Neural Network Output Restriction for Safety-Critical Fault Detection

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current methods for automated safety monitoring in motor vehicles and other technical units are complex, time-consuming, and often require expert intervention due to the high degree of mathematical complexity and the need for accurate fault identification, leading to unreliable automated fault identification results.

Innovation Solution

A method and system utilizing a neural network with dynamic restrictions, where the output values are constrained within a predefined value range using a softmax function, ensuring reliability and simplicity in generating safety-critical output values by modifying the neural network architecture to include additional parameters that restrict output values within a specific range, thereby preventing unexpected outputs and enhancing safety monitoring.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If complex calculation algorithms and analysis chains are used for safety monitoring, then measurement precision and fault detection capability are improved, but device complexity and time consumption increase significantly

Engineering Contradiction:

Improvefault detection precisionVSAvoidalgorithm complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and isolates the critical safety monitoring function from the complex neural network system. By using an output restriction function that independently constrains neural network outputs to predefined safe value ranges, the system separates the complex learning task from the safety-critical decision task, maintaining detection precision while reducing algorithmic complexity in safety operations

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies preliminary action by pre-defining safe value ranges and output constraints before the neural network operates. The restriction function φ(y) is configured in advance with knowledge about safe operating parameters, so that during runtime, the system only needs to apply the pre-computed restrictions rather than performing complex real-time safety analysis

Inventive Principle:

Principle #10Preliminary action

2Reliability

If expert intervention is required for fault identification, then reliability of fault assessment is improved, but productivity and response time deteriorate

Engineering Contradiction:

Improvefault assessment reliabilityVSAvoidfault identification speed

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The patent implements self-service by enabling the system to automatically perform safety monitoring and fault identification without requiring expert intervention. The output restriction function automatically confines neural network outputs to safe ranges and identifies faults when outputs exceed these ranges, allowing the system to serve itself in safety-critical decisions while maintaining reliability through mathematically guaranteed constraints

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent uses feedback mechanisms where the restricted output values are continuously monitored and compared against safe ranges. When outputs approach or exceed predefined thresholds, the system provides immediate feedback through automated fault identification, eliminating the need for delayed expert review while maintaining assessment reliability through continuous validation

Inventive Principle:

Principle #23Feedback

3Productivity

If automated fault identification is implemented without restrictions, then productivity is improved, but reliability deteriorates due to unexpected outputs and inaccurate results

Engineering Contradiction:

Improveautomated fault identification speedVSAvoidfault identification accuracy

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent applies beforehand cushioning by pre-configuring the output restriction function with safe value ranges and boundary conditions. This cushioning layer of constraints is placed between the neural network's potentially unreliable outputs and the final fault identification decision, preventing unexpected outputs before they can cause reliability issues while allowing automated operation to proceed at full speed

Inventive Principle:

Principle #11Beforehand cushioning (Prior cushioning)

Data Source

PatentUS11481614B2Method and system for generating safety-critical output values of an entity

Publication Date: 2022.10.25 DR ING H C F PORSCHE AG

AI summary

A method for generating one or more safety-critical output values (y) of an entity, including determining at least one parameter (s) of the entity, whereby the parameter describes at least one state and/or at least one feature of the entity. At least one delimited value range (C(s)) for the parameter (s) is set. At least one measured value (x) for the determined parameter (s) by at least one sensor is ascertained. The measured value (x) from the sensor to at least one processor and/or at least one data storage unit is transmitted. At least one output value (y) is calculated using the processor from the measured value (x) by way of a software application including a neural network (NN) with dynamic restrictions which outputs the output value (y) within the set value range (C(s)) for the determined parameter (s).