More than 100 years after Frederick Winslow Taylor founded ergonomics, human assembly activities are still largely observed and measured manually. The consequences are a lack of information, incomplete data sets, observation biases and a lack of system perspectives in manual assembly processes. Considering that over 70% of assembly activities worldwide are still carried out by human labor, the Industrial Internet of Things has a huge, man-made blind spot here.

The status quo can be changed by technologies that revolutionize the methods of time and motion studies and eliminate the shortcomings mentioned above. Drishti, together with some of the world's leading manufacturing companies, including those based in Europe, is playing a pioneering role here.

Figure 2: A view of the variability at workstations: For each data point, the corresponding video can be called up with one click. Drishti offers a comprehensive analysis of standardized work steps in real time, for example in response to the questions: Where are capacity bottlenecks? Where is waste occurring?

© Drishti

Numerous features have been introduced to overcome these challenges:

• An almost completely unsupervised learning function enables accurate data generation with smaller data sets and a 10-fold reduction in training time compared to other methods.
• Thanks to 3D Convolution, Artificial Intelligence not only considers individual frames, but also the time aspect between them. In this way, Drishti achieves significantly higher accuracy than other solutions.
• Using occlusion detection with memory capability, the AI anticipates what is happening behind an object that is obscuring parts of the camera's field of view. For example, if a normally visible tool is obscured, the AI deduces from the movement of the visible arm holding the tool what is happening in the non-visible area of the image.
• To create tangible added value for customers, a fully integrated system architecture, a video pipeline and a data science engine were created.

Videos on root cause analysis

Image 3: Edge-based masking in the cameras ensures the protection of personal data.

© Drishti

The deployment of the solution begins with the identification of a use case at the customer's premises. After the project kick-off, the cameras are installed and connected to Drishti's cloud via a secure VPN connection. This can also be done by the customer's employees on site. In the following weeks, the AI is trained to recognize the process steps of the selected assembly lines. In close cooperation with the customer, the employees learn how to use the tools in their daily work to achieve their goals. The first successes in the customer's continuous improvement processes can usually be quantified within a few weeks of the start of the project.

Concrete benefit

The existing processes for investigating customer complaints are time-consuming and expensive. Since Drishti's database is connected to the customer's MES, the assembly process for each individual product can be traced in detail at a later date and the videos are easy to find. This reduces both the time and cost of investigating customer complaints: A German automotive supplier, for example, speaks of an average saving of the equivalent of 10,000 US dollars in one of its plants in Mexico.

Another positive aspect is the increase in productivity through continuous process improvements. The technology is always 'on' and therefore provides insights into production processes with human involvement on a scale and at a speed that was previously difficult to imagine. The data obtained on individual work steps and cycle times is processed clearly and systematically by the AI. Customers can quickly identify the causes of waste and define measures to eliminate them. One example: German automotive supplier Hella deliberately used Drishti on a line that was considered to be optimized in order to test the performance of the technology. Using conventional methods, it was no longer possible to identify new potential for improvement. With the help of Drishti, increases in throughput of up to 7% and system availability of 4% were achieved within ten weeks.