Neuromorphic anomaly detector in the edge for predictive maintenance

The main objective of NeuroCPS4Maintenance is to develop and demonstrate a neuromorphic anomaly detector in the edge that is robust against concept drift, alerts of failures beforehand and provides a fast and real-time response for predictive maintenance applications in high demanding industrial scenarios (industrial stamp press.

PartnerRoleExpertise
ITCLExperiment leaderITCL is a non-profit Technology centre with a strong expertise in Neuromorphic Computing for time series analysis applied to predictive maintenance applications and Deep Learning for industrial applications.
MUVU TechnologiesDIHDIHBU is a fully operational DIH focused on Industry 4.0., suitable to coordinate the experiment and to incorporate its industrial members to the use cases, as well as to disseminate the first results.
IntigiaTechnology providerIntigia is a start-up specialized in implementing complex algorithms on embedded systems, especially on FPGAs and SoCs, for safety critical real time applications.

Solution

The main objective of NeuroCPS4Maintenance is to develop and demonstrate a neuromorphic anomaly detector in the edge that is robust against concept drift, alerts of failures beforehand and provides a fast and real-time response for predictive maintenance applications in high demanding industrial scenarios (industrial stamp press.

NeuroCPS4Maintenance aim to overcome the difficulties that SMEs faces to deploy predictive maintenance solutions due to lack of datasets and cybersecurity concerns.

The common goal of the project is to solve these difficulties by generating new solutions that can be applied in a short period of time, through specific applications for each type of machine and maintenance problems.

This consortium can be a first seed within DIH4CPS for the development of solutions based on neuromorphic processing in the edge for predictive maintenance replicable throughout Europe, and a first step for new use cases in robotics applied to other non-industrial sectors

  • ITCL, technical leader of the consortium, will work on the design and development of the time series analysis algorithms based on deep learning.
  • DIBHU will coordinate the diffusion of the project, define the use cases, and provide the application partners for the experiment.
  • Intigia will be the technical partner in charge of implementing these algorithms on FPGAs and deploy them on the field.

Within the planned experiment, it will be possible to allocate costs for each resource and its lifetime: the overhead cost of the machine over time and the cost of a variable for a given operation, the cost of labor per unit time, overhead and operational and other costs. This will allow carrying out design and optimization steps and minimize costs. This provides further business optimization opportunities for the end-user (4S 2000) and will support the company to make informed decisions on production orders with tight deadlines.

Expected results

The development and demonstration of the neuromorphic processor will make extensive usage of CPES technologies.  It will be develop the LSTM-drift algorithm and the hardware accelerators to implement it in real time and deploy the prototype in an industrial press (relevant environment), where its components can be validated.

The results of all this will be:

  1. A hardware prototype of a neuromorphic processor based on SoCs capable of sensing physical variables of the asset, interact with it to avoid failures and display health information.
  2. An ensemble Long Short-Term Memory (LSTM) for anomaly detection that adapts to concept drift implemented in real time on the neuromorphic processor.
  3. A real time estimation of the health status of the asset and generation of alarms that are accessible from the internet and displayed in insightful dashboards.
  4. Proof of concept applied to industrial stamp presses that operates in real time, detects failures beforehand and it is able to issue alerts or stop the machine before failures.
  5. Evaluation of the performance of the system in terms of accuracy, time to failure, robustness against changes in the machine over time and power consumption.

This neuromorphic anomaly detector will be the only new product developed in the experiment. It will be based on the previous works of ITCL and Intigia on LSTM and on edge computing on Zynq platform, respectively.  During the experiment we will develop the LSTM-drift algorithm and the hardware accelerators to implement it in real time and deploy the prototype in an industrial press (relevant environment).

Impact

The innovation capacity of this neuromorphic anomaly detector prototype will favour enabling further technical solutions in predictive maintenance in high demanding industrial environments since it will open up the possibility of future applications of the system to other types of machines or robotic units, especially in advanced manufacturing processes, giving rise to applications with high market potential.