Название:
Автор:
Жанр:
Серия:
Издательство:
href="#ulink_c0302b1d-037d-50e1-9c0c-8d142bbeca72">Table 1.7 Possible limits of advanced technologies and technical risks.
| Technological limits | Descriptions and mitigations |
|---|---|
| Production time velocity | Computational cost of sensing and actuation processes must follow the in‐line production timing. Hardware and data computing can introduce relevant delays for real‐time actuation. Cloud data processing or the use of GPUs can be possible solutions |
| AI training dataset | The training dataset is particularly important in order to define high performance AI models. The training dataset must be constructed with a high quantity of data. In the initial monitoring of the production process, there are few sensor data and the training dataset can provide incorrect outputs. A solution can be the creation of a training dataset based on the creation of artificial data constructed by the initial ones. The dataset must be built with criteria avoiding the creation of redundant data; for example, reproducing in bench test similar data having the features of the production ones |
| Sensor detection and transmission | The Internet of Things sensor network monitoring production must be designed to detect and to transmit efficiently the signals. The full set of sensors able to monitor a whole production line must be allocated in the precise part of the production line mainly where the production is critical. The sensor sampling frequency is a particularly important parameter to provide enough data to process. In this scenario, a well‐structured local area network, based on the efficient solutions as the optical fiber architecture, typically increases the performance of the data transmission system. Another aspect to improve for data transmission is the correct programming of the time windows synchronizing all sensors |
| Sensor calibration | The sensor calibration is fundamental for the specific application field and machine to control. The calibration must be performed according to the production layout. A correct sensor calibration provides good data to process. In order to calibrate a production layout correctly, it is useful to simulate before all the signals are involved in the whole production line |
| Noise | Strong and anomalous vibrations, heat sources, inhomogeneous light conditions, acoustic noises, pollution, and electromagnetic induced fields represent possible causes of incorrect measurements. The noise is decreased by means of electronic boards and amplifiers properly filtering and cleaning the sensor signals |
| Image segmentation techniques | Image segmentation techniques can be inefficient for the specific defect to detect. AI techniques are adopted as alternative approaches for a further quasi real‐time or post‐processing checking. The processing modality is selected based on the production process to control. If the check is performed at the end of the product processing, the post‐processing modality is adopted, thus avoiding the problem of the computational cost |
| Camera resolution | Camera resolution can be inadequate for image processing. The optical components to combine with the camera are very important for the detection improvement. Angle of view, field of view, and focal lens are the parameters and elements to verify before the choice and installation of the camera system in the production line. The production layout and the characteristics of the production define the optical requirements of the camera system. A correct design of the camera system also takes into account the pollution of the working environment which in general disturbs the image acquisition |
| IRT | For an IR thermal camera, as well as the optical parameters, the thermal resolution is also important (a good parameter is <20 mK at +30 °C). Heat sources or reflective surfaces near the detection object increase the measurement signal noise and can saturate the camera chip. The correct positioning and orientation of the thermal camera is important to avoid the problem generated by the reflective surfaces |
| Defect classification | Defect classification is typically performed by AI algorithms. The training model is structured to detect particular typologies of defects according to product features. The correct formulation of the training dataset plays an import role for correct defect detection and classification. For this purpose, it is necessary to know and pre‐classify defects, thus optimizing the AI training model to apply |
| Traceability | The traceability can be inaccurate mainly for indoor applications. The choice of the correct technology is a function of the production layout and production processes. Possible correctly positioned gates reading tags avoid the problem associated with the correct indoor positioning |
| Computational cost of smart sensors | The computational cost in smart hardware technologies must be adapted to the production velocity. For this reason, local data processing can be performed by high performance processors embedded in smart devices, or by GPUs |
| Big data system | Big data cannot be connected and synchronized with the ERP database systems. A typical condition is to implement a big data system into a database management system |
| Big data analytics | The analytics tool can fail if there are empty records or wrong data stored in the big data. The definition of criteria to store and filter data in the big data system ensures the correct creation of the dataset to use for AI modeling |
| Actuation | An incorrect PLC configuration generates a failure in sensing and actuation processing. Preliminary simulations, checks, and tests of the PLC integrated in the whole production line ensures the best actuation according to the specific feedback systems |
| AR | AR smart glasses can have a low image resolution and low acceleration sensing according to the application. Standard smart glasses integrate an IR thermal camera having a low thermal resolution. The alternative solution is to use a high‐resolution thermal camera wireless interconnected with smartphone cardboard systems avoiding the problem of low thermal resolution |
| Unstructured data | Many data are not structured and consequently cannot be transferred by the enterprise service bus/bus network or by the wireless system. The association roles, used to extract information from unstructured data, provide more information for the digitized knowledge gain |
| In‐line parameter setting by AI | The AI engine can fail providing wrong outputs for the real‐time machine parameter calibration and actuation. The AI must be properly checked before its introduction in the production processes |
| Automodification of parameter setting | The auto‐adaptive system able to change the machine parameter setting can be inaccurate by providing wrong operations and wrong scheduling actions. The automated modification of the parameters can be performed by AI algorithms which must be accurately tested for the specific dataset |
AI, artificial intelligence; AR, augmented reality; GPU, graphics processing unit; IR, infrared; IRT, infrared thermography; PLC, programmable logic controller.
An important parameter to check the measurement accuracy of monitoring sensors, is the signal‐to‐noise ratio (SNR) defined as:
(1.2)
where PSignal is the average signal strength, PNoise is the noise strength, ASignal is the
