The INFER research programme will focus on pervasively adaptive software systems for the development of an open, modular software platform for predictive modelling applicable in different industries and a next generation of adaptive soft sensors for on-line prediction, monitoring and control in the process industry. The application of smart adaptive systems will contribute to the operational excellence of the European process industry while the jointly developed software platform will translate the latest academic research results into a commercial software product and therefore strengthen the European competitiveness. Within this research programme, 14 industrial researchers will get the opportunity to gather knowledge in academia, 6 academic researchers will absorb knowledge in industry and 3 new experienced researchers will be recruited.

Motivation and background

The issues named below clearly indicate that although there are several soft sensors, data mining packages and general purpose data analysis tools on the market, none of them meets the requirements and objectives of this project and the product proposed here, which would be able to resolve the problems mentioned below.

Soft sensors

There is a very strong demand for adaptive predictive systems in the process industry, where predictive models are called soft sensors. The range of tasks they fulfill is very broad; the most dominant application area being the prediction of process variables, which can be determined either at low sampling rates or through off-line analysis only. These variables are very important for the process control and management as they are often related to the critical aspects of the process. Thus it is of great interest to deliver additional information at higher sampling rate and/or at lower financial cost, which is exactly the role of the soft sensors. Other important application fields of soft sensors are those of process monitoring and fault detection. The role of process monitoring soft sensors is to build multivariate features based on the historic data, which are relevant for the description of the process state. By presenting the predicted process state or the multivariate features, the soft sensor can support the process operators and allow them to make faster and more objective decisions.

Despite the high number of publications dealing with soft sensor applications, unaddressed pressing issues of the soft sensor development and maintenance remain, which prohibited wider spread of practical soft sensor applications in the process industry. A large part of the issues can be attributed to the data upon which the soft sensors are built. This data often shows varying quality (having outliers, missing values, measurement noise, data co-linearity, drifts), which makes the straight-forward application of common predictive models difficult. Currently, the most common approach to deal with these problems is obtaining as much of process knowledge as possible and incorporating it into the model. The process knowledge is usually applied to select important variables and steady states of the data, to remove data outliers, etc. However, it is problematic since this knowledge differs from one process to another and as such has to be acquired for each new soft sensor to be built. After acquiring the knowledge, it has to be manually incorporated into the models. It is not difficult to demonstrate that such ad-hoc approach is very expensive and a significant obstacle to wider soft sensors applicability. Another problematic fact for practical soft sensor applications is related to their run-time maintenance. After successful launch of the soft sensor, one can often observe a gradual deterioration of its performance. This is caused by such phenomena as varying quality of the input raw materials, changes in the catalyst activity, abrasion of mechanical components, external environment changes or the change of the operational state of the process. These variations affect the data influencing the operation of the softsensor. Usually after some time, the performance of the model reaches an unacceptable level and it has to be retrained or even rebuilt from scratch.

Although soft sensors have been developed for at least two decades and their importance has been constantly reiterated, they have not made a breakthrough as commercial products in the process industry. This is mainly due to the issues discussed above and the lack of suitable tools on the market.

Predictive modelling

There are several freely accessible general purpose data mining (for example WEKA, PRTools) and intelligent data analysis software packages and libraries on the market which could be used to develop soft sensors, but one of their main drawbacks is that advanced knowledge of how to select and configure available algorithms is required. A number of commercial data mining/predictive modelling software packages is also available. These tools automate some steps of the modelling process (e.g. data pre-processing, handling of missing values or even model complexity selection) thus reducing required expertise of the user. Most of them are however either front-ends for a single data mining/machine learning technique. All these tools have one thing in common: generated models are static and the lack of full adaptability implies the need of their periodic manual tuning or redesign.