Operation Time Prediction
Operation Time Prediction is a tool that uses machine learning
to forecast machine run times for shop order operations. This helps you gain
insights into potential delays and improve production efficiency by highlighting
factors that impact the standard pace of production. This information is presented
in lobby elements and can be utilized for operational insights and decision-making.
The goal is to assist those involved in planning and running operations by providing
insights that can prevent problems before they happen and improve production
by identifying factors that affect the usual pace of production. A trained algorithm
also generates model statistics and dataset profiles, which can be useful for
gaining insights into the dataset and factors contributing to deviations in
shop order operation machine run times.
Basic Data Setup
To get the predicted operation machine run time on the Operation
Time Prediction lobby page, the machine learning model needs to
be trained and activated. This needs to be done on the Solution
Manager/Automation and Optimization/Machine Learning/Machine Learning Models
page. After training, the model needs to be set to Active status to enable the
use of the prediction model. The model can be retrained based on business requirements
to ensure the most accurate predictions using recent transactions.
AI Model Identifiers
- AI Model Name: OP_MACHINE_RUN_TIME_PREDICTION
- AI Model Use Case: op-machine-run-time-prediction
- Use Types:
- Trainable
- Batch prediction
Lobby – Operation Time Prediction
The Operation Time Prediction lobby provides an overview of the predictions and
includes the following elements:
- Operation Time Prediction – List Element: Displays shop order
operation information based on predictions. It compares the predicted
operation machine run time with the planned machine run time in hours. Each
row can be expanded to view the operation details on the Shop
Order Operations page.
Lobby filter values can be applied to limit deviations, such as delays in hours
or percentages of the planned time.
- Accumulated Time Offset per Work Center – Bar Chart: Shows the accumulated
predicted offset per Work Center. Note that if you filter on threshold values
in the lobby filter, the summation can be inaccurate.
- Maximum Predicted Offset – Counter: Displays the maximum delay in
terms of hours.
- Maximum Predicted Offset Percent – Counter: Displays the maximum
delay in terms of percentages of the planned time.
Lobby Filter Options
- Site: The site where predictions have been made (this should be the
same as set in the trainable model parameters).
- Past Days: The number of days prior to the current date for which historical
predictions should be displayed.
- Time Fence: The horizon extending from the
current date into the future that should be displayed (note that a limiting
factor here is the prediction horizon used in the scheduled prediction background
job).
- Work Center: Enables filtering on specific Work Centers.
- Shop Order: Enables filtering on specific shop order numbers.
- Part No: Allows filtering on specific part numbers.
- Order Status: Allows filtering on order status.
- Predicted Time Offset: Allows filtering delays greater than the given
number of hours.
- Offset Percent: Allows filtering delays greater than the given percentage
value.
Data Source
- Ml Planned So Operations: This queries stored predictions and
links them to corresponding shop order operations.
Conditional Format
It is recommended to update the conditional formatting to highlight operation
delay according to your own definitions.
Data Used in the Algorithm
The following attributes are considered when training the algorithm and making
predictions:
- Part routing operation average machine run time factor during the training
interval period
- Inventory part ABC class
- Inventory part accounting group
- Inventory part product code
- Operation Quantity (scrapped and used/planned)
- Operation Work Center resource
- Operation start month
- Operation start day of the week
- Operation start time of day
- Number of prior shop order operations
- Number of shop order operations left
- Work Center OEE
- Number of operation components
- Operation crew size
- Operation labor class
- Operation qualification profile
- Number of operations guidelines
- Number of operations analysis data points
- Number of operations tools
Based on this information, the algorithm predicts how the operation run time
factor will deviate. This, together with a pre-processed average machine run time
and the planned operation quantity, is used to compare the planned and predicted
operation machine run times.
Processing Example
- Training Example: Operation 20 of routing alternate 1 and revision
2 of part x123 has 12 records for the selected training period. The average
machine runtime factor of these records is 0.88 h/qty. In this example one registered
operation ID, with a start and stop clocking and qty, has an individual machine
run time factor of 0.90 h/qty. The training target input to the
model for this specific record would be 0.90 - 0.88 = 0.02. In this instance,
the model will consider that the features for this record contribute to a machine
run time factor offset of 0.02 hours per quantity.
