Intelligent Test Diagnosis Method for Airborne Circuit Boards Based on Multi-Branch Deep Neural Network

Abstract

The present invention relates to the field of automatic testing and fault diagnosis of circuit boards, specifically an intelligent testing and diagnosis method for airborne circuit boards based on a multi-branch deep neural network. The specific steps are as follows: S01, establishment of a fault database; S02, data preprocessing; S03, forward network inference; S04, reverse parameter update; S05, fusion of multi-branch prediction results; A flying probe tester is used for testing and data acquisition of the circuit board, and a data enhancement algorithm is designed according to the changes in component characteristics to complete the construction of the database; The designed deep neural network uses different convolutional kernels and pooling kernels at different feature layers to focus on features of different dimensions; The designed multi-branch deep neural network takes the VI characteristic curves of multiple nodes as input, improving the accuracy and robustness of the test and diagnosis results, especially reducing the recognition errors caused by incomplete features generated by human factors and parameter changes of components caused by environmental changes.

Description

Technical Field

[0001] The present invention relates to the field of automatic testing and fault diagnosis of circuit boards, and specifically to an intelligent testing and diagnosis method for airborne circuit boards based on a multi-branch deep neural network. Background Technique

[0002] With the rapid development of electronic technology, while the functions of modern avionics equipment are increasingly enhanced, its design integration and complexity are constantly increasing, making the requirements for functional testing and fault diagnosis higher and higher. The traditional testing and diagnosis method relying on professional personnel equipped with electronic measuring instruments strongly depends on the professional level of testers and has a high time cost, unable to meet the production and maintenance requirements of modern products. Therefore, it is urgent to study an intelligent airborne circuit board diagnosis method to provide guidance for test and maintenance personnel to complete test and fault location with the least cost and the fastest speed.

[0003] For the automatic testing and intelligent diagnosis of circuit boards, the current mainstream method is to collect the volt-ampere (VI) characteristic curves of circuit board nodes for feature extraction in the time domain, frequency domain or time-frequency domain, and then apply machine learning algorithms such as BP neural network, support vector machine, etc. for classification and recognition. Compared with the traditional manual method, although the testing and fault location efficiency is greatly improved, the accuracy is greatly affected by design features and is more sensitive to noise effects such as temperature changes. Summary of the Invention

[0004] To solve the above problems, the present invention proposes an intelligent testing and diagnosis method for airborne circuit boards based on a multi-branch deep neural network.

[0005] The intelligent testing and diagnosis method for airborne circuit boards based on a multi-branch deep neural network specifically includes the following steps:

[0006] S01. Establishment of a fault database:

[0007] a. Build an experimental circuit board with similar functions, add typical fault types to the normal circuit one by one, and use a flying probe tester to collect the VI curves of each key node of the circuit when the fault occurs;

[0008] b. Construct a data matrix of size M×N×2 to represent the VI curve response of the circuit board in the current state, and use the one-hot method for fault numbering and storage to form fault data;

[0009] c. Use the Monte Carlo simulation method to perform 50-100 simulations, which are used as training and testing together with the real collected data;

[0010] S02. Data preprocessing:

[0011] a. The constructed VI data matrix has different distribution ranges for data in different feature dimensions and needs to be preprocessed before being fed into the neural network for processing.

[0012] b. The mean - variance normalization method shown in Equation (1) is used to preprocess the voltage and current data respectively, so that the processed data x scale conforms to the standard normal distribution, where the mean x mean and the standard deviation x std are statistically obtained from the training data.

[0013]

[0014] S03. Network forward inference:

[0015] a. Based on the classic image classification network VGG16, a multi - branch deep neural network is constructed.

[0016] b. Using the convolutional layer for local feature calculation and the pooling layer for feature dimensionality reduction and enhancement as basic operation units, the feature extraction module of the network contains 13 convolutional layers and 5 pooling layers.

[0017] c. Matching the specific configuration of network parameters, for the input VI response matrix of size 67×100×2, the size of the output matrix after each layer of calculation.

