weights are regarded as learnable frequency filters for different hidden dimensions. The filter size is k and it maintains the same dimension as the input data, then there is: ; where, denotes the filtered feature, denotes the learnable filter, denotes the frequency domain signal, denotes the dot product operation, denotes that the dimension of the output signal is, , denotes the number of input feature vectors, Indicates the length of the input feature vector; Furthermore, the element-by-element multiplication between the converted data and the filter is expressed as: ; ; in, represents the amplitude signal, represents the learnable amplitude weight, represents the phase signal, represents the learnable phase weight; The transformed amplitude and phase information is multiplied by learnable amplitude and phase weights to globally adjust the frequency of the signal according to the task requirements. By enhancing the frequencies of interest and attenuating irrelevant frequencies, the different frequency characteristics of the time series data are extracted. The reconstructed complex signal is then converted back to the time domain through the inverse Fourier transform. S200, obtaining aircraft avionics system fault information vectors to form a training sample set; S300, using the training sample set to train the multi-scale frequency attention Transformer model to obtain a trained multi-scale frequency attention Transformer model; in the multi-scale frequency attention Transformer model, using the Adam optimizer to perform model training to accelerate model convergence and improve optimization effect; S400, the application phase, inputs the aircraft avionics system information into the trained multi-scale frequency attention Transformer model and outputs the fault information of the aircraft avionics system. Example

[0029] The following will be combined Figures 1 to 6 The technical solution of the present invention is further described in the contents of Figures 1 to 6 .

