Hardwired MLP Calculation Core for Real-Time Engine Control Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing data-based models, such as Gaussian process models, require high computing capacity to achieve sufficient accuracy and often fail to meet real-time calculation requirements in control devices for complex systems like internal combustion engines, necessitating a more efficient hardware-based solution.

Innovation Solution

A hard-wired model calculation unit specifically designed for multi-layer perceptron models with a variable number of neurons, utilizing a hardware computing core that calculates output variables using input variables, weighting matrices, and offset values, transforming results with activation functions, allowing for real-time calculations with reduced software load.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If data-based models with high number of support points are used to achieve sufficient modeling accuracy, then modeling precision is improved, but computing capacity requirement increases

Engineering Contradiction:
Improvemodeling accuracyVSAvoidcomputing capacity requirement
Core Design Contradiction:
Measurement precisionVSPower

Solution Approach 1:

The patent segments the data-based model into multiple localized basis functions centered at support points, allowing the global modeling task to be divided into independent local calculations. This segmentation enables the use of fewer support points while maintaining accuracy, as each basis function contributes only to its local region rather than requiring global computation across all support points.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial action by using a limited number of localized basis functions only where needed in the input space, rather than computing contributions from all support points for every prediction. This reduces the computing capacity requirement while maintaining sufficient modeling accuracy in the regions of interest.

Inventive Principle:
Principle #16Partial or excessive action

2Measurement precision

If data-based models with high computing capacity are used to achieve sufficient accuracy, then modeling precision is improved, but real-time calculation capability deteriorates

Engineering Contradiction:
Improvemodeling accuracyVSAvoidreal-time calculation capability
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

By segmenting the model into localized basis functions, the patent enables parallel computation of independent local contributions, significantly improving real-time calculation capability while maintaining modeling accuracy. Each basis function can be evaluated independently and simultaneously.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses pre-computed basis function parameters that are copied and applied repeatedly for different input points, avoiding redundant calculations. This copying approach maintains accuracy while dramatically improving computation speed for real-time applications.

Inventive Principle:
Principle #26Copying

3Measurement precision

If a large number of support points are used in data-based models, then modeling accuracy is improved, but device complexity increases

Engineering Contradiction:
Improvemodeling accuracyVSAvoidmodel structure complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the complex global model into simple localized basis functions, reducing device complexity while maintaining accuracy. Each segment is computationally simple but collectively they provide high-fidelity modeling through their localized contributions.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP3542318B1Model-computation-unit and controller for the optional calcualtion of a rbf-model, a gausians-process-model and a mlp-model
Publication Date: 2023.11.08 ROBERT BOSCH GMBH
  • EP3542318B1 patent drawingFigure 1~2
  • EP3542318B1 patent drawingFigure 3
  • EP3542318B1 patent drawingFigure 4a~4d

AI summary

The invention relates to a model calculation unit for calculating a neuron layer of a multi-layered perceptron model, comprising a hardwired calculation core formed in hardware for calculating a fixedly preset computing algorithm in coupled functional blocks. The calculation core is designed to calculate an output variable for each neuron in a neuron layer of a multi-layered perceptron model having a number of neurons, depending on one or more input variables of an input variable vector, on a weighting matrix with weighting factors and an offset value predefined for each neuron. A sum of the values of the input variables, which are weighted with a weighting factor determined by the neuron and the input variable, and the offset value predefined for the neuron is calculated for each neuron, and the result is transformed with an activation function to obtain the output variable for the neuron.