嵌入式可编程逻辑控制器算法中英文翻译.doc

外文文献翻译 Advanced control algorithms embedded in a programmableAbstractThis paper presents an innovative self-tuning nonlinear controller ASPECT (advanced control algorithms for programmable logic controllers). It is intended for the control of highly nonlinear processes whose properties change radically over its range of operation, and includes three advanced control algorithms. It is designed using the concepts of agent-based systems, applied with the aim of automating some of the configuration tasks. The process is represented by a set of low-order local linear models whose parameters are identified using an online learning procedure. This procedure combines model identification with pre- and post identification steps to provide reliable operation. The controller monitors and evaluates the control performance of the closed-loop system. The controller was implemented on a programmable logic controller (PLC). The performance is illustrated on a field test application for control of pressure on a hydraulic valves 2005 Elsevier Ltd. All rights reserved.Keywords: Control engineering; Fuzzy modelling; Industrial control; Model-based control; Nonlinear control; Programmable logic controllers; Self tuning regulators1. IntroductionModern control theory offers many control methods to achieve more efficient control of nonlinear processes than provided by conventional linear methods, taking advantage of more accurate process models (Bequette, 1991; Henson & Seborg, 1997; Murray-Smith & Johansen, 1997). Surveys (Takatsu, Itoh, & Araki,1998; Seborg, 1999) indicate that while there is a considerable and growing market for advanced controllers, relatively few vendors offer turn-key products. Excellent results of advanced control concepts, based on fuzzy parameter scheduling (Tan, Hang, & Chai,1997; Babus ka, Oosterhoff, Oudshoorn, & Bruijn,2002), multiple-model control (Dougherty & Cooper,2003; Gundala, Hoo, & Piovoso, 2000), and adaptive control (Henson & Seborg, 1994; Ha¨ gland & A strom,2000), have been reported in the literature. However, there are several restrictions for applying these methods in industrialapplications, as summarized below:（1）Because of the diversity of real-life problems, a single nonlinear control method has a relatively narrow field of application. Therefore, more flexible methods or a toolbox of methods are required in industry.（2）New methods are usually not available in a ready-to use industrial form. Custom design requires considerable effort, time and money.（3）The hardware requirements are relatively high, due to the complexity of implementation and computational demands.（4）The complexity of tuning (Babus ka et al., 2002) and maintenance makes the methods unattractive to nonspecialised engineers.（5）The reliability of nonlinear modelling is often in question.（6）Many nonlinear processes can be controlled using the well-known and industrially proven PID controller. A considerable direct performance increase (financial gain) is demanded when replacing a conventional control system with an advanced one. The maintenance costs of an inadequate conventional control solution may be less obvious. The aim of this work is to design an advanced controller that addresses some of the aforementioned problems by using the concepts of agent-based systems (ABS) (Wooldridge & Jennings, 1995). The main purpose is to simplify controller configuration by partial automation of the commissioning procedure, which is typically performed by the control engineer. ABS solve difficult problems by assigning tasks to networked software agents. The software agents are characterized by properties such as autonomy (operation without direct intervention of humans), social ability (interaction with other agents), reactivity (perception and response to the environment), pro-activeness (goal-directed behaviour,taking the initiative), etc. This work does not address issues of ABS theory, but rather the application of the basic concepts of ABS to the field of process systems engineering. In this context, a number of limits have to be considered. For example: initiative is restricted, a high degree of reliability and predictability is demanded, insight into the problem domain is limited to the sensor readings, specific hardware platforms are used, etc. The ASPECT controller is an efficient and user-friendly engineering tool for implementation of parameter-scheduling control in the process industry. The commissioning of the controller is simplified by automatic experimentation and tuning. A distinguishing feature of the controller is that the algorithms are adapted for implementation on PLC or open controller Industrial hardware platforms. The controller parameters are automatically tuned from a nonlinear process model. The model is obtained from operating process signals by experimental modelling,using a novel online learning procedure. This procedure is based on model identification using the local learning approach (Murray-Smith & Johansen,1997, p. 188). The measurement data are processed batch-wise. Additional steps are performed before and after identification in order to improve the reliability of modelling, compared to adaptive methods that use recursive identification continuously (Ha¨ gland & A strom,2000).The nonlinear model comprises a set of local lowered linear models, each of which is valid over a specified operating region. The active local model(s) is selected using a configured scheduling variable. The controller is specifically designed for single-input, single output processes that may include a measured disturbance used for feed-forward. Additionally, the application range of the controller depends on the selected control algorithm. A modular structure of the controller permits use of a range of control algorithms that are most suitable for different processes. The controller monitors the resulting control performance and reacts to detected irregularities. The controller comprises the run-time module (RTM) and the configuration tool (CT). The RTM runs on a PLC, performing all the main functionality of real-time control, online learning and control performance monitoring. The CT, used on a personal computer (PC) during the initial configuration phase, simplifies the configuration procedure by providing guidance and default parameter values. The outline of the paper is as follows: Section 2 presents an overview of the RTM structure and describes its most important modules; Section 3 gives a brief description of the CT; and finally, Section 4 describes the application of the controller to a pilot plant where it is used for control of the pressure difference on a hydraulic valve in a valve test apparatus.2. Run-Time ModuleThe RTM of the ASPECT controller comprises a set of modules, linked in the form of a multi-agent system. Fig. 1 shows an overview of the RTM and its main modules: the signal pre-processing agent (SPA), the online learning agent (OLA), the model information agent (MIA), the control algorithm agent (CAA), the control performance monitor (CPM), and the operation supervisor (OS).2.1. Multi-faceted model (MFM)The ASPECT controller is based on the multi-faceted model concept proposed by Stephanopoulus, Henning, and Leone (1990) and incorporates several model forms required by the CAA and the OLA. Specifically, the MFM includes a set of local first- and second-order discrete-time linear