1. Introduction
In recent years, global science and technology have been developing rapidly, and electronic equipment, as a core component in the field of science and technology, is widely used in electric power, industry, transportation, aerospace and many other fields. As an important part of electronic equipment, the reliability of analog circuits directly affects the safe and stable operation of electronic equipment and even the entire system [
1]. In the mixed digital–analog circuits of a typical electronic device, the analog part of the circuit, which accounts for only 20% of the size, causes more than 80% of the total number of malfunctions [
2]. Therefore, it is of great significance to improve the level of analog circuit fault diagnosis to ensure the safe and stable operation of electronic equipment.
According to the failure severity of components, common analog circuit faults are usually divided into two categories: hard and soft faults. Soft faults indicate that the parameter values of components in analog circuits deviate from the allowable tolerance range under the influence of time and environmental conditions, resulting in circuit performance degradation and other situations [
3], hard faults indicate that the circuit is short-circuited or open-circuited due to the structural deformation of the component or the parameter values exceed the limits in the extreme [
4]. Analog circuit soft faults, also known as gradual faults, are faults that can be identified and predicted by prior testing or real-time monitoring. Therefore, how to design efficient fault information acquisition and diagnosis strategies to achieve accurate identification and localization of analog circuit soft faults, and to reduce the probability of occurrence of hard faults due to the accumulation of soft faults that are not investigated in time, has always been a hot spot and a difficult point in the field of analog circuit fault diagnosis research.
With the rapid development of machine learning and deep learning, a large number of optimization algorithms have been introduced into the field of analog circuit fault diagnosis [
5,
6,
7,
8]. Among them, the data-driven analysis method focuses on the data itself, is not limited by circuit linearity or nonlinearity, does not need to establish a very accurate mathematical model, and can determine the analog circuit fault situation by training historical data and testing real-time data [
9]. Therefore, data-driven analog circuit fault diagnosis methods have gradually become the mainstream method in this field, and its optimization problem has also become a research hotspot. Commonly used data-driven methods include support vector machine (SVM) [
10,
11,
12], convolutional neural network (CNN) [
13,
14,
15], deep belief network (DBN) [
16], autoencoder (AE) [
17,
18], and so on. Yuan et al. [
11] used an improved whale optimization algorithm (WOA) to optimize the parameters of the support vector machine to improve the fault diagnosis accuracy. Liang et al. [
13] proposed a multi-scale convolutional neural network with a selective kernel (MSCNN-SK) and used multi-scale average difference sequences as inputs to the fault diagnosis model. Ting et al. [
16] used ensemble empirical mode decomposition (EEMD) to decompose the intermittent fault signal into multiple intrinsic modal functions (IMFs), feature extraction and optimization were performed on the IMFs, and, finally, the feature set was inputted into a DBN for fault diagnosis. Yang et al. [
18] constructed a fault diagnosis model based on an end-to-end denoising autoencoder (EEDAE) and improved the generalization ability of the model by adding noise to the data. In the above studies, the research on improving the fault diagnosis accuracy of analog circuits mostly focuses on the optimization of the model, while less consideration is given to the research on increasing the fault information in the sampled data.
In the current research on fault diagnosis of analog circuits, the output port is commonly used as the measurement node. However, due to the high degree of nonlinearity and the large number of components in analog circuits, some fault features may be weakened or lost during transmission, making them difficult to extract through the output port. The testability analysis of analog circuits is an important tool to improve the fault identification of circuits [
19], in the case of a circuit with a fixed topology, the selection of the measurement nodes is one of the key elements of the circuit testability analysis, which has an important impact on the circuit fault feature extraction [
20]. Zhao et al. [
21] calculated the isolation probability between faults by determining the fault ambiguity gap and then found the measurement node with the maximum fault isolation capability. Cui et al. [
22] proposed a new technique called an extended fault dictionary based on the consideration of component tolerance and redefined the fault isolation rules. As a large number of scholars continue to study the problem of analog circuit measurement node selection, various algorithms have been introduced, such as greedy randomized adaptive search procedure (GRASP) [
23], depth-first search (DFS) algorithm [
24], and so on. The above research on the selection of analog circuit measurement nodes has achieved certain results, but it has not considered the two objectives of the smallest possible number of measurement nodes and the largest possible amount of fault information in combination. In order to solve the above deficiencies, an analog circuit fault diagnosis method that combines a multi-objective selection of measurement nodes is proposed in this paper. Based on the construction of a fault data set from multiple measurement nodes, the research on analog circuit soft-fault diagnosis method is carried out.