[0018] d. In the initial stage of feature extraction, the 1st, 2nd, 3rd, 4th convolutional layers and the 1st, 2nd pooling layers use convolutional kernels of size 1×3 and pooling kernels of size 1×2 respectively, and the remaining convolutional layers all use the widely used convolutional kernels of size 3×3 to fuse the VI features between different nodes.

[0019] e. The pooling layer uses pooling kernels of size 2×2 to simultaneously reduce the size of the VI dimension and the data size of the test node dimension.

[0020] f. The number of nodes in the fully - connected layers FCcls2 and FCloc2 is set according to the possible number of fault categories and the maximum number of positions corresponding to each type of fault, and the Softmax function is used to convert the original output of the network into prediction probabilities in the range of [0, 1] to represent the probabilities of various faults occurring on the circuit board or the probabilities of faults occurring at several corresponding positions.

[0021] S04. Parameter back - update: The process of updating parameters according to the calculated network loss. Given the training samples (x, y, l), where the fault category label y is an 18 - dimensional vector and the fault position number label l is a 15 - dimensional vector. Specifically:

[0022] a. If it is category i, then yi = 1, otherwise yi = 0. If a fault occurs at position i, then li = 1, otherwise li = 0.

[0023] b. Assuming the outputs of the two branches of the network are pcls and ploc respectively, the total loss of the network can be obtained based on the cross-entropy classification loss function as follows:

[0024]

[0025] where K cls = 18 and K loc = 15 are the number of circuit board state types and the maximum position number in the fault state respectively;

[0026] S05. Multi-branch prediction result fusion:

[0027] a. For the input test data, first determine the fault type by selecting the maximum value among the probabilities of various faults output by the network, and then select the fault position with the highest probability;

[0028] b. Combining the two can obtain the fault state and fault position of the current circuit board.

[0029] In step b of S01, the data matrix includes M, which is the number of circuit board test nodes, N, which is the number of VI curve sampling points, Vij and Iij, which represent the values of the jth sampling point on the VI curve at the ith node of the circuit board.

[0030] In step c of S01, the Monte Carlo simulation method simulates and the real collected data are divided into a training set and a test set according to a ratio of 7:3 to prepare for the learning and performance testing of the deep neural network parameters.

[0031] The forward inference of the network in step S03 is a process of sending the preprocessed data into the built deep neural network for feature extraction and mapping to obtain the network output.

[0032] In step b of S03, each convolutional layer is followed by a default activation layer for feature non-linear mapping, and also includes an activation function of the Relu function f(x) = max(0, x) that can prevent gradient disappearance, and output modules for two branches that respectively predict the circuit board fault category and fault position. Each branch contains 2 fully connected layers for further extraction of features and mapping of features to class probabilities.

[0033] In step c of S03, the matrix includes Ck and Pk, which respectively represent the kth convolutional layer and pooling layer, FCclsk and FClock, which respectively represent the kth fully connected layer in the class prediction branch and fault localization branch, ks, which represents the size of the convolutional kernel or pooling kernel, and kn, which represents the number of convolutional kernels.

[0034] In step S04, for each iteration of training, according to the calculated loss, the network parameters are updated using the momentum gradient descent method with a learning rate of 0.001 and a momentum coefficient of 0.9, that is, Momentum SGD.

[0035] The beneficial effects of the present invention are as follows: A flying probe tester is used for testing and data acquisition of circuit boards, and a data augmentation algorithm is designed according to the changes in component characteristics to complete the construction of the database; the designed deep neural network uses different convolutional kernels and pooling kernels in different feature layers to focus on features in different dimensions; the designed multi-branch deep neural network takes the VI characteristic curves of multiple nodes as inputs, improving the accuracy and robustness of test and diagnosis results, especially reducing the recognition errors caused by incomplete features generated by human factors and parameter changes of components caused by environmental changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present invention will be further described below with reference to the drawings and embodiments.

[0037] Figure 1 is a schematic flowchart of the present invention;

[0038] Figure 2 is a schematic diagram of the VI response matrix of the circuit board of the present invention;

[0039] Figure 3 is a schematic diagram of the structure of the multi-branch deep neural network of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the present invention will be further described below.

