Metrics

Metrics are the elements of a subject you want to monitor. All metric types - ingested, calculated and simulated - have several visualization and visibility options:

Ingested metric

An ingested metric receives data from a device, sensor or other external system. Example: a sensor that is sending values for both temperature and humidity.

Measurement filtering

In certain situations, you may not have the ability to filter the data that you wish to send from your device to the Blockbax Platform. This can result in a lot of measurements that add little value to the insights you wish to obtain. In order to optimize costs we allow you to define a filtering that determines when a measurement is processed, or skipped.

There are currently 3 options to perform filtering on measurements:

If any of these filters have a positive match. The measurement will be retained. Expand the options below to find all the possible configurations.

Calculated metric

A calculated metric can be computed from the values of one or more metrics and properties within a subject. A simple example will be an electrical station for which you want to compute the wattage by multiplying voltage with amperage (this is made visible in the examples).

The main elements of a calculation are the name (1) and type (2) together with the arguments (3) that can be used in the calculation (4).

Arguments

In a calculation, metrics and properties of type Number, Text and Location within the subject type can be used. These metrics and properties are referred to as the arguments of the calculation.

For metrics configured as argument, it is possible to use their value and/or date in the computation. For example, if we add a Wattage metric as argument with alias wattage, we can use its value with wattage.number and its date in milliseconds with wattage.date.

For properties configured as argument, it is possible to use their value. For example, if we add a property Height with alias height, we can use its value with height.number.

The field name for the value of the metric or property differs per type:

• Numeric metrics and properties: .number
• Textual metrics and properties: .text
• Location metrics and properties: .latitude, .longitude and .altitude

The value of numeric and textual metrics and properties can also be retrieved with .value, but this is deprecated in favor of .number and .text.

Max age explained

A calculation is always triggered when a measurement is received on a metric that is used in the calculation. Next, measurements of the other metrics are retrieved based on the max age that is configured. Here are some examples:

Calculation

The calculation is defined using the editor in a user-friendly language. The language allows you to:

• define variables;
• apply basic arithmetic operations;
• use functions for more advanced operations;
• save and load values in between consecutive calculations;
• use conditionals;
• output the result of your calculation.

Last but not least, formulating a calculation is getting more convenient when you use the editor features and watch the examples.

Define variables

For your convenience and readability of your calculation it is possible to split your calculation into smaller calculations. Each of the smaller calculations may be saved in a variable definition. An example of this is wattage = voltage.number * amperage.number in the example at the start of this section. This allows referring to the result of the computation voltage.number * amperage.number via the name wattage.

Arithmetic operations

All basic arithmetic operations can be applied to values in your calculation:

Multiplication and division take precedence over addition and subtraction, as per the standard arithmetic rules. Any (decimal) number values, variable definitions, loaded values or input argument values can be used in arithmetic. Operations can also be chained. Some examples:

load sum with default 0
a = voltage.number * 5.5
b = a - 7 + 8 / 2
c = sum + b
save c as sum
d = (c + 10) * 5
output d

Functions

Many default computations are available as functions. Functions can be used anywhere in your computation where a numeric value is expected, and can take as input any literal value, input argument value, loaded value or arithmetic. For example: compute and output the maximum value of a metric argument, with alias inputMetric, by using the max function:

load maxValue
newMaxValue = max(maxValue, inputMetric.number)
save newMaxValue as maxValue
output newMaxValue

Save and load values

It is possible to save and load values between consecutive evaluations of the calculation. A numeric value can be saved at the end of a calculation, so that it can be loaded the next time the calculation is evaluated. A maximum of 10 values can be saved and if a value exceeds 8 decimal places, it will be rounded off. For example: we want to compare the value of a metric to its previous value.

load previousValue
difference = bitcoin.number - previousValue
save bitcoin.number as previousValue
output difference

Values must be loaded at the beginning of a calculation with a name that that is later used to save the value. See load previousValue will create a variable with name previousValue. If a variable is loaded, a value must be saved in it at some point during the calculation. See save bitcoin.number as previousValue which saves the current bitcoin metric value in previousValue to be used by the next evaluation of the calculation.

It is possible to use the value of a loaded variable to update its own value. For example when we are keeping a sum:

load bitcoinSum with default 0
newSum = bitcoinSum + bitcoin.number
save newSum as bitcoinSum
output newSum

In this case a default must be provided when loading the variable, see load valueSum with default 0. This default value is used when the load variable does not yet have a value, if this default was not present it would never be possible to set a value initially.

Conditionals

Within the calculated metric editor you are able to make use of conditionals to extend the ability to perform operations based on certain conditions. This is achieved by using if, then and else statements. To further extend the ability to make use of conditionals in the platform, you can combine conditions using and / or statements.

if <condition> then <value1> else <value2>

voltage.number == 230

soilType.text != "Peat moss"

temperature.number > 25

<condition1> and <condition2>

if <condition1> and <condition2> then <value1> else <value2>

<condition1> or <condition2>

if <condition1> or <condition2> then <value1> else <value2>

setpoint = if hasValue(setpointManual.number)
           then setpointManual.number
           else setpointSchedule.number

Output

Each calculation must specify exactly one output value at the end of the calculation as follows: output <value>. In this case <value> can be any (decimal) number value, arithmetic, variable definition, loaded value or argument value.

