To know the prediction accuracy, the predicted values have to be validated by capacity testing. All references to the accuracy of our forecasts are verified using actual market data samples.
-
Number of sampling data.
-
We use data sets of 1000 DUTs.
-
The sample should be randomly selected from a population with a full range of low to high OCVs.
-
These data sets must contain 1) data measured for 125 seconds with our recipe, and 2) a real capacity Q [Ah] value obtained from capacity testing.
-
-
Modeling.
-
Half of the dataset is used as “training data” to create the
algorithm, while the other half is used as “test data” to verify its accuracy. -
The vertical axis is the pre-screening and the horizontal axis is the capacity testing.
-
When in the introduction and at any time during a review, information can be obtained for setting threshold values.
-
-
-
-
Validation of the model.
-
When used in prescreening, the threshold value is used as the basis for predicting good and bad DUTs. After verification by actual capacity testing, a four-quadrant cell is generated depending on whether the prediction was right or wrong.
-
If the prediction is good and the actual is also good, then it is a true positive (TP), and if the prediction is bad and the actual is also bad, then it is a true negative (TN), and these two indicators should be as close to 1.00 as possible.
-
Similarly, if the predicted good but actually bad is a false positive (FP), and the predicted bad but actually good is a false negative (FN), the two indices should be as close to 0.00 as possible. False positives (FP) mean that the costs of carrying out capacity tests (logistics costs, labor costs, depreciation of equipment, etc.) are wasted. False negatives (FN) result in lost sales opportunities as the batteries are recycled as scrap when they could be sold as batteries.
-
In day-to-day operation, only those exceeding the thresholds set in the pre-screening are sent to capacity testing.
-