Chapter 1 Introduction
1.6 Contents overview
This dissertation is organized as follows: the on-line time delay estimation design is presented in Chapter 2. Chapter 3 introduces the robust NCS design with the implementation of RTT. Chapter 4 presents the RTT technique and multi-rate design applied to wireless NCS with unpredictable delay. In Chapter 5, the model-free PDC scheme is introduced for SISO control systems. Chapter 6 describes the proposed PDC for MIMO control systems. Finally, conclusions and recommendations for further research are provided in Chapter 7.
Chapter 2
On-line Time-delay Estimation Design
In real applications, a remote control system is generally an integration of a commercial network for message transmission and an industrial network to control the remote hardware through a communication gateway. As the delay in a commercial network Ethernet is significantly time varying depending on the number of end users, the delay is estimated in this study by processing the on-line measurement of RTT between the application layers of the remote node and the client node. This research proposes a remote NCS structure by implementing the on-line time delay estimator with an adaptive Smith predictor because the induced time delay in NCS degrades its stability and performance. The adaptive Smith predictor scheme is developed by directly applying the estimated time delay to deal with network-induced delays. NCS is thus simplified as the desirable closed-loop system with an additional pure time delay
.
To prove the feasibility of the proposed remote control system, the developed design has been applied to an AC 400 W servo motor tested in 15 Km distance. Experimental results indicate that significantly improved stability and motion accuracy can be reliably achieved by applying the proposed approach.2.1 Introduction
Due to the rapid development of data communication network technologies in the Internet, real-time network control applications have increasingly gained attentions.
These applications include tele-operations, remote mobile robots, and factory automation, which are organized by wiring connections among control system devices through network resources. The popularity of network control applications is obvious because they can be conveniently and systematically maintained in an industry (Kaplan, 2001). NCS is one of the newly developed technologies in modern industrial applications. It has potential applications by simply interconnecting all sensors, actuators, and controllers through networks (Lian et al., 2001). The introduction of network technologies provides easy maintenance and expandability for control system
collision. Network scheduling has been studied to cope with these problems. Another concern is that NCS performance may become unstable because network delay is stochastic in nature and it is difficult to directly apply linear delay-time system analysis.
The total network-induced delay, both in the controller and the actuator, may present a bound or random format depending on the network protocols, which may seriously degrade NCS performance.
Recently, the use of NCS to deal with band-limited channels, time delays, and packet loss has been widely studied mainly for the improvement of communication protocols and controller design (Baillieul and Antsaklis, 2007; Hespanha et al., 2007;
Zampieri, 2008). With proper communication protocols, the enhancement of transmission technology provides guaranteed quality of service (QoS) for real-time applications (Grenier and Navet, 2008). A sufficient condition ensuring robust stability of NCS was presented by Chen et al. (2007). Tatikonda et al. (2004) formulated a linear discrete-time control problem with a noiseless digital communication link, and provided the role of information patterns and control policy knowledge. Zai et al.
(2002) used an average dwell time for discrete switched systems to obtain conditions where the stability of NCS is guaranteed. Network-induced delay is one of the most important issues of NCS. Different methodologies have been proposed to deal with the delay effect within the process control loop. Considering both known and constant process delays with noise, a minimum variance control law (Jain and Lakshminarayanan, 2005) and a step-by-step tuning procedure (Goradia et al., 2005) were developed separately to obtain achievable PI performance for linear SISO time delay processes. Furthermore, extension of the abovementioned approaches was then developed to the MIMO system (Jain and Lakshminarayanan, 2007). A solution of the minimum variance control law for linear time-variant processes has been derived in a transfer function form (Huang, 2002). Lian et al. (2002) identified several components of the time delay of network protocols and control dynamics and they determined an acceptable working range of the sampling period in NCS. The feedback gain of a memoryless controller and the maximum allowable delay can be derived by solving a set of linear matrix inequalities (Yue et al. 2004). A design method of time-delayed control systems based on the concept of network disturbance and communication
disturbance observer (CDOB) without the knowledge of the delay-time model was also proposed (Natori and Ohnishi, 2008).
Most of the abovementioned research results are limited to constant delays or less time-varying delays, which are not true in real network environments. In this research, time-based time delay analysis of NCS is provided to explain how it affects network systems. By applying the proposed adaptive Smith predictor based on the on-line time delay estimation, satisfactory control performance of NCS can be obtained even as the time delay increases significantly over integrated commercial and industrial networks.