[0030] In a first aspect, this embodiment provides an aircraft avionics system fault diagnosis method, comprising: S100, improving the Transformer model to obtain a multi-scale frequency attention Transformer model; the multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism into its Transformer encoder architecture; the multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and implements data representation; wherein the multi-scale frequency attention mechanism performs calculations in the frequency domain, including: Converting an input time domain signal into a frequency domain signal, wherein the frequency domain signal contains amplitude and phase information; The encoder-decoder architecture is adopted to compress the channel representation vector into a hidden layer vector with a dimension of ; where, represents the number of input feature vectors, represents the length of the input feature vector; The hidden layer vector is decoded back to the original dimension to obtain the output channel weight vector; among them, the encoding operation is performed by the first convolutional layer, and the decoding operation is performed by the second convolutional layer; the first convolutional layer combines the ReLU function to provide a non-linear transformation function; the second convolutional layer uses the Sigmoid function to map the value to the range between 0 and 1; A learnable soft threshold parameter and a constant of 0 are introduced in the deep convolutional network based on the Fourier space to improve the model's attention to important information; The activation function tanh is used to map the eigenvalue to any range value, and the frequency domain signal is restored to the time domain signal through the inverse Fourier transform; A residual coefficient is introduced to adjust the distribution of attention; the residual connection enables the network to easily learn the identity mapping through the direct path across layers; The learnable parameter of the residual coefficient is introduced, enabling the model to effectively utilize the residual connection, learn the optimal coefficient, and enhance the model's expressive ability; Furthermore, the multi-scale frequency attention Transformer model includes a time-domain dynamic convolutional layer, and the time-domain dynamic convolutional layer predicts convolutional filters according to the features of each sample; the time-domain dynamic convolutional layer has multiple convolutional kernels; among them, the multiple convolutional kernels share the same kernel size and input / output dimension; the multiple convolutional kernels are aggregated by using attention weights. First, global average pooling is used to compress the global spatial information, and two fully connected layers and the softmax function are used to generate the normalized attention weights of the multiple convolutional kernels for the input vibration signal; the aggregation process of the multiple convolutional kernels is described as: ; ; ; where, y represents the extracted feature, represents the activation function, is the aggregated weight, represents the aggregated bias sharing the same attention weight, represents the input feature, represents the k th weight of the convolutional kernel, represents the k th bias of the convolutional kernel, represents the kThe attention score of a convolutional kernel, , ; Furthermore, after aggregating multiple convolutional kernels, convolutional kernels of multiple different scales are used to filter out signal components at different levels from complex signals. The multivariate signal components of the input data are encoded and coupled with global and local correlations; for multiple different convolutional kernels, denoted as , there are: ; wherein, represents the signal output by applying the i th convolutional kernel, represents convolution, y represents the original input signal, represents the i th convolutional kernel, represents the n th output value, represents the k rd original input signal, represents the n - k th value in the convolutional kernel; Furthermore, the multi-scale frequency attention Transformer model includes a global frequency encoding layer; the global frequency encoding layer transforms the input signal from the time domain to the frequency domain and captures long-range dependencies through a weight sharing mechanism in the frequency domain dependency, deconstructing the signal into amplitude and phase information of different frequencies, enabling the multi-scale frequency attention Transformer model to understand data from a wide perspective; Furthermore, the global frequency encoding layer deconstructs the signal into amplitude and phase information of different frequencies, including: for the input data, C represents the number of input feature vectors, L represents the length of the input feature vector. Using the one-dimensional fast Fourier transform to convert the time-domain signal into a frequency-domain signal, there are: ; wherein, represents the frequency-domain signal after Fourier transform, represents that the dimension of the output frequency-domain signal is , represents the number of input feature vectors, represents the length of the input feature vector, represents the result of the one-dimensional fast Fourier transform, which is a set of complex variables containing amplitude and phase information, extracted from the signal through the following formula: ; ; wherein, Represents the amplitude information of the frequency-domain signal, Represents the real part of the frequency-domain signal, Represents the imaginary part of the frequency-domain signal, Represents the phase information of the frequency-domain signal; In the global frequency encoding layer, a learnable filter is introduced for frequency-domain convolution. For two different sets of learnable weights, they are respectively denoted as and . The two sets of learnable weights are regarded as learnable frequency filters for different hidden dimensions. The filter size is k, and it maintains the same dimension as the input data. Then there is: ; where Represents the filtered feature, Represents the learnable filter, Represents the frequency-domain signal, Represents the dot product operation, Represents that the dimension of the output signal is , Represents the number of input feature vectors, Represents the length of the input feature vector; Furthermore, the element-wise multiplication between the transformed data and the filter is expressed as: ; ; where Represents the amplitude signal, Represents the learnable amplitude weight, Represents the phase signal, Represents the learnable phase weight; Multiply the obtained amplitude and phase information after transformation by the learnable amplitude and phase weights, globally adjust the frequency of the signal according to the task requirements, fully extract the different frequency feature information of the time series data by enhancing the frequencies of interest and attenuating the irrelevant frequencies, and then convert the reconstructed complex signal back to the time domain through the inverse Fourier transform: ; where Represents the time-domain feature after the inverse Fourier transform, Represents the result of performing the one-dimensional inverse fast Fourier transform on , Represents the feature of the signal after filtering in the frequency domain, Represents that the dimension of the output signal is , Represents the number of input feature vectors, Represents the length of the input feature vector, stands for one-dimensional Fourier inversion, or one-dimensional inverse fast Fourier transform. This process recovers the original time-domain signal from a complex signal encoded with global frequencies, helping the model preserve both local and global information. However, the model currently lacks the ability to extract local information. To enhance this capability, this embodiment introduces a multi-scale frequency attention mechanism, which aims to enable the model to focus more closely on local features in the input signal. S200, obtaining aircraft avionics system fault information vectors to form a training sample set; Furthermore, the above-mentioned aircraft avionics system fault information vector is the environment and system status data; specifically, in modern avionics systems (i.e., aircraft avionics systems), environmental perception and system status monitoring are crucial to flight safety and efficiency. Aircraft avionics systems can integrate a variety of high-performance sensors to collect environmental and system status data in real time, thereby achieving accurate perception of the aircraft's surroundings and accurate identification of their own status; these sensors include weather radars for monitoring meteorological conditions, GPS receivers for determining aircraft position and navigation, strain sensors for detecting the health of aircraft structures, and temperature and pressure sensors for monitoring aircraft engine performance; the multivariate signal samples generated by these sensors can be represented as a set ,in, Indicates the number of sensor signals, Indicates i sensor signals; due to sensor degradation or contamination and interference, sensor signals will have a variety of different failure modes. Through a large amount of signal acquisition and fault injection work, a sensor fault diagnosis dataset can be obtained, where D represents a dataset, N is the number of data samples, , indicating the i A multi-sensor signal, its corresponding fault label is , Represents a collection of multiple sensor signals, The dimension of the sensor signal is ; Further, preprocess the data collected by the sensors; in the data preprocessing stage, first clean and preliminarily process the collected raw data; specifically, the data can be normalized, and the data is scaled to a certain range; the continuous time series data is segmented into fixed-length segments for model processing; usually, the sliding window technique is used, and the window length is determined according to specific applications and data characteristics; the above process helps to capture the dynamic changes of the sensor signals, while reducing the amount of data and improving the computational efficiency of the model; identify and process outliers through statistical analysis or machine learning methods; for outliers, they may be caused by sensor failures, data transmission errors, etc., and these outliers will have a negative impact on model training. Algorithms such as threshold-based methods, clustering analysis, or isolation forests can be used to detect outliers and process them by replacement or deletion; S300. Use the training sample set to train the multi-scale frequency attention Transformer model to obtain a trained multi-scale frequency attention Transformer model; in the multi-scale frequency attention Transformer model, the Adam optimizer is used for model training to accelerate model convergence and improve the optimization effect; S400. In the application stage, input the avionics system information of the aircraft into the trained multi-scale frequency attention Transformer model to output the fault information of the avionics system of the aircraft; Further, please refer to Figure 2 , in the avionics system of an aircraft with a complex noise environment, a large amount of noise may hinder the identification of the fault mode characteristics of avionics equipment, which poses a major challenge to the health monitoring of avionics systems or equipment. In the above technical solution of the present invention, the time-frequency integration method is used to effectively suppress the influence of noise on feature extraction. The core design of the technical solution of the present invention aims to construct a model that integrates the time domain and the frequency domain to improve the efficiency of the network in capturing the correlation between frequency features. As Figure 2As shown, the overall architecture of the multi-scale frequency attention Transformer model of the present invention includes a global frequency encoding layer, a multi-scale frequency attention mechanism, and a temporal dynamic convolution layer. The temporal dynamic convolution dynamically adjusts the convolution kernel weights based on changes in input data features, thereby enhancing the model's adaptability to different input patterns and refining its data representation. The Transformer encoder layer comprises two main modules: a global frequency encoding layer and a multi-scale frequency attention mechanism (MSFA mechanism). These modules work together to facilitate accurate feature extraction and modeling. The global frequency encoding layer coordinates the fusion of multi-scale global information to reveal fault signatures, while the MSFA mechanism discerns the relevance and importance of different frequency components, focusing on relevant frequency information and enhancing the model's ability to discern local data features. By seamlessly integrating global and local information, the model is able to skillfully extract and model features across various scales and layers, thereby improving the model's accuracy and resilience. The temporal dynamic convolution, multi-scale frequency attention mechanism, and global frequency encoding layer of the present invention are described in detail below.

