SMART IOT SENSOR ANOMALY DETECTION AND MONITORING SYSTEM

Abstract

The Smart IoT Sensor Anomaly Detection and Monitoring System is designed to observe readings from a network of low-cost sensors and to identify abnormal behaviour as soon as it appears. Sensor networks are now used in homes, laboratories, cold-storage rooms, water tanks, industrial sheds, and electrical panels to measure quantities such as temperature, humidity, gas concentration, vibration, voltage, and current. These networks produce a continuous stream of values, and a fault in the monitored equipment or in the sensor itself usually shows up first as an unusual pattern in that stream. The proposed system aims to detect such patterns automatically rather than relying on a person to watch graphs throughout the day. The hardware side of the system is built around microcontroller nodes based on the ESP32 and similar boards, which read analog and digital sensors, perform basic filtering, and transmit the readings over Wi-Fi using the MQTT protocol. A central Python service receives these messages, stores them in a time-series database, and prepares them for analysis. The data is cleaned to remove missing packets, duplicate timestamps, and obvious spikes caused by electrical noise, and it is then resampled to a uniform interval so that readings from different nodes can be compared on the same time axis. For anomaly detection, the system compares several machine learning approaches implemented with scikit-learn. Statistical baselines such as rolling z-score and interquartile range limits are compared with Isolation Forest, One-Class Support Vector Machine, and Local Outlier Factor models. Features are extracted from sliding windows of the signal, including the mean, standard deviation, rate of change, peakto-peak value, and simple frequency-domain energy. The models are evaluated on recorded data in which faults such as sensor drift, stuck values, sudden spikes, and gradual overheating have been labelled, using precision, recall, F1-score, and detection delay as the main measures. A web dashboard presents the live state of every node, the recent history of each sensor, the anomaly score computed by the selected model, and a log of alerts. When the anomaly score crosses a configurable threshold, the system raises an alert on the dashboard and can send a notification by e-mail or messaging service. The dashboard also shows the battery voltage and received signal strength of each node, which helps the user separate genuine process faults from problems caused by weak power supply or poor wireless coverage.