Fusing Causality, Reasoning, and Learning for Fault Management and Diagnosis

This talk presents an online data and knowledge based diagnosis method. It leverages decision trees in which decisions are made based on diagnosis meta knowledge, namely knowledge about the properties of diagnosis indicators. This knowledge is used at the level of each node to set a symbolic classification problem that brings out discriminating functions. This results in a multivariate decision tree that produces a compact model for diagnosis. The use of decision trees increases the explicability of the results found, all the more so as one discovers the explicit formal expressions of diagnosis indicators in the process. The method has been tested on static systems. On the well-known polybox, the three diagnosis indicators known as analytical redundancy relations, that are generally computed from the model, are found.

Runtime Verification (RV) is a lightweight verification technique that aims at checking whether a run of a system under scrutiny (SUS) satisfies or violates a given correctness specification. The tutorial first gave an overview about the general framework of RV, and the techniques to synthesize run-time monitors that can be efficiently executed in combination with the SUS. Then, we will cover the relationship between RV and the field of Fault Detection and Isolation (FDI). In FDI, runtime monitors are built taking into account models of the SUS, in order to monitor the occurrence of internal (faulty) conditions that are not directly observable.

In recent years, data-driven artificial intelligence (AI) has acquired relevance in diagnostics. Traditional methodologies are being complemented and, in some cases, supplanted by AI-powered solutions, suggesting a possible paradigm shift. AI algorithms have demonstrated remarkable capabilities in analyzing vast amounts of data with speed and precision, enabling early detection of anomalies and predictive insights into potential faults.

The necessity of monitoring the condition of devices with diverse characteristics, each of which may exhibit unique features, behaviors, and failure modes, makes it challenging to develop a universally applicable monitoring solution based on AI that usually necessitates extensive training data. This is even more true when considering classical predictive maintenance tasks such as Remaining Useful Life (RUL) estimation, with the typical scarcity of data covering the life of the monitored system until failure. Consequently, a critical need arises for methodologies that enable these algorithms to effectively perform on unseen cases, ensuring their reliability and accuracy in practical scenarios.

Domain adaptation techniques try to solve this problem by bridging the gap between the source domain (where the model is trained) and the target domain (where it is deployed), allowing for effective knowledge transfer and adaptation to specific contexts.

This contribution presents a comprehensive, modular, and scalable solution for data-driven diagnosis and prognosis that integrates deep learning algorithms and adversarial domain adaptation to permit transfer learning and increase generalization. The pipeline starts with the data acquisition and preprocessing steps, with a configurable feature extraction phase that produces a compressed latent representation of the input samples, computed using feature extraction techniques from multiple channels of raw signals. Additional features are calculated from the latent space of deep autoencoders. This latent space is deliberately composed of only a few variables, enforcing the compression of the information contained in the input. Domain adaptation based on adversarial learning uses the features extracted from all the samples available for the source and only the first few samples from the target system. This is to align the Deep Learning regressor responsible for estimating the health index of the target necessary to compute the RUL.

In our solution, the setup includes a communication system via MQTT, enabling an online data stream for real-time monitoring and maintenance. The overall infrastructure was deployed in an industrial setting and tested in a real-time experiment, demonstrating the validity of the proposed approach.

This talk described two big ideas:

1. 1.

   Analogy plays key roles in human diagnosis It provides reasoning from experience and detection of novel situations Analogical generalization constructs probabilistic relational schema Provides a source of priors for model-based diagnosis Analogical learning is incremental, inspectable, and data/training efficient

2. 2.

   Cognitive architectures need diagnosis

   Goal: Software social organisms instead of tools Systems need to manage their own learning Achieving agency needs internal diagnostic capabilities How to build software that never blue-screens?

The real world is made up of cyber-physical systems – we want them not to fail, to be invisible. How can we improve the resilience of our CPSs? Recently, a space craft designer said to me “all failures are failures of imagination.” By that he meant that it’s the designers’ responsibility to imagine all the things that could possibly go wrong with the space craft and design around or compensate for them. With the advances in AI including planning and ML we can now put AIs inside of our systems which can address the designer’s unknown unknowns. To make significant advances we need to define what resilience is and how to measure it. I will describe a number of definitions of resilience. In this talk I will describe a comprehensive approach to achieving certain types of resilience with examples ranging from printers to unmanned ships.

A diagnosis system can be described as a function that uses observations from the monitored system to compute diagnoses. Because of its industrial and scientific relevance, the fault diagnosis problem has been approached in many different communities. A popular approach is data-driven fault diagnosis which refers to methods that use historical data from different fault scenarios to learn the relation between observations and diagnoses. Compared to model-based diagnosis, which uses physically based models that have a long theoretical foundation, data-driven fault diagnosis is often treated as a general classification problem. This presentation has looked at data-driven fault diagnosis from a model-based diagnosis perspective. It is shown that central model-based concepts, like redundancy, can be interpreted in a data-driven framework and it is also illustrated how these ideas can be used to develop new data-driven fault diagnosis methods.

