How PRECOG evaluates your data

The timeframe in which the data was collected, arranged from the oldest to the most recent entry. It can be numerical values or date and time stamps. The data can be aggregated to cover longer periods, for example measurements taken every minute can be aggregated into days. Whenever possible, this should be done at the raw data level.

In general, consistent time intervals such as days, hours or minutes are recommended, which can then be aggregated for longer-term analysis. In special cases, such as complex outdoor mechanical assets where the measurements may be noisy and can mask underlying issues, supply as many high-quality reference datasets as possible, ideally at least 10 from any historical period. The last 6 records should also contain the most recent values, measured at consistent intervals or aggregated.

The measured value of the asset or parameter that is monitored. The input file has to contain at least 6 records.

The maximum or minimum threshold that the physical characteristic should not exceed, depending on the type of indicator. If a higher value is better, the target is the lowest acceptable value. If a lower value is better, the target is the highest allowable value.

If the manufacturer does not specify the target, choose it from test results so that it falls outside the upper and lower intervals marked in blue on the chart. This avoids frequent alerts.

The target can change over the whole period of the data and can affect the prediction. Changes in the target can represent the influence of other physical factors on the asset, such as temperature, rain or other environmental effects. These adjustments are often used for outdoor devices, where seasons and weather conditions may affect the asset.

The target column also specifies how many periods are needed for longer-term predictions. This can be done by adding the wanted periods with target values and tolerances but without actual values. The recommendation is 6 to 10 periods. More than 14 periods is not recommended, because the uncertainty grows exponentially and the accuracy drops significantly. Any additional target values must match the target value of the last actual measured value. Going backward through the periods, the more target values match the last one, the more accurate the long-term forecast is.

The permissible deviation from the target value. It is a narrow threshold within which the equipment still operates, but beyond which the operation may cause damage. The tolerance is added to or subtracted from the target, depending on whether the indicator type is higher is better or lower is better.

A binary indicator, TRUE or FALSE, that depends on the nature of the indicator. If a higher value means better performance (also called green, yellow, red or GYR), the column is TRUE. If a lower value means better performance (red, yellow, green or RYG), the column is FALSE.

The most important result of the analysis is the lamp. It shows the summary result of more than 14 different algorithms, which can predict warnings or indicate critical alerts for the indicator. Further information is in the details section.

The longer-term prediction uses several prediction algorithms to estimate the overall performance probability of an indicator over an extended period. The short-term prediction relies on many algorithms as well, which is why the lamp gives suitable predictions.

The longer-term prediction assesses the probability of an indicator’s performance even if it has not yet been fulfilled. The number of periods can be adjusted with the target that is set. If the probability of meeting the target is below 70%, the risk is high.

Our AI is trained on more than 7 trillion measurement data points from several industries and rests on 15 years of engineering and mathematical development.

At least 6. The maximum is a reasonable number.

The target is, for example, the power output recommended by the manufacturer, or it can be defined during testing. The tolerance is a narrow border where the equipment still works but the operation is harmful. If the target is the limit at which the device gets damaged, the tolerance has to be zero.

The algorithm runs on one data type at a time. The best practice is to run it for each data type. If any of them returns a yellow or red light, the whole evaluation is considered yellow or red. To represent the effect of different physical aspects such as seasons, rain or temperature, the target can be changed. This makes sense mostly for outdoor devices.

You can use whatever data you prefer, as long as it is consistent.

Yes. In general, consistent time intervals such as days, hours or minutes are recommended, which can then be aggregated for longer-term analysis. In special cases, such as complex outdoor mechanical assets where the measurements may be noisy and can mask underlying issues, supply as many high-quality reference datasets as possible, ideally at least 10 from any historical period. The last 6 records should also include the most recent values, measured at consistent intervals or aggregated.

A project typically takes 2 to 4 months per equipment type. This applies if you start at the lowest maturity level and want to reach the highest. After the roll-out there is a further 2 months of observation and adjustment.