Use Artificial Intelligence to Predict the Duration to the Next Equipment Downtime’s artificial intelligence algorithms analyze equipment performance 24×7 and thus help employees to increase equipment productivity. From the library of available algorithms, in this article we present the AI ​​algorithm “Subsequent Loss Prediction”, which predicts the time until the next equipment downtime.

The technological basis of the function is the Long Short Term Memory (LSTM) algorithm, which, to put it colloquially, has a short-term memory. This memory is used to analyze the data situation just before the standstill in order to identify a similarity with other situations from the past.

The prediction question is presented as a classification problem for the algorithm. So we form classes, e.g. in 10-minute steps, and expect the algorithm to answer the probability that, based on its knowledge from short-term memory, the next disruption in the cluster will be 0 to 10 minutes, 10 to 20 minutes, etc. from this point in time.

For training purposes, historical data in the time series database is used, with the general rule being that more training data produces a higher precision of the prediction. From experience we can say that at least 3 months of data history is required for basic training.

The training is equipment-specific. So-called graphics processors (GPUs), which are used in consumer products such as gaming PCs, are used for this. GPUs are specially designed for this type of calculation and are therefore so fast that the model can be regularly retrained. Since changes are the order of the day in the industrial environment – new products, other employees, etc. – we at rely on retraining the algorithm once a week.

For the training, the algorithm is provided with different parameters, including the current reason for the loss, which are then processed over several levels in the LSTM model.

The accuracy of the algorithm for the respective system is already known immediately after the training – it is calculated by the training infrastructure as a by-product. Depending on the existing data quality and the configured algorithm parameters, an average prediction accuracy of 80+% can be achieved for this question.

Image: Results of different training runs

Unfortunately, due to the great influence of the supplied data quality, a forecast quality cannot be predicted even by an experienced data scientist before the project starts. This information is only available once the algorithm has been configured in detail and a training run has been performed.

This type of forecast is used, for example, for resource planning in maintenance. This can be done by a maintenance worker already heading towards the equipment when the prognosis is < 10 minutes until the next downtime or, in another interpretation, that the maintenance worker can leave the equipment after a repair if the next downtime > 40 minutes is predicted. The information is then displayed, for example, on an Andon board or tablet to enter the reason for the loss.

Are you interested in these or other prediction questions in industrial manufacturing? Feel free to contact us.