DE102016216948A1 - Model calculation unit and control unit for calculating a neuron layer of a multilayer perceptron model with optional input and output transformation - Google Patents

Abstract

The invention relates to a model calculation unit for selectively calculating a layer of a multi-layer perceptron model and at least one further data-based function model with a hardwired computation core embodied in hardware for calculating a fixed computation algorithm in coupled function blocks, the computation core having a state machine and operation blocks, wherein the state machine comprises the Arithmetic operation for calculating the layer of the multi-layer perceptron model and the at least one further data-based function model, wherein the state machine is further configured to selectable before the calculation of the layer of the perceptron model or the calculation of the at least one other data-based function model an input transformation of input variables and / or after the calculation of the layer of the perceptron model or the calculation of the at least one further data-based n function model to make an output transformation of output variables.

Description

Technical area

The invention relates to the calculation of functional models in a separate hardwired model calculation unit, in particular for the calculation of multilayer perceptron models.

Technical background

Functions of controls of technical systems, e.g. Internal combustion engines, electric drives, battery storage and the like, are often implemented with models that represent a mathematical image of the real system. However, in physical models, especially in complex contexts, the required computational accuracy is lacking, and with today's computing capabilities, it is usually difficult to compute such models within the real-time requirements required for a controller. For such cases, it is envisaged to use data-based models which describe relationships between an output variable and input variables exclusively on the basis of training data obtained with the aid of a test stand or the like. In particular, data-based models are suitable for modeling complex contexts in which multiple inputs between which interrelations exist are properly considered in the model. In addition, modeling with the help of data-based models offers the possibility to supplement the model by adding individual input variables.

Data-based function models are usually based on a large number of support points in order to achieve sufficient modeling accuracy for the respective application. Due to the high number of nodes, a high computational capacity is needed to compute a model value with a data-based functional model, such as a Gaussian process model. Therefore, in order to be able to calculate such a data-based functional model in a controller application in real time, model calculation units based on a hardware design can be provided.

Disclosure of the invention

According to the invention, a model calculation unit for calculating a layer of a multilayer perceptron model according to claim 1 as well as a control device and a use of the control device according to one of the independent claims are provided.

Further embodiments are specified in the dependent claims.

The above model calculation unit provides an embodiment that makes it possible to calculate a neuron layer of a multilayer perceptron model (MLP model) with a variable number of neurons or at least one other data-based function model.

According to a first aspect, a model calculation unit for selectively calculating a neuron layer of a multi-layered perceptron model and at least one further data-based function model with a hardwired computational kernel designed to compute a fixed computation algorithm in coupled function blocks, wherein the computational core comprises a state machine and operation blocks State machine specifying the arithmetic operation for computing the neuron layer of the multi-layered perceptron model and the at least one other data-based function model, the state machine being further configured to select an input transformation of input quantities and / or before the calculation of the neuron layer of the perceptron model or the computation of the at least one further data-based function model. or after the calculation of the neuron layer of the perceptron model or the calculation de s perform at least one additional data-based function model an output transformation of output variables.

One idea of the above model calculation unit is to form these separately in a calculation kernel in a control unit for calculating a neuron layer of a multilayer perceptron model and at least one further data-based function model in hardware structures. In this way, a substantially hard-wired hardware circuit can be provided for the realization of functions, which makes it possible to calculate one or more neuron layers of a multilayer perceptron model and thereby only a very small computational load in a software-controlled system Microprocessor of a controller to effect. By the hardware acceleration provided by the model calculation unit, a multilayer perceptron model or other data-based functional model can be calculated in real time, so that the use of such a model becomes interesting for control applications for internal combustion engines in automobiles.

Due to the possibility of making or skipping an input transformation and / or output transformation of input variables or output variables, an adaptation of the input variables or the output variables can be carried out, depending on the type of model to be calculated.

The model calculation unit may be equipped with interfaces to calculate the MLP model in layers, so that the number of MLP neuron layers and the number of neurons in each neuron layer can be freely selected. Through this layered division, the parameters, such as the synapse weights, for each neuron layer can be specified separately.

