This paper presents a benchmark problem for fault diagnosis of an internal combustion engine that has been formulated and solved. The objective is to design a diagnosis system using and incomplete model information training data that only contains a limited set of fault realizations. Six different solutions to the benchmark, that were presented at the IFAC Safeprocess symposium 2024, are described and evaluated. The contribution of this paper is the benchmark and the presentation of six different solutions in one paper. The paper is intended to provide a starting point for engineers and researchers who work with fault diagnosis and monitoring of technical systems.

Data-driven fault diagnosis of technical systems use training data from nominal and faulty operation to train machine learning models to detect and classify faults. However, data-driven fault diagnosis is complicated by the fact that training data from faults is scarce. The fault diagnosis task is often treated as a standard classification problem. There is a need for methods to design fault detectors using only nominal data. In model based diagnosis, the ability construct fault detectors depends on analytical redundancy properties. While analytical redundancy is a model property, it describes the diagnosability properties of the system. In this work, the connection between analytical redundancy and the distribution of observations from the system on low-dimensional manifolds in the observation space is studied. It is shown that the intrinsic dimension can be used to identify signal combinations that can be used for constructing residual generators. A data-driven design methodology is proposed where data-driven residual generators candidates are identified using the intrinsic dimension. The method is evaluated using two case studies: a simulated model of a two-tank system and data collected from a fuel injection system. The results demonstrate the ability to diagnose abnormal system behavior and reason about its cause based on selected signal combinations.

Hybrid diagnosis systems combine model-based and data-driven methods to leverage their respective strengths and mitigate individual weaknesses in fault diagnosis. This paper proposes a unified framework for analyzing and designing hybrid diagnosis systems, focusing on the principles underlying the computation of diagnoses from observations. The framework emphasizes the importance of assumptions about fault modes and their manifestations in the system. The proposed architecture supports both fault decoupling and classification techniques, allowing for the flexible integration of model-based residuals and data-driven classifiers. Comparative analysis highlights how classical model-based and pure data-driven systems are special cases within the proposed hybrid framework. The proposed framework emphasizes that the key factor in categorizing fault diagnosis methods is not whether they are model-based or data-driven, but rather their ability to decuple faults which is crucial for rejecting diagnoses when fault training data is limited. Future research directions are suggested to further enhance hybrid fault diagnosis systems.

This paper addresses the training of Neural Ordinary Differential Equations (neural ODEs), and in particular explores the interplay between numerical integration techniques, stability regions, step size, and initialization techniques. It is shown how the choice of integration technique implicitly regularizes the learned model, and how the solver’s corresponding stability region affects training and prediction performance. From this analysis, a stability-informed parameter initialization technique is introduced. The effectiveness of the initialization method is displayed across several learning benchmarks and industrial applications.

Data-driven fault diagnosis of dynamic systems is complicated by incomplete training data, unknown faults, and overlapping classes. Many existing machine learning models and data-driven classifiers are not expected to perform well if training data is not representative of all relevant fault realizations. In this work, a data-driven model, called a flexi-pipe model, is proposed to capture the variability of data in residual space from a few realizations of each fault class. A diagnosis system is developed as an open set classification algorithm that can handle both incomplete training data and overlapping fault classes. Data from different fault scenarios in an engine test bench is used to evaluate the performance of the proposed methods. Results show that the proposed fault class models generalize to new fault realizations when training data only contains a few realizations of each fault class.

Data-driven fault diagnosis requires training data that is representative of the different operating conditions of the system to capture its behavior. If training data is limited, one solution is to incorporate physical insights into machine learning models to improve their effectiveness. However, while previous works show the usefulness of hybrid approaches for isolation of faults, the impact of training data must be taken into consideration when drawing conclusions from data-driven residuals in a consistency-based diagnosis framework. By giving an understanding of the physical interaction between the signals, a hybrid fault diagnosis approach, can enforce model properties of residual generators to isolate faults that are not represented in training data. The objective of this work is to analyze the impact of limited training data when training neural network-based residual generators. It is also investigated how the use of structural information when selecting the network structure is a solution to limited training data and how to ameliorate the performance of hybrid approaches in face of this challenge.

Fault diagnosis is important for automotive systems, e.g., to reduce emissions and improve system reliability. Developing diagnosis systems is complicated by model inaccuracies and limited training data from relevant operating conditions, especially for new products and models. One solution is the use of hybrid fault diagnosis techniques combining model-based and data-driven methods. In this work, data-driven residual generation for fault detection and isolation is investigated for a system injecting urea into the aftertreatment system of a heavy-duty truck. A set of recurrent neural network-based residual generators is designed using a structural model of the system. The performance of this approach is compared to a baseline model-based approach using data collected from a heavy-duty truck during different fault scenarions with promising results.

With trends as IoT and increased connectivity, the availability of data is consistently increasing and its automated processing with, e.g., machine learning becomes more important. This is certainly true for the area of fault diagnostics and prognostics. However, for rare events like faults, the availability of meaningful data will stay inherently sparse making a pure data-driven approach more difficult. In this paper, the question when to use model-based, data-driven techniques, or a combined approach for fault diagnosis is discussed using real-world data of a permanent magnet synchronous machine. Key properties of the different approaches are discussed in a diagnosis context, performance quantified, and benefits of a combined approach are demonstrated.

When optimising the vehicle trajectory and powertrain energy management of hybrid electric vehicles, it is important to include look-ahead information such as road conditions and other traffic. One method for doing so is dynamic programming, but the execution time of such an algorithm on a general purpose CPU is too slow for it to be useable in real time. Significant improvements in execution time can be achieved by utilising parallel computations, for example, using a Field-Programmable Gate Array (FPGA). A tool for automatically converting a vehicle model written in C++ into code that can executed on an FPGA which can be used for dynamic programming-based control is presented in this paper. A vehicle model with a mild-hybrid powertrain is used as a case study to evaluate the developed tool and the output quality and execution time of the resulting hardware. Copyright (C) 2020 The Authors.

By running simulation models on FPGAs, their execution speed can be significantly improved, at the cost of increased development effort. This paper describes a project to develop a tool which converts simulation models written in high level languages into fast FPGA hardware. The tool currently converts code written using custom C++ data types into Verilog. A model of a hybrid electric vehicle is used as a case study, and the resulting hardware runs significantly faster than on a general purpose CPU.