[0041] As Figures 1 to 3 shown, an intelligent test and diagnosis method for airborne circuit boards based on a multi-branch deep neural network has the following specific steps:

[0042] S01. Establishment of a fault database:

[0043] a. Build an experimental circuit board with similar functions, add typical fault types to the normal circuit one by one, and use a flying probe tester to collect the VI curves of each key node of the circuit when the fault occurs;

[0044] b. As Figure 2 shown, a data matrix of size M×N×2 is formed to represent the VI curve response of the circuit board in the current state, and the fault numbers are numbered and saved in one-hot manner to form fault data;

[0045] c. To further enrich the diversity of data, the Monte Carlo simulation method is used to perform 50 - 100 simulations, and then together with the real - collected data, they are used for training and testing;

[0046] S02. Data pre - processing:

[0047] a. For the constructed VI data matrix, data with different feature dimensions have different distribution ranges and need to be pre - processed before being fed into the neural network for processing;

[0048] b. The mean - variance normalization method shown in Equation (1) is used to pre - process the voltage and current data respectively, so that the processed data x scale conforms to the standard normal distribution, where the mean x mean and the standard deviation x std are statistically obtained from the training data;

[0049]

[0050] S03. Network forward inference:

[0051] a. Based on the classic image classification network VGG16, a multi - branch deep neural network is constructed;

[0052] b. Using the convolutional layer for local feature calculation and the pooling layer for feature dimensionality reduction and enhancement as the basic operation units, the feature extraction module of the network contains 13 convolutional layers and 5 pooling layers;

[0053] c. Table 1 shows the specific configuration of the network parameters and the size of the output matrix after each layer's calculation for the input VI response matrix of size 67×100×2;

[0054] Table 1 Parameter configuration of the multi - branch deep neural network

[0055]

[0056]

[0057] d. In the initial stage of feature extraction, to focus on the extraction of features of the same node, the 1st, 2nd, 3rd, and 4th convolutional layers and the 1st and 2nd pooling layers use convolutional kernels of size 1×3 and pooling kernels of size 1×2 respectively, and the remaining convolutional layers all use the widely used convolutional kernels of size 3×3 to fuse the VI features between different nodes;

[0058] e. The pooling layer uses pooling kernels of size 2×2 to simultaneously reduce the sizes of the VI dimension and the test - node dimension data;

[0059] f. The number of nodes in the fully connected layers FCcls2 and FCloc2 is set according to the number of possible fault categories and the maximum number of positions corresponding to each type of fault. The Softmax function is used to convert the original output of the network into predicted probabilities in the range of [0, 1] to represent the probabilities of various types of faults occurring on the circuit board or the probabilities of faults occurring at several corresponding positions.

[0060] S04. Parameter backpropagation update: The process of updating parameters based on the calculated network loss. Given the training samples (x, y, l), where the fault category label y is an 18-dimensional vector and the fault position number label l is a 15-dimensional vector. Specifically:

[0061] a. If it is category i, then yi = 1, otherwise yi = 0. If a fault occurs at position i, then li = 1, otherwise li = 0.

[0062] b. Assuming the outputs of the two branches of the network are pcls and ploc respectively, then based on the cross-entropy classification loss function, the total network loss is:

[0063]

[0064] where, K cls = 18 and K loc = 15 are the number of circuit board state types and the maximum position number in the fault state respectively.

[0065] S05. Multi-branch prediction result fusion:

[0066] a. For the input test data, first determine the fault type by selecting the maximum value among the probabilities of various types of faults output by the network, and then select the fault position with the highest probability.

[0067] b. Combining the two can obtain the fault state and fault position of the current circuit board.

[0068] Specifically, in step S01, a large amount of reliable data is the main reason for the success of the deep learning method. However, the number of faulty circuit boards is generally very limited, while the number of circuit boards in the normal state is relatively large. To establish a circuit board fault data set with sufficient scale, experimental circuit boards with similar functions are built.

[0069] Specifically, in step S'01a, a flying probe tester is used to test and collect data of the circuit board, and a data augmentation algorithm is designed according to the changes in component characteristics to complete the construction of the database. It has strong robustness. The VI characteristic curves of component nodes on the airborne circuit board in different states are collected by the flying probe tester and extended through data augmentation, and a data set with relatively complete features can be obtained.

