Field Device Neural Network Pruning for Fault Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing field devices are unable to execute or train neural networks of sufficient complexity to predict errors effectively, as their maximum complexity limits prevent the detection of field device errors before they occur, impacting normal operation and industrial system performance.
Innovation Solution
A method is developed to generate and train neural networks with initial complexity greater than the field device's maximum capability, reducing complexity by removing neurons until the network can be executed by the field device, allowing for prediction of field device errors and enabling proactive measures to prevent or mitigate errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network with high complexity is trained to predict field device errors, then prediction accuracy is improved, but the field device computer cannot execute the neural network due to complexity limits
Solution Approach 1:
The patent extracts the neural network training and complexity optimization process from the field device computer and performs it externally on a server or cloud platform. The trained model is then deployed to the field device, separating the complex training operation from the resource-constrained execution environment.
Solution Approach 2:
The patent changes the parameters of the neural network (number of neurons, layers, connectivity) to optimize the balance between prediction accuracy and computational complexity. This involves adjusting architectural parameters to fit within the field device's maximum complexity constraints while maintaining effective error prediction capability.
2Ease of operation
If the neural network complexity is reduced to fit field device constraints, then executability is improved, but error prediction capability is lost
Solution Approach 1:
The patent performs preliminary training and model optimization actions on a powerful external system before deploying the neural network to the field device. This preliminary action ensures the model is pre-optimized for both accuracy and complexity constraints, allowing the field device to execute it effectively without needing to perform complex training operations.
Solution Approach 2:
The patent introduces an intermediary system (server or cloud platform) that mediates between the requirement for high prediction accuracy and the field device's complexity constraints. This intermediary performs the complex training and optimization work, then delivers a simplified model that the field device can execute while maintaining prediction capability.
3Ease of manufacture
If a neural network is trained directly on field device data with limited complexity, then training can be performed on the field device, but the trained network cannot effectively predict field device errors
Solution Approach 1:
The patent extracts the training function from the field device computer and relocates it to an external system with sufficient computational resources. This allows the neural network to be trained with the necessary complexity to achieve effective error prediction, while the field device only needs to execute the pre-trained model within its complexity constraints.
Data Source
Figure 1
Figure 2
AI summary
A method for generating a neural network for a field device (2) for predicting field device errors is presented and described, wherein the field device (2) comprises a field device computer (7) for executing neural networks up to a maximum complexity. The invention is based on the objective of providing a method for generating a neural network for predicting field device errors, which can be executed by the field device computer (7) of a field device.The problem is solved by a method wherein a neural network containing neurons for predicting field device errors is generated with an initial complexity greater than the maximum complexity, wherein training data is generated from data about field device errors, wherein the neural network is trained with the training data so that it is trained to predict field device errors, and wherein the initial complexity is reduced at least to the maximum complexity by first removing at least one neuron from the neural network in at least one reduction step and then retraining the reduced neural network so that it is trained to predict field device errors.