[0031] The time-domain dynamic convolution layer is the first key module of the Multi-Scale Frequency Attention Transformer model (MSFA-Trans) in this paper. It is responsible for extracting temporal features from the preprocessed time-domain signal. Unlike traditional convolution layers, this time-domain dynamic convolution layer can adaptively adjust the convolution kernel weights to better capture the temporal dynamic characteristics of the input signal. The traditional filter weight sharing mechanism inherently limits the ability of standard convolution in modeling semantic changes, such as Figure 2 As shown. Although standard convolution usually requires a significant increase in the number of filters in the channel dimension to capture different semantic information, the time-domain dynamic convolution layer introduced in this invention can predict the convolution filter according to the characteristics of each sample separately, which can encode various valuable information correlations in the input signal. Specifically, the time-domain dynamic convolution layer can have K convolution kernels that share the same kernel size and input / output dimensions; they are obtained by using attention weights. For aggregation, we first use global average pooling to compress the global spatial information and use two fully connected layers and softmax as the input vibration signal Generate the normalized attention weights of K convolution kernels, C Indicates the number of channels, L Indicates the data length, represents the adaptive attention weight; The output signal dimension is , denotes the number of input feature vectors, denotes the length of the input feature vector; in the process of aggregating K convolution kernels, in addition to decomposing multi-level signal components, the dynamic adaptation of the input data features using the learned feature representation is also achieved; further, the multivariate signal components of the input data will be encoded and coupled with the global and local correlations.