- Prediction Example: Shop order 202020, Operation 20 of alternate
1 and revision 2 of part x123 is planned to start in 2 days. The plan is to
produce 100 parts, and the machine runtime, from the routing, on the operation
is 0.85 hours per quantity. The planned operation time is 85 hours. During prediction,
the attributes for the planned operations are sent to the trained model. The
returned value is a predicted offset from the calculated average. In this case,
the model reacted to the fact that the planned production start was late in
the evening on a Friday*. Based on this, the returned predicted target value
is an offset of 0.05 hours per quantity. The average machine runtime factor
is 0.88 hours per quantity (stored from the training). This means that the predicted
operation machine run time in this example is 100 (0.88 + 0.05) = 93 hours,
or 8 hours more than the original plan.
*Note that this is a hypothetical assumption and should not be interpreted as
the exact expected behavior of the algorithm.
Data Quality and Prediction Quality
The algorithm is best at handling common patterns with known data. New information
or unprecedented/novel scenarios that are not well reflected in the training set
will not be accurately predicted. If the data used as input is not well maintained
or accurate, the predictions will likely reflect this level of quality.
Configuration
To run the predictions the model must be configured by scheduling training and
prediction intervals.
Step 1 - Training
The algorithm can be trained and scheduled from the Machine Learning Model
or Machine Learning Models page.
Mandatory Parameters:
- SITE: The site used to make predictions.
- TRAINING_RANGE_DAYS: The number of past days to include in the training dataset.
- TRAINING_START_DAYS_AGO: The number of days from the current day
when the training dataset should begin.
Optional Parameter:
- WORK_CENTER: Filters the training dataset to a subset of selected
work centers. Use “;” as a delimiter between work centers.
Step 2 - Predictions Background Job
Predictions can be scheduled in Database Task or Database Tasks using the “Predict
Operation Machine Run Time” task. When this task is executed and finished, the planned
operations within the given parameters will have a predicted operation time.
Mandatory Parameters:
- Site: The site used to make predictions.
- OP_HORIZON_NUMBER_OF_DAYS: The horizon into the future in the number
of days from the time of predictions.
Optional Parameter:
- WORK_CENTER: Filters the prediction dataset to a subset of selected
work centers. Use “;” as a delimiter between work centers.
Optional step - Delete Predictions Background Job
This can be scheduled in Database Task or Database Tasks using the “Mass Delete
ML Shop Order Operation Predictions” task.
Mandatory Parameter:
- REMOVE_ENTRIES_OLDER_THAN_DAYS: predictions older than the given
number of days will be removed.
Scheduling Settings
The algorithm's training schedule can be customized based on specific requirements.
A recommended approach is to train the algorithm biannually using data from the
previous one to two years. However, this frequency can be adjusted according to
the unique needs of different industries and business contexts.
The prediction background job should be scheduled to run outside of regular business
hours, ensuring that results are available at the start of the next workday. This
scheduling can be tailored in various ways depending on the industry and business
context. For instance, one approach is to schedule the job to run daily with a 14-day
forecast. Alternatively, the job can be set to run weekly with a seven-day forecast
horizon.
It is essential to configure the scheduling in a manner that aligns with the
operational workflows and preferences of the end users.
Limitations and Prerequisites
- To train the algorithm, operations must be accurately registered with start
and stop clocking of the machine run time.
- If the planned operation time is very small (within the margin of error
for the prediction) it can produce negative time predictions.
- This model expects that the planned start time of the operation will be
the actual start time. If this is not reflected in the historical data or the
operational way of working, it will impact the quality of the prediction.
- When a shop order operation Quantity or Work Center is changed then a new
prediction job needs to be executed for this change to update in the lobby.
- The model is limited to training and predicting on one site. Training and
predicting on multiple sites are not supported.
- The algorithm does not consider changes in efficiency factors when making
predictions.
- The operations to be predicted must be planned to start on a future date
and must be of the time factor type quantity per hour or hour per quantity.
- Only operations with predictions will be visible in
the lobby. Operations that are not candidates for predictions will not be visible.
- The algorithm only learns during training. New data added after training
will not be considered until a new training is triggered and the new data is
within the training range parameters.
- Implemented changes that affect production time can take a long time
to be reflected in the prediction. This is because the training dataset
will still contain the old production times during training.
- Setup and labor times are not predicted with this algorithm.
- The model uses batch predictions. If the payload is large (greater than
10,000 operation predictions) in one prediction job, it will fail. Consider
splitting the predictions on different Work Centers or lowering the prediction
horizon to avoid this.
- Explainable AI (XAI) in predictions is currently not available.