Post Training Quantization of a Compact 1D CNN for Bearing Fault Diagnosis

This study evaluates the static post-training INT8 quantization in a compact one-dimensional convolutional neural network to diagnose bearing fault. The model has 9,188 trainable parameters, and the four bearing states are classified according to normalized vibration windows. CWRU records at 1 hp, 2 hp and 3 hp are used for training, validation and testing, respectively. Each partition has 472 windows. Five models are trained from scratch, and only training windows are used for calibration. The saved state size is reduced from 38.75 KiB to 18.95 KiB after quantization, a decrease of 51.11%. The accuracy of the test decreases from 99.45% to 96.44%. After quantization, test accuracy varies more across training seeds. Newly introduced errors are concentrated in the outer-race fault classification. The mean CPU latency per window increases from 0.316 milliseconds to 0.623 milliseconds. The test results show storage benefits, but there is no speed improvement in the tested configuration. The shared bearing configuration and limited recording limit the generalization claim. Quantization needs separate evaluation of storage, accuracy, and hardware latency, even if the starting model is relatively compact.