[0032] Global frequency encoding layer; for the field of fault detection, valuable features related to faults are hidden in the frequency-domain clock, so the present invention introduces the fast Fourier transform to encode the fault features in the frequency domain; the one-dimensional discrete Fourier transform is defined as follows: ; where, π represents the pi, k denotes the k th Fourier coefficient, n denotes the n th original transform signal, N denotes that there are a total of N points for discrete Fourier transform, is the input data, denotes the k th discrete Fourier transform coefficient, , j denotes the imaginary unit, , e is the natural constant; The signal can be decomposed into the amplitude and phase information of different frequency components through the one-dimensional discrete Fourier transform; in the time series model, the time series features are represented by a series of discrete features on multiple channels; for the input feature , the one-dimensional discrete Fourier transform processes the feature sequence of each channel in parallel to obtain the frequency-domain feature ; where, Z denotes the set of input features, Z C denotes the C th input feature, denotes that the dimension of the output signal is , denotes the number of input feature vectors, denotes the length of the input feature vector, Z F denotes the frequency-domain feature, denotes the discrete Fourier transform of the original signal Z C ; After the one-dimensional discrete Fourier transform, the digital signal in the frequency domain can be converted back to the time domain using the inverse Fourier transform; the formula for the inverse Fourier transform is as follows: ; where, represents the original time-domain signal, N represents having N discrete signals, n represents the n th discrete signal, represents the k th discrete Fourier coefficient, e is the natural constant, j represents the imaginary symbol, , π represents the pi, k represents the k th discrete Fourier coefficient; It should be noted that in the existing Transformer models, the self-attention layer has many limitations in dealing with long-range dependencies; since the self-attention layer mainly focuses on local context information, it cannot well capture the dependencies between more distant positions; for this reason, the technical solution of the present invention introduces a global frequency encoding layer, aiming to enhance the ability of the Transformer to better capture and process global feature correlations by operating in the frequency domain; the global frequency encoding layer in the present invention essentially transforms the input signal from the time domain to the frequency domain and captures long-range dependencies through the weight sharing mechanism in the frequency domain dependencies; this transformation is achieved through one-dimensional fast Fourier transform, deconstructing the signal into amplitude and phase information of different frequencies; this enables the model to understand the data from a broader perspective rather than just a local perspective, and the specific principle has been elaborated in detail in step S100 of Embodiment 3 of the present invention and will not be repeated here.

[0033] The present invention introduces a multi-scale frequency attention mechanism; the existing Transformer encoder captures long-range dependencies within the input sequence by stacking multiple layers, but this fixed multi-layer structure lacks flexibility and cannot adapt to the characteristics and task requirements of different input sequences. To address this problem, the present invention introduces a multi-scale frequency attention mechanism and integrates it into the traditional encoder architecture, as Figure 4 shown; this mechanism aims to dynamically capture the dependencies of different frequencies within the input sequence to enhance the feature representation, and by performing calculations in the frequency domain, transforms the input signal into amplitude and phase information, thereby achieving a novel data representation; first, perform a signal processing method on the input signal to transform it from the time domain to the frequency domain; then there is: ; where represents the signal after fast Fourier transform, represents the result after one-dimensional fast Fourier transform, represents that the dimension of the output signal is , represents the number of input feature vectors, denotes the length of the input feature vector; in the frequency-domain representation, the input signal is decomposed into amplitude and phase information, enabling the model to better understand the components of the signal at different frequencies; the amplitude information can be represented as , and the phase information can be represented as .

[0034] Furthermore, an encoder-decoder architecture is adopted to compress the channel representation vector into a hidden layer vector with a dimension of , and then these vectors are decoded back to the original dimension to obtain the output channel weight vector; the encoding and decoding operations of the multi-scale frequency attention mechanism are performed by two convolutional layers; the first convolutional layer incorporates the ReLU function to provide a non-linear transformation function, and the second convolutional layer uses the Sigmoid function to map the values to the range between 0 and 1; the mathematical formula is as follows: ; ; where Z represents the compressed data features, W 1 represents the weight coefficient of the first layer of adaptation, represents the amplitude feature of the frequency-domain signal, b 1 represents the bias of the first layer, represents that the dimension of the compressed signal is , represents the multiplication sign, represents the signal features after restoring the compressed data, W 2 represents the weight coefficient of the second layer of adaptation, b 2 represents the bias of the second layer, represents that the dimension of the output frequency-domain signal is , represents the number of input feature vectors, represents the length of the input feature vector.

[0035] It is set that in the frequency domain, the frequency-domain coefficient of the noise component represents the frequency-domain coefficient of the observed signal containing the noise component, and the frequency-domain coefficient of the fault feature signal represents the frequency-domain coefficient of the signal containing the fault-related features. The frequency-domain coefficient of the noise component is set to 1 to explain the existence of noise, and the frequency-domain coefficient of the fault feature signal is set to 0 to assume that there are no fault-related features at the beginning; these coefficients are not fixed but are automatically learned and optimized through the training process of the model; the formula is expressed as: ; where represents the amplitude signal in the frequency-domain signal, represents the frequency-domain coefficient of the noise component, The frequency domain coefficient representing the fault characteristic signal.

