Received April 5, 2015, accepted April 21, 2015, date of publication April 28, 2015, date of current version May 7, 2015. Digital Object Identifier 10.1109/ACCESS.2015.2427291
Optimal Design of SAW Gas Sensing Device by Using Improved Adaptive Neuro-Fuzzy Inference System JINN-TSONG TSAI1 , KAI-YU CHIU2 , AND JYH-HORNG CHOU2,3,4 , (Fellow, IEEE)
1 Department
of Computer Science, National Pingtung University, Pingtung 900, Taiwan of Electrical Engineering, National Kaohsiung First University of Science and Technology, and with the Assembly Department of Advanced Semiconductor Engineering Inc, Kaohsiung 824, Taiwan 3 Department of Electrical Engineering, National Kaohsiung University of Applied Sciences, Kaohsiung 807, Taiwan 4 Department of Healthcare Administration and Medical Informatics, Kaohsiung Medical University, Kaohsiung 807, Taiwan 2 Institute
Corresponding author: J.-H. Chou (
[email protected]) This work was supported by the National Science Council of Taiwan, under Grant NSC 102-2221-E-151-021-MY3, Grant NSC 102-2221-E-153-002, and Grant MOST 103-2221-E-153-004-MY2.
ABSTRACT A Taguchi-based-genetic algorithm (TBGA) is used in an adaptive neuro-fuzzy inference system (ANFIS) to optimize design parameters for surface acoustic wave (SAW) gas sensors. The Taguchi method is used to reduce the number of experiments and collect performance data for an SAW gas sensor. The TBGA has two optimization roles. In the ANFIS, the TBGA selects appropriate membership functions and optimizes both the premise and the consequent parameters by minimizing the performance criterion of the root mean squared error. Another role of the TBGA is optimizing design parameters for an SAW gas sensor. Simulated experimental application of the proposed TBGA-based ANFIS approach showed that, in terms of both resonant frequency shift and precision performance, this systematic design approach obtains far superior results compared with the conventional trial-and-error design methods and other Taguchi-based design methods. INDEX TERMS Adaptive network fuzzy inference system, surface acoustic wave (SAW) gas sensors, Taguchi-genetic algorithm. I. INTRODUCTION
Various applications of surface acoustic wave (SAW) devices have recently received considerable attention due to their sensitivity to small changes in surface composition, such as increased mass resulting from surface adsorption [1]–[5]. The SAW devices comprise a thin ST-cut quartz slice sandwiched between metal electrodes and then coated with sensitive membranes. Interdigital transducers (IDTs) are fabricated over piezoelectric substrates to can excite and receive acoustic waves. Anisimkin and Verona [6] reported that, compared to quartz crystal microbalance sensors, SAWs provide superior resolution because of their higher operating frequencies (100 to 200 MHz). The SAW delay-line devices have attracted the interest of researchers because of their rapid response, small physical size, high sensitivity, low cost, and simple fabrication. The output response of SAW sensors is a linear function of the input signal, which corresponds to the deposited mass [7]. For example, in Anisimkin and Verona [8], a calibration curve for sensor response versus mass concentration was successfully used to optimize mass sensitivity.
420
The mass effect influences the resonant frequency of a SAW sensor. As the resonant frequency shift of a SAW device increases, mass variation decreases. For SAWs, this property is a highly sensitive design parameter. Other design parameters include the electrode thickness, the number of finger pairs, the electrode overlap, the separation distance of the two IDTs on the substrate, and the dimensions of the ST-quartz substrate. In Wu and Chen [7], the use of Taguchi method to optimize design parameters for SAW sensors obtained a 24.55 (%) improvement in resonant frequency shift compared to the conventional trial-and-error design method [7]. However, the effectiveness of the Taguchi method is highly dependent on the selected ranges of SAW parameters. Thus, if only Taguchi method is used for parameter optimization, designers must still perform a trial-and-error procedure to find the optimal ranges of SAW parameters. To solve this problem, this study developed a systematic method for optimizing SAW parameters in terms of resonant frequency shift and sensing precision. A well-established approach to solving nonlinear mapping problems is the adaptive network-based fuzzy inference
2169-3536 2015 IEEE. Translations and content mining are permitted for academic research only. Personal use is also permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
VOLUME 3, 2015
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
system (ANFIS) [9]–[12]. By using a hybrid learning algorithm that combines gradient method with least squares estimate for parameter identification, the ANFIS can perform input-output mapping based on both human knowledge (in the form of fuzzy if-then rules) and the stipulated inputoutput data pairs [10]. However, despite the successful use of ANFIS to solve these nonlinear problems, a remaining issue is identifying the most suitable membership functions while simultaneously optimizing the premise and consequent parameters. Therefore, our research team proposed the use of Taguchi-based genetic algorithm (TBGA) [13], [14] to address the above limitations of the conventional ANFIS and developed a TBGA-based ANFIS approach to solving engineering optimization problems [15]–[19]. A search of the existing literature [7] and industry-related publications (http://www.mirdc.org.tw) reveals only two methods of optimizing SAW parameters in terms of maximum resonant frequency shift: the conventional trial-and-error design method and the Taguchi-based design method. From the industrial application perspective, the objective was to develop a new and effective method of optimizing design parameters for SAW gas-sensing devices. Thus, in the study, the TBGA-based ANFIS is proposed to design parameters for an SAW sensor. The solution proposed in this study is TBGA-based ANFIS. To optimize design parameters for SAW sensors, the proposed approach combines data collection, modeling, and optimization methods [20], including Taguchi method [21], [22], ANFIS, and TBGA. The two roles of the TBGA are creating the best system model and optimizing the design parameters. After first defining the design parameters and problems, Taguchi method is used for experimentation and data collection in performance evaluations of the SAW sensor. Next, a system model is constructed by using ANFIS and TBGA to select appropriate membership functions and to optimize parameters in the premise and consequent parts. Finally, TBGA is used to optimize SAW sensor design parameters in terms of resonant frequency shift and sensing precision. The combined approach to optimizing design parameters for a SAW sensor is then tested and discussed. Simulated experimental comparisons of sensor performance confirm that the proposed systematic method indeed obtains superior design parameters compared to approaches recently reported in the literature. This paper is organized as follows. Section II defines the research issue. The proposed approaches are introduced in Section III. Section IV presents the case study, results, and discussion. Finally, Section V concludes the study.