I present the Rayleigh-Ritz Autoencoder (RRAE) architecture [1] for unsupervised hybrid machine learning. It is suitable for applications where the system being observed by multiple sensors is well modeled as a boundary value problem. The embedding of the admissible functions in the decoder implements a truly physics-informed machine learning architecture. The RRAE provides an exact fulfillment of Neumann, Cauchy, Dirichlet or periodic constraints. Only the encoder needs to be trained; the RRAE is numerically more efficient during training than traditional autoencoders.

We extended the RRAE architecture [2] to distribution-free statistics to achieve stability with respect to non-Gaussian data. This provides consistent results for sensor data with both Gaussian and non-Gaussian perturbations. The necessity for handling non-Gaussian data in sensor applications is documented by the behavior of inclinometer sensors where the perturbations are characterized by Cauchy-Lorentz distribution. In such cases variance does not provide a reliable measure for uncertainty; consequently, 1-norm error measures are investigated thoroughly.

Monitoring is a runtime verification technique that can be used to check whether an execution of a system (trace) satisfies or not a given set of properties. Compared to other formal verification techniques, e.g., model checking, one needs to specify the properties to be monitored, but a complete model of the system is no longer necessary. In the talk (uploaded slides), we first introduce the notion of monitoring and display a simple architecture of a monitoring system. Then, we provide a characterisation of positively and negatively monitorable properties, and we define the safety and cosafety fragments of Linear Temporal Logic (LTL). We complete the picture by showing that monitorability goes behind safety and cosafety LTL fragments, and that there are natural properties which are not monitorable. Next, we proceed by pointing out that monitoring suffers from some significant limitations. In particular, modern systems have such a level of complexity that it is impossible for a system engineer to specify in advance all properties to be monitored, and even minor changes to the system to be monitored can introduce unforeseen bugs. To overcome these limitations, we provide a multi-objective genetic programming algorithm to automatically extend the set of properties to monitor on the basis of the history of failure traces collected over time. The monitor and the learning algorithm are then integrated in a unifying framework, whose distinguishing features are (i) interpretability (the machine learning methods manipulate and produce only formulae, that can be easily inspected by a system engineer), (ii) formal guarantees on monitorability (every formula produced during the learning phase is guaranteed to be monitorable), and (iii) generality (different monitoring and machine learning backends). The framework has been experimentally validated on various public datasets, and the outcomes of the experimentation confirm the effectiveness of the proposed solution.

The transition from sub-symbolic representation in artificial intelligence such as time series to symbolic representations such as expressions in formal logic or in language is crucial to creating resilient cyber physical systems. The first step for this is often the discretization, i.e. the identification of symbolic concepts that come true at certain points in time. The tutorial presents typical discretization algorithms for time series, especially with a focus on engineering and scientific applications. In the last step, a brief overview of possibilities to further develop the identified concepts into causalities is given.

For Seminar 24031, we aimed to invite a diverse audience from a variety of fields connected to the seminar’s focus. For us organizers it was thus important to provide the attendees with some basic knowledge, common grounds concerning well-established techniques in the field, and also a basic context and terminology. In this introductory talk, I focused on providing some brief basics about model-based diagnosis (MBD), also known as consistency-based diagnosis or DX approach. Due to its attractive features, MBD is such a central technology in the scope of this seminar, and I covered the following aspects in my presentation:

• $\mathrm{■}$

  the basic underlying approach of reasoning from first principles

• $\mathrm{■}$

  the standard scenario of explaining some unexpected behavior like a failed test case

• $\mathrm{■}$

  connections to verification tasks and techniques

• $\mathrm{■}$

  the very basic definitions of diagnoses, conflicts and minimal hitting sets

• $\mathrm{■}$

  the impact of using a weak fault model or strong fault models

• $\mathrm{■}$

  two basic algorithmic concepts for MBD: conflict-driven and direct

• $\mathrm{■}$

  MBD’s flexibility in terms of application and deployed algorithm

• $\mathrm{■}$

  diagnosing multiple scenarios at the same time (e.g. like results from a test suite) and the resulting opportunity to characterize a system (when using a representative test suite, e.g., obtained with combinatorial testing)

• $\mathrm{■}$

  diagnosing static scenarios and sequential behavior

• $\mathrm{■}$

  improving the basic algorithmic concepts via algorithmic optimizations (example RC-Tree) and structural, diagnosis problem-specific information (like exploiting its parse tree when diagnosing some LTL description)

• $\mathrm{■}$

  completeness and soundness of MBD in relation to the model and the scenario(s)

• $\mathrm{■}$

  information about which authors of the covered papers participated in the seminar.

