1、附录 A Design of a High Precision Temperature Measurement System 1 Introduction Sensors are one of the most important elements used in many instrumentation circuits. They are used in many industrial applications and take a certain form of input (temperature, pressure, altitude, etc.) and convert it in
2、to readings that can be interpreted. Many types of sensors are nonlinear in nature from which a linear output is desired. There are many different sensors for temperature measurement and thermocouples are the most commonly used . They are preferred in industrial applications due to their low cost, w
3、ide operation range, fast response time and accurate when their peculiarities are understood. Thermocouples have also outputs nonlinearly related to temperature. Therefore, sensor modeling and linearization techniques are necessary. To solve the linearization problem of a sensor, there are generally
4、 two methods proposed. The first one requires nonlinear analog circuit and the second uses numerical methods that are computed by microprocessor or computer . Analog circuits are frequently used for improving the linearity of the sensor characteristics, which implies additional analog hardware and t
5、ypical problems associated to analog circuits such as temperature drift, gain and offset error. Using the second method, sensor nonlinearities can be compensated by means of arithmetic operations, if an accurate sensor model is available (direct computation of the polynomials), or use of a multidime
6、nsional look-up table. Direct computation of the polynomial method is more accurate but takes a long time for computation, while the look-up table method, though faster, is not very accurate .In recent years, application of ANNs has emerged as a promising area of research in the field of instrumenta
7、tion and measurement . It provides a neurocomputing approach for solving complex problems especially in nonlinear system modeling which the network itself is a nonlinear system. This is extremely useful when the area of interest is absolutely nonlinear including the experimental data that is used fo
8、r training. One of the most powerful uses of ANNs is in function approximation (curve fitting) . Interpolation based on ANN provides lower interpolation errors when compared with conventional numerical interpolation .In the work we present here, high precision temperature measurement system based on
9、 ANN approach is proposed. The calibrating data is obtained by Wavetek 9100 calibration unit that is necessary for the training and the testing phases of the ANN. The hardware and software parts of the system are integrated in a VI used for system operation and calibration. The ANN is matched to the
10、 calibrating data by providing a desired final error. The mean square error between calibration and the ANN modeled data is minimized in terms of the structure, number of layers, and number of neurons by the developed software. 2 System Hardware A thermocouple generates a voltage proportional to the
11、 measurement junction temperature at mV levels while the cold junction temperature is constant. In order to make an accurate measurement the cold junction temperature must be known. Figure 1(a) shows the block diagram of the temperature measurement system designed via an ANN in the operation phase.
12、It consists of a thermocouple (type E) exposed to a desired temperature, including signal conditioning circuit with 16-bit analog to digital converter (ADC) and Input/Output interface card interfacing with a computer. The designed signal conditioning circuit has a programmable gain instrumentation a
13、mplifier (PGA204BP) with the gain of 1, 10, 100 and 1000, a 16-bit ADC (AD976A), an AD595 monolithic thermocouple amplifier with cold junction compensation which is configured as a stand-alone Celsius thermometer and a 4 channel analog multiplexer (ADG529A) which select the thermocouple or output of
14、 Celsius thermometer. The AD976A is a high speed, low power 16-bit A/D converter that operates from a single 5V supply. This part contains a successive approximation, switched capacitor ADC, an internal 2.5V reference and a high speed parallel interface. Accuracy of the system depends directly on st
15、ep size of ADC. With a 10V inputs, one LSB of AD976A is 305V. When AD595 is used as a Celsius thermometer, the thermocouple is omitted, and the differential inputs are shunted together to common. In this mode, AD595 generate a voltage with a scale factor of 10mV/C and its output is used for cold jun
16、ction temperature data that the written software is used. Some important characteristics of the AD595 are: operation temperature range -55 to 125C； stability vs. temperature: 0.05C/C and sensitivity: 10mV/C. Output signal of PGA204BP is digitized by AD976A which its output is connected to the I/O in
17、terface card and transferred to a personal computer where data reduction and optimization are implemented.Fig. 1. Measurement system block diagram: (a) operational phase, (b) calibration phase To establish the ANNs weights and biases, during the calibrating phase (ANN training phase), Wavetek 9100 c
