As part of the digitization project carried out for one of our main clients, one of the areas addressed was proximity-based presence control and IoT, as well as real-time tracking of the entry and exit of tools, parts, and equipment across different virtual regions of the production plant.
Once processed, this data allows us to approximately locate each element in real-time and, furthermore, accurately trace the paths they follow within the production flow.
In conclusion, the information provided by the collected data offers a real-time snapshot of the production status from a physical perspective. However, a bigger challenge was posed: Can we use this data to determine if something is going wrong? The answer is yes.
Objective? To implement an AI model that, fed with continuously collected data related to positioning and entries and exits, alerts in real-time when something might be going wrong. Most analysis systems are limited to comparing deviations from a pre-established plan. However, we want a solution capable of discovering hidden patterns.
What data do I want: patterns from time series. The first question we asked, obviously, is what specific data we should use. This choice directly affects the quality of the model, as it is built around that data. Each time an entity enters or exits a specific region, the system stores a real-time record of this event, thus obtaining time series of events that can be exploited to extract models or behavioral patterns. Therefore, the solution lies in implementing a model capable of detecting anomalies in those patterns.
For the construction of the model, the K-means algorithm was used, implemented through the scikit-learn library integrated into the project. The resulting model is a cluster trained with three centroids, so that all points (entry or exit events in the practical scenario) classified by the model into a cluster containing less than 20% of the total points (events) could be considered anomalies.
Very well, but how is it integrated into the production control system? The use of the anomaly detection model by third parties is implemented through the deployment of its services on the system's REST API, providing its real-time results for each request that is made. It is important to highlight that the manual action of the end user can be crucial for the training and continuous improvement of the model in relation to the classification of anomalies.
One step further in process improvement through the use of data.
The solution provided, beyond being a functional element that contributes to process improvement, introduces innovative and differentiating mechanisms that open a new path for progress: the use of data. It is important to highlight that the manual action of the end user can be crucial for the training and continuous improvement of the model regarding anomaly classification.
Beyond the unsupervised automation it proposes, it is considered valuable to provide interfaces that allow a user to manually declare an event or occurrence as anomalous, thus assisting in the categorization process and, therefore, contributing to the continuous improvement of the system.
- AI applied to the production process in a simple, accessible, and efficient way.
- Reuse of data generated by third-party systems that leads to improvements.
- Discovery of hidden bottlenecks, not visible or difficult to detect.
As part of the digitization project carried out for one of our main clients, one of the areas addressed was proximity-based presence control and IoT, as well as real-time tracking of the entry and exit of tools, parts, and equipment across different virtual regions of the production plant.
Once processed, this data allows us to approximately locate each element in real-time and, furthermore, accurately trace the paths they follow within the production flow.
In conclusion, the information provided by the collected data offers a real-time snapshot of the production status from a physical perspective. However, a bigger challenge was posed: Can we use this data to determine if something is going wrong? The answer is yes.
Objective? To implement an AI model that, fed with continuously collected data related to positioning and entries and exits, alerts in real-time when something might be going wrong. Most analysis systems are limited to comparing deviations from a pre-established plan. However, we want a solution capable of discovering hidden patterns.
What data do I want: patterns from time series. The first question we asked, obviously, is what specific data we should use. This choice directly affects the quality of the model, as it is built around that data. Each time an entity enters or exits a specific region, the system stores a real-time record of this event, thus obtaining time series of events that can be exploited to extract models or behavioral patterns. Therefore, the solution lies in implementing a model capable of detecting anomalies in those patterns.
For the construction of the model, the K-means algorithm was used, implemented through the scikit-learn library integrated into the project. The resulting model is a cluster trained with three centroids, so that all points (entry or exit events in the practical scenario) classified by the model into a cluster containing less than 20% of the total points (events) could be considered anomalies.
Very well, but how is it integrated into the production control system? The use of the anomaly detection model by third parties is implemented through the deployment of its services on the system's REST API, providing its real-time results for each request that is made. It is important to highlight that the manual action of the end user can be crucial for the training and continuous improvement of the model in relation to the classification of anomalies.
One step further in process improvement through the use of data.
The solution provided, beyond being a functional element that contributes to process improvement, introduces innovative and differentiating mechanisms that open a new path for progress: the use of data. It is important to highlight that the manual action of the end user can be crucial for the training and continuous improvement of the model regarding anomaly classification.
Beyond the unsupervised automation it proposes, it is considered valuable to provide interfaces that allow a user to manually declare an event or occurrence as anomalous, thus assisting in the categorization process and, therefore, contributing to the continuous improvement of the system.
- AI applied to the production process in a simple, accessible, and efficient way.
- Reuse of data generated by third-party systems that leads to improvements.
- Discovery of hidden bottlenecks, not visible or difficult to detect.