[0036] It should be noted that for fault diagnosis applications, the goal of the multi-scale frequency attention Transformer model of the present invention is to eliminate noise and irrelevant information as much as possible while retaining high-value frequency components related to the fault. The noise interference commonly encountered by aircraft equipment has characteristics similar to Gaussian white noise, and its frequency domain representation is random at all frequencies, but the average value is constant. Therefore, the present invention introduces a learnable soft threshold parameter threshold_min and a constant threshold_0 that is constant to 0 in the deep convolutional network based on Fourier space; the formula is as follows: ; ; in, represents the feature after soft threshold denoising of the attention weight, Indicates the features of the original space without soft threshold denoising, Indicates that the control threshold is always greater than 0, represents a positive natural exponent, represents the negative natural exponential; tanh (•) represents the activation function.

[0037] Furthermore, in order to enhance the model's nonlinear expression ability of features, the activation function tanh is used to map the eigenvalues to a range, and finally the frequency domain signal is restored to the time domain through the inverse Fourier transform; in order to prevent the model from over-relying on local information, the residual coefficient is introduced to adjust the distribution of attention. The residual connection enables the network to easily learn the identity mapping through a direct path across layers, thereby better capturing the key features of the input sequence; the introduction of learnable parameters of the residual coefficient enables the model to effectively utilize the residual connection and learn the optimal coefficient, thereby improving the model's expression ability; the specific formula is as follows: ; in, Indicates that the signal characteristics after processing become the time domain signal, Represents one-dimensional inverse Fourier transform, which converts frequency domain signals into time domain signals, restores the shape and characteristics of the original signal, and facilitates subsequent analysis of the signal in the time domain. represents the calculated attention weight, represents the amplitude characteristics of the frequency domain signal, represents the learnable residual coefficient, represents the amplitude characteristics of the frequency domain signal, Represents the phase characteristics of the frequency domain signal.

[0038] Furthermore, when training and optimizing the above-mentioned multi-scale frequency attention Transformer model, it can be as follows: The fault classifier consists of a convolutional module, an adaptive average pooling layer, and a fully connected layer. Finally, the cross-entropy loss function is used to measure the distance between the categories predicted by the multi-scale frequency attention and the actual distribution, which can be expressed as: ; where represents the calculation of the cross-entropy loss for the predicted labels and the actual labels, N represents the number of samples, K represents the number of categories, P represents the target distribution, Q represents the estimated distribution, represents the k th category and the target distribution of the n th sample, represents the k th category and the estimated distribution of the n th sample; The Adam optimizer is used for model training; The Adam optimizer combines the advantages of momentum and adaptive learning rate, and can effectively accelerate model convergence and improve the optimization effect; This optimization process can be carried out on a personal computer equipped with an NVIDIA GTX 3050 GPU (8GB memory) and an AMD Ryzen 5 5600H processor. During the model training process, the Adam optimization algorithm is adopted, the learning rate is set to 0.0001, and the Batch Size is set to 64.