The proposed NCS has been applied to a remote control system for an AC 400 W servo motor tested in 15 Km distance to verify the proposed design.
2.2 NCS and time-delay measurement
The general NCS in the closed-loop model is shown in Fig. 2.1, where
t and
1t
2 are the time delays induced in the network structure for the controller-to-actuator direction and the sensor-to-controller direction, respectively. Basically, the induced network delay varies according to the network load, scheduling policies, number of nodes, and different protocols. Network delay systems are also different from general linear time delay systems, where there is an assumption that the delay on the former is constant or bounded. NCS with time-varying characteristics makes modeling and design more difficult. The total time delay can be categorized into three classes based on the parts where they occur: (1) the client node, (2) the network channel, and (3) the remote node. Time delay at the client node is mainly in the preprocessing time, which is the sum of the computation, encoding, waiting, total queuing, and blocking time.Network time delay includes the total transmission time of a message and its propagation delay, which depends on the message size, data rate, and length of the network cable. Time delay at the remote node is mainly in the post-processing time, as shown in Fig. 2.1.
Fig. 2.1 The NCS block diagram
Fig. 2.2 The experimental setup
Figure 2.2 shows the structure of the present remote NCS which includes the controller in the remote node and the client for the remote-controlled device or plant.
The client and the remote nodes communicate with each other from a distance through the Ethernet network. The client consists of two parts in the present experimental setup. The first part is the gateway. This is implemented in a computer with USBCAN, which is designed to communicate between the Ethernet network and the CAN bus. The second part is the local servo motor controller implemented on TI TMS320F2812 DSP with a speed-control mode. The data communication protocol adopts the TCP to construct the position loop for the remote control (Cheng et al., 2007). As shown in Fig. 2.2, the communication network can be modeled as the time
delay on the forward-command direction for actuators (t1) and on the feedback direction for sensors (t2). Therefore, the network time delay includes both the total transmission time of a message and the transformation time of the package from CAN to Ethernet data. The total time delay (RTT) can be expressed as tp = t1 + t2 (Fig. 2.2).
Fig. 2.3 The package transition diagram
These two protocols, Ethernet and CAN, cannot communicate with each other directly. Thus, message packages have to be processed through a gateway, as shown in Fig. 2.2. When data are transmitted to the remote node from the local hardware DSP, the type and the transmission data in a data frame should be set up in advance (Fig.
2.3). These data are then included into the CAN package and transmitted to the gateway through the CAN network, as indicated in step 1 of Fig. 2.3. After the gateway has received the package from the CAN network, the data of the CAN package will be included in the Ethernet package and the Ethernet network may thus transmit the package directly (step 2 in Fig. 2.3). When the remote node has received the package from the Ethernet network, part of the CAN package will be extracted from the Ethernet package and the data defined by users can be further obtained. In the end, the data frame will be analyzed and transmitted (step 3 of Fig. 2.3). By following the procedure (1) (2) (3), the message of the local DSP can be transmitted to a remote node. On the contrary, when data are transmitted to the remote node from the DSP, both transmission data in the data frame should be set up in the CAN package.
The CAN message is then included in the Ethernet package and part of the data will be transmitted to the gateway through the Ethernet network (step 4 of Fig. 2.3). After the gateway receives the package from the Ethernet network, the data from the CAN
package will be extracted from the Ethernet package. The CAN network will be utilized to transmit this package to DSP (step 5 of Fig. 2.3). After DSP receives the package from the CAN network, the data frame in the CAN package will be extracted.
This is step 6 in following the procedure (4) (5) (6) shown in Fig. 2.3.
The network time delay for the present experiments includes the following cases:
(1) NCTU Laboratory NCTU Laboratory and (2) NCTU Laboratory Hukuo (the two places are 15 Km apart). The computer used for this network transmission has the following specifications: Intel® Pentium CPU 1.60 GHz, 496 MB of RAM, Realtek RTL8139/810x Family Fast Ethernet NIC Network Card, and Windows XP Professional Version 2002 OS with SP2. The local area network (LAN) is used with the time delay between the application layer of the client and remote nodes. In addition, the RTT measurement is crucial in obtaining accurate delay measurements periodically.