Given information about (a) the desired operating conditions for a system and (b) the tasks the system must carry out, the fault-tolerant control design (FD) task is to design a set of controllers to ensure that conditions (a) and (b) are guaranteed, even when faults and/or external disturbances occur.

Designing control systems from requirements and model specifications is a challenging task, and has been addressed from many perspectives, most notably design optimization and multi-controller tuning. Our approach extends both design optimization and multi-controller tuning. Multi-controller tuning adopts a “divide and conquer” approach, decomposing the system’s operating range into smaller local sub-spaces, each associated with a “local” model. These local models are then combined to create the global system response. The primary advantage of the (MM) approach lies in its simplification of complex modeling through the use of these local models.

We develop an optimisation-based approach to designing systems with fault-tolerance and resilience capabilities, i.e., it enables multi-mode operation. Standard design optimisation approaches assume a single nominal mode; in contrast, we explicitly define an approach that generalises this framework to enable the design of systems that operate in multiple modes. Multi-controller tuning methods typically use methods to compute the set of possible modes, and then tune a controller for each mode. In contrast to multi-controller tuning methods, this new approach does not optimize just performance of individual controllers, but task-centric performance, based on topological transformations from plant spaces to control spaces.

In industrial processes, anomalies in the production equipment may lead to expensive failures. To avoid and avert them, the identification of the right root cause is crucial. Ideally, the search for a root cause is backed by causal information like causal graphs. We presented an extension of a framework that fuses causal graphs with anomaly detection to infer likely root causes in a process setup. The causal graph is required to contain measurement and root cause variables and causal relations with annotations for process steps. The framework uses this graph to compute for each root cause variable which measurement variable it might affect in which process step. Thereby, it considers that a cause must always precede its effect. Independently, an anomaly detection algorithm is performed on given sensor measurements to provide information about anomalies and the corresponding process step. Finally, the framework computes the likelihood of each potential root cause by comparing the results from the graph preprocessing and the anomaly detection using the Jaccard similarity to identify the most likely root cause. We demonstrated the use of this framework on a simulated robotic gripping process. Future research might investigate how to learn the causal graph from the provided data using causal discovery methods and how to apply the framework in an online fashion on given process data.

Health monitoring approaches are usually either model-based or data-based. This work aims at using available data to learn a hybrid model to profit from both the data-based and model-based advantages. The hybrid model is represented under the Heterogeneous Petri Net formalism. The learning method is composed of two steps: the learning of the Discrete Event System (DES) structure using a clustering algorithm (DyClee) and the learning of the continuous system dynamics using two regression algorithms (Support Vector Regression or Random Forest Regression). The method is illustrated with an academic example.

Two distinct and parallel research communities have been working along the lines of the Model-Based Diagnosis approach: the FDI community and the DX community that have evolved in the fields of Automatic Control and Artificial Intelligence, respectively. This talk clarifies and links the concepts and assumptions that underlie the FDI analytical redundancy approach and the DX consistency-based logical approach. A formal framework is proposed to compare the two approaches. It is shown that by adopting the same assumptions regarding fault exoneration, they produce the same diagnostic results.

Space propulsion systems continue to be a significant source of faults in space activities, necessitating dedicated fault management strategies to meet stringent safety requirements. The operational nature of these systems, pushed to the limits of technical feasibility to minimize weight, makes them susceptible to a diverse set of faults, with abnormal behavior having potentially catastrophic consequences. The substantial costs associated with the loss of a launch vehicle or spacecraft underscore the critical need for effective fault detection, diagnosis, and mitigation functionalities. Real-time detection and assessment of faults based on available sensor data are imperative to initiate emergency shutdowns or reconfigurations, further compounded by the constraint of limited computing resources.

The German Aerospace Center (DLR), in collaboration with various partners, has long been engaged in exploring the application of AI methods for the operation of space propulsion systems. While past efforts primarily focused on intelligent control in the absence of faults, recent initiatives, such as the collaboration with Silicon Austria Labs (SAL) within the SUNRISE project, mark the initial strides towards fault management. The SUNRISE project is dedicated to researching dependable sensor concepts for resilient control.

The first segment of this presentation unveils preliminary findings, showcasing how trained virtual sensors can effectively estimate critical quantities, such as combustion mixture ratios, and detect faults with high accuracy. In the second part, we introduce the LUMEN demonstrator engine, currently in development, serving as an ideal test bed for diverse control and diagnosis approaches. This includes intentionally injecting faults into the system and utilizing the platform for generating training data for machine learning algorithms.