[0070] In step S01 b), the data matrix includes M, which is the number of circuit board test nodes, N, which is the number of VI curve sampling points, Vij and Iij, which represent the value of the j-th sampling point on the VI curve at the i-th node of the circuit board.

[0071] In step S01 c), the Monte Carlo simulation method simulates and divides the real collected data into a training set and a test set according to a ratio of 7:3, preparing for the learning and performance testing of the deep neural network parameters.

[0072] Specifically, the designed deep neural network uses different convolutional kernels and pooling kernels in different feature layers to focus on features of different dimensions, with high accuracy. Compared with existing methods, it can automatically extract high-dimensional feature descriptions through the deep neural network for recognition and has stronger discrimination ability for different types of signals.

[0073] Specifically, for the VI data matrix constructed in step S02, data with different feature dimensions have different distribution ranges. If directly sent into the neural network for processing without preprocessing, it is easy to cause gradient dispersion or explosion, resulting in difficult convergence of network training.

[0074] Specifically, in step S02 a), the designed multi-branch deep neural network takes the VI feature curves of multiple nodes as input, predicts the probabilities of various types of faults occurring in the network and the probabilities of faults occurring at the corresponding positions through two output branches respectively, and then obtains the fault state and specific fault position of the circuit board after fusion, which is more efficient and easier to train than a single-branch network that directly predicts all combinations of fault types and their positions.

[0075] The network forward inference in step S03 is a process of sending the preprocessed data into the constructed deep neural network for feature extraction and mapping to obtain the network output. Based on the classic image classification network VGG16, the present invention constructs Figure 3 the multi-branch deep neural network shown.

[0076] In step S03 b), a default activation layer for feature non-linear mapping is followed after each convolutional layer. It also includes an activation function of the Relu function f(x)=max(0,x) that can prevent gradient disappearance, and two output modules for the branches that respectively predict the circuit board fault category and fault position. Each branch contains 2 fully connected layers for further extraction of features and mapping of features to class probabilities.

[0077] In step S03 c), the matrix includes Ck and Pk that respectively represent the k-th convolutional layer and pooling layer, FCclsk and FClock that respectively represent the k-th fully connected layer in the class prediction branch and the fault location branch, ks that represents the size of the convolutional kernel or pooling kernel, and kn that represents the number of convolutional kernels.

[0078] In the step S04, in each iteration training, according to the calculated loss, the network parameters are updated by using the method of Momentum Stochastic Gradient Descent (Momentum SGD) with a learning rate of 0.001 and a momentum coefficient of 0.9.

[0079] Specifically, the reverse update of the network parameters in step S04 is Figure 1 The last step of the network training process shown by the dashed line is the process of updating the parameters according to the calculated network loss.

[0080] Specifically, the multi-branch prediction result fusion in step S05 is Figure 1 The last step of the network testing process shown by the solid line improves the accuracy and robustness of the testing and diagnosis results, especially reduces the recognition errors caused by the incomplete features generated by human factors and the parameter changes of components caused by environmental changes.

