Microsoft Azure

Introduction

The unscheduled downtime of the equipment can be detrimental to any company. It is essential to keep field equipment operational to maximize usage and minimize costly unscheduled downtime and maintenance, safety and environmental risks. The goal of a good predictive maintenance strategy is to extend the life of the equipment and prevent errors. Anomaly detection is a common method because it indicates when a device is behaving differently than it should. Often, anomaly detection solutions are more accurate than a simple rule based on error detection methods. In addition, they are very useful to prevent costly errors or interruptions.

Prepare data

The first step in a predictive maintenance solution is to prepare the data. This includes data ingestion, cleaning and feature engineering. The problems of predictive maintenance usually include data such as the following:

Machine information (for example, engine size, make and model).

Telemetry data (for example, sensor data, such as temperature, pressure, vibration, fluid properties and operating speeds).

Maintenance history and interventions: history of repairs of a machine and records of operation.

Error history of the machine or of a component of interest.

To predict errors, the data must contain examples of correct and incorrect states. The more examples, the better and more generalizable the predictive maintenance models will be. It is also important to have data from both the devices that have failed and those that remain in service. The data may also include readings of equipment that has given an error for the specific problem that you are interested in and of devices that have failed for other reasons. In both cases, the more data you have, the better the solution will be.

Build and train

Many predictive maintenance solutions use multiclass classification models to calculate the remaining useful life of a resource. Use predictive maintenance using multiclass classification when you want to predict two results, that is, a time interval for an error to occur and the probability of error due to one of the many main causes. In addition to choosing the right algorithms, a well-functioning model requires well-adjusted hyperparameters. Among these parameters is, for example, the number of layers of a neural network, which is established before the training process begins. Often, hyperparameters are specified by a data scientist following the trial and error method. They affect the accuracy and performance of the model, and finding the optimal values ​​sometimes involves several iterations.

Each training round generates metrics that are used to evaluate the effectiveness of the model. Precision is the most popular metric that is used to describe the performance of a classifier, although, in predictive maintenance solutions, recovery and F1 scores are often used. Precision is defined as the number of true positives with respect to the number of true positives plus the number of false positives, while the recovery denotes the number of true positives with respect to the number of true positives plus the number of false negatives of the prediction instances. of errors. F1 scores consider both precision and recovery ratings.

Deploy

Once the most effective variant of a model has been identified, it is necessary to implement that model as a web service with a REST connection point. Then, business line applications or analysis software call the model. In the case of predictive maintenance, however, integral architectures usually include real-time machine telemetry, which is compiled by systems such as Azure Event Hubs. Stream Analytics ingests the data and processes it in real time. The processed data is passed to a predictive model web service and the results are displayed in a panel or sent to an alert mechanism that informs technicians or service personnel of possible problems. The ingested data can also be stored in historical databases and combined with external data, such as local databases, to send them back to training examples for modeling. The Internet of Things (IoT) scenarios can have a model implemented in the perimeter so that the detection takes place as close as possible to the event, both in time and space.