[0039] Furthermore, the present invention uses the data collected by a certain university using a 24-pulse autotransformer rectifier. The 24-pulse autotransformer rectifier (Auto-Transformer Rectifier Unit, ATRU) consists of a transformer, a main rectifier bridge module, and three rectifier bridge modules. Its basic topological structure is as Figure 5 shown. The physical photo of the experimental 24-pulse ATRU of the present invention is as Figure 6 shown, and in Figure 6Among them, "transformer" represents the Transformer architecture, and "rectifier bridge modules" represents the rectifier bridge modules. The experimental 24-pulse ATRU adopts a three-phase input design of 115V / 400Hz and consists of a transformer, a main rectifier bridge module, three auxiliary rectifier bridge modules, a radiator, input / output terminals, and internal leads. The experimental 24-pulse ATRU of the present invention includes 12 external terminals, among which three terminals are the input terminals of the ATRU and also the input terminals of the main rectifier bridge module; the remaining terminals are the input terminals of the three auxiliary rectifier bridge modules. The experiment of the present invention was carried out on an experimental ATRU fault simulator, and a total of 15 types of faults were collected. The specific fault types are shown in Table 3. The core of the present invention is to realize the fault diagnosis of the aircraft avionics system. Fault diagnosis is defined as a multi-classification problem. Therefore, the present invention uses Accuracy and F1-Score to measure the performance of the diagnosis model. The larger the values of Accuracy and F1-Score, the better the performance of the model. Then, there is: ; ; ; ; Among them, Accuracy represents the proportion of correctly predicted labels in the total correct label values, Precision represents the proportion of correctly predicted positive label values in all predicted positive label values, Recall represents the proportion of correctly predicted positive label values in all actual positive label values FP , FN , TN , TP respectively represent the number of samples of false positives, false negatives, true negatives, and true positives, F1 - score represents the harmonic mean of precision and recall; in order to deeply explore the influence of the GFE-Layer and MSFA mechanisms on the model performance, three different network architectures were designed and compared in the experiment: CNN-Trans, GFE-Trans, and MSFA-Trans; among them, CNN-Trans only uses convolutional layers and fully connected layers to form a basic deep learning framework; GFE-Trans introduces a global frequency encoding layer (GFE-Layer) on this basis, aiming to improve the feature extraction ability through global frequency domain information interaction; while MSFA-Trans further integrates the MSFA mechanism to enhance the model's attention to multi-scale frequency features and noise suppression ability; through ablation experiments, the contribution of each component to the overall performance was evaluated one by one, so as to provide a basis for model optimization.

[0040] The experimental results demonstrate the performance of each architecture under noise-free conditions. Specifically, the accuracy of CNN-Trans under noise-free conditions is 84.36%. Although this result reflects its basic classification ability to a certain extent, it also reveals its limitations in identifying complex fault features. In contrast, the accuracy of GFE-Trans is significantly improved to 92.84%, fully demonstrating the effectiveness of the GFE-Layer in frequency-domain feature extraction; by converting the signal from the time domain to the frequency domain and using the Fast Fourier Transform (FFT) to capture the global frequency distribution features, more abundant information is provided for the model. More prominently, the accuracy of MSFA-Trans is further improved to 93.96%, indicating that the MSFA mechanism can effectively enhance the model's attention to key frequency components while suppressing noise interference, thus performing excellently in the recognition of complex fault patterns; for the specific experimental results, please refer to Table 1 below. These results not only verify the effectiveness of the GFE-Layer and the MSFA mechanism in improving the model's performance but also provide important references for the selection and optimization of the model in practical applications. By introducing frequency-domain analysis and attention mechanisms, the accuracy and robustness of fault diagnosis can be significantly improved in a noisy environment, providing reliable technical support for the health monitoring and fault diagnosis of aircraft avionics systems.

[0041] Table 1 Experimental Results of GFE-Trans, CNN-Trans, and MSFA-Trans

[0042] Furthermore, in the present invention, MSFA-Trans was compared with seven other existing methods that performed relatively well in the fault diagnosis of avionics systems. These seven existing methods are VGG-16, CNN_Transformer, SECN, FFT-Trans, Signal-Transformer, Mexhat_AlexNet, and Morlet_AlexNet. The prior art discloses Mexhat-Net and Morlet-Net for bearing fault diagnosis, which integrate different wavelet basis functions into the convolutional layer of the first deep model to improve signal decomposition and feature extraction. VGG-16 uses small-sized convolutional kernels and pooling layers to increase the network depth and receptive field, which helps in fault detection. CNN_Transformer utilizes the self-attention mechanism of Transformer to capture long-range dependencies in time series data and combines it with the local feature extraction ability of CNN. SECN effectively addresses distribution differences and noise interference through the combination of sequential embedding and multi-scale convolution. The Signal-Transformer model achieves efficient and interpretable intelligent fault diagnosis under different operating conditions through signal embedding, multi-head self-attention mechanism, and attention visualization method. As shown in Table 2, MSFA-Trans shows obvious advantages; the experimental results under noise-free conditions are shown in Table 2, where MSFA-Trans can stably reach over 93% and can accurately diagnose the fault modes of avionics systems. This excellent performance benefits from the design of MSFA-Trans in the model architecture and its deep mining ability for fault features, making it perform outstandingly in the complex and changing operating environment of avionics systems and providing more powerful technical support for the reliable operation and fault prevention of avionics systems. The present invention further analyzes the performance of MSFA-Trans. The ratio of classifying faults as normal is 6.8% (the number of faults classified as normal / the total number of faults), and the false alarm rate is 4.7% (the number of normal cases classified as faults / the total number of normal cases). This excellent performance benefits from the design of MSFA-Trans in the model architecture and its deep mining ability for fault features, making it stand out in the complex and changing operating environment of avionics systems and providing more powerful technical support for the reliable operation and fault prevention of avionics systems.

