FB_ALY_LinearRegressionFitting

The Linear Regression Fitting function block approximates one variable (the Dependent input) by linear combination of several other variables (Input 01 ... Input 0n). This is done by the incremental stochastic gradient method. At the end of the analysis, the calculated coefficients are written to a file.

The linear combination is given by the following equation:

In each cycle, the values for to are recalculated using the following rule:

This corresponds to the minimization of the squared deviation of the calculated values y (output by the function block as result) from the corresponding input value Dependent. The parameter corresponds to the step size and specifies how strongly the parameters are adjusted. The larger the value, the faster the coefficients approach a local optimum. However, if the value is too large, the algorithm may not converge.

Typically, the Linear Regression Fitting function block is first used to fit the weights for the regression of a target variable. Then, using the Linear Regression Inference function block and the fitted weights, the target variable can be predicted based on the input variables.

Syntax

Definition:

FUNCTION_BLOCK FB_LinearRegressionFitting
VAR_OUTPUT
    ipResultMessage: Tc3_EventLogger.I_TcMessage;
    bError: BOOL;
    bNewResult: BOOL;
    bConfigured: BOOL;
    fResult: LREAL;
    fMSE: LREAL;
    bBusy: BOOL;
    eState: E_TeachState;
END_VAR

Outputs

Methods

State diagram

Due to the asynchronous file access in real time applications this function block needs a state machine to prepare and finish file access.

At startup the function block is in Idle state. To write the incoming data to a file it has to be switched to state Teach. Therefore the method SwitchState_Teach() has to be called once to set the function block to the PendingTeach state. After that the method UpdateState() has to be called until the function block is in state Teach. In this state one or multiple teach cycles can be proceeded. If the function block should not teach additional cycles it can be set to Idle state again. To initiate the state switch the method SwitchState_Idle() has to be called. After that the method UpdateState() has to be called until the function block is in state Idle.

State diagram for the teach procedure of the data:

Sample

VAR
    fbLinearRegressionFitting : FB_ALY_LinearRegressionFitting(nNumChannels := 1);
    bBias : BOOL := TRUE;
    fStepSize : LREAL := 0.01;
    nMiniBatchSize : UDINT := 1;
    tTimeout : TIME := T#5S;
    bInvolveExistingFile : BOOL := TRUE;
    sFilePath : STRING := 'C:\TwinCAT\3.1\Boot\LinearRegressionFitting.tas';
    bConfigure : BOOL := TRUE;
    eState : E_ALY_TeachState := E_ALY_TeachState.Idle;
    bTeach : BOOL;
    bFit : BOOL := TRUE;
    fInputX : LREAL;
    fInputY : LREAL;
    fOut0 : LREAL;
    fOut1 : LREAL;
END_VAR

// Configure algorithm
IF bConfigure THEN
    bConfigure := FALSE;

    fbLinearRegressionFitting.Configure(bBias, fStepSize, nMiniBatchSize, tTimeout, bInvolveExistingFile, sFilePath);
END_IF

// Call algorithm
eState := fbLinearRegressionFitting.eState;
CASE eState OF
E_ALY_TeachState.Idle:
    IF bTeach THEN
        fbLinearRegressionFitting.SwitchState_Teach();
        fbLinearRegressionFitting.UpdateState();
    END_IF    
E_ALY_TeachState.Teach:
    fbLinearRegressionFitting.SetChannelValue(1, fInputX);
    fbLinearRegressionFitting.Call(bFit := bFit, dependent := fInputY);
    fbLinearRegressionFitting.GetChannelOutputValue(0, fOut0);
    fbLinearRegressionFitting.GetChannelOutputValue(1, fOut1);
    IF NOT bTeach THEN
        fbLinearRegressionFitting.SwitchState_Idle();
        fbLinearRegressionFitting.UpdateState();
    END_IF    
E_ALY_TeachState.Pending,
E_ALY_TeachState.PendingIdle,
E_ALY_TeachState.PendingTeach:
    fbLinearRegressionFitting.UpdateState();
    eState := fbLinearRegressionFitting.eState;
END_CASE

Requirements