A confidence fusion-based fault diagnosis method for an airborne radio frequency system filter circuit

Abstract

The application discloses a confidence fusion-based fault diagnosis method for a filter circuit of an airborne radio frequency system, and belongs to the technical field of analog circuit fault diagnosis. Firstly, output end voltage data sets in a healthy state and in each fault state are collected, time domain data is converted into frequency domain data through FFT, and then logarithmic compression is performed on the frequency domain data, so that the time domain data is normalized; a 1D-CNN model is constructed, time domain data is used for fault diagnosis to obtain a time domain probability vector; a Transformer model is constructed, frequency domain data is used for fault diagnosis to obtain a frequency domain probability vector; a fusion probability vector is obtained through weighted fusion, a fault category corresponding to the maximum probability in the fusion probability vector is selected as a diagnosis result; and a training set is used for training, and a test set is used for verifying effectiveness. The application adopts time-frequency dual domain feature extraction to capture multi-dimensional representation of the circuit state, combines 1D-CNN local feature extraction with Transformer global correlation, reduces the consumption of computing resources, and greatly improves the accuracy and robustness of the diagnosis result.

Description

Technical Field

[0001] This invention belongs to the field of analog circuit fault diagnosis technology, specifically relating to a fault diagnosis method for airborne radio frequency system filter circuits based on confidence fusion. Background Technology

[0002] Airborne radio frequency (RF) systems are crucial for aircraft to achieve communication, radar, and navigation functions, and are key systems for information exchange between aircraft and between aircraft and ground stations. Filtering circuits play a vital role in RF systems. During wireless communication, they act as frequency band selectors, eliminating noise interference, filtering harmonics, and suppressing clutter, which is essential for the successful execution of aircraft missions. However, due to the complexity of the aircraft operating environment and the natural aging of components, circuit failures are highly likely. Therefore, research into fault diagnosis methods for airborne RF system filtering circuits is essential to ensure flight safety and reduce the occurrence of dangerous situations.

[0003] Deep learning, with its efficient feature extraction and powerful classification capabilities, has been widely used in fault diagnosis. However, it requires a large number of training samples and significant computational resources, while actual circuit fault samples are relatively few. Furthermore, a single deep learning model often struggles to simultaneously consider both local and global features, easily leading to misdiagnosis. Therefore, a fault diagnosis method that considers both local and global features is needed to ensure the normal operation of airborne radio frequency system filtering circuits. Summary of the Invention

[0004] To overcome the aforementioned problems, this invention proposes a fault diagnosis method for airborne radio frequency system filter circuits based on confidence fusion, belonging to the field of analog circuit fault diagnosis technology. First, output voltage datasets under circuit health and various fault states are collected. The time-domain data is converted to frequency-domain data using FFT, followed by logarithmic compression and normalization. A 1D-CNN model is constructed, using the time-domain data for fault diagnosis to obtain a time-domain probability vector. A Transformer model is then constructed, using the frequency-domain data for fault diagnosis to obtain a frequency-domain probability vector. A fused probability vector is obtained through weighted fusion, and the fault category corresponding to the highest probability in the fused probability vector is selected as the diagnosis result. The method is trained using a training set and validated using a test set. This invention employs time-frequency dual-domain feature extraction to capture multi-dimensional representations of circuit states; combining 1D-CNN local feature extraction with Transformer global correlation reduces computational resource consumption and significantly improves the accuracy and robustness of the diagnostic results.

[0005] A method for fault diagnosis of airborne radio frequency system filter circuits based on confidence fusion, such as Figure 1 As shown, it includes the following steps:

[0006] Step 1: Collect the output voltage data set of the circuit's health status and various fault states;

[0007] Step 2: Process the collected voltage dataset;

[0008] The data processing procedure is as follows:

[0009] Step 2.1: Normalize the time-domain data in the voltage dataset to obtain the voltage time-domain dataset;

[0010] Step 2.2: Perform frequency domain preprocessing on the time domain data in the voltage dataset to obtain the voltage frequency domain dataset;

[0011] Step 2.3: Divide the voltage time domain dataset and the voltage frequency domain dataset into training set and test set in a 7:3 ratio, and add labels for circuit fault states to the training set samples;

[0012] Step 3: Construct a 1D-CNN model (one-dimensional convolutional neural network model), and use the normalized voltage time-domain data to perform fault diagnosis through the 1D-CNN model to obtain the time-domain probability vector;

[0013] Step 4: Construct a Transformer model and use the frequency domain data to perform fault diagnosis through the Transformer model to obtain the frequency domain probability vector;

[0014] Step 5: Use a weighted fusion method to perform decision-level fusion of the time-domain probability vector and the frequency-domain probability vector to obtain a fused probability vector; select the fault category corresponding to the highest probability in the fused probability vector as the final diagnosis result;

[0015] Step 6: The 1D-CNN model, Transformer model, and decision-level fusion constitute a two-branch confidence fusion fault diagnosis model; the two-branch confidence fusion fault diagnosis model is trained using training set samples from the voltage dataset;

[0016] Step 7: Validate the effectiveness of the dual-branch confidence fusion fault diagnosis model using test set samples.

