Elevator Pitch
Smart sensors enable smart production as they enable the transition from big data to smart data. But sensors are only as smart as they are designed. Our toolbox realizes the problem-solution fit. From model-based analysis to hardware prototyping we accompany you on the way to your smart solution.
Description
On the path towards smart factories, automation of production is if central interest. To achieve this, accurate sensors – the senses of the system – are required. However, they mostly transmit their measured values as raw data to a cloud, where data processing is then carried out. This classic approach has disadvantages in terms of latency and data security and a trend towards smart sensors is evolving. Instead of raw data, smart sensors transmit pre-processed signals with relevant information (features) which have been autonomously determined and processed in the sensor. This decentralized approach significantly relieves the network and enables the transition from Big Data to Smart Data. However, in order to realize a smart factory using intelligent sensor technology, it is necessary to think holistically about automation and to combine the possibilities of smart sensor technology, communication and the cloud, to optimally coordinate them with each other and finally to use them. In view of the many implementation options for hardware, software, and automation systems, Fraunhofer IIS/EAS supports companies in identifying the right overall solution. Our toolbox therefore facilitates decision-making and the design of hardware and software building blocks with model-based methods that range from problem to integrated hardware implementation. Our approach combines classical signal processing with artificial intelligence methods, including neural networks. This allows the optimal combination of both processing schemes and enables robust and efficient feature extraction through tailored software and hardware close to the sensor. Conducting the analysis makes the potential of smart sensor technology quantifiable and, at the same time, identifies the associated hardware requirements. Our method and know-how help both users from mechanical and plant engineering as well as the operators of such plants to identify and prototype the appropriate hardware and software structure for implementing the targeted IIoT solution, thus making the senses of the factory smart.
Notes
Addressing the field of edge AI computing, we analyzed off-the-shelf hardware solutions for two different use cases: 1) object detection from video stream and 2) industrial condition monitoring based on time series signals. As part of a development flow for industrial monitoring problems, the analysis covers the feasibility study of necessary hardware. The analysis consists of FFT preprocessing followed by different deep neural networks with: 1) fully-connected layers, 2) Convolution, 3) LSTM. This signal chain was utilized and compared on different hardware to analyze runtime, estimated power consumption, and accuracy. First a performant server was used for general setup and development of signal preprocessing and deep neural network. When completed, it was migrated to different mini computers (partially reduced accuracy) for benchmarking. The conclusion of checking these outcomes against the use cases is that with well-aligned signal preprocessing and deep neural networks, the use cased could be run on fitting mini-computers being in part capable of realizing real-time at power levels so that battery is plausible, too. As follow-up steps we support prototyping of integrated circuit design in order to meet requirements such as size, power consumption or robustness.