Design for diagnosability of multistation manufacturing systems based on sensor allocation optimization

Abstract

System monitoring and diagnosing can be achieved through measuring sensing data. Meanwhile, diagnosability is the capability of a system in terms of diagnosing dimensional variation and root cause identification. In order to obtain enough sensing data, special attention should be taken to the sensor allocation optimization within the framework of ensuring system diagnosability, which can lower sensing cost and reduce time to diagnosis. This paper investigates both theoretically and experimentally the effect of sensor allocation optimization on the diagnosability of a multistation manufacturing system (MMS). The design for diagnosability (DFD) is formulated and implemented in the early design phase. Three indices, namely detectability, locatability, and isolability, are proposed to measure the system diagnosability. A two-step process of diagnosing variation sources is presented to specify the variation transmission between stations and variation diagnosis within a station. An optimal methodology of sensor allocation is then proposed. A typical machining process is demonstrated as an example of the analytical procedure and reference for sensor allocation in practice. Four sensing strategies are conducted for performance comparison. Results indicate that the optimal sensor allocation yields less sensing cost and shorter time to diagnosis without loss of diagnosability.

Introduction

Manufacturing enterprises are facing unpredictable and rapidly changing market competition driven by customer demands. To survive and remain competitive in such an environment, tremendous efforts have been taken to improve product quality. These quality issues have been manifested in advanced manufacturing processes that require frequent reconfigurations to produce different products in various quantities, while keeping high efficiency. Tool condition monitoring has gained considerable importance in the manufacturing industry over the preceding two decades, as it significantly influences process efficiency and machined part quality [1], [2]. Meanwhile, dimensional variation causes quality defects, which are a major cause of faults in manufacturing systems [3]. Consequently, it is necessary to monitor the key product characteristics (KPCs) in such a system to reach a desired quality level [4]. Quality monitoring is the prerequisite for fault diagnosis, which is then performed to identify faults caused by the dimensional variations and root causes based on the states of the KPCs [5]; all these are manifested by abundant sensing data. It is of high priority to ensure that the KPCs are optimally selected, and are measured by sensors in a cost-effective, timely manner.

Measurement scheme and sensor allocation optimization are essential in the early design phase of a multistation manufacturing system (MMS), which utilizes plenty of sensors for multivariate measuring both in-line and off-line. A huge amount of sensing data is helpful for system diagnostic functionality enhancement. However, sensing data acquisition without proper sensor allocation is inapplicable. It has a great significance on some other factors, such as cost, energy consumption, space, and efficacy of the measurement system. Effective use of sensing data for diagnosis depends to a great extent on the optimal design of a measurement system. Therefore, the concept of diagnosability is introduced to assess the degree to which the quality-related faults in a manufacturing system is diagnosable. A design for diagnosability (DFD) structure should be formulated and implemented during the design phase as early as possible. The KPCs must be continuously monitored and measured to ensure that any quality-related faults owing to dimensional variation or other abnormal conditions can be promptly detected and diagnosed. Conversely, an inadequately designed measurement system is apt to generate some irrelevant or even conflicting information, which is incapable of offering preferred diagnosability to identify faults and further root cause the dimensional variation sources.

Research in the fields of sensor allocation and system diagnosability are limited due to the early stage of development. However, several researchers have conceptually investigated, and elaborated on viable approaches, to solve these problems. Sensor allocation optimization and system diagnosability have been addressed in recent literature [6], and they have been applied in the design of manufacturing systems [7], [8], [9]. Liu et al. [10] have proposed an axiomatic design methodology for a manufacturing system based on a diagnosability matrix. Xia and Rao [11] have presented a design approach for sensor and actuator placement in the process industry. Ding et al. [12] have investigated global and local diagnosability due to between-station and across-station variation propagation, and further pointed out that system diagnosability could be calculated based on the rank of the diagnosis matrix. Many effective optimization approaches exist for sensor allocation to determine sufficient information for dynamic analysis, such as the covariance matrix approach [13], [14], the eigensystem realization algorithm [15], the Bayesian approach [16], the effective independence approach [17], the systematic procedure [18], the maximum energy approach [19] as well as the genetic algorithm [20]. However, these approaches, besides being highly computationally intensive, suffer from a major drawback of yielding hardly any insight into why certain sensor locations are preferable to others. Furthermore, recent sensor optimization research has mainly been restricted to single station, and little can be found about the system level or the multistation environment. Yan [21] has developed a method for sensor installation based on the diagnosability analysis. A simulation model in a CAD environment has been developed and the minimal sensor set from a fault signature matrix has been computed. Fisher information matrix has been frequently used in optimization for a single station [22]. Khan and Ceglarek [23] have formalized an approach to find optimal sensor allocation by maximizing the minimum distance between two variation vectors. However, it has failed to consider the variation propagation between various stations. Therefore, a new approach for sensor allocation should be developed for the applications frequently encountered in industrial practice.

As is seen from the literature, no research has yet combined the design for diagnosability methodology of a multistation manufacturing system with the sensor allocation optimization. This research focuses on the impacts of distinctive measurement schemes and the associated sensor allocation strategies. The remainder of the paper is organized as follows: Section 2 provides a methodology for the diagnosability analysis, and three indices are proposed for the design for diagnosability. Section 3 formulates the optimization procedures of the sensor allocation for minimizing the sensing cost and reducing the time to diagnosis without any loss of diagnosability. An industrial example is used to illustrate the application of the proposed approach in Section 4. Finally, Section 5 gives some concluding remarks.