The arithmetic kernel may be configured to calculate an output for each neuron for computing the neuron layer of the multi-layered perceptron model having a number of neurons depending on one or more inputs of an input vector, a weighting matrix with weighting factors, and an offset value for each neuron for each neuron, calculating a sum of the values of the input quantities and the offset value given to the neuron, weighted by a weighting factor determined by the neuron and the input, and transforming the result with an activation function to obtain the output quantity for the neuron.

According to one embodiment, the state machine may be further configured to select, before the calculation of the neuron layer of the perceptron model or the calculation of the at least one further data-based function model, a copying process of the input variables into a further memory area and / or after the calculation of the neuron layer of the perceptron model or the computation of the at least one further data-based function model to perform a copy operation of the output variables in a further memory area.

The arithmetic kernel can be designed to select a type of activation function for the multilayer perceptron model depending on a selection variable (cfg_activation_function) and / or select with the aid of a further selection variable whether a Gaussian process model or an RBF model or a neuron layer of the perceptron model is to be calculated.

The arithmetic core can be formed in a surface area of an integrated component.

According to a further aspect, a method for operating the above model calculation unit is provided, wherein in a calculation of the neuron layer of the perceptron model and / or depending on a selection variable, the input transformation of input variables and / or the output transformation of output variables is skipped.

In another aspect, a controller having a microprocessor and one or more of the above model calculation units is provided.

In particular, the controller may be formed as an integrated circuit.

According to another aspect, a use of the above control device is provided as a control device for controlling an engine system in a motor vehicle.

Brief description of the drawings

Embodiments are explained below with reference to the accompanying drawings. Show it:

1 a schematic representation of a control unit for use for an engine system in a motor vehicle;

2 a schematic representation of a calculation unit as part of the control unit;

3 a schematic representation of a neuron layer of an MLP model; and

4a - 4d Representations of possible activation functions.

Description of embodiments

1 shows an example of a schematic representation of a control device 2 for an engine system 1 with an internal combustion engine 3 as a technical system to be controlled. Alternative systems, such as motor systems with an electric motor or a battery system can be controlled in a comparable manner by such a control device. The control unit 2 includes a microprocessor 21 and a model calculation unit 22 , which may be formed as separate components or in an integrated manner in separate areas on a chip. In particular, the model calculation unit provides 22 a hardware circuit that is structurally from a computational core of the microprocessor 21 can be separated.

The model calculation unit 22 is essentially hardwired and therefore not like the microprocessor 21 adapted to execute a software code and thereby perform a variable predetermined by software function. In other words, in the model calculation unit 22 no processor is provided so that it is not operable by a software code. Focusing on a given model function results in a resource-optimized realization of such a model calculation unit 22 allows. In an integrated structure, the model calculation unit 22 optimized in terms of area, which also enables fast calculations.

The control unit 2 essentially serves to sensor signals S and sensor sizes that of a sensor in the internal combustion engine 3 be detected, and / or process external specifications V and cyclically at fixed predetermined intervals, ie for the control of internal combustion engines periodically within a cycle time, for example, between 1ms and 100ms values, or angle synchronous (synchronous to the position of a crankshaft) in dependence on a crankshaft angle an operated internal combustion engine of one or more corresponding control variables A to the internal combustion engine 3 create, so that it can be operated in a conventional manner.

In 2 is a model calculation unit 22 shown in more detail. The model calculation unit 22 includes a state machine 11 , a store 12 and one or more operation blocks, for example, one or more MAC blocks 13 (MAC: Multiply-ACcumulate, for fixed-point or floating-point calculation) and an activation function calculation block 14 to calculate an activation function. The state machine 11 and the one or more operation blocks 13 . 14 form a calculation kernel ALU of the model calculation unit 22 , The operation blocks may include a multiplier block and an addition block in addition to or as an alternative to the MAC block.

With the help of the state machine 11 may be in memory in an input memory area 12 stored values of input variables are to be offset by repeated loop calculations in order to obtain intermediate quantities or output quantities that are stored in a corresponding output memory area of the memory 12 to be written.

