Abstract
In the last one decade, neural networksbased modeling has been used for computing different performance parameters of microstrip antennas because of learning and generalization features. Most of the created neural models are based on software simulation. As the neural networks show massive parallelism inherently, a parallel hardware needs to be created for creating faster computing machine by taking the advantages of the parallelism of the neural networks. This paper demonstrates a generalized neural networks model created on field programmable gate array (FPGA) based reconfigurable hardware platform for computing different performance parameters of microstrip antennas. Thus, the proposed approach provides a platform for developing lowcost neural networkbased FPGA simulators for microwave applications. Also, the results obtained by this approach are in very good agreement with the measured results available in the literature.
1. Introduction
Low profile, conformable to planar and nonplanar surfaces, most economical, mechanically robust, light weight, and easily mountability are the key advantages of microstrip antennas (MSAs). Because of these addon advantages, the microstrip antennas are widely used in many communication applications. Since the microstrip antenna operates only in the vicinity of the resonant frequency, it needs to be calculated accurately for analyzing the microstrip antennas. Similarly, for designing the microstrip antennas, the physical dimension(s) must also be calculated precisely [1]. There are two conventional ways for analyzing and/or designing the microstrip antennas, analytical methods and numerical methods. The analytical methods provide a good spontaneous explanation for the operation of microstrip antennas. As the analytical methods are based on the physical assumptions for simplifying the radiation mechanism of the microstrip antennas, these methods are not suitable for many structures, where the thickness of the substrate is not very thin. The numerical methods also provide the accurate results but the analysis using these methods leads to the expressions as an integral equation. The choice of test functions and path integrations appears to be more critical without any initial assumption in the final stage of the numerical results. Also, these methods require a new solution for any sort of alteration in the geometry. The problems associated with these conventional methods can be overcome by selecting the appropriate neural network methods [1]. In recent years, artificial neural networks (ANNs) have acquired tremendous utilization in microwave communications because of high classifying power and learning ability of the ANN [2–4]. The ANN model is trained using measured, calculated, and/or simulated patterns. Once the model is trained for a specified error, it returns the results for every infinitesimal change in the input patterns within a fraction of a second. Several neural models are available in the literature [5–8] for computing single performance parameter (resonance frequency or geometric dimensions) of the microstrip patch antennas, respectively. But having individual neural model for each performance parameter becomes sometimes unattractive to include in modern antenna computeraideddesign (CAD) programs. It has recently been overcome by introducing the concept of generalized neural approach for computing different performance parameters, simultaneously [9–12]. The resonance frequencies of rectangular, circular, and triangular MSAs have been computed [9–11] using generalized neural networks model. The resonance frequencies and the physical dimensions of the rectangular MSAs have been computed using generalized neural networks model [12]. Hence, the generalized neural models [9–12] have been used only for computing three different parameters of MSAs, respectively. Recently, authors have also proposed more accurate and very simple generalized neural networks approaches for computing two parameters [13–15], three parameters [16, 17], four parameters [18], and seven parameters [19], respectively.
The neural networks models [2–19] have been realized using software simulation. Since the software simulation involves large number of complex arithmetical operations, therefore it does not have the desired performance as in case of realtime computation [20]. For speedingup the computing process, some sort of hardware needs to be created for training and/or testing algorithm of the neural networks. Replication of neural networks on a dedicated hardware is a challenging task [20]. There are two conventional methods for implementing the algorithm(s) of a neural model on a hardware platform, either by using the hardwired technology or reconfigurable hardware approach. Further, there are two ways for using the hardwired technology, either by creating the application specific integrated circuits (ASICs) or by using a group of individual components using printed circuit board (PCB) technology. ASICs are very fast and efficient but they cannot be altered after fabrication. This forces a redesign and a refabrication of the chip for the requirement of any kind of modification and, therefore, becomes an expensive process. PCBbased approach is also inflexible, as it requires redesign and the replacement of the board in the event of changes in the application [21]. Presently, reconfigurable hardware is being widely used as a technique for projecting and prototyping the hardware because it allows fast designing and prototyping [22]. Reconfigurable hardware platforms have a balance between hardware and software solutions exclusively, as they have the programmability of software with performance capacity approaching that of a custom hardware implementation [22]. Because of several attractive features like low cost as compared to MPLDs (mask programmable logic devices), easy implementation and reprogrammability, the FPGAs are being widely used for creating reconfigurable hardware of neural networks for different applications [23–27] but the literature on this approach for microwave applications domain is limited [28–31]. To et al. [28] have described prototyping of a neuroadaptive smart antenna beamforming algorithm using hardwaresoftware approach by implementing the RBF neural network on FPGA platform. For prototyping strategy, they have used three steps, implementing a simulation model in MATLAB software, translating it into generic C/C++ working model with an external matrix arithmetic library, and, finally, implementing this C/C++ model on the Altera “APEX FPGA EP20K200E” embedded processor platform [28]. Al Zuraiqi et al. [29] have designed a neural network FPGA controller for reconfigurable antennas and the optimized weights, biases, and network topology have been configured on Xilinx’s FPGA. Recently, Ghayoula et al. [30] have explored prototyping and implementing concept of neural networks on FPGA platform for designing phased antenna array. In this purpose, they have optimized the neural model with 17 neurons in the input layer and 8 neurons in the output layer. They have then implemented the optimized model with 8bit precision using Xilinx 8.1i, simulated with ModelSim SE 6.0, and downloaded and tested on Xilinx “XC3S500”. Fournier et al. [31] have implemented the neural networks model on FPGA platform for beam steering of an array of four microstrip patch antennas. The feeding structure of Butler matrix has been optimized using the momentum feature of advance design system (ADS) for 2.4 GHz frequency. For controlling the beam angle, they have performed the training of the ANN model in MATLAB Simulink; coding for the trained model has been created using Xilinx system generator and, finally, it is downloaded on Xilinx’ FPGA board [31]. The beauty of the present work lies in creating an FPGAbased reconfigurable hardware for a generalized neural model which is capable of computing seven different performance parameters. The training and/or testing algorithms of the neural networks can be implemented in reconfigurable hardware. In the proposed work, training of the model is done offline in personal computing machine and the testing algorithm is implemented on Xilinx’s FPGA board, XC3S500E.