FIGURE 1. Schematic of surface acoustic wave sensor.
Rayleigh waves. The ST-cut quartz crystal is cut at a 42.75◦ angle to the Y-axis as described in [7]. The operating resonant frequency of the device is highly dependent on the IDT period. The efficiency of the transducer is highest when the acoustic wavelength of the sensor matches the IDT period. B. SAW MASS EFFECT
The SAW sensor is highly sensitive to mass changes on the surface [1]. The original resonant frequency decreases when a small mass is deposited on the surface of the ST-cut substrate, and the frequency decrease is proportional to the deposited mass. Fig. 2 shows that a uniform IDT with period λ is implemented on the SAW substrate where λ is the wavelength. In Fig. 2, w is the electrode overlap, and L is the separation distance. The distance between adjacent electrodes is λ/2. Both finger electrode spacing d and electrode width are λ/4. The resonant frequency f0 of the SAW device is [7] fo =
vo , λ
(2.1)
where the Rayleigh velocity ν0 is 3158 (m/s).
II. PROBLEM DESCRIPTION A. SAW GAS SENSOR
FIGURE 2. Structure of interdigital transducer.
The SAW delay-line type sensors have numerous mass-sensing applications. Fig. 1 shows the SAW sensor, which consists of two IDTs fabricated on an ST-cut quartz piezoelectric substrate that excites and receives
Mass sensitivity is a key consideration when designing a SAW sensor capable of precise measurement. Gas-sensing sensitivity S is given by S = dR/dn where n is the gas concentration (ppm) and R is the sensor response. As described
VOLUME 3, 2015
421
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
in [7], response R is defined as R=
1v 1f 1m = = K f fo , v fo A
(2.2)
where ν is the acoustic-wave phase velocity, Kf is the mass sensitivity constant, and A is the area of the film. The resonant frequency shift 1f is obtained by 1f = fo − fm ,
(2.3)
where fm is the resonant frequency of the SAW with a deposited mass and fo is the original resonant frequency. C. DESIGN OBJECTIVE
The design objective is to maximize the resonant frequency shift 1f such that the mass sensitivity of the SAW sensor is increased. As in Wu and Chen [7], the following five SAW parameters are selected for optimization such that resonant frequency shift 1f is maximized: (i) parameter EF: number of electrode finger pairs, (ii) parameter SD: the separation distance L (mm) between two IDTs on the substrate, (iii) parameter EO: electrode overlap w (mm), (iv) parameter ET: electrode thickness te (mm), and (v) parameter DI: dimensions (mm) of the ST-cut quartz substrate.
procedure composes a forward pass and a backward pass. In a forward pass, a training set of inputs is offered to the ANFIS and neuron outputs are computed by the layer-bylayer basis. The consequent parameters are determined by the least squares algorithm. The parameters are changed in the learning process according to the membership functions. The parameter computation and adjustment are updated by a gradient vector. For a given set of parameters, how well the fuzzy inference system models the input/output data can be measured by the gradient vector. When the gradient vector is achieved, any of optimization routines can be used to amend the parameters so as to reduce error measure, which is the sum of the squared difference between desired and actual outputs. The ANFIS uses the typical membership functions, for example, triangular, trapezoidal, Gaussian, and bell types. The precondition parameters of membership function and the consequent parameters are determined by training via hybrid learning algorithm. However, how to determine the most suitable membership functions and simultaneously find the optimal premise and consequent parameters is a research challenge. Therefore, the optimization method should be included in the ANFIS for deciding the appropriate membership functions and discovering optimal both premise and consequent parameters.
