4.5.3. Energy Efficiency

A simple metric that can be used to measure the software energy efficiency is presented in [123] and is shown in Equation (4). In Edge-AI, the useful work performed can be defined as the number of model inferences. As a result, Energy Efficiency can be measured as the number of inferences per Joule.

$$\text{Energy Efficiency} = \frac{\text{Useful Work Performed}}{\text{Energy Consumption}} = \frac{\text{Number of Infected}}{\text{Energy Consumption}} \tag{4}$$

## 4.5.4. Execution Time

This metric represents the execution time of a specific task in the ML process to obtain a valid result, which may include, model training or model inference [123], and are measured in seconds, being typically referred as the "training time" and "inference time", respectively.

#### **5. Cross-Analysis of G-IoT and Edge-AI: Key Findings**

Although Edge-AI G-IoT system deployment in real-world applications has already started, the research and development are still undergoing, and some issues compromise its wider acceptance, of which we highlight: trustworthiness (e.g., algorithm transparency, traceability, privacy, and data integrity); capacity (e.g., communications bandwidth and coverage, hardware constraints such as power and computational power, security in edge distributed architectures); heterogeneity (e.g., dealing with distinct data sources and formats as well as adapting with a variety of operational, technical, and human requirements); and scale (e.g., inadequate volume of publicly available data, high-quality data required to effectively simulate the physical world's complexity). In addition, the cross-analysis of the G-IoT and Edge-AI literature allows for obtaining the following key findings that can be useful for future developers and researchers:


#### **6. Application Case: Developing a Smart Workshop**

#### *6.1. Workshop Characterization and Edge-AI System Main Goals*

To illustrate the concepts described in the previous sections, it was selected a practical Industry 5.0 use case in a real-world scenario. Specifically, the selected Industry 5.0 scenario consists in an industrial workshop that looks for improving operator safety through IIoT sensors/actuators and Edge-AI. The chosen scenario is based on the previous work of the authors [125–127], which participated in a Joint Research Unit together with one of the largest shipbuilders in the world (Navantia). The specific scenario is the pipe workshop that such a company owns in its shipyard in Ferrol (Spain). The workshop manufactures pipes as follows:

1. First, raw pipes are stored in the Reception Area (shown in Figure 6a). Thus, they are collected by the workers as they are needed. If the pipes are delivered with dirt or grease, then, before being stored in the Reception Area, they are cleaned in the Cleaning Area (in Figure 6b). Operators need to keep away from the Cleaning Area unless authorized because of the presence of dangerous chemical products (e.g., chloridric acid, caustic soda) and water that is pressurized and hot.

(**a**) Reception Area.

(**b**) Cleaning Area.

**Figure 6.** Relevant areas of the workshop.

2. Second, every pipe is first cut in the Cutting Area according to the required dimensions. Really powerful (and dangerous) mechanical and plasma saws (shown in Figure 7a,b) are used in the Cutting Area. It is important to note that pipes are moved from the Reception Area to the Cutting Area (or from one area to any other area) by stacking them on pallets, which are carried by big gantries installed in the ceiling of the workshop (several pallets can be seen on the foreground of Figure 7b).

(**a**) Mechanical saws in the Cutting Area. **Figure 7.** Saws of the Cutting Area.

(**b**) Plasma saw in the Cutting Area.


(**a**) Outbound storage area. **Figure 8.** Main storage areas.

(**b**) Another outbound storage area.

Figure 9 depicts the main areas of the workshop floor map and shows the position of the IIoT cameras that monitor the presence of the workers. In addition, the dashed semicircles indicate the estimation of the field of view of such cameras. Specifically, Figure 9 shows 18 distinct areas of the factory floor that are equipped with cameras for continuous monitoring (24 h a day, 7 days a week) of a complete manufacturing process. Note that, in this specific application case, images should be neither transmitted nor recorded in the cloud, not only due to bandwidth and connectivity limitations, but also due to the impositions of the General Regulation on Data Protection (GDPR) in force.

**Figure 9.** Floor map of the smart workshop.

The objective of the proposed solution is to harness "visual wake words" in order to detect the presence of the workers with the help of cameras and then lock or unlock the deployed industrial devices and machinery, and automate the available security mechanisms. For instance, industrial robot arms or cutting machines can harm a worker during their operation when safety distance is not respected. Thus, the system takes advantage of the proposed mist AI-enabled architecture (described next in Section 6.2) to achieve two specific application goals:


Besides the mentioned goals, it is important to note that the proposed system impacts different circular economy aspects:

