Digital transformation: predictive maintenance with analytics
By Dr.-Ing. Rolf Sohrmann, VP Global Segment Management, Weidmüller
Wednesday, 08 April, 2020
Digitalisation offers machine builders and manufacturers opportunities to increase their productivity and differentiate themselves from competition through data-driven services.
Nowadays, machinery and plant are optimised to the extent that they can guarantee comparably low downtimes and the high quality of end product. However, new market requirements, such as customised production and increasingly shorter product lifecycles mean that manufacturing companies are faced with the task of further optimising processes in order to hold their own and advance in the face of international competition. Digitalisation offers machine builders and manufacturing companies new opportunities to increase their own productivity levels and differentiate themselves from the competition through data-driven services.
Artificial intelligence (AI) and the resulting data analysis methods open up a new approach to problems and product developments. They also form the basis for service-oriented business models. Predictive maintenance is a solution that is already available today and that creates considerable added value by reliably predicting machine behaviour.
Knowing what is going to happen: predicting errors with analytics
The existing sensors in machinery and plant record various measured values, such as temperature, pressure, energy consumption and vibrations. This data can create clarity with regard to individual procedures as well as the entire process — provided that the company is able to evaluate all the relevant information. Self-learning AI modules use recorded data to automatically generate models that are used as a reference for real-time operation. With the right approach, it is possible to detect even minor deviations from a machine’s normal behaviour that may point towards future problems. Depending on the complexity of the machines, the desired predictive quality and the forecast lead time, 95% prediction accuracy and 24-hour forecast lead times can be achieved through the use of trained analytics models.
To successfully implement predictive maintenance, a consistent and collaborative project workflow is required, from data acquisition via analytic modules to visualisation. Figure 1 explains the interaction between the individual process steps.
Within the scope of the implementation, a decision must be made as to whether the different processes are to take place locally or in the cloud. Hybrid solutions are also possible.
The data required for the AI model are acquired directly at the machine. The existing infrastructure is used to record and transfer the required data. If necessary, this can be supplemented by additional measuring points and interfaces. To optimise computing power and the amount of data, data selection and processing should be done locally if possible.
The self-learning model design based on the recorded data can take place either onsite or in the cloud. The same applies to the model application and visualisation. The decision regarding the point at which the modules should be executed for a predictive maintenance solution depends on the existing infrastructures, the datasets, security requirements and costs. Generally, users opt for hybrid solutions, with the proportions of local and cloud-based applications varying considerably. An analytics project should be conducted within a well-organised workflow. In practice, there is a proven five-stage procedure (see Figure 2). At the start of the project, the focus is on analysing the problem and defining the objective. During this phase, developers and users define, among other things, which specific failures it should be possible to predict. During the subsequent exploration phase, checks are carried out to see whether a defined error can be detected on the basis of the collected measured values or if a higher data quality is required. With the proof of concept (PoC), a statistical model is developed to automatically detect the error. Thus the technical and economic feasibility is checked on the basis of the previously recorded data (offline analysis). During the pilot phase, a functioning prototype is executed on an IoT platform in real time (online analysis). In doing so, findings and experience values are collected, which are implemented in the final analytics solution during the last phase. This solution can be applied to an unlimited number of machines of the same type.
The success prospects of an analytics project can and should be re-evaluated after each phase, to ensure that the final solution meets the objectives defined at the start. The duration of such a project is reasonable: two to three months are usually required from the start of the project through to the successful PoC. In order for the project workflow to be successful, the data scientists must have a deep understanding of the machine construction technologies. There must also be close exchange with the future user.
However, the creation of reliable analysis processes alone does not generate sufficient economic added value. Thanks to analytics solutions such as predictive maintenance, machine manufacturers can benefit from new business models in the service area and gain a competitive edge through new functions. The economic aspects with the creation of a cost/benefit analysis are also part of the project schedule.
Step by step towards predictive maintenance
In order to be able to predict machine behaviour and the quality of the end product through predictive maintenance, a number of steps need to be taken.