III. TBGA-BASED ANFIS APPROACH
The TBGA-based ANFIS is used to produce a system model and to optimize parameters for a SAW sensor design. A system model is created by using ANFIS and the TBGA to select appropriate membership functions and to optimize parameters in the premise and consequent parts. Based on the system model of a SAW sensor, the TBGA is used to optimize the design parameters such that the resonant frequency shift is maximized and the sensing precision is increased. The steps of the TBGA-based ANFIS approach are described below. A. SYSTEM MODEL CREATED USING ANFIS AND TBGA 1) ADAPTIVE NEURO-FUZZY INFERENCE SYSTEM
The ANFIS was proposed by Jang [9], and the Sugeno fuzzy model was utilized for a systematic method to create fuzzy rules from a given set of input-output data. This ANFIS structural design includes a five-layer feed-forward neural network. Layer 1 is the fuzzification layer that takes the inputs and performs fuzzification to decide the membership degrees for each input based on the given fuzzy membership function. Layer 2 is the rule layer, in which a rule node collects inputs from the respective fuzzification nodes and computes the firing strength of the rule. Layer 3 is called the normalized layer, which evaluates the ratio of the firing strength of a given rule to the sum of firing strengths of all rules. Layer 4 is the defuzzification layer which also receives initial inputs and gives the consequent parameters of the rule. Layer 5 is a single node that computes the overall output as the sum of all incoming signals. Hybrid learning algorithm, which integrates least squares and gradient descent algorithm, is employed for ANFIS training. In ANFIS training algorithm, each epoch on the hybrid 422
2) ANFIS STRUCTURE FOR A SAW SENSOR
The proposed approach fuses ANFIS and TBGA to obtain a system model of a SAW sensor, which is then used to optimize the design parameters for a SAW sensor. As described in Section II-C, the five parameters that have the largest impact on the SAW sensor performance are EF, SD, EO, ET, and DI. Therefore, the input variables for the ANFIS are design parameters EF, SD, EO, ET, and DI, and the output variable is resonant frequency shift 1f . A representative set of n fuzzy if-then rules can be stated as Rl : IF EF is Ag and SD is Bh and EO is Ci and ET is Dj and DI is Ek , THEN yl = pl × EF + ql × SD + rl × EO + sl × ET + tl × DI + ul ,
(3.1)
where = 1, 2, . . . , n) denotes the l-th implication. The terms Ag , Bh , Ci , Dj , and Ek , (g, h, i, j, k = 1, 2 and n = 32 (i.e., 25 )) are the linguistic terms of the precondition part with membership functions µAg (EF), µBh (SD), µCi (EO), µDj (ET), and µEk (DI), respectively. The triangular, bell-shaped, and Gaussian membership functions are used in the ANFIS to establish the SAW system model in the study. The yl is the output value. The consequent parameters are pl , ql , rl , sl , tl, and ul . The ANFIS output inferred from Eq. (3.1) is Rl (l
y=
n X
¯ l ( pl × EF + ql × SD + rl × EO + sl × ET W
l=1
+ tl × DI + ul )
(3.2) VOLUME 3, 2015
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
where y denotes the inferred ANFIS output, and n P ¯ l = Wl / Wi, in which Wl = µAg (EF)×µBh (SD)× W i=1
µCi (EO)×µDj (ET)×µEk (DI) (l = 1, 2, . . . , n and g, h, i, j, k = 1, 2). Let the precondition parameters of µAg (EF), µBh (SD), µCi (EO), µDj (ET), and µEk (DI) in the premise part be {aAg , bAg }, {aBh , bBh }, {aCi , bCi }, {aDj , bDj }, and {aEk , bEk }, respectively. For instance, if a Gaussian membership function is selected for the µAg (EF), then aAg and bAg are the center and the width of the Gaussian membership function, respectively. For performance assessment of the ANFIS model, observed and predicated values are compared in terms of root mean squared error (RMSE) [9]. Therefore, after all training processes are completed, both the precondition and consequent parameters can be obtained by minimizing the RMSE criterion as shown below: M X (Rm − ym )2 J= α
"
# 12 ,
(3.3)
m=1
where M is the number of training data items, Rm is the actual 1f value, and ym is the predicted 1f value. In Eq. (3.3), performance criterion J is function of the parameter set {aAg , bAg , aBh , bBh , aCi , bCi , aDj , bDj , aEk , bEk , pl , ql , rl , sl , tl , ul }, (l = 1, 2, . . . , n and g, h, i, j, k = 1, 2), that is, J = f (aAg , bAg , aBh , bBh , aCi , bCi , aDj , bDj , aEk , bEk , pl , ql , rl , sl , tl , ul ),
Step 5: Choose a two-level OA Lγ (2γ −1 ) for the matrix experiments where γ represents the experiment number and where γ − 1 is the column number of an OA. Step 6: Simultaneously select two random chromosomes for use in matrix experiments. Step 7: Use Eq. (3.3) to compute the fitness value of each experiment in the OA. Step 8: Compute the effects of the factors and find the better chromosomes. Step 9: Based on the results of Step 8, find the best chromosome. Step 10: Repeat Steps 6-9 until the expected number meets (1/4) × Ps × pc . Step 11: Obtain the population by Taguchi method. Step 12: Use mutation rate pm to perform a mutation operation. Step 13: Obtain the offspring population. Step 14: Sort the fitness values in increasing order among the offspring and parent populations. Step 15: Choose the better Ps chromosomes as the parents for the next generation. Step 16: If the current generation number C_no meets the default generation number G_no, i.e., the stopping criterion, go to Step 17. Otherwise, set C_no = C_no + 1 and repeat Steps 3-16. Step 17: Evaluate and record the RMSE, give other membership function types, and repeat Steps 2-16. If the minimum RMSE is obtained, go to Step 18. Step 18: Show the best chromosome and its fitness value.
