Using time-series signals to recognize vibration or acoustic patterns is a highly convenient and effective approach for diagnosing faults in machinery or detecting abnormalities in human beings. The aim of this study is to explore methods for identifying acoustic or vibrational modes using time series datas to capture more comprehensive features for mechanical and human fault detection. In this paper, we propose a novel CNN-based approach with MFCC and attention mechanism, we introduce two innovative components on CNN: MFCC fusion feature and multi-head self-attention mechanism (MH-SAM). MFCC feature extraction is performed on the acoustic or vibration data and the extracted features are then used as inputs to the model for fault detection tasks. MH-SAM enables the network to focus on different parts of data, enhancing its ability to identify anomalies effectively. We perform tests on three publicly accessible acoustic or vibration datasets to confirm the efficacy of our proposed algorithms and models.