models with time delay and offset, which are specified by a given scheduling variable s(k).The model equation of first order local models is (1)while the model equation of second order models is（2）where k is the discrete time index, j is the number of the local model, y(k) is the process output signal, u(k) is the process input signal, v(k) is the optional measured disturbance signal (MD), du is the delay in the model branch from u to y, dv is the delay in the model branch from v to y, and ai,j, bi,j, ci,j and rj are the parameters of the jut local model. The set of local models can be interpreted as a TakagiSurgeon fuzzy model, which in the case of a second order model can be expressed in the following form: （3）Where bj( k) is the value of the membership function of the jut local model at the current value of the scheduling variable s(k). Normalized triangular membership functions are used, as illustrated in Fig. 2.The scheduling variable s(k) is calculated using coefficients kr, ky, ku, and kv, using the weighted sum (4)The coefficients are configured by the engineer according to the nature of the process nonlinearity.2.2. Online Learning Agent (OLA)The OLA scans the buffer of recent real-time signals, prepared by the SPA, and estimates the parameters of the local models that are excited by the signals. The most recently derived parameters are submitted to the MIA only when they pass the verification test and are proved to be better than the existing set. The OLA is invoked upon demand from the OS or autonomously, when an interval of the process signals with sufficient excitation is available for processing. It processes the signals batch-wise and using the local learning approach. An advantage of the batch-wise concept is that the decision on whether to adapt the model is not performed in real-time but following a delay that allows for inspection of the identification result before it is applied. Thus, better means for control over data selection is provided. A problem of distribution of the computation time required for identification appears with batch-wise processing of data (opposed to the online recursive processing that is typically used in adaptive controllers).This problem is resolved using a multi-tasking operation system. Since the OLA typically requires considerably more computation than the real-time control algorithm, it runs in the background as a low-priority task. The following procedure, illustrated in Fig. 3, is executed when the OLA is invoked.2.2.1. Copy signal bufferThe buffer of the real-time signals is maintained by the SPA. When the OLA is invoked, the relevant section of the buffer is copied for further processing.2.2.2. Excitation checkA quick excitation check is performed at the start, so that processing of the signals is performed only when they contain excitation. If the standard deviations of the signals r(k), y(k), u(k), and v(k) in the active buffer are below their thresholds, the execution is cancelled.2.2.3. Copy active MFM from MIAThe online learning procedure always compares the newly identified local models with the previous set of parameters. Therefore, the active MFM is copied from the MIA where it is stored. A default set of model parameters is used for the local models that have not yet been identified (see Section 2.3).2.2.4. Select local modelsA local model is selected if the sum of its membership functions bj(k) over the active buffer normalized by the active buffer length exceeds a given threshold. Only the selected local models are included in further processing.2.2.5. IdentificationThe local model parameters are identified using the fuzzy instrumental variables (FIV) identification method developed by Blaz ic et al. (2003). It is an extension of the linear instrumental variables identification procedure (Ljung, 1987) for the specified MFM, based on the local learning approach (Murray-Smith & Johansen, 1997). The local learning approach is based on the assumption that the parameters of all local models will not be estimated in a single regression operation. Compared to the global approach it is less prone to the problems of ill-conditioning and local minima. This method is well suited to the needs of industrial operation (intuitiveness, gradual building of the nonlinear model, modest computational demands). It enables inventory of the local models that are not estimated properly due to insufficient excitation. It is efficient and reliable in early configuration stages, when all local models have not been estimated yet. On the other hand, the convergence in the vicinity of the optimum is slow. Therefore, it is likely to yield a worse model fit than methods employing nonlinear optimisation.The following briefly describes the procedure. Model identification is performed for each selected local model (denoted by the index j) separately. The initial estimated parameter vector is copied from the active MFM, and the covariance matrix is initialized to 105 I (identity matrix). The FLS (fuzzy least-squares) estimates, and are obtained using weighted least-squares identification, with bj(k) used for weighting. The calculation is performed recursively to avoid matrix inversion. The FIV (fuzzy instrumental variables) estimates, and are calculated using weighted instrumental variables identification. In order to prevent result degradation by noise, adead zone is used in each step of FIV and FLS recursiveEstimation. The vector of parameters and the covariance matrix are updated only if the absolute weighted difference between the process output and its prediction is above the configured noise threshold.In case of lack of excitation in the branch from u to y or in the model branch from v to y (or when measured disturbance is not present at all), variants of the method with reduced parameter estimate vectors are used.2.2.6. Verification/validationThis step is performed by comparing the simulation output of a selected local model with the actual process output in the proximity of the local model position. The normalized sum of mean square errors (MSEj) is calculated. The proximity is defined by the membership functions bj. For each of the selected local models, this step is carried out with three sets of model parameters: and the set with the lowest MSEj is selected. Then, global verification is performed by comparing the simulation output of the fuzzy model including the selected set with the actual process output. The normalized sum of mean square errors (MSEG) is calculated. If the global verification result is improved compared to the initial fuzzy model, the selected set is sent to the MIA as the result of online learning, otherwise the original set remains in use. For each processed local model, the MIA receives the MSEj, which serves as a confidence index, and a flag indicating whether the model is new or not.2.2.7. Model structure estimationTwo model structure estimation units are also included in the OLA. The dead-time unit (DTU) estimates the process time delay. The membership function unit (MFU) suggests whether a new local model should be inserted. It estimates an additional local model in the middle of the interval between the two neighboring local models that are the most excited. The model is submitted to the MIA if the global validation of the resulting fuzzy model is sufficiently improved, compared to the original fuzzy model.2.3. Model Information Agent (