(3.4) B. PARAMETER DESIGN USING TBGA
where J is a nonlinear function of parameter variables. The problem considered in this study is how to minimize J . The detailed steps in optimizing parameters of the premise and consequent parts of membership functions are shown below. Step 1: Define input parameters and outputs. Input: population size Ps , crossover rate pc , mutation rate pm , and number of generations G_no. Output: the design parameter set {aAg , bAg , aBh , bBh , aCi , bCi , aDj , bDj , aEk , bEk , pl , ql , rl , sl , tl , ul } obtained by minimizing the value of J in Eq. (3.3). Step 2: Generate initial population. Compute fitness values for the initial population by using J in Eq. (3.3) as the fitness function of the TBGA. Randomly reproduce the initial population for chromosomes in the form {aAg , bAg , aBh , bBh , aCi , bCi , aDj , bDj , aEk , bEk , pl , ql , rl , sl , tl , ul } for the problem. Membership function types are given beforehand. Set current generation number C_no = 1. Step 3: Use roulette wheel method to perform a selection operation [23]. Step 4: Use crossover rate pc in a crossover operation [13], [14]. VOLUME 3, 2015
The TBGA was proposed by Tsai et al. [13] to solve global optimization problems which have continuous variables. The TBGA integrates a conventional genetic algorithm (CGA) [23], [24] and the Taguchi method. In the TBGA, the use of the Taguchi method is set between the crossover and mutation of a CGA. By applying two major tools, signal-to-noise ratio (SNR) and orthogonal array (OA), the ability of systematic reasoning is integrated into the crossovers to systematically choose the better genes to accomplish crossover, and hence improve the GA. As a result, the TBGA can be statistically sound, more robust, and quickly convergent. The detailed description of the Taguchi method can be disclosed in the books given by Taguchi et al. [21]. In addition, the detailed TBGA can be disclosed in the works of Tsai et al. [13], [14]. In this work, the TBGA is used to optimize the five design parameters of the SAW sensor EF, SD, EO, ET, and DI. The Taguchi method uses an OA to study a large number of decision variables with a small number of experiments. The best combinations of decision variables are determined by the OA and by the SNR. The key concept of the Taguchi method is to maximize the SNRs, which are used as a performance measure, by using the OA to run a partial set of experiments. This work uses a two-level OA with Q variables, where Q is 423
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
TABLE 1. Experimental results of 1f as a training data set.
the number of design variables, each of which has two levels. When establishing an OA with two levels of Q variables, Lγ (2γ −1 ) represents γ − 1 columns, γ separate experiments correspond to γ rows, γ = 2k where k is a positive integer (k > 1) and Q ≤ γ −1. If Q < γ −1, only the front Q columns are used, and the other γ − 1 − Q columns are ignored. The SNR (ηi ), which is measured in decibels, refers to the mean-square-deviation in the objective function. According ! τ 1 P 1 to the Taguchi definition, ηi = −10 log τ for the 2 j=1
zij
larger-the-better characteristic, and ηi = −10 log
1 τ
τ P j=1
! z2ij
for the smaller-the-better characteristic where i = 1, 2, . . . , γ and where γ is the number of experiments for a two-level OA Lγ (2γ −1 ). Let zij denote the experimental output value of the ith row of a two-level OA Lγ (2γ −1 ) where j = 1, 2, . . . , τ and where τ is the number of experiments for the ith row of the OA. When building the response table, the effects of different variables can be defined as follows: Eβl = sum of ηi for variable β at level l,
(3.5)
where i is the experiment number, β is the variable name, and l is the level number. 424
IV. PRACTICAL EXAMPLE AND IMPLEMENTATION
A practical application of the method in engineering design is demonstrated in the problem of optimizing the design parameters for a SAW gas-sensing device. A 20 MHz SAW is chosen for the following example. Consider a SAW under the same conditions given in [7]: the SAW wavelength is 156 (µm), the ST-quartz substrate thickness is 1 (mm), and the IDT finger width is 39 (µm), additionally, CO2 gas (volume, 1 × 10−8 (l); mass, 1.7864 × 10−8 (g)) is deposited on the SAW device surface. First, a L18 (21 ×37 ) OA of Taguchi method is used to perform experiments and to accumulate data indicating the quality of the SAW sensor. The 1f performance of a SAW sensor is governed by five design parameters: EF, SD, EO, ET, and DI. For the five considered design parameters (EF, SD, EO, ET, and DI) and one experimental output (1f ), the specified control factors and their respective level settings to meet the specified requirements of engineering were selected based on the specifications of the original design (EF: 20 (pairs), SD: 4 (mm), EO: 4.8 (mm), ET: 1.4 × 10−5 (mm), and DI: 9 × 7 × 1 (mm)). Table 1 shows the simulated experimental results obtained for the L18 (21 × 37 ) OA table in [7], which are used to model the SAW ANFIS system in this study. Additionally, the test data, which also meet the specified
VOLUME 3, 2015
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
TABLE 2. Experimental results of 1f as a testing data set.
