As you all should know Brainchip and MYWAI announced they were...

As you all should know Brainchip and MYWAI announced they were partnering for the following purpose:

Laguna Hills, Calif. – January 16, 2023 – BrainChip Holdings Ltd (ASX: BRN, OTCQX: BRCHF, ADR: BCHPY), the world’s first commercial producer of ultra-low power, fully digital, event-based, neuromorphic AI IP, and MYWAI, the leading AIoT solution provider for Edge intelligence in the European Union (EU), today announced a strategic partnership to deliver next-generation Edge AI solutions leveraging neuromorphic compute.

The solutions will leverage BrainChip’s Akida™, which processes and learns from sensor data with unparalleled efficiency, precision, and economy of energy. It incorporates MYWAI’s AIoT Platform for EaaS (Equipment as a Service), which can stream, process, and prepare multimodal sensor data (e.g. time series, audio, vision, touch data) for Edge AI, as well as manage the Machine Learning Engineering for Production (MLOps) workflows, the Digital Ledger Technology (DLT) certification of the data and outcomes, in line with the new EU regulations for trustworthy AI.

The partnership is expected to accelerate the adoption of Edge AI in the industrial and robotic sectors and generate significant value for both companies and their customers. Developing and delivering robotic solutions for various industries, such as manufacturing, logistics, energy management and healthcare will be the focus of this partnership and allows customers to deploy AI at the very Edge of their equipment, achieving unprecedented performance, efficiency, and security.

“We believe in helping businesses reach new heights by adding intelligence to processes and machinery at the very Edge using Generative AI on the cloud and trustworthy AI at the Edge,” said Fabrizio Cardinali, CEO of MYWAI. “By integrating BrainChip’s Akida with our EaaS platform, we can enable our customers to optimize their processes and machinery with efficient AI, delivering intelligence where needed, when needed.”

Just recently the CEO of MYWAI during a discussion on LinkedIn wrote the following:

After a little bit of further research I found the following in a 236 page EU report:

Annex 5: RAISE - Robots as anintelligent services ecosystem

1. Basic information Name ofdemonstration

RAISE (Robots as an IntelligentServices Ecosystem)

Main objective

The objective of the RAISETMproject was to turn the manufacturing servitization trend to the benefit ofrobotic manufacturers developing a Robots-as-a-Service (RaaS) platformdemonstrating the possibility for 3rd party providers to offer value-adding services(e.g. prognostic maintenance, production intelligence, insurance, and paymentservices ..) to Industrial Robots manufacturers by means of new openinteroperability standards (e.g. the Robotics open standard by OPC UA).

Short description

The demonstrator was based onthe Equipment as a Service (EaaS) platform of prime proposer MYWAI SRL customenhanced for a world-leading Robots Manufacturer, Mitsubishi ElectricTM, with the engagement of Italianand Lithuanian Insurtech and Neurocomputing start-ups YOLO SRL andNeurotechnologijos UAB.

Owner of the demonstrator

MYWAI SRL

Responsible person

CEO MYWAI Fabrizio Cardinali, [email protected]

NACE

C28.4.9 - Manufacture of othermachine tools

Keywords

Robotics, Machine Learning, IoT- Cybersecurity - Artificial Intelligence - Predictive Maintenance – Revamping,Iot, artificial neural networks.

2. Innovation

Benefits for the users

The RAISETM Project enablesrobotic machine tool vendors to manage their Robotic fleet as an IntelligentServices EcoSystem using AI, IIoT and DLTs at the very edge of IndustrialRobots working in Real World today and in the Productive Metaverse tomorrow.Todate the RAISETM project has piloted delivering advanced Servicetech (e.g.prescriptive and prognostic maintenance), Insurtech (e.g. parametric insurancesand warranty extension) and Fintech (e.g. pay per use, pay per outcome)services to Robotic Workforces using the standard OPC UATM Robotics InformationModel to interface Robots and their digital twins to the MYWAI TM EaaS(Equipment as a Service) Platform.

Innovation

The RAISE platform supports thedelivery of Artificial Intelligence based Prognostic Maintenance and ProductionQuality Control at the very edge of Industry 4.0 machine via Time Series SmartData Labelling, MLOPS AI Pipeline Build UP followed by Edge, Fog or CloudDelivery of developed algorithms also supporting chip based delivery of neuralnetworks

Risks and limitations

The RAISE demonstrator helpsrobotic machine tool vendors to move towards a product as a service model inorder to cope with the slow down and rocky performance during pandemic and warperiod

Technology readiness level

7 - System model in operationalenvironment

3. Exploitation Sectors ofapplication

Robotics.

Potential sectors of application

Maintenance, Quality Control

Patents / Licenses / Copyrights

Hardware / Software

Hardware:

The Demonstrator was based on a proprietary edge computer and IMU sensors by MYWAI

Software:

d on the MYWAI Equipment as aService platform provided by MYWAI

4. Media Photos

Video

Project Intro Video

HTTPS://YOUTU.BE/SDZHJNKGM9W

Commented Video

HTTPS://YOUTU.BE/CS-BOCTRW5Q

5. Modules

MYWAY 4 Robotics AI algorithmscatalogue

Main functionalities

This module consists of acatalogue of AI algorithms for the analysis of sensory data coming fromindustrial equipment. In particular, the presented algorithms allow tohighlight deviations from standard behaviours and identify recurrent patterns.All the presented algorithms have been tested on triaxial data coming from aseries of Inertial Measurement Units (IMUs) attached to a robotic manipulator,but their interfaces are modular, and they can be adapted to other applicationsor sensors.