Technically, the Windows Forms Timer component in the operating system is single threaded and is limited to an accuracy of 55 ms. A higher resolution performance counter of the DSP timer with an accuracy of 1 ms is used to measure network delay between the remote and the client nodes. We measured the time delay from two different clients within the NCTU Laboratory, and from two different clients located each in the NCTU Laboratory and Hukuo, separately, as shown in Fig. 2.4 (a) and (b).
The delay time in the integrated Ethernet and the CAN Bus within a 20 ms sampling period was measured, as shown in Fig. 2.4. Only a very small time delay (around 3–15
ms) was recorded because the transmission speed of the intranet was at 100 Mbps and
there was only a relatively short route within the NCTU Laboratory. From the NCTU Laboratory to Hukuo, the delay time increases because the transmission procedure takes more routes and switches. Experimental results as shown in Fig. 2.4 indicate that the application environment greatly affects the induced delay time in NCS. Moreover, as distance increases, the delay time of a network increases as more nodes are involved.0 2 4 6 8 10 12 14 16 18 20
Fig. 2.4 Measured Internet delays (a) NCTU Lab - NCTU Lab and (b) NCTU Lab - Hukuo
2.3 Adaptive Smith predictor
The communication network can be modeled as the time delay on the forward-command direction for the actuator and on the feedback direction for the sensor as shown below:
Fig. 2.5 The simplified block diagram of NCS
In Fig. 2.5, t1 is the command delay time, t2 is the feedback delay time, and Gc(s) is the controller. Gp(s) denotes the transfer function of the real plant model without the delay time. The transfer function from input r to output y is obtained as follows:
s 2001; Peng et al., 2004). Since the time delay in the Internet can be measured between sending and receiving a packet, the delay in a closed-loop NCS can be well-compensated by applying the Smith predictor, as shown in Fig. 2.6.
(a)
(b)
Fig. 2.6 The system with the Smith predictor (a) the original system and (b) the equivalent system
Fig. 2.7 The block diagram of the adaptive Smith predictor with a PI controller
The nominal delay time and system model adopted for the Smith predictor is tm and
ˆ ( )p
G s , respectively. In ideal conditions, Gˆ ( )p s Gp( )s and tm= tp, the block diagram in Fig. 2.6(a) can be simplified into Fig. 2.6(b) with an additional pure time delay term applied for the Smith predictor. In this study, the delay time is estimated from the real-time measured RTT for the Smith predictor. To cope with significant variation in the delay time due to network transmission, an adaptive method is proposed for the present remote control systems with the integration of the Smith predictor, the PI controller, and the real-time delay estimation, as shown in Fig. 2.7.
2.3.1 On-line estimation of the delay time
A method for estimating the delay time within the Internet for the NCS
architecture with a combination of the time-driven and event-driven processes is proposed in this section. The designed control algorithm is realized on the present network by integrating both the Ethernet and the CAN bus with a serial data communications bus in-between. Technically, the standard CAN bus transmits only 8 bytes per frame. However, the minimum data length to realize the proposed RTT measurement is 9 bytes. A programming method wherein messages will be divided into two parts and each part will be sent at each half sampling period through the CAN network is proposed here, as shown in Fig. 2.8.Fig. 2.8 The CAN data frame in the proposed NCS for measuring RTT
To illustrate the estimation of the induced network time delay from the measurement of RTT, the NCS transmission is shown in Fig. 2.9. At the beginning of the sampling period, the clock-driven sensor node transmits the sampling data to the controller node. By assuming the sensor-to-controller delay as t2 for this setup, the event-driven controller node uses the sensor data to compute the control signal and then transmits it to the actuator node. By assuming the controller-to-actuator delay as
t
1, the time-driven transmission is applied. The measurement of RTT is adopted due to its easy implementation and the fact that no clock synchronization is required because all computations are operated in the same device. The RTT measurement is crucial in providing accurate delay measurements periodically. A higher resolution performance counter of the DSP timer is used to measure the network delay between the remote and client nodes, as shown in Fig. 2.5. A real-time method for estimating the delay time in the Internet is proposed with all measurements for counters, indices, and delays denoted as Sc : sending counter
Rc : receiving counter
N : number of packets
i
: index of a sequence number, i 1, 2, . . . , .N t ip[ ] : round-trip delay measurement of the i-th packet, using counter and
i
1,2, . . . , N.
t i
p[ ]S i
c[ ]R i
c[ ] An example of message transmission based on a 20 ms sampling time is shown in Fig. 2.9(b). If the time delay is less than one sampling time, its effect on control performance is one-sample delay. Moreover, the first frame is in normal transmission.The second frame is sent 20 ms later and a packet is received at 68 ms. The corresponding RTT is 48 ms. There is no data frame received at the sampling times of
40 and 60 ms. This phenomenon is called vacant sampling (Halevi and Ray, 1998).