The main contributions of this paper are summarized as follows:
- (1)
Aiming at the status quo that some fault information in analog circuits cannot be extracted from the sampled data of output ports, a measurement node selection method based on the multi-objective optimization algorithm is proposed in this paper, which designs the number of measurement nodes and the number of fault types as two objective functions of the multiple objective beluga whale optimization (MOBWO) algorithm in order to achieve the selection of optimal measurement node set, thereby increasing the effective fault information in the sampled data.
- (2)
Aiming at the weak soft-fault characteristics of analog circuits, an analog circuit fault diagnosis method based on a stacked denoising autoencoder (SDAE) optimized by adaptive particle swarm optimization (APSO) is proposed in this paper. By constructing a parameter optimization module in the SDAE model, APSO is enabled to optimize the hyperparameter combinations of SDAE based on the fitness values, thereby enhancing the model’s learning capability and improving fault diagnosis accuracy.
- (3)
For the methods proposed in this paper, the four-opamp biquad high-pass (FOBHP) filter circuit and the leapfrog filter circuit are used as the objects to conduct measurement node selection experiments and fault diagnosis experiments. The experimental results demonstrate that the method proposed in this paper can effectively improve the accuracy of analog circuit fault diagnosis and outperforms other comparison methods under the same conditions.
2. Measurement Node Selection Method
2.1. A Framework for Measurement Node Selection Method
The measurement node selection method is based on the calculation of measurement node fault discrimination, and then the MOBWO algorithm is used to obtain the optimal set of test points for analog circuits. The block diagram of the test node selection is elaborated in
Figure 1.
The process of constructing the fault information table is as follows. Firstly, a three-layer wavelet packet transform (WPT) is carried out on the original signal, and the energy of the third-layer wavelet packet is calculated as the sample features. Secondly, linear dimensionality reduction is carried out on the energy feature data by principal component analysis (PCA) to obtain the feature vector that contains most of the effective information. Again, the K-nearest neighbors (KNN) algorithm is utilized to classify and predict any two types of faults under each measurement node to obtain the differentiation between faults. Then, the differentiation degree of each measurement node for each type of fault category information is calculated by combining formulae, and the measurement point fault information table is constructed. Finally, the differentiation degree of each measurement node for each type of fault is calculated by combining the formula, and the measurement node fault information table is constructed. Based on the measurement node fault information table, the MOBWO algorithm is utilized for the selection of measurement nodes under multi-objective scenarios, and the trade-off between the number of measurement nodes and the fault information is realized through the design of individual coding, genetic operator and fitness target, and, finally, the optimal set of measurement points is obtained.
2.2. Constructing a Fault Information Table for Measurement Nodes
As can be seen from
Section 2.1, the core of constructing a fault information table for measurement nodes is to compute the differentiation between each of the two types of faults. Compared with the integer coding table, the fault differentiation can more accurately reflect the isolation relationship between different faults under the same measurement node, as well as the amount of various types of fault information contained in each measurement node. Therefore, in this paper, the KNN algorithm is used to calculate the differentiation between any two types of faults under each measurement node.