Deploying systems requires to show that the system complies with its requirements and specifications. Hence, quality assurance is an essential part of any system development, which also holds for systems with attached resilience functionality utilizing model-based reasoning. In my talk, I discuss the general challenge of quality assurance applied to systems with monitoring and diagnosis functionality and the importance of residual risks. In particular, I mention which faults that come when introducing monitoring and diagnosis have to be considered and their effects on residual risks. Afterward, I present early work on using testing for quality assurance for models used for diagnosis and challenges. In particular, I discuss the challenge of coming up with methods for checking the quality of the tests and its consequences. Finally, I summarize the findings and present open challenges.

The breakout session on “Large Language Models for Root Cause Diagnosis” featured a comprehensive exploration and discussion on the potential and challenges of using Large Language Models in the topics of the “DX – Principles of Diagnosis” community. The central aspects that were discussed, revolved around (i) the models themselves and their current and potential future capabilities, (ii) the training data for training and refining LLMs for diagnosis tasks, (iii) potential application areas, as well as (iv) current, and (v) future trends and topics that should be monitored or covered by the DX community. As a result, the attendees agreed on writing a position paper, which will capture the current potential and drawbacks of LLMs within DX domains.

The beginning of the session included a brief introduction into the principles of LLMs. Introducing the capability of state-of-the art LLMs and their capability on simple diagnostic benchmark problems like the Polybox. Subsequently, the prerequisites for current LLMs were discussed to effectively perform diagnoses, including necessary data, semantics, and specialized training. The capabilities of current LLMs and LLM-ensembles were discussed, highlighting the capability of formulating programming code for simple, testable and traceable reasoning, as well as the capability of understanding image data, like circuits or technical drawings, for an easier and more precise recognition of concepts of diagnosable systems. Extending the capability of pre-trained LLMs for diagnosis by fine-tuning them on diagnosis problems, as well ensemble approaches of different Expert-LLMs for different diagnostic tasks, were highlighted as low-hanging fruits. Lastly, the capability of continuous training of LLMs for their application on changing systems was identified as a challenge.

Regarding the training data for LLMs that perform diagnostic tasks, a broad field was identified, reaching from image data, to time-series data from system observations, natural language, technical documentations, or structured knowledge graphs. The discussion came to the consensus that for training from ground up general world-knowledge should be included, whereas for fine-tuning models only specific information would be needed.

Identifying causality with LLMs was considered as a fundamental aspect of LLMs for fostering applications in diagnosis, like root cause analysis or root cause identification. Additionally, the ability to capture causality and contradiction was identified as a major aspect that should not only be considered and researched on a phenomenological level which considers LLMs as black-boxes, but also by looking into the functioning of the models.

Current trends and topics that should be monitored by the DX community revolve around the short term perspective of current LLMs in the application field of diagnosis. This includes understanding the immediate capabilities and limitation of current models, as well as application scenarios in which LLMs could already perform diagnostic tasks or at least pose as a component within a diagnosis framework.

Future trends and topics include the long term perspective of LLMs in the DX context. These trends revolve around enhancing the accuracy of LLMs and should be driven by autonomy research for structuring the requirements for LLMs in diagnostic roles.

The session concluded in writing a statement paper toward the current state of LLMs in diagnostic applications, highlighting aspects that LLMs already can achieve, and limitation they occur. The claims will be proven with empirical evaluations, like testing LLMs for creating causal graphs or evaluating on standard diagnosis problems.

In respect of these specific questions we saw that the application and scenario play a significant role when aiming to assess what would be “good” and “bad” behavior. The same goes for the question of whether we would assess the performance of a system in a local or a global context, an example for the latter being a system of resilient systems context. To this end we identified a set of such relevant scenarios ranking from an energy management scenario at a local home, via the operation of an electric grid, via agents/robots in a collaborative disaster or military scenario, to supply chain management. We also discussed and converged to a definition of resilience that would tailor to all the expressed needs. Enabled by our discussions, we identified also some follow-up actions:

• $\mathrm{■}$

  Authoring a white paper on resilience by a group of the attendees of this seminar (in 2024)

• $\mathrm{■}$

  Submitting a proposal for a follow-up Dagstuhl Seminar proposal that focuses on resilience, and where we will invite scientists from more relevant fields as well as relevant stakeholders (agency, psychology and societal sciences, security & safety, law and public regulations, …)

Please note that the discussions led in an interdisciplinary context at Dagstuhl will also contribute to the evolution of the Int. Workshop on Principles of Diagnosis to International Conference on Principles of Diagnosis and Resilient Systems that we are implementing in 2024.