Hardwired Model Calculation Unit for Real-Time RBF and MLP Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing control devices in motor vehicles face challenges in calculating complex mathematical models, such as multilayer perceptron and RBF models, within real-time requirements due to high computational demands, especially when dealing with numerous support points, which require significant computing capacity.
Innovation Solution
A hard-wired model calculation unit with a computing core designed to selectively calculate RBF or neuron layer models, utilizing coupled function blocks, a state machine, and specific memory areas for support points and weighting matrices, allowing for variable neuron numbers and activation functions, enabling efficient hardware-based calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data-driven functional models (Gaussian process model, multilayer perceptron model) are used to achieve sufficient modeling accuracy for complex relationships, then modeling accuracy is improved, but computing power requirements increase significantly
Solution Approach 1:
The patent segments the computational workload by separating the computationally intensive model calculations into dedicated hardware-based model calculation units (22) that operate independently from the microprocessor (21). This segmentation allows complex data-driven models to be calculated with high accuracy while offloading the computational burden from the main processor, thus resolving the contradiction between modeling accuracy and computing power requirements.
2Productivity
If hardware-based model calculation units are used to compute data-driven functional models in real time, then real-time calculation capability is improved, but device complexity increases
Solution Approach 1:
The patent implements a universal model calculation unit (22) that can calculate multiple types of data-driven functional models (Gaussian process models, RBF models, multilayer perceptron models) using a single hardware architecture. This multi-functional design achieves real-time calculation capability for various complex models while avoiding the need for separate dedicated hardware for each model type, thus resolving the contradiction between real-time calculation capability and device complexity.
3Speed
If a fixed predefined calculation algorithm is used in hard-wired computing core, then calculation speed is improved, but adaptability decreases
Solution Approach 1:
The patent employs a dynamic configuration mechanism where the model calculation unit (22) can be selectively activated or deactivated based on operational requirements. The unit includes configurable components such as the RBF model calculation with adjustable support points and length scales, and the MLP model calculation with configurable neuron layers and activation functions. This dynamic configurability allows the hard-wired computing core to maintain high calculation speed while adapting to different model types and computational needs.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
The invention relates to a model-calculating unit (22) for selectively calculating an RBF model or a neuron layer of a multilayer perceptron model, comprising a hardwired calculation core (11, 13, 14) formed in hardware for calculating a fixedly defined computing algorithm in coupled functional blocks, wherein the calculation core (11, 13, 14) is designed to calculate, for an RBF model, an output value (y) in accordance with one or more input values of an input value vector (ut), with node points (V), with length scales (L), and with parameters defined for each node point, wherein the calculation core (11, 13, 14) is also designed to calculate, for the neuron layer of the multilayer perceptron model having a number of neurons (20), an output value for each neuron (20) in accordance with the one or more input values of the input value vector (ut), with a weighting matrix having weighting factors (V0…p7-1,0…p6-1), and with an offset value defined for each neuron (20).