KNN is a supervised learning method based on distance metrics and majority decision making, which requires labeled samples as training data. KNN takes all the samples of known categories as a reference, calculates the distance between the sample to be categorized and all the known samples, and selects the K-nearest known samples from it. According to the voting law of minority–majority, the one that belongs to the largest proportion of categories among the K-nearest neighbor samples is the category of the sample to be classified. A commonly used distance metric is Euclidean Distance [
25], and the Euclidean Distance between point
and point
in
n-dimensional space is shown in Equation (1):
The process of using KNN to find fault differentiation is as follows: calculate the Euclidean Distance between the unknown samples in the test set and all the known samples in the training set by using Equation (1) and discriminate the fault types according to the “nearest neighbor” decision rule of the KNN algorithm, and the ratio of the number of correctly categorized samples in the test set to the total number of samples of the two types of faults is the fault differentiation, as shown in Equation (2), where
l and
m represent a certain two types of fault types. Let the total number of fault types under a measurement node be
M and the number of test samples contained in each fault type be
. Then, the fault sample vector for categories
l and
m is
, and the prediction result after classification by the KNN algorithm is
. Comparing the prediction result with the actual label
of the two types of fault samples, the fault discrimination
FD is obtained as follows:
Table 1 shows an example of the differentiation of various types of faults for a certain measurement node. The table is a symmetric matrix, where the rows and columns all represent fault categories, and
F0 represents the case when the circuit is in a normal state. The numerical values represent the degree of differentiation between the two fault categories, the larger value means that the two faults are differentiated to a higher degree under the current measurement node, where “1” means that the two faults can be completely isolated, and the rest of the values indicate that there are different degrees of confusion between the two fault characteristics.
The table of fault differentiation for a single measurement point can reflect the differentiation between a certain type of fault and all other faults. For example, in
Table 1, the differentiation between fault
F1 and normal state
F0 is 0.500, the differentiation between
F1 and
F2 is 1.000, and the differentiation between
F1 and
F4 is 0.925, which indicates that
F1 and
F2 can be completely isolated, and the differentiation between
F1 and
F0 and
F4 is significantly different, although they cannot be completely isolated. Therefore, it can be found that the fault differentiation table can reflect the differentiation relationship between the faults more finely.
Let
denote the value corresponding to the row
and column
in the fault differentiation table of measurement node
n1, i.e., the differentiation between fault
and fault
. Then, the information of a certain type of fault contained in measurement node
n1 can be calculated according to Equation (3):
where
is the information of class
fault contained in measurement node
n1, determined by the differentiation between this fault and normal data, as well as other fault data. Where
denotes the differentiation between class
faults and normal state
F0,
denotes the summation of the differentiation between class
faults and other classes of faults, and
N denotes the total number of fault classes containing
F0. The weight coefficients
and
both take the value 0.5.
From the above, the fault information of each measurement node in the circuit can be obtained by Equation (3); however, for complex circuits, the number of measurement nodes included in the optimal measurement node set may be more than one, so the fault information of the measurement node set consists of the maximum value of the corresponding fault information of all measurement nodes. Let the fault information of measurement nodes
n1 and
n2 be
,
, then the fault information of measurement node set {
n1,
n2} is
, as shown in
Table 2.
2.3. Measurement Node Selection Based on MOBWO Algorithm
Multiple objective beluga whale optimization (MOBWO) is a multi-objective optimization algorithm proposed by integrating Non-dominated Sorting Genetic Algorithm-II (NSGA-II) on the basis of beluga whale optimization (BWO) [
26,
27]. The MOBWO algorithm mainly contains the steps of fast non-dominated sorting, crowding distance calculation, child population generation, parent–child merger, and elite strategy selection. Among them, fast non-dominated sorting and crowding distance calculation are the core mechanisms of NSGA-II, which together ensure that the multi-objective optimization algorithm can effectively find the Pareto optimal solution. And BWO, as a genetic operator in the algorithm, updates the positions of individuals in the preferred parent population through the three phases of exploration, exploitation and whale fall to obtain the offspring population. Where the formula for updating individual positions in the exploration stage is shown in Equation (4), the formula for updating individual positions in the exploitation stage is shown in Equation (5), and the formula for updating individual positions in the whale-fall stage is shown in Equation (6).
