Bearing Fault Diagnosis via Incremental Learning Based on the Repeated Replay Using Memory Indexing (R-REMIND) Method

Abstract

In recent years, deep-learning schemes have been widely and successfully used to diagnose bearing faults. However, as operating conditions change, the distribution of new data may differ from that of previously learned data. Training using only old data cannot guarantee good performance when handling new data, and vice versa. Here, we present an incremental learning scheme based on the Repeated Replay using Memory Indexing (R-REMIND) method for bearing fault diagnosis. R-REMIND can learn new information under various working conditions while retaining older information. First, we use a feature extraction network similar to the Inception-v4 neural network to collect bearing vibration data. Second, we encode the features by product quantization and store the features in indices. Finally, the parameters of the feature extraction and classification networks are updated using real and reconstructed features, and the model did not forget old information. The experiment results show that the R-REMIND model exhibits continuous learning ability with no catastrophic forgetting during sequential tasks.