18、alibration unit, with the accuracy of 0.006%+4.16V in the range of 000.000mV to 320.000mV, is connected to the terminals of analog multiplexer to generate tabled thermocouple voltages as shown in Figure 1 (b). This voltage is used as the input of the ANN, and thermocouple temperature without cold ju
19、nction compensation is the output of the ANN. In the operation phase (Figure 1(a), in order to make the cold junction compensation, data taken from Celsius thermometer output is used. The output value of ANN is shifted by the environment temperature that is obtained by Celsius thermometer. Then this
20、 value is displayed on the VI as the thermocouple temperature.The developed VI is used to acquire the data for ANN training phase and to show the calculated temperature in the operation phase. Figure 2 shows the front panel of the VI. The main features associated with this instrument are: display of
21、 the measured temperature and corresponding output voltage from conditioning circuit for collectingthe data in the calibrating phase and actual temperature with cold junction compensation in the operation phase. The system is controlled by the software written in both operation and calibration phase
22、s. 3 Artificial Neural Network ANNs are based on the mechanism of the biologically inspired brain model. ANNs are feed-forward networks and universal approximators. They are trained and learned through experience not from programming. They are formed by interconnections of simple processing elements
23、, or neurons with adjustable weights, which constitute the neural structure and are organized in layers. Each artificial neuron has weighted inputs, summation and activation functions and output. The behaviour of the overall ANN depends upon the operations mentioned on the artificial neurons, the le
24、arning rule and the architecture of the network. During the training (learning), the weights between the neurons are adjusted according to some criterion (The mean square error between the target output and the measured value for all the training set falls below a predetermined threshold) or the max
25、imum allowable number of epochs is reached. Although the training is a time consuming process, it can be done beforehand, offline. The trained neural network is then tested using data was previously unseen during training. MLPs are the simplest and most commonly used neural network architectures . T
26、hey consists of input, output and one or more hidden layers with a predefined number of neurons. The neurons in the input layer only act as buffers for distributing the input signals xi to neurons in the hidden layer. Each neuron j in the hidden layer sums up its input signals xi, after weighting th
27、em with the strengths of the respective connections wji from the input layer and computes its output yj as a function f of the sum, namelywhere f is one of the activation functions used in ANN architecture.Training a neural network consists of adjusting the network weights using different learning a
28、lgorithms. A learning algorithm gives wji(t) in the weight of a connection between neurons i and j at time t. The weights are then updated according to the following formula:There are many available learning algorithms in the literature . The algorithms used to train ANNs in this study are Levenberg
29、Marquardt (LM) , BroydenFletcherGoldfarbShanno (BFGS) , Bayesian Regularization (BR) , Conjugate gradient backpropagation with Fletcher-Reeves updates (CGF) , and Resilient back-propagation (RP) algorithms. Neural LinearizationIn this paper, the multilayered perceptron (MLP) neural network architect
30、ure is used as a neural linearizer. The proposed technique involves an ANN to evaluate the thermocouple temperature (ANN output) when thermocouple output voltage is given as input. Training the ANN with the use of mentioned learning algorithm to calculate the temperature involves presenting it with
31、different sets of input values and corresponding measured values. Differences between the target output and the actual output of the ANN are evaluated by the learning algorithm to adapt the weights using equations (1) and (2). The experimental data taken from thermocouple data sheets are used in thi
32、s investigation. These data sheets are prepared for a particular junction temperature (usually 0C). The ANN is trained with 80 thermocouple temperatures that is uniformly distributed between -200 and 1000C which is obtained in the calibration phase. However the performance of the final network with
33、the training set is not an unbiased estimate of its performance on the universe of possible inputs, and an independent test set is required to evaluate the network performance after training. Therefore, the other data set of 20 thermocouple temperatures that is uniformly distributed between -200 and 1000C, is used in the test process.The input and output
