Failure Rate Estimation Using a Cox Proportional Hazards Deep Neural Network

Purpose: A deep neural network (DNN) is a network that is configured to perform nonlinear operations several times with multiple hidden layers in artificial neural networks[1]. In reliability analysis, the failure rate is an important measure of the reliability of a system or component. This paper reports the use of a DNN to estimate the failure rate for complex outcome data with probabilistic and statistical models such as Cox proportional hazards models.<br/>Methods: DNNs are the most appropriate statistical model for estimating failure rates, and artificial neural networks have been applied to failure rate modeling. We estimated the hazard function (failure rate) using DeepSurv, a DNN model. We evaluated the cumulative reference risk function using the Breslow estimator and then employed it to determine the hazard (failure) rate function.<br/>Results: We consider the data set provided by task[15] (left truncated and right truncated survival time data sets) and archived at https://github.com/jiaxiang-cheng/Random-Weighted-Bootstrap-with-Weibull. The sample size is 18,000.<br/>Conclusion: The risk score and the manufacturer's (average) risk score were calculated. The risk score tends to decrease approximately as the number increases. That is, the risk score gradually decreases with increasing failure time. Upon estimating the hazard (failure rate) functions for 10 manufacturers, high and low values are observed for manufacturers 2 and 8, respectively.