Editor features

The editor has some smart features which makes it easy to use, but these features are not visible from the outside. The features are listed in the table and showed in the video below the table.

Examples

This section show some examples to make calculated metrics concrete and explain the settings.

Wattage calculation

The wattage will only be calculated when when voltage and amperage are received with the same timestamp. This is due to zero seconds max age setting for both arguments.

Heat index

The heat index will be calculated with the last known temperature measurement when the humidity measurement is received and with a humidity measurement not older than 10 seconds when a temperature measurement is received.

Speed

The speed is calculated with the last known location which contains the latitude, longitude and date values. This example includes save and load values, functions and arithmetic operations which all together shows the power and possibilities of a calculated metric.

Load Factor

In order to calculate the load factor, there are various conditions that need to be met in order to create a representative calculated metric. In this case, we are calculating the load factor when rP and qP are both positive, else a 0 value is returned as output for the calculated metric. This example shows the ability to use conditionals.

Simulated metric

A simulated metric generates data based on the configuration. Example: a two hour sinus.

Forecasted metric

A forecasted metric allows you to forecast future values of an existing metric. To achieve this, a model to forecast future values is created under the hood using artificial intelligence (AI), specifically deep learning. The model is created by training on historical data to learn patterns and trends. A new incoming measurement on the metric to forecast will instantly be inferred by the model, resulting in a new measurement on the forecasted metric with its forecasted future value and future date. Conceptually, it is thus quite similar to a calculated metric.

The main elements of a forecasted metric are the name (1), the metric and time in the future to forecast (2), and its training data, which can be existing metrics (3) or properties (4).

Training data

One of the key success criteria for getting accurate forecasts is the training data we use to train a model. By experimenting with different training data, you can significantly improve your model. Once you select new data and save the metric, it will immediately start training on this new data; it will only accept and use models if they result in improved forecast accuracy. The metrics you use as training data should have enough historical measurements (e.g., if you want to forecast one week in the future, the metrics should have a significant amount of measurements that are more than one week old).

To make data more useful for a model, it can be beneficial not just to add the existing metrics to the training data but also to construct new metrics to make them more useful for the model. This part of developing a model is called feature engineering, which is often overlooked while being vital to success. For example, you could use the rate of increase or decrease as a feature instead of simply the value of a metric. We could also engineer features based on the date and time, such as whether it is day or night or a working day or not. In order to achieve this, you can for instance create new calculated metrics and add them as training data.

Forecast accuracy

To monitor the accuracy of your forecasts, two key graphs are available for evaluating averaged hourly forecast accuracy over the past week. The first graph (1) compares actual measurements with forecasted values, providing insight into how well your forecast matches real data. The second graph (2) displays absolute errors, offering a clear indication of forecast accuracy. You can choose to view aggregated accuracy metrics for all subjects or focus on a specific subject (3).

Immediately following the initial training, these graphs are populated with forecasts from one week prior, providing an immediate assessment of the model’s performance. Note that data from this period is not used for training, ensuring that it gives a good, i.e., unbiased, indication of the expected accuracy.

Recent trainings

After initial training, any changes to inputs trigger a new training. By default, the platform also retrains a model automatically every two months to ensure new observations are taken into account. Models that achieve higher accuracy are accepted and used automatically, while models that are not accepted are only shown in the recent trainings overview. The accuracy difference is determined by using the active model and the newly trained model to make forecasts for test data that has not been used for training.

A green check mark indicates an accepted model, while a red cross indicates that it is not accepted (1). The version of the model is displayed, with active models marked by an active badge (2). The creation date of each model is also displayed (3). Training duration information includes training times for completed models and a progress bar for models currently in training (4). The training data contains a history of the input metrics and properties used for each training version, which users can hover over to see the composition of input metrics and properties (5). The accuracy comparison shows the difference in absolute error compared to the previous active model, with positive values indicating an improvement in error by that amount (6). Lastly, users can enable notifications for updates on training completion (7).

If altering the model’s inputs doesn’t improve accuracy, the changes can be easily undone.

Explore

Users can use the Blockbax Explorer to analyze and compare forecasted metrics with actual observed values. When a forecasted metric is selected, the Explorer suggests comparing it with the target metric. Likewise, selecting the target metric offers a comparison with its forecasts.

Conceptually, we consider a forecast to be measured (i.e., determined) at the same time as the measurement that triggers the forecast. In the Explorer, you can thus easily match the triggering measurement with the resulting forecast. Of course, in most cases, you are more interested in aligning with the time a metric forecasts for. This is enabled by default when you select a forecasted metric in the Explorer (click the circled button in the image below to align using measurement time).

Case studies

We compiled blogs featuring case studies that can inspire your own use cases.

Customer case

Forecasting water quality using AI

October 23, 2024 · 12 min read