The state machine 11 is designed to compute a single neuron layer of a multilayer perceptron model. The state machine 11 can be described by the following pseudocode:

p7:

maximaler Indexwert für die Eingangsgrößen des Eingangsgrößenvektors

p8:

minimaler Indexwert bzw. vorgegebener Startindex für die Anzahl der Neuronen

p6:

maximaler Indexwert für die Anzahl der Neuronen

p3:

Offsetwert

p1, p2:

Variablen für die Eingangstransformation

p4, p5:

Variablen für die Ausgangstransformation

With

p7:

maximum index value for the input variables of the input variable vector

p8:

minimum index value or default start index for the number of neurons

p6:

maximum index value for the number of neurons

p3:

offset value

p1, p2:

Variables for the input transformation

p4, p5:

Variables for the output transformation

With the aid of the above pseudocode, the following calculation can be carried out for each neuron of the neuron layer to be calculated: y [j] = act (p3 [j] + Σp7-1 / k = 0v _{j, k} · ut [k] for j = 0 ... p6 - 1

This represents a computation for a neuron layer of a multilayer perceptron model as shown in FIG 3 is shown.

3 shows a neuron layer of several neurons 20 to which values of input variables of an input variable vector ut ₀ ... ut _{p6-1 are} supplied. The values of the input variables are weighted by means of a corresponding predefined weighting matrix of weighting values, which are provided as weighting factors v _{0... P7-1,0... P6-1} . The weighting is generally done by multiplicatively applying the associated weighting factors v _{0 ... p7-1.0 ... p6-1} , but can also _load the values of the input variable vector in another way.

The sums of the weighted values of the input vector ₀ ... ut ut _P6-1 are each acted upon with an offset value O O ₀ ... _P6-1, in particular applied additive. The result is transformed with a given activation function "act". The result is a corresponding value of an output vector y ₀ ... y _p6-1 . By providing the offset value for each neuron, there is another degree of freedom for modeling.

By setting the running variable p6, the number of neurons 20 be set to be calculated neuron layer. A multilayer perceptron model may be constructed by using the values of the output vector y ₀ ... Y _{p6-1 of} a neuron _layer as the input vector for calculation of a subsequent neuron _layer in the model calculation unit 22 so that the number of neuron layers of the multi-layered perceptron model is repeated by invoking the function according to the above pseudocode or by repeatedly calling the model calculation unit 22 can be realized with different parameters.

An input and / or output transformation of the input variables of the input variable vector or the output variables of the output vector can be carried out by means of the standardization variables p1 and p2 or p4 and p5 given for each neuron.

The layered calculation of the MLP model enables a lean design of the model calculation unit 22 , so that their space requirement in integrated construction is low. Nevertheless, the model calculation unit allows 22 a calculation of a multilayer perceptron model in a simple manner by repatriation or redefinition of the values of the output variables of the output vector as input variables of an input vector for the calculation of another neuron layer.

The activation function "act" can be one of several activation functions provided by the activation function calculation block 14 the model calculation unit 22 can be calculated. The activation functions can be used, for example, a kink function, a sigmoid function, a hyperbaric hyperbolic function or a linear function as described in US Pat 4a to 4d are shown accordingly.

By the single-layered construction of the neuron model, which is realized by the above pseudocode, it is also possible by simple modification in addition to a neuron layer of an MLP model also a Gauss process model or an RBF model (RBF: Radial Basis Function) to calculate. For this purpose, the weighting values are applied not multiplicatively to the values of the input variable, but additive or subtractive. Furthermore, the square distance is calculated, which is weighted with predetermined length scales L [k]. Furthermore, an exponential function is selected as the activation function for the RBF model. Thus, a Gaussian process model can be used accordingly y = Σp6-1 / j = 0p3 [j] * exp (-Σp7-1 / k = 0L [k] * (-vj _{, k} + ut [k]) ² ) by a modification of the pseudo code optionally calculated as follows.

It can be seen that when the loop function is performed by the variable cfg_mlp, a case distinction can be made. If cfg_mlp = 1, the calculation of the neuron layer is selected and with cfg_activation_function = 0 ... 3 the type of the activation functions described above can be selected.

If cfg_mlp = 0, a Gaussian process model or an RBF model is calculated instead of the MLP perceptron model. Here is a selection of the activation function is not necessary, since they are always calculated with the exponential function. In this way it is possible to use the model calculation unit 22 to use both for the calculation of Gaussian process models and RBF models as well as for the calculation of a neuron layer of an MLP model and thereby require only a small footprint in integrated state machine architecture.

