Developing a battery-powered industrial sensor that is retrofitted to a machine, for example to automatically detect maintenance requirements with the help of an AI application and avoid unplanned machine downtimes, is a complex multidisciplinary task. In some cases, completely different hardware and software topics as well as data science and project management tasks need to be addressed.

For the sensor as a functional communication endpoint, a single-chip microcontroller with relatively low memory resources and embedded firmware with measurement data acquisition, wireless communication functions and high-performance power management are required. Both in the cloud as an additional end point and directly on site - i.e. at the edge - special software modules are required that extract valuable information from the sensor data and make automatic decisions, for example to adapt machine operation to the respective condition, order the necessary spare parts and set a maintenance date for operators and service managers.

As resource-limited wireless sensors for battery operation should only have a direct connection to the cloud in exceptional cases for various reasons (e.g. energy requirements, cybersecurity, complex communication protocols, etc.), external gateway functions are also required as a link. They enable protocol and data conversions as well as management tasks such as over-the-air software updates (OTA) and sensor data pre-processing through to AI-based edge data analyses. A highly secure and trustworthy end-to-end communication connection can be established between the two endpoints, i.e. the wireless sensing endpoint and the cloud endpoint, using a digital signature process based on an asymmetric cryptosystem. It guarantees legal certainty with regard to decisions made automatically and the costs incurred as a result.

Intelligent human behaviour requires the correct recognition of a specific situation in order to act in a context-related and targeted manner. If this procedure is automated with the help of sensors and AI algorithms for condition monitoring, significant efficiency improvements for machines and systems are possible.

Data analysis and decision-making

The implementation of a condition monitoring application with wireless sensors and artificial intelligence (AI) requires coordinated software modules for three different platforms: the wireless sensor endpoint, an edge gateway and the cloud. The software for the wireless sensor plus various edge gateway components are usually tied to specific hardware. Due to the complexity, it is advisable to use a pre-developed technology stack, such as the 'WSEI/154A' from SSV.

© SSV Software Systems

A suitable AI building block for data analysis and decision-making in industrial sensor applications is 'TensorFlow'. Originally developed by
This open source framework for data flow-oriented programming, originally developed by Google, enables highly complex machine learning applications based on artificial neural networks (so-called deep learning applications). TensorFlow is suitable for two tasks in a condition monitoring application with wireless sensors, gateway and cloud:

1. modeling in the cloud endpoint: In a learning phase, a machine learning model is first generated. For this purpose, a classification or regression algorithm is configured in the form of a neural network and equipped with the corresponding model parameters, the so-called trainable parameters, using previously prepared training data in an iterative learning phase. This learning phase for deep learning model generation is controlled by a back-propagation algorithm.

2. inference (model use) within the gateway functions: Continuous inference operation is used to obtain information from the data of a sensing endpoint. For this purpose, sensor data is periodically sent to a gateway and analyzed via a TensorFlow inference runtime environment using the model. The result of this machine learning inference is suitable for automated decision-making, but also for logging machine and system statuses in a database.

As a rule, deep learning modeling with TensorFlow requires a very powerful computing platform, which is usually available in the cloud. The inference operation of a condition monitoring application, on the other hand, can be carried out with relatively low computing power. An embedded system - such as an ARM Cortex A5-based embedded Linux module with low-power features - is already sufficient for machine learning inference with wireless sensing data. A full TensorFlow installation is also not required on such a gateway platform, only a 'TensorFlow Lite' runtime environment.

'TensorFlow Lite' uses a portable interpreter for inference to perform a classification or regression calculation of the most recently received sensor data using the machine learning model. To enable the TensorFlow Lite interpreter to use the model generated in the cloud for inference on resource-limited hardware, it is first converted into a special format; the workflow required for this is also known as 'TinyML'.

Diverse communication scenario

The typical data transmission scenario of a condition monitoring application with wireless sensors and edge data analysis via 'TensorFlow' fulfills various communication tasks. There are a total of four data transmission paths: two each between the wireless sensing endpoint and the gateway and between the gateway and a cloud endpoint. In addition to software updates to the sensor, new machine learning models are transferred from the cloud to the edge gateway as required.

© SSV Software Systems

In a condition monitoring application with wireless sensors and edge data analysis via TensorFlow, there are various communication tasks. There are a total of four data transmission paths: two each between the wireless sensing endpoint and the gateway and the gateway and a cloud endpoint:

1. sensor data from the Sensing Endpoint to the Gateway: first of all, this is the classic interface for passing on the digital sensor readings to other systems - in this case a Wireless Sensor Gateway. The behavior of this interface can be influenced by updates, which are transmitted in the opposite direction, with various parameters. It must therefore be possible to switch the sensor between at least two system operating states: firstly, sensor data for data acquisition for TensorFlow modeling in the cloud (training data acquisition) and secondly, sensor data for the periodic generation of feature vectors that are suitable for the inference mode of the gateway (machine learning inference).

2. software updates from the gateway to the sensor: In addition to a complete or differential sensor software update, this communication relationship also enables the transmission of new configuration data, for example to change the parameter values of individual sensor operating modes (new interval times for sensor data required for machine learning inference) or to trigger the switchover between the modeling and inference operating states described above.

3. sensor data and inference results from the gateway to the cloud endpoint: The data received from the sensing endpoint can be forwarded to the cloud endpoint unchanged or pre-processed. Alternatively, when using TensorFlow and TensorFlow Lite for data analysis, only the results of the TensorFlow Lite inference phase are sent to the cloud. To do this, the gateway periodically performs an inference with the machine learning model generated in the cloud and forwards the output data of the model usage.

4. software updates from the cloud endpoint to the gateway: All files for gateway and sensing endpoint updates are stored on an update server in the cloud and downloaded from there to the gateway. The files on this server contain configuration data, software updates and new machine learning models that were previously generated by another learning phase. As such over-the-air updates are classified as high-risk in terms of security, state-of-the-art end-to-end IT security with a public key infrastructure (PKI) and the corresponding experts for operations (i.e. a DevSecOp team) are essential.

Complex, but valuable

At first glance, such a condition monitoring solution seems relatively complex. The question might arise as to whether it would be possible to simply continue streaming sensor data to a cloud in order to visualize it in a dashboard?

The difference definitely lies in the quality of the results: When it comes to data analysis for pattern recognition, AI algorithms are now better than most humans. Ultimately, it's like autonomous driving: The technology required for this also seems a little daunting at first glance due to its complexity, but the potential user benefits are fascinating.

Wireless Sensor Edge Intelligence

Klaus-Dieter Walter is CEO of SSV Software Systems in Hanover.

© SSV Software Systems

Countless IoT sensor applications stream raw data to the cloud in order to use the data processing options available there. In addition to security concerns, this central solution approach in the industrial environment also has functional disadvantages due to bandwidth, latency and availability problems. A cobot voice/gesture interface for collaboration between humans and robots, quality assurance via machine vision, condition monitoring with real-time anomaly detection and driverless transport systems (AGVs) cannot be implemented with a simple sensor-to-cloud connection. Additional on-site data evaluation is required here.

To simplify the development of AI-based wireless sensor edge applications, SSV has developed the WSEI/154A, an end-to-end technology stack with 868 MHz radio technology in accordance with IEEE 802.15.4. As a sensing endpoint, an evaluation board with an ARM Cortex M0+ SoC and integrated sub-GHz radio transceiver as well as RIOT as the embedded operating system are included in the scope of delivery.

The numerous gateway software functions are tailored to a Debian Linux and are supplemented by a pre-certified radio hardware as a solder-on module. Various Jupyter notebooks are part of the technology stack for cloud use. They support MQTT communication with the gateway and the training of neural networks with TensorFlow. The machine learning models generated in the process can be used in edge inference mode on resource-limited gateway hardware. The WSEI/154A enables OEMs to implement data-based solutions with wireless sensors, various communication protocols, machine learning, PKI-based cybersecurity including authentication and secure over-the-air (OTA) software updates right into the sensor without having to build up in-depth specialist knowledge in the individual subject areas beforehand.