In particular the first set of AI algorithmsdeveloped and shared enable the intelligent, unsupervised detection of:

• Robotic Joints fault detection

• Robotic Movement patternsearching

• Robotic Multivariate crossrelated fault detection

Technical specifications

The module groups threedifferent algorithms for the analysis of industrial machinery activities:

• Radial Basis Function NeuralNetwork for fault detection

• Dynamic Time Warping forpattern searching

• Principal Component Analysisfor fault detection

Radial Basis Function NeuralNetwork for fault detection

This algorithm is distributed astwo functions one to train the Radial Basis Function Neural Network (RBF- NN)and the second one to monitor the machinery behavior. In this algorithm theRBF-NN is used as a regressor to predict the sensor information at the nexttime instant given the information at the previous time stamp, as depicted inFigure 1. For this reason, it is necessary to train the RBF-NN to correctlypredict the sensor information by providing an example of the expectedbehavior. Therefore, the training function gets as input a multidimensionalarray containing the sensor data stream and the network parameters (i.e.,number of neurons and a value, aka gamma, controlling the level ofnon-linearity of the radial basis activation function). The sensor array has adimension of NxL, where L is the number of features (e.g., for a triaxialaccelerometer is going to be equal to tree) and N is the number of time stampscollected. Notice that, since the RBF-NN parameters are computed in closed formand only a single repetition of the expected behavior is necessary to train thenetwork, this step is carried out in a very limited amount of time. A vectorcontaining the network parameter resulting from the training process is thensaved on file. Similarly, the second function gets as input a multidimensionalarray, containing the data that should be monitored, a threshold, and loadsfrom file the RBF-NN parameters. The RBF-NN processes each time stamp up to N-1 generating a prediction and matches the prediction with the real value usingand Euclidean distances. If for a time stamp the prediction error is higherthan the threshold received, that timestamp is reported as an anomaly.Therefore, the module output is binary array of length N-1 specifying if eachtime stamp presents an anomaly. An example of fault detection applied to the IMUdata acquired from the robotic arm is shown

in Figure 2.

Dynamic Time Warping forpattern searching

This algorithm is composed of asingle function. This function gets as input two multidimensional arrays, onecontaining data associated to the reference pattern and a second one containingthe whole window of data where the search should be performed. Furthermore, thefunction gets as input two parameters, specified as percentage values: thelevel of desired correspondence between the reference pattern and possiblematches, and a value controlling the overlap between two consecutive windows.Notice that, the higher the overlap value, the more computationally intensivethe function becomes, as more iterations of the search are performed, but witha higher accuracy in successfully identifying correspondences.

The reference pattern has a sizeof NxL, where N is the sample length and L is the number of features. Instead,the second array has a size of MxL, where M is the dimension of the full windowon which the search is run. The module inspections the data using a slidingwindow of length N and uses Dynamic Time Warping (DTW) to measure the distanceof the data contained in the windows with the reference pattern. If thedistance is under the given threshold (which, in turn, is based on the desiredlevel of correspondence), the window is recognized as a match with thereference pattern. The module returns a binary array of length M where eachsample is specified if they belong to the reference pattern or not. Examples ofthe execution of the presented algorithm are shown in Figure 3 and Figure 4.

Principal ComponentAnalysis for fault detection

This algorithm is composed of asingle function. This function gets as input a single multidimensional arraycontaining the data to process, the maximum number of components that thealgorithm should consider and a threshold value. The multidimensional array ofdimension NxL, where L is the number of features (e.g., for a triaxialaccelerometer is going to be equal to tree) and N is the number of time stampscollected, is transformed using the Principal Component Analysis (PCA). PCA isa transformation process for multivariate data usually used for featurereduction. The components computed by the PCA are ranked according to the datavariation that they represent. This mechanism implies that, especially when arepetitive behavior is present, the first few principal components representthe behavior while the other contains only noise. We use this mechanismmonitoring the last components extracted from the PCA to identify theoccurrence of faults. This is performed thresholding the less representativecomponents. Therefore, the module output is binary array of length N specifyingif each time stamp presents an anomaly or not. The working mechanism of thisalgorithm is illustrated in Figure 5 and 6.

Inputs and outputs

Radial Basis Function NeuralNetwork for fault detection

The training function gets asinput an NxM array and the network parameters (number of neurons and avalue,akagamma,controllingthelevelofnon-linearityoftheradialbasisactivationfunction).Thetraining function writes on file the learned network parameters whose sizedepends on the number of neurons specified. The monitoring function gets asinput and NxM array, a threshold and the trained parameters previously saved onfile. The monitoring function returns a binary array of length N-1 specifyingif each time stamp represents an anomaly or not. Therefore, the function outputis binary array of length N specifying if each time stamp presents an anomalyor not.

Dynamic Time Warping forpattern searching

The algorithm gets as input anNxL array representing the pattern thought, an MxL array consisting of the fullwindow of data and two percentage values: correspondence and overlap. Then itreturns a binary array of length M specifying if each time stamp belongs to thereference pattern or not.

Principal Component Analysis forfault detection

The algorithm gets as input anNxL array containing the data to monitor, the number of principal components tocompute and a threshold. Then it returns a binary array of length N specifyingif each time stamp presents an anomaly or not.

Formats and standards

All the algorithms Pythonimplementation uses standard libraries and are available at the followinglink: HTTPS://GITHUB.COM/THEENGINEROOM-UNIGE/MYWAY4ROBOTICS- REPO

Owner (organization)

KNOWHEDGE SRL UNIVERSITY OFGENOA

This project hasreceived funding from the European Union's Horizon 2020 research and innovationprogramme under grant agreement No 825196

https://trinityrobotics.eu/wp-content/uploads/2023/06/D4.3.-Catalogue-of-use-case-demonstrations-2_compressed.pdf

Read together you might form an opinion that the engagement between Brainchip and MYWAI is flourishing.

My opinion only DYOR
Fact Finder