requirements of engineering and are shown in Table 2, were generated by simulated experiments for validating the SAW system model obtained by the ANFIS. The SAW system model is then established by using the ANFIS with TBGA to select the appropriate membership functions and to optimize both premise and consequent parameters by minimizing RMSE, which is used as a performance criterion. Fig. 3 shows the five inputs and one output used for TBGA training in the ANFIS architecture. Table 3 shows the RMSE for different membership functions used to build the SAW system ANFIS model. These data show that the Gaussian membership function used in the ANFIS to establish the SAW system model outperforms the triangular and bell-shaped membership functions in terms of training and test errors. The RMSE values of the Gaussian membership function on the training and testing sets are 0.001 and 1.13, respectively. Finally, the TBGA is used to optimize the design parameters based on the system model of a SAW sensor. Thus, the best design parameters are optimized as follows: the number of finger electrode pairs is 10, the separation distance between the two IDTs on the substrate is 3.5 (mm), the electrode overlap is 5 (mm), the electrode thickness is 1.4×10−5 (mm), and the ST-quartz substrate dimensions are 7 × 7 × 1 (mm). For comparison, Table 4 presents the best design parameters obtained by the conventional trial-and-error design method, VOLUME 3, 2015
the Taguchi design method [7], GA-based ANFIS method, PSO-based ANFIS method, and the proposed TBGA-based ANFIS method. Those best design parameters, shown in Table 4, were used by experiments to analyze their respective resonant frequency shift 1f and precision shown in Table 5. Table 5 shows the best design results from different design methods that the resonant frequency shift 1f obtained by the proposed TBGA-based ANFIS method is 39.88 (Hz), which is larger than 28.96 (Hz) and 36.07 (Hz) obtained by the trial-and-error method and the Taguchi method, respectively. That is, the incremental increase in 1f is 10.92 (Hz) and 3.81 (Hz) compared to the trial-and-error method and Taguchi method, respectively. The precision of the designed SAW reaches 4.2 × 10−10 (g/Hz), which is 48% higher and 17% higher improvement compared to the trial-and-error method and Taguchi method, respectively. Additionally, the resonant frequency shift 1f of 36.07 (Hz) obtained by the Taguchi design method is larger than 28.96 (Hz) obtained by the trial-and-error design method. That is, the incremental increase in 1f is 7.11 larger. The precision of the SAW obtained by the Taguchi design method is 4.9×10−10 (g/Hz), which is 27% higher improvement compared to the trial-anderror design method. The GA with ANFIS (GA-based ANFIS) and particle swarm optimization [25] with ANFIS (PSO-based ANFIS) were further used to optimize design parameters for a SAW 425
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
FIGURE 3. Adaptive neuro-fuzzy inference system architecture with five inputs and one output. TABLE 3. Comparison of root mean square error (RMSE) for different membership functions used to build the adaptive neuro-fuzzy inference system model for the SAW system.
TABLE 4. Comparison of best design parameters obtained by different optimization approaches.
gas-sensing device. Tables 4-5 show the best design parameters of best performances obtained by the GA-based ANFIS and by the PSO-based ANFIS, respectively. The comparisons showed that the best resonant shift 1f values obtained by the GA-based ANFIS (38.81 (Hz)) and by the PSO-based ANFIS (39.20 (Hz)) are lower than that obtained by the TBGA-based ANFIS (39.88 (Hz)). The best 426
precisions of the SAWs designed by the GA-based ANFIS (4.7 × 10−10 (g/Hz)) and by the PSO-based ANFIS (4.5 × 10−10 (g/Hz)) are also lower than that of SAWs designed using the proposed TBGA-based ANFIS (4.2 × 10−10 (g/Hz)). Notably, after experimental validation of the superior performance of the proposed TBGA-based ANFIS method, the proposed method was licensed to VOLUME 3, 2015
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
TABLE 5. Comparison of best design results from different design methods.
TABLE 6. Comparison results of GA-based ANFIS, PSO-based ANFIS, and proposed TBGA-based ANFIS methods in terms of best, average, and standard deviation (SD).
