Circuit Reliability Prediction Based On Dual Autoencoder Model

As the feature size of digital logic circuits is developing towards nanometer and deep submicron levels,and circuit integration is rapidly increasing,people's demand for high-reliability circuits is increasingly urgent.Accurately predicting circuit reliability,especially for circuits in the design stage,can greatly reduce production costs and is of great significance to promoting the development of the IC industry.The methods of evaluating circuit reliability are mainly divided into sampling-based or analysis-based methods.The latter has higher prediction accuracy,but consumes more time and is not suitable for circuit reliability evaluation in the design stage.The deep learning model has the ability to quickly evaluate the circuit reliability in the design stage,and the time complexity is not affected by the circuit scale,so it has certain development prospects.But at present,this method still has the problem of low prediction accuracy,and at the same time,it is also difficult to standardize the circuit input vector features of indeterminate length and different dimensions.In view of the above problems,this paper uses different hash algorithms to standardize the features of the circuit input vector,and proposes a dual autoencoder model to evaluate the circuit reliability,which effectively improves the prediction accuracy.The main research contents of this paper are as follows(1)Creating feature datasets for circuit reliability assessment.Firstly,the corresponding features are selected according to the decisive factors affecting the reliability of the gate-level circuit,and the circuit reliability value calculated by the E-PTM model is used as the data label.Select 74 series circuits,ISCAS85 reference circuits and some circuits in EPFL reference circuits as sample circuits,analyze their circuit netlists,and decompose them into circuit sub-netlists with only one output terminal.Extract relevant circuit topology information from the circuit subnet list,and combine the randomly generated circuit input vector,failure probability value and label value to create a data set for circuit reliability assessment.(2)Research on the standardization method of circuit input vector features based on RSHash algorithm.Aiming at the problem that the circuit input vector features have indeterminate lengths and different dimensions,a hash algorithm is used to map them and normalize them into fixed-length features.First,the collision rate of different hash algorithms is analyzed,and 4algorithms with better scores are selected for feature standardization.Secondly,based on the data set created by the method in research content 1 for circuit reliability evaluation,the performance of the collision rate and distribution characteristics is analyzed and compared,and the RSHash algorithm with the best performance is selected.At the same time,in order to control the hash map range,select a specific prime value to take the remainder of the hash result.Finally,it is applied to the feature normalization of the circuit input vector and a comparative experiment based on the DAN model is carried out.(3)Designing a dual autoencoder model for gate-level circuit reliability evaluation.Through the parallel connection of two stacked autoencoders with different activation functions,the advantages of both sigmoid and tanh activation functions are combined to capture different characteristic information of sample data.At the same time,in view of the limited ability of block strategy to deal with large-scale circuit input vectors,a feature standardization method of circuit input vector based on CRC20 fusion block strategy is proposed.On the dataset processed by this method,a comparative experiment between the dual autoencoder model and the DAN model is carried out.Then,the grid search method is used to search for the optimal network structure and the dual autoencoder model with related parameter values.Based on this model,it is applied to sample circuits of various scales,and compared with the DAN model and the MC method.Finally,the dual autoencoder model is further optimized by adding noise or adding sparsity constraints.