[0017] Furthermore, in step 1, the fault states include resistor value drift, capacitor failure, and operational amplifier saturation. Resistor value drift, capacitor failure, and operational amplifier saturation all refer to the failure of a single component. Resistor value drift refers to the rise or fall of the value of a single resistor component by more than 10%. Capacitor failure refers to a short circuit in a single capacitor component. Operational amplifier saturation refers to the phenomenon that the output of the operational amplifier cannot continue to follow the input change after the input signal amplitude is too large, the driving capability of the operational amplifier is insufficient, or the supply voltage of the operational amplifier drops.

[0018] Furthermore, in step 2, the normalization equation is as follows:

[0019] (1);

[0020] in, This refers to the time-domain data signal in the voltage dataset; It is the minimum value in the voltage dataset; The maximum value in the voltage dataset; The normalized voltage time-domain data signal, The value ranges from 0 to 1;

[0021] The frequency domain preprocessing process is as follows:

[0022] Step 2.2.1: Use Fast Fourier Transform (FFT) to convert the time-domain data into frequency-domain data;

[0023] The window applied in the Fast Fourier Transform is the Hanning window. The Fast Fourier Transform is performed on the windowed signal to obtain the frequency domain data in complex form. The amplitude spectrum of the voltage frequency domain dataset is obtained by squaring the real and imaginary parts of the frequency domain data and then summing and taking the square root.

[0024] Step 2.2.2: Perform logarithmic compression and standardization on the amplitude spectrum of the voltage frequency domain dataset;

[0025] First, logarithmic compression is performed to enhance the weak signal, resulting in compressed data L[k]:

[0026] (2);

[0027] Where M[k] represents the amplitude spectrum, ; , Used to prevent zero values;

[0028] Then, the compressed data L[k] is standardized to obtain the voltage frequency domain data F[k]:

[0029] (3);

[0030] in, Let L[k] be the mean of the compressed data; The standard deviation of the compressed data is given; the frequency domain data F[k] is used as the input to the Transformer model.

[0031] Furthermore, in step 3, the 1D-CNN model includes an input layer and a classification output layer; as follows: Figure 2As shown, the input layer sequentially includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, and a third max pooling layer; the kernel size of the three convolutional layers is 5, and the pooling window size of each max pooling layer is 2; the third max pooling layer is connected to the classification output layer; the classification output layer includes two fully connected layers, the first fully connected layer is used to receive the fused feature information, the second fully connected layer uses the softmax function for fault classification, and a dropout layer with a dropout rate of 0.5 is added between the two fully connected layers to prevent the model from overfitting.

[0032] Furthermore, in step 4, the Transformer model is a lightweight model, such as... Figure 3 As shown, the Transformer model sequentially includes an input layer, an embedding layer, a positional encoding layer, a layer with added CLS classification tokens, a Transformer encoder layer, and a linear classification output layer. The Transformer encoder layer is a three-layer encoder, consisting of a first encoder layer, a second encoder layer, and a third encoder layer. The classification output layer is a two-layer fully connected layer, consisting of a first fully connected layer and a second fully connected layer. Each Transformer encoder layer has a four-head self-attention mechanism, a 256-dimensional feedforward network, and a GELU activation function. The Transformer encoder layer is trained stably through residual connections and layer normalization, extracts the state vector with added CLS tokens, and inputs it to the linear classification output layer for fault classification after layer normalization.