the Metal Industries Research and Development Centre (MIRDC, http://www.mirdc.org.tw) for use in optimizing various process parameters. Table 6 shows GA-based ANFIS, PSO-based ANFIS, and proposed TBGA-based ANFIS methods in terms of best, average, and standard deviation in 20 independent runs. The comparisons showed that the average resonant shift 1f values obtained by the GA-based ANFIS (38.05 (Hz)) and by the PSO-based ANFIS (38.76 (Hz)) are lower than that obtained by the TBGA-based ANFIS (39.68 (Hz)). The average precisions of the SAWs designed by the GA-based ANFIS (4.8 × 10−10 (g/Hz)) and by the PSO-based ANFIS (4.58 × 10−10 (g/Hz)) are also lower than that of SAWs designed using the proposed TBGA-based ANFIS (4.24 × 10−10 (g/Hz)). The standard deviations in 1f and precision obtained by the proposed TBGA-based ANFIS are smaller than those obtained by GA-based ANFIS and by the PSO-based ANFIS. The t-test value of 1f is 7.68 (p < 0.000001) in TBGA-based ANFIS vs. PSO-based ANFIS, and 12.28 (p < 0.000001) in TBGA-based ANFIS vs. GA-based ANFIS. The t-test value of precision is 22.61 (p < 0.000001) in TBGA-based ANFIS vs. PSO-based ANFIS, and 34.14 (p < 0.000001) in TBGA-based ANFIS vs. GA-based ANFIS. The tests mean that there is a significant difference between TBGA-based ANFIS and the other two algorithms. The comparisons demonstrate that the proposed TBGA-based ANFIS is significantly superior to both PSO-based ANFIS and GA-based ANFIS. Though, in Table 6, only few decimal points differ, this differences result as having great significance for the SAW sensor, because the SAW sensor must be designed to be highly sensitive to small mass changes on the surface. From Table 6, it can be seen that the proposed systematic method can obtain best design result of having VOLUME 3, 2015
highly sensitivity. That is, the experiments demonstrated that the proposed systematic approach is an effective tool for optimizing SAW design parameters in terms of resonant frequency shift and sensing precision. Additional advantages of the proposed approach for designing parameters of a SAW gas-sensing device are discussed below. The main mechanism of the TBGA is its combination of CGA and Taguchi method. In macro space, the CGA provides a powerful global exploration capability. Experiments show that, when using the systematic reasoning capability of the Taguchi method, a new chromosome produced by a matrix experiment is better than both of its parents [13], [14]. That is, potential chromosomes in micro space can be exploited. In micro space, the use of an OA with SNR as a systematic reasoning mechanism enhances the performance of the TBGA and accelerates convergence to the global solution. Since the TBGA provides a good balance of enhanced exploration and exploitation capabilities, it effectively optimizes the premise and consequent parameters for ANFIS and the design parameters for a SAW gas-sensing device. Since our previous works [13], [14] showed that the TBGA achieves optimal or near-optimal solutions, we further integrated the TBGA in ANFIS to obtain the proposed TBGA-based ANFIS approach to solving design parameter optimization problems in SAW gas-sensing devices. The results obtained by the proposed TBGA-based ANFIS approach were then compared with those obtained by the conventional trial-and-error design method, the Taguchi design method, the GA-based ANFIS, and the PSO-based ANFIS. Table 5 presents the performance comparison results, which confirm that the proposed TBGA-based ANFIS approach indeed obtains superior solutions to the design parameter optimization problem in SAW gas sensing devices. 427
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
In conclusion, a current search of the existing literature [7] and industry-related publications (http://www.mirdc.org.tw) reveals only two approaches to optimizing SAW parameters in terms of resonant frequency shift and sensing precision: the conventional trial-and-error approach and the Taguchi-based approach. In terms of practical industrial applications, this study provides an effective method of optimizing designs for SAW gas-sensing devices. To our knowledge, this study is the first to apply a soft-computing solution to this problem. The TBGA-based ANFIS increases the resonant frequency shift and the sensing precision by applying soft computing principles and by using Taguchi method to reduce the number of experiments. For optimizing the design parameters of a SAW gas-sensing device, the TBGA-based ANFIS is indeed a more effective alternative compared to the conventional trial-and-error design method, the Taguchi design method, the GA-based ANFIS, and the PSO-based ANFIS. Although, from the theoretical perspective, the proposed approach appears to be a relatively minor innovation of ANFIS, the simulated experiments in this study confirmed that the proposed TBGA-based ANFIS approach substantially improves resonant frequency shift and sensing precision. Therefore, the proposed TBGA-based ANFIS method has immediate real-world applications and can promote the further transfer of soft-computing-based technology from academia to industry. For example, the Metal Industries Research and Development Centre has been licensed to use the proposed TBGA-based ANFIS method, instead of the conventional non-systematic approaches currently used, to optimize various process parameters. The proposed design method also meets the two standard criteria applied by IEEE Society technical journals to define an innovative engineering application. First, the proposed method is a novel soft-computing-based approach to solving the design parameter optimization problem in SAW gassensing devices. The simulated experimental results discussed above indicate that the performance of the proposed approach is indeed comparable to that of conventional approaches. Second, the proposed method of using fuzzy models (i.e., a new alternative approach for designing optimal SAW gas-sensing devices) outperforms techniques currently used in the industry, which are not soft-computing-based (e.g., Taguchi method and the conventional trial-and-error method). Our team has worked on applications of intelligent evolutionary algorithms in optimal system modeling for many years. Therefore, we have studied many related researches and hope to share them to interested researchers. Additionally, our story entitled ‘‘Applications of intelligent evolutionary algorithms in optimal system modeling and mechanical design’’ got the winning one of 2014 CI Industrial Application Success Story. Our industrial application story has been reported in the IEEE Computational Intelligence Society website (http://cis.ieee.org/industrynews-success-stories.html).
