Rolling bearings are the vital joints of industrial equipment and its anomaly detection is a prerequisite for predictive maintenance. This study addresses the anomaly detection of rolling bearings under complex working conditions, focusing on the challenges of limited sample data, difficulties in feature extraction, and significant noise interference. To tackle the anomaly detection issue in rolling bearings, we construct a model based on Feature Transformation, Quantized Convolution, and Bidirectional Long Short-Term Memory (FT-QNN-Bi-LSTM). The FT-QNN-Bi-LSTM model demonstrates a strong capability for extracting spatiotemporal features from single time point, facilitating early anomaly detection in bearings. It achieves over 95% accuracy and provides more precise detection points on the Xi’an Jiaotong University bearing dataset compared to the existing methods.