[0033] Furthermore, in step 5, the decision-level fusion process is as follows:

[0034] Step 5.1: Calculate the confidence level of the temporal probability vector output by the 1D-CNN model. :

[0035] (4);

[0036] in, This represents the temporal probability vector output by the 1D-CNN model. The information entropy represents the temporal probability vector output by the 1D-CNN model. The formula for calculating information entropy is as follows:

[0037] (5);

[0038] Where n represents the total number of elements in the time-domain probability vector; Represents the elements in the time-domain probability vector. ;

[0039] Similarly, the confidence level of the frequency domain probability vector is calculated using the method in step 5.1. ;

[0040] Step 5.2: Calculate the confidence level of the time-domain probability vector. confidence level of frequency domain probability vector The fusion weights are used to obtain the weights of the time-domain probability vector. Weights of the frequency domain probability vector ;

[0041] The fusion weights are calculated as follows:

[0042] (6);

[0043] (7);

[0044] Step 5.3, weighting the time-domain probability vector Weights of the frequency domain probability vector Weighted fusion is performed to obtain the fusion probability vector.

[0045] The technical effects of this invention are as follows:

[0046] This invention employs time-frequency dual-domain feature extraction to capture multi-dimensional representations of circuit states, and utilizes confidence fusion to achieve decision-level fusion for fault diagnosis of airborne radio frequency system filter circuits. The dual-branch structure complements the ability of 1D-CNN to extract local features with the ability of Transformer to correlate global features, overcoming the problems of deep learning methods being unable to take into account both local and global features and consuming large computational resources, thus greatly improving the accuracy and robustness of diagnostic results. Attached Figure Description

[0047] Figure 1 This is an overall flowchart of the present invention;

[0048] Figure 2 This is a schematic diagram of the temporal branch 1D-CNN model in this invention;

[0049] Figure 3 This is a schematic diagram of the frequency domain branch Transformer model in this invention. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0051] A method for fault diagnosis of airborne radio frequency system filter circuits based on confidence fusion, such as Figure 1 As shown, it includes the following steps:

[0052] Step 1: Collect the output voltage data set of the circuit's health status and various fault states;

[0053] The fault states include resistor drift, capacitor failure, and operational amplifier saturation. Resistor drift, capacitor failure, and operational amplifier saturation all refer to the failure of a single component. Resistor drift refers to the rise or fall of the value of a single resistor component exceeding 10%. Capacitor failure refers to a short circuit in a single capacitor component. Operational amplifier saturation refers to the phenomenon that the output of the operational amplifier cannot continue to follow the input changes after exceeding a certain value due to excessive input signal amplitude, insufficient operational amplifier driving capability, or a drop in operational amplifier supply voltage.

[0054] Step 2: Process the collected voltage dataset;

[0055] The data processing procedure is as follows:

[0056] Step 2.1: Normalize the time-domain data in the voltage dataset to obtain the voltage time-domain dataset;

[0057] The normalization equation is as follows:

[0058] (1);

[0059] in, This refers to the time-domain data signal in the voltage dataset; It is the minimum value in the voltage dataset; The maximum value in the voltage dataset; The normalized voltage time-domain data signal, The value ranges from 0 to 1;

[0060] Step 2.2: Perform frequency domain preprocessing on the time domain data in the voltage dataset to obtain the voltage frequency domain dataset;

[0061] The frequency domain preprocessing process is as follows:

[0062] Step 2.2.1: Use Fast Fourier Transform (FFT) to convert the time-domain data into frequency-domain data;

[0063] The window applied in the Fast Fourier Transform is the Hanning window. The Fast Fourier Transform is performed on the windowed signal to obtain the frequency domain data in complex form. The amplitude spectrum of the voltage frequency domain dataset is obtained by squaring the real and imaginary parts of the frequency domain data and then summing and taking the square root.

[0064] Step 2.2.2: Perform logarithmic compression and standardization on the amplitude spectrum of the voltage frequency domain dataset;

[0065] First, logarithmic compression is performed to enhance the weak signal, resulting in compressed data L[k]:

[0066] (2);

[0067] Where M[k] represents the amplitude spectrum, ; , Used to prevent zero values;

[0068] Then, the compressed data L[k] is standardized to obtain the voltage frequency domain data F[k]:

[0069] (3);

[0070] in, Let L[k] be the mean of the compressed data; The standard deviation of the compressed data is given; the frequency domain data F[k] is used as the input to the Transformer model.

[0071] Step 2.3: Divide the voltage time domain dataset and the voltage frequency domain dataset into training set and test set in a 7:3 ratio, and add labels for circuit fault states to the training set samples;

[0072] Step 3: Construct a 1D-CNN model (one-dimensional convolutional neural network model), and use the normalized voltage time-domain data to perform fault diagnosis through the 1D-CNN model to obtain the time-domain probability vector;

[0073] The 1D-CNN model includes an input layer and a classification output layer; as follows: Figure 2 As shown, the input layer sequentially includes a first convolutional layer, a first max pooling layer, a second convolutional layer, a second max pooling layer, a third convolutional layer, and a third max pooling layer; the kernel size of the three convolutional layers is 5, and the pooling window size of each max pooling layer is 2; the third max pooling layer is connected to the classification output layer; the classification output layer includes two fully connected layers, the first fully connected layer is used to receive the fused feature information, the second fully connected layer uses the softmax function for fault classification, and a dropout layer with a dropout rate of 0.5 is added between the two fully connected layers to prevent the model from overfitting.

