Toward Explainable Deep Neural Network Based Anomaly Detection

Explainable Deep Learning for Fault Prognostics in Complex Systems: A Particle Accelerator Use-Case

“…The synthetic data is generated with a time t br ∼ N (µ = 14.61d, σ = 14.61d) between randomly firing precursors, S R l , l = 1, 2, ..., n rand , a time t bp ∼ N (µ = 1d, σ = 1d/24) between deterministic precursors S p , a time t pe ∼ N (µ = 10d/24, σ = 1d/24) between deterministic precursors S p and infrastructure failures S F s , and a time t ep ∼ N (µ = 36.525d, σ = 36.525d) between infrastructure failure S F s and deterministic precursors S p with d being a day of 24 hours. The data is generated for a time range of 2.7 years and n rand being [1,2,4,8,16,32,64,128,256,512].…”

“…The framework is applied with sampling times δt = [2h, 3h] (h for hours), an input range n i = 40, a lead-time t p = 0, an output range of n o = [1,2,3,4], a sub-sampling target ratio p 0,targ = 0.8, and a class '1' neighbourhood coverage n cov = 2. We split the data set with a splitting time t split chosen so that 80 percent of the data-set are used for training and model-selection and 20 percent for final testing.…”

“…Therefore, the noise robustness experiment was repeated with the randomly firing systems being replaced by real-world data containing 8 LASER alarm signals, 27 interlock signals, and 8 beam destination signals from the PSB. The framework was applied to the data with sampling times δt = [2h, 3h] (h for hours), an input range n i = 40, a lead-time t p = 0, an output range of n o = [1,2,3,4], a sub-sampling target ratio p 0,targ = 0.8, and a class '1' neighbourhood coverage n cov = 2. The same model validation strategy is chosen as for the synthetic data experiment.…”

“…The framework is applied with sampling times δt = [10min, 30min, 2h] (h for hours, min for minutes), input ranges n i = [16,32,64], lead-times t p = [0, 1], output ranges of n o = [1,2,4,16], a sub-sampling target ratio p 0,targ = 0.8, and a class '1' neighbourhood coverage n cov = 2. The same model validation strategy is chosen as for the synthetic data experiment.…”

“…They noted that the data set used was not representative for real world scenarios. Amarsinghe et al [2] detected Denial of Service attacks in computer networks using a Deep Neural Network and a so called Layer-wise Relevance Propagation (LRP) introduced by Bach et al [4] to highlight relevant inputs for classification decisions. High classification accuracy was reported and explanations were intuitive to interpret.…”