[0081] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. What is described in the above embodiments and the specification is only the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent test and diagnosis method for airborne circuit boards based on a multi-branch deep neural network, characterized in that: The specific steps are as follows: S01. Establishment of the fault database: a. Build an experimental circuit board with similar functions, add typical fault types to the normal circuit one by one, and use a flying probe tester to collect the VI curves of each key node of the circuit when a fault occurs; b. Construct a data matrix of size M×N×2, represent the VI curve response of the current state circuit board, and use the one-hot method for fault numbering and storage to form fault data; c. Use the Monte Carlo simulation method to perform 50 - 100 simulations, which together with the real collected data are used for training and testing; S02. Data preprocessing: a. For the constructed VI data matrix, data with different feature dimensions have different distribution ranges and need to be preprocessed before being sent to the neural network for processing; b. Preprocess the voltage and current data respectively using the mean-variance normalization method shown in Equation (1) so that the processed data x scale conforms to the standard normal distribution, where the mean x mean and the standard deviation x std are statistically obtained from the training data; S03. Forward inference of the network: a. Based on the classic image classification network VGG16, construct a multi-branch deep neural network; b. Using the convolutional layer for local feature calculation and the pooling layer for feature dimensionality reduction and enhancement as the basic operation units, the feature extraction module of the network contains 13 convolutional layers and 5 pooling layers; c. According to the specific configuration of the network parameters, for the input VI response matrix of size 67×100×2, calculate the size of the output matrix after passing through each layer; d. In the initial stage of feature extraction, the 1st, 2nd, 3rd, and 4th convolutional layers and the 1st and 2nd pooling layers respectively use convolutional kernels of size 1×3 and pooling kernels of size 1×2, and the remaining convolutional layers all use the widely used convolutional kernels of size 3×3 to fuse the VI features between different nodes; e. The pooling layer uses pooling kernels of size 2×2 to simultaneously reduce the sizes of the VI dimension and the test node dimension data; f. The number of nodes in the fully connected layers FCcls2 and FCloc2 is set according to the possible number of fault categories and the maximum number of positions corresponding to each type of fault, and the Softmax function is used to convert the original output of the network into a prediction probability in the range of [0,1] to represent the probability of various faults occurring on the circuit board or the probability of faults occurring at several corresponding positions; S04. Parameter backpropagation and update: The process of updating the parameters according to the calculated network loss. Given the training samples (x, y, l), where the fault category label y is an 18-dimensional vector and the fault position number label l is a 15-dimensional vector. Specifically: a. If it is category i, then yi = 1, otherwise yi = 0. If a fault occurs at position i, then li = 1, otherwise li = 0; b. Assuming the outputs of the two branches of the network are pcls and ploc respectively, then based on the cross-entropy classification loss function, the total network loss is: Among them, K cls = 18 and K loc = 15 are the number of circuit board state types and the maximum position number in the fault state, respectively; S05. Fusion of multi-branch prediction results: a. For the input test data, first determine the fault type by selecting the maximum value of the probabilities of various faults occurring output by the network, and then select the fault position with the highest probability; b. Combining the two can obtain the fault state and fault position of the current circuit board.

2. The airborne circuit board intelligent test and diagnosis method based on a multi-branch deep neural network according to claim 1, wherein: In step b of S01, the data matrix includes M, which is the number of circuit board test nodes, N, which is the number of VI curve sampling points, Vij and Iij, which represent the value of the jth sampling point on the VI curve at the ith node of the circuit board.

3. The intelligent test and diagnosis method for airborne circuit boards based on a multi-branch deep neural network according to claim 1, wherein: In step S01(c), the Monte Carlo simulation method is used to simulate and divide the real collected data into a training set and a test set according to a ratio of 7:3, preparing for the learning and performance testing of the deep neural network parameters.

4. The airborne circuit board intelligent test and diagnosis method based on a multi-branch deep neural network according to claim 1, characterized in that: The forward inference of the network in step S03 is a process of sending the preprocessed data into the constructed deep neural network for feature extraction and mapping to obtain the network output.

5. The intelligent test and diagnosis method for airborne circuit boards based on a multi-branch deep neural network according to claim 1, characterized in that: In step S03(b), a default activation layer for feature non-linear mapping is followed after each convolutional layer. It also includes an activation function of the Relu function f(x) = max(0, x) that can prevent gradient disappearance, and two output modules for predicting the circuit board fault category and the fault location branch respectively. Each branch contains 2 fully connected layers for further extraction of features and mapping from features to class probabilities.

6. The intelligent test and diagnosis method for airborne circuit boards based on a multi-branch deep neural network according to claim 1, characterized in that: In step S03(c), the matrix includes Ck and Pk representing the k-th convolutional layer and pooling layer respectively, FCclsk and FClock representing the k-th fully connected layer in the class prediction branch and the fault location branch respectively, ks representing the size of the convolutional kernel or pooling kernel, and kn representing the number of convolutional kernels.

7. The intelligent test and diagnosis method for airborne circuit boards based on a multi-branch deep neural network according to claim 1, characterized in that: In step S04, in each iteration of training, according to the calculated loss, the network parameters are updated using the momentum gradient descent method with a learning rate of 0.001 and a momentum coefficient of 0.9, that is, the Momentum SGD method.

Images