[0074] Step 4: Construct a Transformer model and use the frequency domain data to perform fault diagnosis through the Transformer model to obtain the frequency domain probability vector;

[0075] The Transformer model is a lightweight model, such as... Figure 3As shown, the Transformer model sequentially includes an input layer, an embedding layer, a positional encoding layer, a layer with added CLS classification tokens, a Transformer encoder layer, and a linear classification output layer. The Transformer encoder layer is a three-layer encoder, consisting of a first encoder layer, a second encoder layer, and a third encoder layer. The classification output layer is a two-layer fully connected layer, consisting of a first fully connected layer and a second fully connected layer. Each Transformer encoder layer has a four-head self-attention mechanism, a 256-dimensional feedforward network, and a GELU activation function. The Transformer encoder layer is trained stably through residual connections and layer normalization, extracts the state vector with added CLS tokens, and inputs it to the linear classification output layer for fault classification after layer normalization.

[0076] Step 5: Use a weighted fusion method to perform decision-level fusion of the time-domain probability vector and the frequency-domain probability vector to obtain a fused probability vector; select the fault category corresponding to the highest probability in the fused probability vector as the final diagnosis result;

[0077] The decision-level fusion process is as follows:

[0078] Step 5.1: Calculate the confidence level of the temporal probability vector output by the 1D-CNN model. :

[0079] (4);

[0080] in, This represents the temporal probability vector output by the 1D-CNN model. The information entropy of the temporal probability vector output by the 1D-CNN model is calculated using the following formula:

[0081] (5);

[0082] Where n represents the total number of elements in the time-domain probability vector; Represents the elements in the time-domain probability vector. ;

[0083] Similarly, the confidence level of the frequency domain probability vector is calculated using the method in step 5.1. ;

[0084] Step 5.2: Calculate the confidence level of the time-domain probability vector. confidence level of frequency domain probability vector The fusion weights are used to obtain the weights of the time-domain probability vector. Weights of the frequency domain probability vector ;

[0085] The fusion weights are calculated as follows:

[0086] (6);

[0087] (7);

[0088] Weights of the time-domain probability vector Weights of the frequency domain probability vector Weighted fusion is performed to obtain the fusion probability vector.

[0089] Step 6: The 1D-CNN model, Transformer model, and decision-level fusion constitute a two-branch confidence fusion fault diagnosis model; the two-branch confidence fusion fault diagnosis model is trained using training set samples from the voltage dataset;

[0090] Step 7: Validate the effectiveness of the dual-branch confidence fusion fault diagnosis model using test set samples.

Claims

1. A method for diagnosing the fault of an airborne radio frequency system filter circuit based on confidence fusion, characterized in that, The filter circuit fault diagnosis method comprises the following steps: Step 1, collect the output voltage data set of the circuit in the healthy state and each fault state; Step 2, data processing on the collected voltage data set; The data processing process is as follows: Step 2.1, normalize the time domain data in the voltage data set to obtain the voltage time domain data set; Step 2.2, pre-process the time domain data in the voltage data set in the frequency domain to obtain the voltage frequency domain data set; Step 2.3, divide the voltage time domain data set and the voltage frequency domain data set into a training set and a test set according to a ratio of 7:3, and add labels of circuit fault states to the training set samples; Step 3, construct a 1D-CNN model, and use the normalized voltage time domain data to perform fault diagnosis through the 1D-CNN model to obtain a time domain probability vector; Step 4, construct a Transformer model, and use the frequency domain data to perform fault diagnosis through the Transformer model to obtain a frequency domain probability vector; Step 5, use a weighted fusion method to perform decision-level fusion on the time domain probability vector and the frequency domain probability vector to obtain a fusion probability vector; select the fault category corresponding to the maximum probability in the fusion probability vector as the final diagnosis result; The process of decision-level fusion is as follows: Step 5.1, Calculate the confidence of the time-domain probability vector output by the 1D-CNN model : （4）； wherein, a time domain probability vector representing the output of the 1D-CNN model; an information entropy of the time domain probability vector representing the output of the 1D-CNN model, the information entropy being calculated according to the following formula: （5）； wherein n represents the total number of elements in the time-domain probability vector; representing an element in the time-domain probability vector, ; Likewise, the confidence of the frequency domain probability vector is calculated using the method of step 5.1 ; Step 5.2, calculate the confidence of the time-domain probability vector respectively and the confidence of the frequency-domain probability vector , get the weight of the time-domain probability vector and the weight of the frequency-domain probability vector ; The fusion weight is calculated as follows: （6）； （7）； Step 5.

