A Variational Bayesian Inference with Small Dataset for High-Precision Infrared Thermal Imaging

Recently infrared thermal imaging with high-precision has been widely investigated to detect abnormal thermal defects caused by current leakage, component heating and temperature fluctuation etc. However, two major factors seriously limit the imaging precision: measure uncertainty and small dataset. Therefore, this paper presents a variational Bayesian inference (VBI) with sparsity-enforcing prior so as to solve the above challenges. Though the sampled information is incomplete and measurement error are not trivial, especially in fault diagnosis of electrical power system, the advantages of VBI are that Bayesian cost function will still combine prior model (of abnormal thermal source) and likelihood model (of measurement errors) together to regulate the uncertainty from both physics and measurements. Meanwhile, sparsity priors of irregular temperature features will not only embody the health characteristics of electrical device, but also reduce the dimension of prior model parameters. Such sparsity priors will be modelled in terms of the sparsity-enforcing distributions. Indeed, most of signals can be sparsely represented in certain transformed domain and key knowledge of signatures will be deeply studied using sparse samples rather than condensed data. More importantly, both the prior model and cost function will be trained on-line and updated according to actual and historical data, instead of replying only on definite mechanisms or empirical rules which are hard to be adapted in industry application. Moreover, the proposed method can calibrate the uncertainty of conventional thermal radiation model by updating the model hyper-parameters and latent variables. Through a moderate infrared sensor, results of thermal imaging on ABB controllers and Schneider switches etc. confirm that proposed VBI has the advantages of high-precision(<; 0.95 °C ), far-field detection(>2.0m) and fast inspection (<; .8s for 160x120 pixels), and it is cost-effective to fast detect abnormal thermal sources.