428
V. CONCLUSIONS
The major contribution of this study is a systematic design procedure for optimizing the parameters of a SAW gas-sensing device by integrating Taguchi method, ANFIS, and TBGA. An OA L18 (21 × 37 ) is used to collect experimental data representing the SAW gas-sensing device. The TBGA then trains the five-input, one-output ANFIS architecture to represent the SAW gas-sensing system and explores the better design parameters. The best resonant frequency shift 1f obtained by the proposed TBGA-based ANFIS (39.88 Hz) is far superior to those obtained by the trial-and-error design method (28.96 Hz) and by the Taguchi design method (36.07 Hz). Precision reaches 4.2 × 10−10 (g/Hz), which is 48% and 17% higher improvement compared to the trial-and-error design method and the Taguchi design method, respectively. The TBGA-based ANFIS also obtains a higher resonant frequency shift 1f (39.88 Hz) compared to the GA-based ANFIS (38.81 Hz) and the PSO-based ANFIS (39.20 Hz). Finally, comparisons of SAW designs show that the proposed TBGA-based ANFIS achieves superior precision compared to GA-based ANFIS and PSO-based ANFIS. Overall, the comparison results show that the proposed TBGA-based ANFIS method has superior parameter optimization performance and obtains superior SAW sensor designs in terms of both resonant frequency shift and precision. ACKNOWLEDGEMENTS
The authors thank Professor Der-Ho Wu, National Pingtung University of Science and Technology, for providing the experimental data. REFERENCES [1] A. Abdollahi, Z. Jiang, and S. A. Arabshahi, ‘‘Evaluation on mass sensitivity of SAW sensors for different piezoelectric materials using finiteelement analysis,’’ IEEE Trans. Ultrason., Ferroelectr., Freq. Control, vol. 54, no. 12, pp. 2446–2455, Dec. 2007. [2] P. Panek, ‘‘Time-interval measurement based on SAW filter excitation,’’ IEEE Trans. Instrum. Meas., vol. 57, no. 11, pp. 2582–2588, Nov. 2008. [3] S. Krishnamurthy, M. Z. Atashbar, and B. J. Bazuin, ‘‘Burst transceiver unit for wireless passive SAW sensing system,’’ IEEE Trans. Instrum. Meas., vol. 58, no. 10, pp. 3746–3753, Oct. 2009. [4] P. Zheng, D. W. Greve, and I. J. Oppenheim, ‘‘Langasite surface acoustic wave gas sensors: Modeling and verification,’’ IEEE Trans. Ultrason., Ferroelectr., Freq. Control, vol. 60, no. 3, pp. 579–586, Mar. 2013. [5] H. Zhang and J. A. Kosinski, ‘‘Analysis of contributions of nonlinear material constants to stress-induced velocity shifts of quartz and langasite surface acoustic wave resonators,’’ IEEE Trans. Ultrason., Ferroelectr., Freq. Control, vol. 60, no. 5, pp. 975–985, May 2013. [6] V. I. Anisimkin and E. Verona, ‘‘New properties of SAW gas sensing,’’ IEEE Trans. Ultrason., Ferroelectr., Freq. Control, vol. 45, no. 5, pp. 1347–1354, Sep. 1998. [7] D. H. Wu and H. H. Chen, ‘‘Application of Taguchi robust design method to SAW mass sensing device,’’ IEEE Trans. Ultrason., Ferroelectr., Freq. Control, vol. 52, no. 12, pp. 2403–2410, Dec. 2005. [8] V. I. Anisimkin and E. Verona, ‘‘New capabilities for optimizing SAW gas sensors,’’ IEEE Trans. Ultrason., Ferroelectr., Freq. Control, vol. 48, no. 5, pp. 1413–1418, Sep. 2001. [9] J.-S. R. Jang, ‘‘ANFIS: Adaptive-network-based fuzzy inference system,’’ IEEE Trans. Syst., Man, Cybern., vol. 23, no. 3, pp. 665–685, May/Jun. 1993. [10] J.-S. R. Jang, C. T. Sun, and E. Mizutani, Neuro-Fuzzy and Soft Computing. Taipei, Taiwan: Pearson Education, 2004.
VOLUME 3, 2015
J.-T. Tsai et al.: Optimal Design of SAW Gas Sensing Device Using Improved ANFIS
[11] B. M. Dehkordi, M. Moallem, and A. Parsapour, ‘‘Predicting foaming slag quality in electric arc furnace using power quality indices and fuzzy method,’’ IEEE Trans. Instrum. Meas., vol. 60, no. 12, pp. 3845–3852, Dec. 2011. [12] M. Forouzanfar, H. R. Dajani, V. Z. Groza, M. Bolic, and S. Rajan, ‘‘Feature-based neural network approach for oscillometric blood pressure estimation,’’ IEEE Trans. Instrum. Meas., vol. 60, no. 8, pp. 2786–2796, Aug. 2011. [13] J.-T. Tsai, T.-K. Liu, and J.-H. Chou, ‘‘Hybrid Taguchi-genetic algorithm for global numerical optimization,’’ IEEE Trans. Evol. Comput., vol. 8, no. 4, pp. 365–377, Aug. 2004. [14] J.-T. Tsai, J.-H. Chou, and T.-K. Liu, ‘‘Tuning the structure and parameters of a neural network by using hybrid Taguchi-genetic algorithm,’’ IEEE Trans. Neural Netw., vol. 17, no. 1, pp. 69–80, Jan. 2006. [15] W.-H. Ho, J.-T. Tsai, B.-T. Lin, and J.-H. Chou, ‘‘Adaptive network-based fuzzy inference system for prediction of surface roughness in end milling process using hybrid Taguchi-genetic learning algorithm,’’ Expert Syst. Appl., vol. 36, no. 2, pp. 3216–3222, 2009. [16] W.-H. Ho, J.-T. Tsai, and J.-H. Chou, ‘‘Robust quadratic-optimal control of TS-fuzzy-model-based dynamic systems with both elemental parametric uncertainties and norm-bounded approximation error,’’ IEEE Trans. Fuzzy Syst., vol. 17, no. 3, pp. 518–531, Jun. 2009. [17] W.-H. Ho, J.-X. Chen, I.