Petersime: Cycle Similarity Analysis & Set-Point Discipline

In essence, Petersime’s machines replicate the biological incubation and hatching processes. To do this, two types of machines, located on the premises of their customers, are used: a setter or incubator and a hatcher. In a first stage, the eggs are placed in the setter. The setter aims to artificially create conditions that resemble the natural incubation process, by adjusting the angle and position of the egg and by regulating several environmental parameters like:

• Humidity
• Temperature
• Air pressure
• CO₂

Before moving the eggs to the hatcher, a viability check is performed, where non-viable eggs are removed. Similarly to the setter, the hatcher replicates the biological hatching process by adjusting environmental parameters. Five days after hatching, another viability check is performed, allowing to progress of only the most viable chicks.

Throughout the 21-day process (18 days incubation, 3 days hatching), many different actions have to be taken, as there typically is quite some variation in the natural processes as well. This holds many challenges with respect to controllability. The optimal parameter settings are determined by incubation specialists, and it is vital that these machines faithfully respect the pre-programmed setpoints, within a certain margin.

Despite having a well-thought-out setpoint program, sometimes, the machines give alerts either because the setpoints are not respected within the acceptable margins or because external factors disturb the process, for example, opening a door of the machine. Both can lead to a suboptimal incubation and hatching process, leading to non-viable chicks, hence a suboptimal yield for Petersime’s customers. It’s important to make a distinction between two types of setpoint breaches, each can have a significant impact on the incubation and hatching process. On the one hand, a setpoint breach might be a short yet extreme peak. On the other hand, a setpoint breach might be less extreme, but prolonged. As a hypothetical example, temperature might move from 38°C to 45°C for a few minutes or to 40°C for several hours.

Each scenario can have a detrimental impact on the process. To enhance its hatchery processes, Petersime acknowledges there is a need to gain insights into setpoint breaches on two levels. On a first level, Petersime wants to identify setpoint breaches in its machinery. However, merely identifying them is not sufficient; but understanding where the breaches originate from is vital to solving the related issues and improving the performance of the machines

To do so, Petersime partnered up with Faktion to build an AI-driven application to monitor the machines' setpoint discipline, gain insights into the root cause of setpoint breaches, and optimise the incubation and hatching processes.

The proposed solution entailed leveraging clustering algorithms to discover patterns in the sensor data. The clustering was performed on two levels: first, on the actual sensor values, but also on the difference between the setpoint and actual values. Clustering the setpoint breaches not only allows for looking for machines with similar breaches and hence suboptimal performance, but also allows for gaining insights into what is causing these breaches. As the machines are very sophisticated, finding the root cause of the problem in individual machines can be a complex matter; hence, bringing similar breaches together can reveal the root cause of the problem.

As an example, it might be that a certain machine at a certain location gives an alert due to temperatures not keeping up with the predefined setpoints. The clustering algorithm allows for looking for other machines in other locations that experience the same kind of problem. Being able to link these machines paves the way for additional analysis to find the root cause of the problem by comparing machines on a variety of criteria. These criteria can be, among others, type of machine, year of construction, time since last maintenance, position of the machine, incubation and hatching process settings, etc.

We started off by setting up data pre-processing pipelines to efficiently ingest and prepare data for analysis, which required complex data engineering steps. Subsequently, we performed a data exploration analysis followed by a Proof-of-Concept to demonstrate the performance of our solution. In the data exploration stage, data quality and potential were assessed, and a first analysis of expected sensor values during the setter and hatcher cycles was conducted. The deviations between performance and expectation were quantified, allowing for the identification of anomalies.

Through an extensive featurisation analysis of all cycles, relevant features were extracted from the cycle data. This featurisation analysis, together with dimensionality reduction techniques, allowed to transform complex cycle data into a meaningful, structured format that could be analysed in a systematic way.

Subsequently, we assessed cycle similarity by comparing the extracted features across different cycles and machines. This comparison has two applications. On the other hand, we could easily detect abnormal cycles that were not behaving according to their predefined setpoint program (outliers). On the other hand, by leveraging clustering techniques, similar deviant cycles and machines could be grouped together, which allowed for further investigation of the root cause of the setpoint deviation.

All insights from our analyses were combined in a highly interactive visualisation tool. This tool allowed Petersime’s incubation specialists and sales engineers to continuously monitor the behaviour of setters and hatchers, the detection of outliers, and the identification of causal anomalies. Moreover, Petersime was able to find an explanation for the outliers in the data and to map them back to historical events, for example downtimes, which provided immediate business value.

In the subsequent phases, our focus will shift towards further optimising or fine-tuning incubation points, together with expanding the insights derived from the PoC to include the hatchers in addition to the setters. This balanced approach of delivering business value quickly while simultaneously working towards long-term goals effectively combines immediate benefits with an eye on the future. Ultimately, Petersime’s customers would be able to use the visualisation tool and benefit from the insights.