In Equation (4), denotes the position of the individual on the dimension at the next iteration; denotes the position of the individual on the random dimension under the current iteration; and denotes the position of the random individual on the random dimension under the current iteration. In Equation (6), is the step size of the whale fall.
In the MOBWO-based measurement node selection method, a population of size
N represents
N possible combinations of measurement nodes, e.g., the combination
of measurement nodes is
, where
is the total number of measurement points. In the genetic algorithm, the commonly used population initialization method is to randomly generate random numbers between [0, 1]; however, there are only two states of measurement nodes, either selected or unselected, so the individuals are encoded through the dichotomous method in this paper. If
in the measurement node combination
, then it means that the measurement node
is selected, and vice versa is not selected. Also based on the analysis of the analog circuit measurement node selection problem, two optimization objectives of the MOBWO algorithm are designed, which are the number of measurement nodes in the measurement node set and the number of fault types in the fault information, as in Equations (7) and (8).
In Equation (7), denotes the number of measurement nodes in the measurement node combination . In Equation (8), denotes the fault information of the class fault under the measurement node set , and denotes the number of types of fault information contained in , where Q is the basis for the determination of the fault discrimination ability of the measurement node set, which can be adjusted according to the actual situation, and, in this paper, based on the detailed analysis of the fault differentiation in multiple circuits, the value is taken as 0.85.
The flowchart of the measurement node selection method based on the MOBWO algorithm is shown in
Figure 2.
Firstly, an initial population P0 of size N is randomly generated, and two objective function values are calculated for each individual based on the fault information table of measurement nodes. Secondly, based on these two values, the initial population P0 is subjected to fast non-dominated sorting, crowding distance calculation and ranking, and the individual with the smallest rank or the largest crowding distance when the rank is the same is selected as the parent population by binary tournament selection. Next, the parent population Pt is input into the BWO algorithm, and the child population Qt is generated by the three position update formulas, and then the two populations Pt and Qt are merged to form the population Rt of size 2N. Subsequently, the population Rt is subjected to the fast non-dominated sorting, crowding distance calculation, and ranking, and is retained to generate the next-generation parent by elite selection strategy, and then the process is repeated until the number of iterations reaches a set maximum.
5. Conclusions
Analog circuits, as an important part of electronic equipment, are widely used in electric power, industry, transportation, medical, aerospace and many other fields, and their reliability directly affects the safe and stable operation of electronic equipment and even the whole system. Therefore, it is of great significance to improve the fault diagnosis level of analog circuits to reduce the maintenance cost of electronic equipment and to reduce the potential hazards brought by soft faults in circuits. In this paper, the optimal measurement node set selection and intelligent fault diagnosis are combined, and on the basis of constructing a fault data set from multiple measurement nodes, combined with a deep learning framework, the research on the analog circuit fault diagnosis method based on the APSO-SDAE is carried out.
- (1)
Aiming at the status quo that some fault information in analog circuits cannot be extracted from the sampled data of output ports, a test point selection method based on a multi-objective optimization algorithm is proposed. Calculating the fault differentiation to construct the fault information table of measurement nodes, and designing the number of measurement nodes and the number of fault information types as the two objective functions of the multiple objective beluga whale optimization (MOBWO) algorithm in order to achieve the optimal measurement node set selection. In the measurement node selection experiments of two standard circuits, the measurement node set obtained by this method is superior to the comparison methods.
- (2)
Aiming at the weakness of soft-fault features for analog circuits, a fault diagnosis method based on adaptive particle swarm optimization and stacked denoising autoencoder (APSO-SDAE) is proposed. The APSO parameter optimization module is designed to optimize the SDAE model hyperparameter combinations based on fitness values to improve the fault classification capability. In the soft-fault diagnosis experiments of two standard circuits, the method achieves diagnostic accuracies of 97.87% and 98.41%, respectively, which is about 4% higher than that of the traditional DAE model in dealing with the same fault diagnosis problem, which fully demonstrates that the proposed method is able to identify the soft-fault states of analog circuits effectively.