Anomaly Detection for Autonomous Vehicle using Variable Generative Adversarial Transformer with Fuzzy Based Histogram Oriented Gradients

Abstract

Multivariate temporal anomaly detection is important in fields such as AIOps and medical intelligence. Time series, on the other hand, frequently lack labels and frequently exhibit anomalies when compared to regular sequences. Multivariate time series exhibit high nonlinear correlations and significant differences across distributions. It is challenging to model time series accurately because of these problems. Presently available models employ a GAN framework to handle problems with missing data and simulation time, but they show little interest in mitigating significant changes that are erratically distributed throughout a number of time series. In this research, we present a Variable Generative Adversarial Transformer (VGAT) for WGAN-based smart road anomaly detection development and unobservable detection. Replicators and discriminators are built on training regions and transformers. Firstly, by directly connecting replicated samples, VGAT functions as a replicator generator, simplifying the model and increasing efficiency. Second, regular sequences are compressed using VGAT's replicator with the aid of Soft-DTW procedures. The SPOT technique, which is based on Extreme Value Theory (EVT), independently determines the threshold after VGAT generates outliers, raising the likelihood that other data in the example has been altered. Performance (SOTA) is enhanced with VGAT when compared to benchmarks across several databases. The designed system is centered on identifying unusual road signs, such portholes and anti-speeding jaws, and keeping traffic moving when such signals are identified. We suggest FLC (Modified Fuzzy Logic Control) and HOG (Histogram Oriented Gradients) in this work. Moreover, promising outcomes demonstrate that the suggested HOG algorithm performs better than alternative methods in identifying holes and bubbles quickly. With a WADI index F1 score of 68.19%, VGAT demonstrates a noteworthy advancement above SOTA models at this time.