[0043] Table 2 Experimental Results of VGG-16, CNN_Transformer, Mexhat_AlexNet, Secn, Signal_Transformer, Msfa-Trans, Fft-Trans, and Morlet_AlexNet

[0044] Table 3 Fault Types

[0045] In a second aspect, this embodiment provides an aircraft avionics system fault diagnosis system. The Transformer model is improved to obtain a multi-scale frequency attention Transformer model. The multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism in the Transformer encoder architecture. The multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and realizes data representation. The aircraft avionics system fault diagnosis system includes: A training module, configured to: obtain a training sample set composed of aircraft avionics system fault information vectors; train the multi-scale frequency attention Transformer model using the training sample set to obtain a trained multi-scale frequency attention Transformer model; during the training process of the multi-scale frequency attention Transformer model, the Adam optimizer is used for model training to accelerate model convergence and improve the optimization effect; An application module, configured to: input aircraft avionics system information into the trained multi-scale frequency attention Transformer model and output the fault information of the aircraft avionics system.

[0046] In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in a suitable manner in any one or more embodiments or examples.

[0047] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A fault diagnosis method for an aircraft avionics system, characterized in that, Improve the Transformer model to obtain a multi-scale frequency attention Transformer model; the multi-scale frequency attention Transformer model integrates a multi-scale frequency attention mechanism in its Transformer encoder architecture; the multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and realizes data representation; The aircraft avionics system fault diagnosis method includes: In the training stage, obtain the fault information vectors of the aircraft avionics system to form a training sample set; Use the training sample set to train the multi-scale frequency attention Transformer model to obtain a trained multi-scale frequency attention Transformer model; in the multi-scale frequency attention Transformer model, use the Adam optimizer for model training to accelerate model convergence and improve the optimization effect; In the application stage, input the aircraft avionics system information into the trained multi-scale frequency attention Transformer model, and output the fault information of the aircraft avionics system.

2. The aircraft avionics system fault diagnosis method according to claim 1, characterized in that: The multi-scale frequency attention mechanism performs calculations in the frequency domain, including: Transform the input time-domain signal into a frequency-domain signal, and the frequency-domain signal contains amplitude and phase information; The encoder-decoder architecture is used to compress the channel representation vector into a dimension of The hidden layer vector of ; where, represents the number of input feature vectors, Indicates the length of the input feature vector; Decode the hidden layer vector back to the original dimension to obtain an output channel weight vector; among them, the encoding operation is performed by the first convolutional layer, and the decoding operation is performed by the second convolutional layer; the first convolutional layer provides a non-linear transformation function in combination with the ReLU function; the second convolutional layer uses the Sigmoid function to map the value to the range between 0 and 1; Introduce a learnable soft threshold parameter and a constant of 0 in the deep convolutional network based on the Fourier space to improve the model's attention to important information; Use the activation function tanh to map the eigenvalue to any range value, and restore the frequency-domain signal to the time-domain signal through the inverse Fourier transform.

3. The method for diagnosing faults in the aircraft avionics system according to claim 2, wherein, The multi-scale frequency attention mechanism performs calculations in the frequency domain, and also includes: Introduce a residual coefficient to adjust the distribution of attention; the residual connection enables the network to easily learn the identity mapping through the direct path across layers; Introduce learnable parameters of the residual coefficient, so that the model can effectively utilize the residual connection, learn the optimal coefficient, and improve the model's expression ability.

4. The method for fault diagnosis of an aircraft avionics system according to claim 3, wherein The multi-scale frequency attention Transformer model includes a time-domain dynamic convolutional layer; The time-domain dynamic convolutional layer predicts the convolutional filter according to the characteristics of each sample; The time-domain dynamic convolutional layer has multiple convolutional kernels; Among them, multiple convolutional kernels share the same kernel size and input / output dimensions; multiple convolutional kernels are aggregated by using attention weights. First, use global average pooling to compress the global spatial information, and use two fully connected layers and the softmax function to generate normalized attention weights for multiple convolutional kernels of the input vibration signal.