The first step on the way towards predictive maintenance is the importing, evaluation and illustration of the relevant data. Readability and quality are checked using sample data. In parallel, the recording frequencies are checked and adjusted, if necessary. The checked data are transferred in time series and should subsequently be illustrated in informative diagrams. The visualisation of all relevant data series in one single image enables initial findings on correlations between the sensor data.
Learning patterns and recognising anomalies
During operation, an analytics solution compares the previously automatically learned reference models with the real-time data of the machines. The AI modules are able to detect even minor deviations from the learned normal behaviour during operation. Due to the simultaneous consideration of all sensor data, errors and faults are determined that are not registered through rules-based systems. Future problems with the machine can be identified at an early stage.
Anomalies are evaluated according to the level of deviation, which is expressed via the anomaly index. Smaller deviations can be suppressed using a definable threshold for the anomaly index calculated by the system, so that messages are only issued for more serious discrepancies. Additionally, an accumulation of minor errors can trigger a message. Using anomaly detection, errors and abnormal behaviour can be detected earlier than with rules-based condition monitoring. The system also recognises such errors that are not usually registered by condition monitoring systems. As a result, machine operators and service technicians are able to rectify errors at an early stage and ensure the seamless operation of the production processes.
The detected errors are mapped in an anomaly sample and split into categories (important and unimportant). This classification is carried out in collaboration with data scientists and machine experts. The samples classified as important are assigned to reasons for failure, which are automatically detected in the system application. This means there is no need to spend lots of time searching for the source of an error when it occurs. Quicker diagnosis leads to noticeably shorter downtimes and thus optimised production performance and reduced costs. Furthermore, the user can adapt and supplement the classification. As a result, the systems improve over time and predict errors with greater precision.
Traditionally, maintenance intervals according to operating hours or quantities are applied. However, with this comes the risk of unnecessary service interventions or unexpected machine failures. Therefore, maintenance costs are a considerable factor that is difficult to calculate in relation to the overall operating costs of machinery and plant. With a predictive maintenance solution, the intelligent analysis of data generates knowledge of future errors or undesired states. This enables maintenance to be planned on a needs basis and to reduce the service costs (see Figure 3).
Depending on the requirement and available infrastructure, predictive maintenance software is able to display forecasted errors directly on the machine, in the machine builder’s service centre or the operator’s control room. The time it takes until a message regarding an anticipated error is issued varies with the cause of error and the anomaly classification. By notifying the responsible area, repair and maintenance processes can be shortened and planned in a targeted manner.
Due to the seamless monitoring of sensor, condition and process data, the quality of a product can be predicted across the individual production steps, therefore increasing the quality. Detailed reference models are developed for this purpose on the basis of historical data. These models are also used to identify previously undetected problems. The software enables quality problems to be identified during an ongoing process. By adjusting the production parameters early on, this secures the quality of the end product and reduces the scrap rate. As this requires interventions on the machine control and the operating procedures, it is carried out in close coordination with the machine builder and operator.
Existing machine data has the potential to generate considerable added value for machine builders and manufacturing companies. The key here is the profitable analysis and use of this data with an industrial analytics solution. There is no standard concept that suits all users — individual solutions are required that are developed in close coordination with the future user. The successful implementation of analytics projects requires that certain key criteria are met:
- Areas of application should be selected on the basis of a cost/benefit analysis at the start of the project.
- The machine know-how of the analytics experts is combined with the application knowledge of the user in order to develop a needs-based solution.
- The data infrastructure and data platform must be analysed, defined and supplemented, if necessary, in order to guarantee an end-to-end process.
- It should be possible to adapt the software and the user interface to the user’s requirements.
- The planning and execution take place collaboratively with the involvement of the analytics provider, the machine builder and the machine operator.
- The aim should be to establish a long-term partnership between the user and the developer in order to generate long-term added value from the analytics solution.
Specific examples from industry demonstrate that the use of analytics solutions, such as predictive maintenance, results in new business models, improves service concepts, generates additional revenue and reduces costs. Furthermore, previously undetected findings with regard to machine behaviour are leading to new product developments and the optimised use of sensors.
The ever-expanding overlap between IT and OT networks is forcing many organisations to assess and...
Would you ask a software vendor how easy it is to rip out and replace their software?
Python is now increasingly being used in the industrial automation and embedded systems world.