Two data messages (2 and 3) arrive in the same sampling period. However, only the most recent data message is used while the other data are discarded. This is referred as message rejection (Halevi and Ray, 1998; Chow and Tipsuwan, 2001). For messages 4–8, all data arrive sequentially at each sampling point, although the exact receiving time varies slightly. This occurrence is similar to delayed transmission. In summary, the delay time of NCS can be modeled using four phenomena: normal transmission, vacant sampling, message rejection, and delayed transmission. The time delay tm adopted for the adaptive Smith predictor is estimated from the measured RTT (tp) with the following rules:
(1) Normal transmission:
When the time delay is less than one sampling period, its delay effect is negligible and the measured RTT is directly adopted as tm.
(2) Vacant sampling:
When the data message is not received before occurrence of the next sampling period, the previous measured RTT added with one sampling period is recognized as the current estimation of the delay time tm.
(3) Message rejection:
When more than two data messages arrive at the same sampling period, only the most recently measured RTT is adopted as tm and all the previous measured data are discarded.
(4) Delayed transmission:
The continuously measured RTT is the estimated time delay and is directly adopted for the time delay compensation.
(a)
(b)
Fig. 2.9 The illustrative example for the time-delay estimation: (a) the architecture of the proposed RTT measurement, and (b) the four transmitted models for the RTT and the time delay estimation
2.3.2 Adaptive Smith predictor design
Fig. 2.10 shows the block diagram of the network control system with a time delay estimator. The total time of the command delay time and the feedback delay time is tp. The Smith predictor is proposed as a control structure to compensate for the delay time in NCS (Vrecko et al., 2001; Peng et al., 2004). As shown in Fig. 2.10, Gˆ sp( ) is the nominal model of the system without the delay time. The transfer function for the system with the adaptive Smith predictor is expressed as follows:
1
When Gˆp(s)Gp(s) and
t
m t
p, then the Eq. (2.3) simply becomes1
( ) ( ) ( )
( ) 1 ( ) ( )
c p t s
c p
G s G s
Y s e
R s G s G s
(2-5) Equation (2-5) shows that the transfer function when combined with the delay time and the system model transforms to two simple parts as the adaptive Smith predictor is adopted. The first part is the transfer function of the system without time delay, while the other is pure time delay. The equivalent block diagram of Eq. (2-5) is also shown in Fig. 2.6(b). Here, the system presents the same closed-loop system but only with the pure command (forward) delay time as t1. In this case, the adaptive Smith predictor is applied because the network delay is significant and the nominal value of the delay time is adopted directly from the estimated value tm from the measured RTT.
Fig. 2.10 The control structure with the adaptive Smith predictor
2.4 Experimental results
The experimental setup was implemented to verify the effect of time delay induced by the network. To apply a remote control system on an AC 400W servo motor, both the proposed adaptive Smith predictor control method and the on-line time delay estimation algorithm were implemented efficiently on the DSP micro-controller. The position control loop is located on the remote/client site. Due to the high encoder gain of 10000 P/R, coefficients of the PI controller are tuned as
0001 .
0
Kp andKi 0.00000001.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Command Actual output response Estimated Model respone
Fig. 2.11 Experimental results for system identification
The system identification result of the speed-control loop from the pseudo-random binary signal (PRBS) response for the present AC permanent magnet synchronous motor is shown in Fig. 2.11. The open positional loop is identified as
adaptive Smith predictor with the PI controller. For the client, the sampling time of the experiments was 20 ms with a square-wave command. The upper/lower commands of 30,000/15,000 pulses were provided. As the delay time increases (Fig. 2.12(d)), simulation results indicate that the control performance of the proposed adaptive Smith predictor presents the best performance compared with the PI controller and the Smith predictor. Experiments were also set up with different sites to test the proposed design.
The delay time within the NCTU Laboratory is much smaller than the sampling time.
The delay time within the NCTU Laboratory is much smaller than the sampling time.