Optimization of constructing the neuromorphic fault dictionary for testing and diagnostics of analog ICs

Fault diagnostics of integrated circuits (IC) is an important stage of the production cycle, providing the detection of the faults location during testing and ensuring the high quality of the ICs batch at the output. Testing and diagnostics of analog ICs represent significant functional complexity compared to digital ICs. In many respects this is facilitated by the continuous nature of the signals being processed, the complex nonlinear dependences of the input and output signals, the tolerances on the parameters of internal components that randomly affect the deviations of the parameters and the transfer function from the nominal values, the sensitivity of the output characteristics to the deviations of internal and external parameters, possibility of appearance not only catastrophic, but also parametric faults, the lack of effective models for analog faults, etc. There are two approaches to functional testing of analog ICs: specification-driven methods based on compliance with the specification and data-driven methods based on the fault simulation. The second approach is widely used because it provides the solution to the task of not only testing, but also faults diagnostics. The dimension of the fault dictionary (FD) and the mechanism for sequential matching in the FD structure are the weaknesses of this method. The development of machine learning tools has opened the possibility of transition to neuromorphic FD, operating in an associative mode, the architecture of which is not sensitive to the number of considered faults. The choice of the parameters that will be used to train the neural network is an important task, the result of which has a significant influence on the convergence of the training process, its duration and the quality of the fault coverage. The scopes of the proposed work is to search for and study the method of selecting the essential characteristics of the output response for the circuit under test, which reduce the computational and time costs for training the neuromorphic FD (NFD) without reducing the coverage of the considered faults. To achieve this goal, the following tasks are defined and solved: representation of a continuous analog signal in the time domain by discrete values in the frequency domain based on the wavelet transform (WT); reducing the dimension of the matrix of WT-coefficients applied for training the neural network, using the principal components analysis; software implementation of the proposed method for selecting the essential characteristics of output responses; experimental study of indices of training the NFD with the use of input vectors of different lengths; analysis of the obtained results. In the result of the wavelet transform the continuous output response is represented by a tuple of the scaling factors a and the shifts b: x = [a; b], and the set of considered responses is a matrix of WT-coefficients. The application of the principal component analysis (PCA) provides a reduction in the number of coefficients used for training the neural network PCA. PCA is implemented by an iterative procedure in which new principal components (PC) are added sequentially one after another. It is important to determine when to stop this process, and to take a sufficient number of principal components. With a small number of PCs, the aggregate sample of data will be incomplete, at large number an overvaluation arises. The experimental studies were carried out for the analog filter circuit on operational amplifier using the developed software . The obtained results have demonstrated the effectiveness of the proposed method of selecting the essential characteristics of output responses, which is reflected in the reduction of time and computational costs for training. The resultant NFD provides the fault coverage up to 100%.

Keywords

Authors

References