introduction
The complexity and diversity of circuit faults, the large dispersion of component parameters, the limited accessibility of test nodes, and the widespread nonlinearity mean that circuit fault diagnosis, both theoretically and methodologically, is still quite far from practical application. In recent years, with the rapid development of nonlinear system theory, circuit fault diagnosis theory and methods have also made significant progress. As systems become more complex and the requirements for system safety and reliability increase, the need for fault prediction and diagnosis technology is becoming increasingly urgent.
Circuit fault diagnosis has been widely applied in communications, military, automatic control, home appliances, measuring instruments, and other fields. Methods for diagnosing analog circuit faults include expert systems and artificial neural networks. The circuit cutting method typically involves cutting the circuit before and after a suspected component and testing the current and voltage signals at key points to determine if a fault has occurred. However, this method is not only cumbersome to test, but also requires cutting many components to pinpoint the faulty component because it cannot predict which component is faulty. In many cases, such destructive diagnosis is not permissible, especially in searching for faults in critical instrument circuits or operating electromechanical equipment. Furthermore, random interference caused by unknown uncertainties can cause circuits to exhibit complex nonlinear characteristics. Artificial neural network systems, due to their strong nonlinear mapping, online learning, and associative memory capabilities, have become an important research method for handling fault diagnosis problems. This paper introduces CPN neural network information fusion technology into circuit fault search. By testing the current, temperature, and voltage generated by electronic components during circuit operation, CPN neural networks are used to model analog circuit faults and fuse multiple information, thereby accurately locating the faulty component.
1. CPN Network Operating Principle
1.1 CPN Network Model
The backpropagation neural network, also known as the reverse propagation network, mainly consists of three layers: an input layer, a Kohonen layer, and a Grossberg layer. Figure 1 shows the structure of the CPN neural network. The first layer is the input layer, which feeds the input vectors of the neurons to the second layer. The second layer is the Kohonen layer, which performs unsupervised learning, operating on a "winner-takes-all" basis. The network generates the winning neurons in the competitive layer according to the SOM learning rule and adjusts the connection weights W of the corresponding Kohonen layer according to this rule. The third layer is the output layer, which mainly implements the Outstar method proposed by Grossberg. The Grossberg layer performs supervised training, obtaining the actual outputs of each output neuron according to the basic competitive network learning rules, and correcting the connection weights V of the Grossberg layer according to the supervised error correction method to achieve class representation. The number of neurons in the input layer, n, is equal to the number of sensors.
Number of diagnostic elements. In this system, the number of neurons in the Kohonen layer is 8. The number of neurons in the Grossberg layer, m, is equal to the number of diagnostic elements.
Figure 1. CPN neural network structure diagram
1.2 Implementation steps of the improved CPN network algorithm
① Data preprocessing stage: Normalize the input value Pk.
Table 1. CPN Neural Network Mode Samples
Serial Number | sensor | Input membership degree | Output value | ||||||
A1 | A2 | A3 | A4 | A 1 | A 2 | A 3 | A 4 | ||
1 | Voltage | 0.51 | 0.26 | 0.01 | 0.18 | 1 | 0 | 0 | 0 |
temperature | 0.70 | 0.17 | 0.01 | 0.11 | |||||
2 | Voltage | 0.03 | 0.71 | 0.13 | 0.16 | 0 | 1 | 0 | 0 |
temperature | 0.08 | 0.65 | 0.12 | 0.07 | |||||
3 | Voltage | 0.32 | 0.21 | 0.39 | 0.05 | 0 | 0 | 1 | 0 |
temperature | 0.01 | 0.01 | 0.77 | 0.10 | |||||
4 | Voltage | 0.29 | 0.24 | 0.02 | 0.46 | 0 | 0 | 0 | 1 |
temperature | 0.09 | 0.05 | 0.18 | 0.59 | |||||
Table 2. CPN Neural Network Fusion Diagnostic Results
Serial Number | sensor | Input membership degree | Diagnostic results | |||
A1 | A2 | A3 | A4 | |||
1 | Voltage | 0.513 | 0.214 | 0.020 | 0.193 | uncertain |
temperature | 0.683 | 0.181 | 0.102 | 0.025 | uncertain | |
Fusion | 0.803 | 0.052 | 0.033 | 0.081 | A1 fault | |
2 | Voltage | 0.033 | 0.701 | 0.215 | 0.051 | uncertain |
temperature | 0.075 | 0.631 | 0.134 | 0.068 | uncertain | |
Fusion | 0.013 | 0.789 | 0.034 | 0.105 | A2 fault | |
3 | Voltage | 0.314 | 0.222 | 0.389 | 0.039 | uncertain |
temperature | 0.000 | 0.000 | 0.800 | 0.101 | uncertain | |
Fusion | 0.024 | 0.049 | 0.853 | 0.063 | A3 fault | |
4 | Voltage | 0.202 | 0.214 | 0.081 | 0.434 | uncertain |
temperature | 0.079 | 0.081 | 0.194 | 0.558 | uncertain | |
Fusion | 0.053 | 0.128 | 0.051 | 0.732 | A4 fault | |
5. Conclusion
This paper combines the Grossberg layer, which performs the alien algorithm, with the Kohonen layer, which performs self-organizing mapping, to obtain a CPN network similar to a BP neural network. This solves the training problems of multi-level networks and overcomes the limitations of single-level networks. By optimizing the initial weight setting rules of the CPN network, the limitations of overly strict input vector restrictions can be effectively overcome, thus avoiding the significant impact of initial weights and sample input order on network learning. The improved CPN network exhibits more stable performance and a wider range of applications. Optimizing the CPN network increases the computational speed of the algorithm, resulting in higher diagnostic accuracy and faster convergence, making it highly suitable for fault diagnosis of nonlinear circuits.
For more information, please visit the Instrumentation Channel.
