We start with the need to study and analyze the behavior of the manufacturing process of specific items, using only the manufacturing times of each. Is it possible to obtain useful information with only the manufacturing times? Of course. Manufacturing time is a clear indicator of many situations, but primarily it is a variable that is heavily affected by the potential anomalies that occur in any area of the process.
Whether an operator faces a problem executing an operation due to inadequate training, or if there is a logistical issue in dispatching the necessary materials to carry out the task, the manufacturing time will be directly affected.
However, the manufacturing time tends to be as close as possible to the estimated manufacturing time for each task. This makes the distribution of manufacturing times stationary by nature, which makes it ideal for detecting anomalies, trends, and fluctuations, as well as for predicting future values.
The goal is to analyze and try to explain anomalies and trends in the manufacturing process of certain items, using only manufacturing time as the main variable. At the same time, this analysis aims to enable predictions.
Determining analogous work units. The first problem to solve is determining which work units are analogous to each other, so they can be compared. In manufacturing industry, the concept of a production order executed on a specific item is commonly used. However, a production order is usually made up of a set of operations, those that must be performed on the material. In this case, it is the operations that determine the work unit, not the production orders. It is the execution time of the same operation over time that we can study as a time series.
Autoregressive analysis and stochastic methods. There are a multitude of methods for autoregressive analysis (AR), but focusing on stochastic methods, there are a number of AR-based models that are very useful, easy to implement and usually give good results for this type of studies. These are the Autoregression Models with Moving Averages (ARMA), which add a moving average effect model to the basic autoregression model; the Autoregression Models Integrated with Moving Averages (ARIMA), which include a component for the correction of non-stationary series; or the Seasonal Autoregression Models Integrated with Moving Averages (SARIMA), which include the possibility of studying series that fluctuate over different seasons.
A solution: Box-Jenkins methodology. The solution to our problem lies in applying the Box-Jenkins methodology, based on identifying candidate models and estimating their parameters, as well as analyzing the goodness of fit of the model through the study of residuals, known as white noise.
The study of time series is not only useful for obtaining value prediction models but also for anomaly detection. In our case study, a method was developed where, for each operation under analysis, the estimated manufacturing time for the next piece is obtained.
But at the same time, for each observed value, and depending on the difference between that value and the one predicted by the model, a very versatile mechanism for anomaly detection can be established.
In our case, this method was used to implement a mechanism where, once the actual manufacturing time of a part is recorded and compared with the estimated time, it is considered an anomaly or not depending on whether the difference is greater than the standard.
But perhaps the most important value obtained in this case was achieving methods that, using very common data, such as the recorded manufacturing times of a part, provide very useful information for decision-making.
We start with the need to study and analyze the behavior of the manufacturing process of specific items, using only the manufacturing times of each. Is it possible to obtain useful information with only the manufacturing times? Of course. Manufacturing time is a clear indicator of many situations, but primarily it is a variable that is heavily affected by the potential anomalies that occur in any area of the process.
Whether an operator faces a problem executing an operation due to inadequate training, or if there is a logistical issue in dispatching the necessary materials to carry out the task, the manufacturing time will be directly affected.
However, the manufacturing time tends to be as close as possible to the estimated manufacturing time for each task. This makes the distribution of manufacturing times stationary by nature, which makes it ideal for detecting anomalies, trends, and fluctuations, as well as for predicting future values.
The goal is to analyze and try to explain anomalies and trends in the manufacturing process of certain items, using only manufacturing time as the main variable. At the same time, this analysis aims to enable predictions.
Determining analogous work units. The first problem to solve is determining which work units are analogous to each other, so they can be compared. In manufacturing industry, the concept of a production order executed on a specific item is commonly used. However, a production order is usually made up of a set of operations, those that must be performed on the material. In this case, it is the operations that determine the work unit, not the production orders. It is the execution time of the same operation over time that we can study as a time series.
Autoregressive analysis and stochastic methods. There are a multitude of methods for autoregressive analysis (AR), but focusing on stochastic methods, there are a number of AR-based models that are very useful, easy to implement and usually give good results for this type of studies. These are the Autoregression Models with Moving Averages (ARMA), which add a moving average effect model to the basic autoregression model; the Autoregression Models Integrated with Moving Averages (ARIMA), which include a component for the correction of non-stationary series; or the Seasonal Autoregression Models Integrated with Moving Averages (SARIMA), which include the possibility of studying series that fluctuate over different seasons.
A solution: Box-Jenkins methodology. The solution to our problem lies in applying the Box-Jenkins methodology, based on identifying candidate models and estimating their parameters, as well as analyzing the goodness of fit of the model through the study of residuals, known as white noise.
The study of time series is not only useful for obtaining value prediction models but also for anomaly detection. In our case study, a method was developed where, for each operation under analysis, the estimated manufacturing time for the next piece is obtained.
But at the same time, for each observed value, and depending on the difference between that value and the one predicted by the model, a very versatile mechanism for anomaly detection can be established.
In our case, this method was used to implement a mechanism where, once the actual manufacturing time of a part is recorded and compared with the estimated time, it is considered an anomaly or not depending on whether the difference is greater than the standard.
But perhaps the most important value obtained in this case was achieving methods that, using very common data, such as the recorded manufacturing times of a part, provide very useful information for decision-making.