3. Weighting of the time-domain probability vector and the frequency-domain probability vector Weighted fusion is performed to obtain a fused probability vector; Step 6, the 1D-CNN model and the Transformer model constitute a double-branch confidence fusion fault diagnosis model after decision-level fusion; and the double-branch confidence fusion fault diagnosis model is trained by using the training set samples in the voltage data set; Step 7, verify the effectiveness of the double-branch confidence fusion fault diagnosis model by using the test set samples.

2. The method according to claim 1, wherein, In step 1, the fault states include resistance value drift, capacitor failure and operational amplifier saturation, and each of the resistance value drift, capacitor failure and operational amplifier saturation is a single component failure; the resistance value drift refers to that the value of a single resistance component increases or decreases by more than 10%, the capacitor failure refers to that a single capacitor element is short-circuited, and the operational amplifier saturation refers to that the input signal amplitude is too large, the operational amplifier driving capacity is insufficient, or the operational amplifier supply voltage decreases, so that the output of the operational amplifier cannot continue to follow the input change after being greater than a predetermined value.

3. The method according to claim 1, wherein, In step 2.1, the normalization equation of the normalization processing is as follows: （1）； wherein, is a time domain data signal in the voltage data set; is a minimum value in the voltage data set; is a maximum value in the voltage data set; is a normalized voltage time domain data signal, has a value range between 0 and 1.

4. The method according to claim 1, wherein, In step 2.2, the frequency domain preprocessing process is as follows: Step 2.2.1, use fast Fourier transform to convert time domain data into frequency domain data; The window added by the fast Fourier transform is the Hanning window, the complex frequency domain data is obtained by performing fast Fourier transform on the windowed signal; the amplitude spectrum of the voltage frequency domain data set is obtained by squaring and summing the real part and the imaginary part of the frequency domain data, respectively, and then taking the root; Step 2.2.2, logarithmic compression and standardization are performed on the amplitude spectrum of the voltage frequency domain data set; First, logarithmic compression is performed to enhance weak signals, and the compressed data L[k] is obtained: （2）； where M[k] represents the amplitude spectrum, ; , for preventing zero values; Then, the normalized data F[k] is obtained by standardizing the compressed data L[k]: （3）； wherein, is the mean of the compressed data L[k]; is the standard deviation of the compressed data; the frequency domain data F[k] is input to the Transformer model.

5. The method according to claim 1, wherein, In step 3, the 1D-CNN model comprises an input layer and a classification output layer; the input layer comprises a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a third convolutional layer and a third max-pooling layer in sequence; the convolutional kernel size of the three convolutional layers is 5, and the pooling window size of each max-pooling layer is 2; the third max-pooling layer is connected to the classification output layer; the classification output layer comprises two fully connected layers, the first fully connected layer is used to receive fused feature information, the second fully connected layer uses a softmax function for fault classification, and a dropout layer with a dropout rate of 0.5 is added between the two fully connected layers to prevent model overfitting.

6. The method according to claim 1, wherein, In step 4, the Transformer model is a lightweight model, and the Transformer model comprises an input layer, an embedding layer, position encoding, an added CLS classification token, a Transformer encoder layer and a linear classification output layer in sequence; the Transformer encoder layer is a three-layer encoder comprising a first-layer encoder, a second-layer encoder and a third-layer encoder in sequence; the classification output layer is a two-layer fully connected layer comprising a first fully connected layer and a second fully connected layer in sequence; each layer of the Transformer encoder has a four-head self-attention mechanism, a 256-dimensional feedforward network and a GELU activation function; the Transformer encoder layer is stabilized by residual connection and layer normalization for training, extracts a state vector of the added CLS classification token, and inputs the state vector to the linear classification output layer after layer normalization for fault classification.

7. An electronic device, comprising: comprise: one or more processors; a memory; one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs are configured to perform the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program code, and the program code can be invoked and executed by a processor to perform the method according to any one of claims 1 to 6.

Images