Despite the immense performance improvement of deep learning-based object detection, the state-of-the-art object detection systems are still prone to misdetections. This work presents a method to predict such misdetections at run-time by using a small network, referred to as ConsensusNet, to learn the correlation patterns or consensus of neighboring detections before non-maximum suppression (NMS). Based on such correlations, ConsensusNet predicts if there are misdetection failures. The proposed method is experimentally evaluated considering single person class from COCO dataset and using YOLOv3 as the object detection system. It shows the proposed method can achieve accuracy of 84.6% and the performance measured in other metrics are also promising. To the best of our knowledge, ConsensusNet is the first network reported for predicting misdetections in object detection.

We study the decentralized control of discrete-event systems withmultiplelocal specifications. Only a subset of events occur at a local site, and a local specification as well as control/observation capabilities of a local supervisor are defined with respect to such local events. The goal of control is to ensure that the executed behavior at each local site is as desired, i.e., there are multiple specifications, one for each site. We show that the control problem for multiple local specifications is different from that of a single global specification, and present a necessary and sufficient condition for the existence of decentralized supervisors for enforcing the given multiple local specifications. The synthesis of decentralized control for enforcing multiple local specifications is also presented. We also specialize our results to the case of concurrent plants (one which is composed of several local subplants), which offers certain computational savings. The results are illustrated through a simple manufacturing system example.

We study diagnosis of timed discrete-event systems (TDESs) modeled as timed-automata. Earlier works on diagnosis of TDESs assumed that a diagnoser has partial observation of events but can measure (or observe) time with arbitrary precision. In practice, however, time can only be measured with finite precision. We model the finite precision observability of time using a digital-clock that measures time discretely by executing ticks. For the diagnosis purposes, the set of nonfaulty timed-traces is specified as another timed-automaton that is deterministic, generalizing the forms of nonfaulty specifications considered in the earlier works. We show that the set of timed-traces observed using a digital-clock with finite precision is regular, i.e., can be represented using a finite (untimed) automaton. We show that the verification of diagnosability (ability to detect the execution of a faulty timed-trace within a bounded time delay) as well as the offline synthesis of a diagnoser are decidable by reducing these problems to the untimed setting. The reduction of the diagnosis problem to the untimed setting also suggests an effective method for the offline computation of a diagnoser as well as its online implementation for diagnosis.

More and more advanced features such as adaptive cruise control, collision avoidance, and stability control are being implemented in vehicles. These features are usually implemented as distributed CAN (controller area network) systems right now. In a CAN system, normally there is no clock synchronization among the ECU (electronic control unit) nodes connected by the CAN bus. Without synchronization, the clocks of those ECU nodes could drift away from each other. A typical clock drift rate of 30 ppm (parts per million) could cause a clock to drift by 108 milliseconds in one hour. The clock drift this large could cause problems in those advanced vehicle control systems. In fact, a sequence of messages could get lost in a distributed CAN system due to the combination of clock drift, transmission jitter, and finite buffer size. To solve the above problem, instead of performing high overhead clock synchronization in the CAN system, this paper provides an economical solution for the prevention of the above message loss problem. The idea is to synchronize the task activations on different ECU nodes and such synchronizations are performed only when necessary. An analysis method is developed for the determination of the synchronization frequency; and an algorithm is provided for the task activation synchronization.

Allocating software components and messages to computing devices and their interconnecting networks is a great challenge in vehicle control system development. The allocation decisions are usually made at an early development stage without the implementation details of the computing platform. To produce low-cost products with correct vehicle control, the software allocation must facilitate the final system scheduling with better resource utilization and the presence of platform implementation uncertainties. In this paper, we present an allocation method to improve the resource utilization and scheduling while keeping the decisions independent of delayed platform implementation. Our method determines the allocation using a genetic algorithm with an optimization function combining the peak usage request and the usage request density to minimize the potential contentions on shared resources. The preliminary evaluation using the representative synthetic systems has shown the effectiveness and applicability of the method.