5. The aircraft avionics system fault diagnosis method according to claim 4, characterized in that: The aggregation process of the multiple convolutional kernels is described as: ； ； ； Among them, y represents the extracted feature, represents the activation function, is the aggregated weight, represents the aggregated bias sharing the same attention weight, represents the input feature, represents the k weight of the th convolutional kernel, k represents the bias of the th convolutional kernel, k represents the attention score of the , .

6. The aircraft avionics system fault diagnosis method according to claim 5, characterized in that: After aggregating multiple convolutional kernels, convolutional kernels of multiple different scales are used to filter out signal components at different levels from complex signals. The multivariate signal components of the input data are encoded and coupled with global and local correlations; for multiple different convolutional kernels, denoted as , there is: ； in, Indicates application i The signal output by the convolution kernel is represents convolution, y represents the original input signal, Indicates the i convolution kernels, Indicates the n output values, Indicates the k The original input signal, Represents the convolution kernel n-k values.

7. The method for fault diagnosis of the aircraft avionics system according to claim 6, characterized in that, The multi-scale frequency attention Transformer model includes a global frequency encoding layer; The global frequency encoding layer transforms the input signal from the time domain to the frequency domain, captures long-range dependencies through the weight sharing mechanism in the frequency domain dependencies, decomposes the signal into amplitude and phase information of different frequencies, enabling the multi-scale frequency attention Transformer model to understand the data from a wide perspective.

8. The method for fault diagnosis of an aircraft avionics system according to claim 7, characterized in that, The decomposing the signal into amplitude and phase information of different frequencies includes: Using the global frequency encoding layer, for the input data, C represents the number of input feature vectors, L represents the length of the input feature vector. Using the one-dimensional fast Fourier transform to convert the time-domain signal into a frequency-domain signal, we have: ； Among them, represents the frequency-domain signal after Fourier transform, indicates that the dimension of the output frequency-domain signal is , represents the number of input feature vectors, represents the length of the input feature vectors, represents the result of one-dimensional fast Fourier transform, which is a set of complex variables containing amplitude and phase information and is extracted from the signal by the following formula: ； ； Among them, represents the amplitude information of the frequency-domain signal, represents the real part of the frequency-domain signal, represents the imaginary part of the frequency-domain signal, represents the phase information of the frequency-domain signal; A learnable filter is introduced into the global frequency coding layer to perform frequency domain convolution. For two different sets of learnable weights, they are respectively denoted as and , the two sets of learnable weights act as learnable frequency filters for different hidden dimensions, and the filter size is k , and keep the same dimension as the input data, then: ; among them, represents the filtered feature, represents the learnable filter, represents the frequency-domain signal, represents the dot product operation, represents that the dimension of the output signal is , represents the number of input feature vectors, represents the length of the input feature vector.

9. The aircraft avionics system fault diagnosis method according to claim 8, characterized in that: The element-wise multiplication between the transformed data and the filter is expressed as: ； ； in, represents the amplitude signal, represents the learnable amplitude weight, represents the phase signal, represents the learnable phase weight; Multiply the obtained amplitude and phase information by learnable amplitude and phase weights to globally adjust the frequency of the signal according to the task requirements. By enhancing the frequencies of interest and attenuating the irrelevant frequencies, different frequency feature information of the time series data is extracted. Subsequently, the reconstructed complex signal is transformed back to the time domain through the inverse Fourier transform.

10. An aircraft avionics system fault diagnosis system, characterized in that, The multi-scale frequency attention Transformer model is obtained by improving the Transformer model. The multi-scale frequency attention Transformer model integrates the multi-scale frequency attention mechanism in the Transformer encoder architecture. The multi-scale frequency attention mechanism performs calculations in the frequency domain, transforms the input signal into amplitude and phase information, and realizes data representation. The aircraft avionics system fault diagnosis system includes: A training module, configured to: obtain a training sample set composed of aircraft avionics system fault information vectors; train the multi-scale frequency attention Transformer model using the training sample set to obtain a trained multi-scale frequency attention Transformer model; during the training process of the multi-scale frequency attention Transformer model, the Adam optimizer is used for model training to accelerate model convergence and improve the optimization effect; An application module, configured to: input the aircraft avionics system information into the trained multi-scale frequency attention Transformer model and output the fault information of the aircraft avionics system.

Images