A predictive maintenance solution based on the condition monitoring of a machine generally comprises three elements. Firstly, an Internet of Things (IoT) network for collecting data about (the condition of) the machine and its environment. Secondly, domain knowledge of the performance profile and failure modes of the machine and how to sense these. Finally, analytics to translate sensor data into information, i.e. predictions of machine condition. For OEMs, this last element usually poses the biggest challenge.
If you want to get started with predictive maintenance, you need a way to generate predictions. Predictions of the moment that machine performance will degrade due to the malfunctioning of certain components, or the moment that a machine will fail altogether. To be able to make these predictions, you need domain knowledge about the machine and the environment in which it operates. That knowledge tells you which indicators to look for and the questions you should ask: what are the different ways in which a machine will exhibit degraded performance or even fail; which signals should you observe for this; and which sensors do you need for that purpose? You then have to be able to read out these sensors and, using an IoT network, collect the necessary data for an analysis that will deliver the intended predictions.
To complete a predictive maintenance solution, you have to set up the analytics. A first idea may be to define a mathematical model for this, based on a description of the machine’s behavior derived from fundamental physics and engineering principles. Rules can be distilled from this model for recognizing when behavior deviates from the normal pattern. For example, vibrations, which occur in a machine naturally, can rise above a certain amplitude or pop up in unexpected frequency regions. This may indicate the abnormal behavior that is characteristic of a failure mode.
The advantage of such a rule-based approach is that it is derived from a description of physical reality, and therefore understandable, and thus helps facilitate reasoning about cause and effect. However, it also comes with a disadvantage. The model-based description often fails to capture reality sufficiently to make reliable predictions possible. One reason for this is that maintenance involves elusive phenomena such as wear and tear, which are very hard to describe with sufficient depth in a manageable model. In short, the rule-based approach is not Columbus’ egg.
As an alternative, machine learning can be used to produce a model that takes sensor data input to generate predictions without recourse to explicit physical knowledge. Machine learning comes in a variety of implementations, one of which is the so-called neural network. To some extent, a neural network mimics the workings of the human brain. It consists of a network of nodes, the artificial neurons, in which each connection between neurons is provided with a specific weight that has either a positive (excitatory) or a negative (inhibitory) value. Sensor data is fed as input to the model, which then has to generate a prediction about the behavior of the machine (“normal” or “abnormal” in the simplest, binary version) as its output.
The weights in a neural network can be modified, which influences the output of the model given a certain input. The model is first “trained” using so-called labelled data, meaning a set of known input and output data: sensor readings and the associated behavior, respectively. The set of weights, defining the model, is optimized to reproduce the known output. Then, in the real application, the model uses real-life sensor data to generate new output data, i.e. predictions.
Machine learning also comes with a disadvantage. The relation between the input and output of a neural network is not represented by an understandable model, but by the set of weights, which means that reasoning about it is not straightforward. The big advantage, however, is that accurate predictions can be generated provided the model has been optimized with a high-quality training data set.
Domain knowledge is still required, to identify the critical condition parameters that have to be sensed. But domain knowledge is no longer required to build a realistic model. Instead, data scientists come in to build the machine learning model – and this can be used for diverse applications, such as condition monitoring.
Condition monitoring for predictive maintenance can be achieved in various implementations, one of which is anomaly detection. This concerns relevant physical quantities identified by domain experts, such as temperature, pressure, flow, and frequency spectra of natural vibrations. Their sensor readings are used as input to the neural network, which generates an output that is a measure of the (ab)normal functioning of the machine. If this measure exceeds a certain threshold value, as determined during the training phase, then abnormal anomalous behavior is detected. Expert judgement may still be required to assess whether a detected anomaly is real or just an artefact, for example due to sensor malfunctioning.
Another implementation of condition monitoring is the prediction of future sensor data based on their history, i.e. involving the processing of time-series sensor data. This calls for the use of a recurrent neural network, which is characterized by connections between nodes having a temporal direction. A specific form of such a recurrent neural network architecture is a long short-term memory network, which is particularly suited for producing predictions based on time-series data. If the real sensor data differ from the predicted values, this may signal a potential malfunctioning.
We have seen that machine learning offers multiple options for condition monitoring to facilitate predictive maintenance. A successful implementation requires the combination of IoT expertise, domain knowledge and data science.
Do you want to start with predictive maintenance and need an introduction to analytics? Get in touch with us.
Grab your e-book copy today and get started: E-book: Predictive Maintenance from Scratch