While input and output transformation is necessary for the calculation of Gaussian process models, RBF models, or the like, this is not mandatory for the calculation of a neuron layer of a perceptron model. Therefore, it may be sufficient to perform a simple copy operation of the input and output variables instead of the input and / or output transformation. If the input variables and output variables are already in the memory area provided for this purpose, the copying operation can also be dispensed with.

Therefore, for the input transformation, cfg_skip_input_scaling indicating whether to perform an input transformation and an input size copy cfg_copy_input indicating whether the provided inputs should be copied to another memory area or whether the provided inputs are transformed in the further memory area is controlled by an input transformation indication should be provided.

The pseudocode for the input transformation is as follows:

Similarly, for the output transformation, cfg_skip_output_scaling indicating whether an output transformation should be made and an output size copy cfg_copy_output indicating whether the provided outputs should be copied to another memory area or whether the provided outputs are transformed in the further memory area is controlled by an output transform indication should be provided.

The pseudocode for the output transformation is as follows:

On the whole, this makes it possible optionally to perform an input transformation and / or output transformation and / or a copying process of the input variables into a further memory area and a copying process of output variables into a further memory area in accordance with the input transformation information cfg_skip_input_scaling, the input variable copy cfg_copy_input, the output transformation information cfg_skip_output_scaling and the output variable copy cfg_copy_output , As a result, in particular when calculating a neuron layer of a perceptron model, the input and / or output transformation can be skipped so that the calculation as a whole can be accelerated.

Claims (9)

Model calculation unit ( 22 ) for selectively calculating a neuron layer of a multi-layered perceptron model and at least one further data-based functional model with a hardwired computational core embodied in hardware ( 18 ) for calculating a fixed calculation algorithm in coupled function blocks, wherein the calculation kernel is a state machine ( 11 ) and operation blocks ( 13 . 14 ), wherein the state machine ( 11 ) predetermines the arithmetic operation for calculating the neuron layer of the multi-layer perceptron model and the at least one further data-based function model, wherein the state machine ( 11 ) is further configured to select, before the calculation of the neuron layer of the perceptron model or the calculation of the at least one further data-based function model, an input transformation of input variables and / or after the calculation of the neuron layer of the perceptron model or the calculation of the at least one further data-based function model an output transformation of output variables make. Model calculation unit ( 22 ) according to claim 1, wherein the arithmetic kernel ( 18 ) is adapted to be used for a neuron layer of a multilayer perceptron model with a number of neurons ( 20 ) depending on one or more input variables of an input variable vector (ut), of a weighting matrix with weighting factors (v _{j, k} ) and one for each neuron ( 20 ) has an output variable (y [j]) for each neuron ( 20 ), where for each neuron ( 20 ) a sum of the one by the neuron ( 20 ) and the input quantity determined weighting factor (v _{j, k} ) weighted values of the input quantities with the neuron ( 20 ) is applied and the result is transformed with an activation function (act) in order to determine the output quantity (y [j]) for the neuron ( 20 ) to obtain. Model calculation unit ( 22 ) according to claim 1 or 2, wherein the state machine ( 11 ) is further configured to be selectable before the calculation of the neuron layer of the perceptron model or the calculation of the at least one further data-based function model, a copying process of the input variables in a further memory area and / or after the calculation of the neuron layer of the perceptron model or the calculation of the at least one further data-based function model make a copy of the output variables in another memory area. Model calculation unit ( 22 ) according to one of claims 1 to 3, wherein the arithmetic kernel is designed to select a type of activation function for the multilayer perceptron model depending on a selection variable (cfg_activation_function) and / or to select with the aid of a further selection variable whether a Gaussian process model or an RBF model or a neuron layer of the perceptron model is to be calculated. Model calculation unit ( 22 ) according to one of claims 1 to 4, wherein the arithmetic core is formed in a surface area of an integrated module. Method for operating a model calculation unit ( 22 ) according to one of claims 1 to 5, wherein depending on a calculation of the neuron layer of the perceptron model and / or depending on a selection variable, the input transformation of input variables and / or the output transformation of output variables is skipped. Control unit with a microprocessor and one or more model calculation units according to one of claims 1 to 5. Control unit according to claim 7, wherein the control unit ( 2 ) is formed as an integrated circuit. Use of the control device according to claim 7 or 8 as a control device for controlling an engine system ( 1 ) in a motor vehicle.

Images