|
Title:
|
Machine Learning for Fault and Anomaly Detection in IoT Sensor Networks: An Experimental Comparison of Classical Classifiers Under Edge Constraints
|
|
Author:
|
Dheerdhwaj Barak and Khushwant Singh
|
|
Keywords:
|
Machine Learning, Fault Detection, IoT, Sensor Networks, Anomaly Detection.
|
|
Abstract:
|
Deployments of the Internet of Things (IoT) produce flowing streams of sensor measurements and network telemetry, and much of this data is eventually utilized to answer the following one simple question: Is this device normal? A controlled experiment was conducted on this question and the results are reported herein. An artificially deployed smart-building is used to create a dataset with labeled sensor windows, corresponding to four conditions: normal operation, gradual sensor drift, a frozen (stuck) sensor, and a network flood. The results of seven classical learners are compared using 5-fold stratified cross validation: logistic regression, k-nearest neighbours, a decision tree, a random forest, gradient boosting, a support vector machine and a small multilayer perceptron. As well as accuracy, we evaluate the quantities that are important when the hardware is constrained: single sample inference latency, serialised model size, and degradation in the presence of sensor noise and missing values. Logistic regression achieved the best macro F1 (91.75%) and was the smallest model (approx. 1.8 KB) but this was just a bit behind random forest (91.46%) and gradient boosting (91.41%). The random forest used about 3.5 MB, which is nearly 2000 times as much storage. When noise was added to the model with a standard deviation of half the features the loss of macro-F1 point between 5 and 13 points for all models. We believe that instead of a ranking for accuracy, model selection for fault detection in IoT should be considered as a multi-objective trade-off, and robustness testing should be given the same weight as cross-validation. In this paper, the researchers used the data collected from the Internet of Things (IoT) to build a model that is robust against faults. The researchers utilized data from the Internet of Things (IoT) to develop a fault-tolerant model.
|
|
Download Paper:
|
|
|