We propose a two-tiered hierarchical approach for detecting faults in embedded control software during their runtime operation: The observed behavior is monitored against the appropriate specifications at two different levels, namely, the software level and the controlled-system level. (The additional controlled- system level monitoring safeguards against any possible incompleteness at the software level monitoring.) A software fault is immediately detected when an observed behavior is rejected by a software level monitor. In contrast, when a system level monitor rejects an observed behavior it indicates a system level failure, and an additional isolation step is required to conclude whether a software fault occurred. This is done by tracking the executed behavior in the system model comprising of the models for the software and those for the nonfaulty hardware components: An acceptance by such a model indicates the presence of a software fault. The design of both the software-level and system-level monitors is modular and hence scalable (there exists one monitor for each property), and further the monitors are constructed directly from the property specifications and do not require any software or system model. Such models are required only for the fault isolation step when the detection occurs at the system level. We use input-output extended finite automata (I/O- EFA) for software as well as system level modeling, and also for modeling the property monitors. Note since the control changes only at the discrete times when the system/environment states are sampled, the controlled- system has a discrete-time hybrid dynamics which can be modeled as an I/O-EFA.

End-to-end latency of messages is an important design parameter that needs to be within specified bounds for the correct functioning of distributed real-time control systems. In this paper we give a formal definition of end-to-end latency, and use this as the basis for checking whether a stipulated deadline is violated within a bounded time. For unbounded verification, we model the system as a set of communicating Timed Automata, and perform reachability analysis. The proposed method takes into account the drift of clocks which is shown to affect the latency appreciably. The method has been tested on a medium sized automotive example.

In our earlier work, we introduced a state-based approach for the diagnosis of repeatedly occurring failures in discrete event systems (DESs). Since temporal logic provides a simpler way of specifying system properties; in this paper, a temporal-logic-based approach for diagnosing the occurrence of a repeated number of failures is developed. Linear-time temporal-logic (LTL) formulae are used to represent the specifications of DESs. Notions of prediagnosability for failures and diagnosability for repeated failures are introduced in the setting of temporal logic. A polynomial algorithm for the test of prediagnosability for failures is provided. The diagnosis problem for repeated failures in the temporal-logic setting is reduced to one in a state-based setting, and so the prior results of a state-based repeated failure diagnosis can be applied. Finally, a simple example is given for illustration. Note to Practitioners-Certain failures in a system are repeatable, such as routing errors in a manufacturing system. A theory for the diagnosis of such failures was presented in an earlier work of Jiang et al. The present paper uses temporal logic to specify such failures. It turns out that repeatable failures can be specified as violations of invariant properties (i.e., properties that must always hold). Given an invariant property that the system must always satisfy, an algorithm is presented to refine the system model and label those states of the refined system where the property is violated. The problem of repeated diagnosis then requires determining, within a bounded delay, each time a "failure-state" is visited. For this analysis, the existing theory developed by Jiang et al. is used.

We introduce the notion of repeated failure diagnosability for diagnosing the occurrence of a repeated number of failures in discrete event systems. This generalizes the earlier notion of diagnosability that was used to diagnose the occurrence of a failure, but from which the information regarding the multiplicity of the occurrence of the failure could not be obtained. It is possible that in some systems the same type of failure repeats a multiple number of times. It is desirable to have a diagnoser which not only diagnoses that such a failure has occurred, but also determines the number of times the failure has occurred. To aid such analysis we introduce the notions of K-diagnosability (K failures diagnosability), [1, K]-diagnosability (1 through K failures diagnosability), and [1, /spl infin/]-diagnosability (1 through /spl infin/ failures diagnosability). Here the first (resp., last) notion is the weakest (resp., strongest) of all three and the earlier notion of diagnosability is the same as that of K-diagnosability or that of [1, K]-diagnosability with K=1. We give polynomial algorithms for checking these various notions of repeated failure diagnosability and also present a procedure of polynomial complexity for the online diagnosis of repeated failures.

The decentralized supervisory control problem of discrete event systems under partial observation is studied in this paper. The main result of the paper is a necessary and sufficient condition for the existence of decentralized supervisors for ensuring. That the controlled behavior of the system lies in a given range. The contribution of the paper is (1) our setting of decentralized control generalizes the prior ones; (2) we present an alternative approach for solving the decentralized control problem, which leads to computational saving for concurrent systems and certain other systems; and (3) our generalized formulation and its solution lets us extend several of the existing results reported previously. The results of our paper are illustrated by an example of a simple manufacturing system.