Leakage-Aware Evaluation of Fault-State Detectors for a Low-Cost ESP32 Multi-Sensor Battery Safety Monitor

Abstract Low-cost microcontroller monitors for lithium-ion battery safety are frequently reported with near-perfect machine-learning accuracy, yet the datasets behind such figures are seldom audited and are usually split at random. This study documents an ESP32 prototype with a component cost below 2,000 BDT (16.19 USD) that samples cell voltage, motor-branch current, cell temperature and MQ-2 gas-sensor output every 2 s, applies a three-level threshold rule and drives a protection relay, and audits the 5,186-row dataset recorded on it. The dataset joins four phase recordings, each carrying a single label. The firmware rule reproduces 97.13% of the labels, temperature alone separates the three states, only 24.7% of temperature values fall on the output grid of the DS18B20 sensor, and 48.7% of current samples are positive in a branch that should conduct in one direction. Eight detectors (the firmware rule, five tabular learners, a one-dimensional convolutional network and a long short-term memory network) were compared under random, blocked and blocked five-fold splits with a 60-row embargo over five seeds. With temperature and gas available, every detector reached at least 0.994 macro-F1 under the blocked split, so no learner improved on the 126-byte rule. With voltage and current only, random splitting raised macro-F1 by 0.183 to 0.271 over blocked splitting (paired t-test, p ≤ 0.002). Compiled for the ESP32, the trained tabular detectors occupied 205 to 29,256 bytes. Recording-aware splits and a dataset audit should precede any accuracy claim for such monitors.