Residual generation synthesis for sensor fault diagnosis in multi-agent systems.* C´ esar T. Mart´ ınez-Villegas Centre de Recherche en Automatique de Nancy Universit´ e de Lorraine, CNRS UMR 7039 F-54506 Vandoeuvre-les-Nancy, France martine21@univ-lorraine.fr Lizeth Torres C´ atedras CONACYT Instituto de Ingeniera, UNAM 04510 Coyoacan, Mexico ftorreso@iingen.unam.mx Didier Theilliol Centre de Recherche en Automatique de Nancy Universit´ e de Lorraine, CNRS UMR 7039 F-54506 Vandoeuvre-les-Nancy, France didier.theilliol@univ-lorraine.fr Abstract—This paper presents a novel approach for the gen- eration of residual in order to diagnose (detect and identify) sensor faults in a particular class of network based on multi-agent system (MAS). Specifically, the network under consideration is composed of local subsystems with nonlinear dynamics which are interconnected through a second-order consensus control law over one of their local states. The proposed methodology takes advantage of a model of the interconnection dynamics rather than nonlinear local models of each subsystem. Simulation results are provided to illustrate the use of the proposed scheme on a network of planar take-off and landing vehicles (PVTOL) with consensus over the horizontal coordinate. Index Terms—Sensor faults, fault detection and isolation, multi-agent systems, diagnosis, consensus control. I. I NTRODUCTION Nowadays, large-scale networks are equipped with numer- ous sensors to collect high amounts of information about the network’s state. However, the increase in the number of sensors has brought new challenges in fault diagnosis tasks since each new sensor in the network represents a fault source. Sensor faults might affect not only systems designed to perform state estimation and monitoring tasks [1], but they can be transmitted and propagated to any system through measured information. This represents a common issue that affects the performance of the so-called interconnected systems. In general terms, an interconnected system is a type of dynamic system that is made up of a set of local sub- systems that shares information among them in the form of interconnection signals or variables. A particular case of interconnected systems exists when the information shared among subsystems contains local outputs (outputs of local subsystems). In that case, a sensor fault would be transmitted to all the subsystems that are connected to a “faulty” one regardless of their “local” health state (faulty or fault-free). Different methodologies have been developed to address sensor fault diagnosis (SFD) in different ways, nonetheless, many of them are based on the idea of redundancy, which in the sensing context is the capability to measure or estimate some variables by using different mechanisms or devices [2], [3]. In this regard, a common classification of methods for *This work was supported by Consejo Nacional de Ciencia y Tecnolog´ ıa, M´ exico (CONACYT) through a scholarship jointly with french government. the sensor diagnosis is related to the nature of the redundancy used for: physical redundancy or analytical redundancy [2]. The sensor fault diagnosis methodology presented in this paper is based on analytical redundancy. Analytical redun- dancy involves the use of mathematical, symbolic or quali- tative representations of the system, which can be used for SFD methods to perform the sensor diagnosis by comparing some data or characteristics of the system itself, to their equivalent data or characteristics of the model [4], [5]. Among the analytical-redundancy methods, observer-based approaches have been studied and successfully tested by the control- oriented fault detection and identification (FDI) community [6], [7]. However, some particular challenges arise when they are used for interconnected networks, for instance, they need to deal with a high amount of dynamic variables, with the complexity of the network architecture and with the physical distribution of the system components. For this reason, recent SFD methods for interconnected networks are based on non- centralized architectures instead of centralized schemes in order to simplify the overall diagnosis task into several smaller and less complex problems [7]. Model choice and system decomposition are two key ele- ments of observer-based SFD of interconnected system design. These key elements have received attention from the control- oriented FDI community in works using linear [8], [9], nonlin- ear continuous-time models considering a complete measured state vector [10], [11] later extended to the incomplete output vector case (input-output case) [12], or nonlinear discrete- time models [13] with different assumptions that limit the freedom of system decomposition choice. Nonetheless, in some practical situations, the model/system decomposition representation is not unique since many different models or decomposition can represent the system’s dynamical behavior. The idea of using first and second-order models based on the interconnection structure of the network has been used successfully to perform control, estimation and diagnosis ob- jectives. In [14] first order interconnected system models have been used to perform team centroid tracking and formation control, as well as overall state estimation and diagnosis on a distributed fashion. [15] describes a decentralized fault diagno- sis methodology for a fist-order interconnected network where some of the agents are driven by external signals (independent 2018 UKACC 12th International Conference on Control (CONTROL) Sheffield, UK, 5-7 Sept 2018 978-1-5386-2864-5/18/$31.00 ©2018 IEEE 389