-N. Lee, and H.-C. Su, ‘‘An ANFIS-based model for predicting adequacy of vancomycin regimen using improved genetic algorithm,’’ Expert Syst. Appl., vol. 38, no. 10, pp. 13050–13056, 2011. [18] W.-H. Ho, J.-T. Tsai, and H.-Y. Wang, ‘‘Neural fuzzy network model with evolutionary learning algorithm for mycological study of foodborne fungi,’’ Int. J. Innov. Comput., Inf. Control, vol. 8, pp. 4565–4577, Jul. 2012. [19] I.-T. Chen, J.-T. Tsai, C.-F. Wen, and W.-H. Ho, ‘‘Artificial neural network with hybrid Taguchi-genetic algorithm for nonlinear MIMO model of machining processes,’’ Int. J. Innov. Comput., Inf. Control, vol. 9, no. 4, pp. 1455–1475, 2013. [20] W.-H. Ho, J.-T. Tsai, G.-M. Hsu, and J.-H. Chou, ‘‘Process parameters optimization: A design study for TiO2 thin film of vacuum sputtering process,’’ IEEE Trans. Autom. Sci. Eng., vol. 7, no. 1, pp. 143–146, Jan. 2010. [21] G. Taguchi, S. Chowdhury, and S. Taguchi, Robust Engineering. New York, NY, USA: McGraw-Hill, 2000. [22] K. Y. Chan and T. S. Dillon, ‘‘On-road sensor configuration design for traffic flow prediction using fuzzy neural networks and Taguchi method,’’ IEEE Trans. Instrum. Meas., vol. 62, no. 1, pp. 50–59, Jan. 2013. [23] D. E. Goldberg, Genetic Algorithms in Search, Optimization and Machine Learning. Reading, MA, USA: Addison-Wesley, 1989. [24] C.-H. Wu, I.-S. Lin, M.-L. Wei, and T.-Y. Cheng, ‘‘Target position estimation by genetic expression programming for mobile robots with vision sensors,’’ IEEE Trans. Instrum. Meas., vol. 62, no. 12, pp. 3218–3230, Dec. 2013. [25] J. Kennedy and R. C. Eberhart, ‘‘Particle swarm optimization,’’ in Proc. IEEE Int. Conf. Neural Netw., vol. 4. Nov./Dec. 1995, pp. 1942–1948.
JINN-TSONG TSAI received the B.S. and M.S. degrees in mechanical and electromechanical engineering from National Sun Yat-sen University, Taiwan, in 1986 and 1988, respectively, and the Ph.D. degree in engineering science and technology from the National Kaohsiung First University of Science and Technology, Taiwan, in 2004. He was a Lecturer with the Vehicle Engineering Department, Chung Cheng Institute of Technology, Taiwan, from 1988 to 1990. From 1990 to 2004, he was a Researcher and the Chief of the Automation Control Section with the Metal Industries Research and Development Center, Taiwan. From 2004 to 2006, he was an Assistant Professor with the Medical Information Management Department, Kaohsiung Medical University, Kaohsiung, Taiwan. From 2006 to 2014, he was an Assistant Professor and Associate Professor with the Department of Computer Science, National Pingtung University of Education, Pingtung, Taiwan. He is currently a Professor with the Department of Computer Science, National Pingtung University, Pingtung. His research interests include evolutionary computation, intelligent control and systems, neural networks, and quality engineering.
VOLUME 3, 2015
KAI-YU CHIU received the M.S. degree in electrical engineering from the National Kaohsiung First University of Science and Technology, Kaohsiung, Taiwan, in 2014. He is currently a Process Engineer with the Assembly Department of Advanced Semiconductor Engineering Inc., Kaohsiung. His research and major interests include intelligent systems and control, computational intelligence and methods, and quality engineering.
JYH-HORNG CHOU (SM’04–F’15) received the B.S. and M.S. degrees in engineering science from National Cheng Kung University, Tainan, Taiwan, in 1981 and 1983, respectively, and the Ph.D. degree in mechatronic engineering from National Sun Yat-sen University, Kaohsiung, Taiwan, in 1988. He is currently the Chair Professor with the Electrical Engineering Department, National Kaohsiung University of Applied Sciences, Taiwan. He has co-authored four books, and authored over 260 refereed journal papers. He also holds six patents. His research and teaching interests include intelligent systems and control, computational intelligence and methods, automation technology, robust control, and robust optimization. He is a fellow of the Institution of Engineering and Technology, the Chinese Automatic Control Society, the Chinese Institute of Automation Engineer, and the Chinese Society of Mechanical Engineers. He was a recipient of the 2011 Distinguished Research Award from the National Science Council of Taiwan, the 2012 IEEE Outstanding Technical Achievement Award from the IEEE Tainan Section, the 2014 Distinguished Research Award from the Ministry of Science and Technology of Taiwan, the Research Award and the Excellent Research Award from the National Science Council of Taiwan 14 times, and numerous academic awards/honors from various societies. Based on the IEEE Computational Intelligence Society (CIS) evaluation, his Industrial Application Success Story has received the 2014 Winner of Highest Rank, thus being selected to become the first, and only internationally industrial success story being reported on the IEEE CIS website in 2014.
429
