We have thousands of sensors in our manufacturing plant and we’re trying to move from reactive to predictive maintenance. I’m looking into using Recurrent Neural Networks (RNNs) or LSTMs to analyze the time-series data. Has anyone had success with this, and what were the biggest data hurdles you faced?
3 answers
Predicting hardware failure with LSTMs is highly effective because they are great at capturing long-term dependencies in sensor data. The biggest hurdle we faced in 2023 was "data imbalance." Since machines rarely fail, 99% of your data represents healthy operations. If you just feed that into a model, it will learn to predict "no failure" every time and be 99% accurate but useless. You need to use techniques like SMOTE for oversampling the failure cases or use an Anomaly Detection approach instead of straight classification. We found that training the model to recognize "normal" and flagging anything that deviates significantly worked much better than trying to predict specific failure types.
How are you handling the latency of sending all that sensor data to the cloud for processing? Are you doing any "Edge AI" where the model runs locally on the gateway to give instant alerts, or is it all centralized?
Focus on your feature engineering. Raw sensor data is noisy. Applying Fourier transforms to look at vibration frequencies often provides better "signals" for the neural network than just raw timestamps.
Absolutely, Susan. Pre-processing the data into the frequency domain made our model's accuracy jump from 75% to nearly 92% in our recent pilot project.
Thomas, we actually use a hybrid approach. We do light-weight anomaly detection at the edge using quantized models. If an anomaly is detected, the raw high-frequency data is then pushed to our central Deep Learning model in the cloud for a more detailed root-cause analysis. This saves us about 70% on bandwidth costs while still giving us the deep insights we need for planning.