2. Generation of Patterns
A microstrip antenna, in its simplest configuration, consists of a radiating conductive patch on one side of a dielectric substrate of relative permittivity “” and of thickness “” having a ground plane on the other side [1]. Figure 1 illustrates the geometry for three different shapes of microstrip antennas, that is, rectangular microstrip antenna (RMSA), circular microstrip antenna (CMSA), and triangular microstrip antenna (TMSA), respectively. The side view of three different patches is shown in Figure 1(a), whereas top views for RMSA, CMSA, and TMSA are mentioned in Figures 1(b)–1(d), respectively.
(a)
(b)
(c)
(d)
For analyzing a microstrip antenna (MSA), the resonance frequency for these geometries can be easily calculated, if the physical dimensions, relative permittivity, dielectric thickness, and mode of propagation are given [32–44]. Total 81 measured patterns (46 for RMSA, 20 for CMSA, and 15 for TMSA) are taken as training and testing patterns for the neural networks modeling. The 46 patterns for RMSAs, 20 patterns for CMSAs, and 15 patterns for TMSAs are mentioned in Tables 1, 2, and 3, respectively.
For designing the MSAs, the physical dimensions of the patch can be determined if the resonance frequency, thickness, and relative permittivity of the substrate and mode of propagation are given [32–44]. This is mentioned in Table 4. The total of 92 patterns (46 for each dimension) for designing the RMSAs can be created using Tables 4 and 1, simultaneously. Similarly, total of 20 patterns for designing the RMSAs can be created using Tables 4 and 2, simultaneously. Similar approach is used for creating 15 patterns for designing TMSAs using Tables 4 and 3, simultaneously. These total 208 generated patterns (81 analyzing patterns + 127 designing patterns) are used for software implementation of training algorithm and hardware implementation of testing algorithm to be discussed in Sections 3 and 4, respectively.
3. Proposed Neural Networks Modeling
A radial basis function (RBF) neural network consists of threelayer feedforward neural network with entirely different roles. The input layer is made up of source nodes which connect the network to its outside environment. The input layer does not accomplish any process but simply buffers the data. The second layer, that is, the hidden layer, applies a multivariate Gaussian nonlinear transformation from the input space to the hidden space [45]. The output layer is linear, supplying the response of the network to the patterns applied to the input layer. As far as learning (or training) of the neural network is concerned, the RBF neural network is much faster than the multilayered perceptron (MLP) neural network. It is so because the learning process in RBF neural network has two stages and both of the stages are made more efficient by using appropriate learning algorithm. This is the prime reason of using RBF neural network instead of MLP neural network in the present work. 208 measured patterns discussed in Section 2 are used for training and testing of RBF neural network. The strategy for implementing the training algorithm for RBF neural networks model is being discussed in this section, whereas hardware implementation of testing algorithm is to be discussed in Section 4.
3.1. Training Algorithm: Software Implementation
The inputoutput patterns for analyzing and designing the RMSAs, CMSAs, and TMSAs are discussed in Section 2 and mentioned in Tables 1, 2, 3, and 4. It is clear from these tables that the input pattern in all cases is fivedimensional, which is, in general, being represented by five variables ,, , , and , respectively. Seven different cases, computation of resonance frequency, width and length of RMSAs, computation of resonance frequency and radius of CMSAs, and the computation of resonance frequency and sidelength are being included in a single RBF neural network model. To distinguish these different cases, an additional variable, “”, is included in the existing fivedimensional patterns. Here, = 1, 2, and 3 represented the computation of resonance frequency, width, and length of RMSAs, = 4 and 5 represented the resonance frequency and radius of CMSAs, and, finally, = 6 and 7 indicateed the resonance frequency and side length of the TMSAs, respectively. This sixdimensional input pattern, that is, [], is used as the training pattern and testing pattern for computing each performance parameter of the microstrip antennas, respectively.
An RBF neural network model with six input nodes, 45 hidden nodes, and single output node is illustrated in Figure 2, in which weight matrices at hidden nodes and at output node are designated as []_{6 × 45} and []_{1 × 45,} whereas the bias matrices at these nodes are represented as []_{45 × 1} and []_{1 × 1}, respectively.
Initially, some random values are assigned for the weight and bias values and these values are optimized using LevenbergMarquardt (LM) training algorithm [46] and the codes for implementing the training algorithm of the proposed model is written using MATLAB software on a personal computing machine with system configuration, Dell Optiplex 780 Core 2 Duo CPU E8400, 3.0 GHz with 4.0 GB RAM. The basic approach for training the model can be understood by the flow diagram shown in Figure 3, in which the structural configuration of the model is selected as . For the applied input pattern, some random numbers between +1 and −1 are assigned to the weights and biases and the output of the model is computed corresponding to that input pattern. Some arbitrary parameters required for training of the neural model like mean square error, learning rate, momentum coefficient, and spread value have also been taken as 5 × 10^{−7}, 0.1, 0.5, and 0.5, respectively. The error between the calculated and the measured result is then computed and, according to this computed error, all the weights and biases are updated with the help of LM training algorithm. This updating process is carried out after presenting each set of input pattern, until the calculated accuracy of the model is estimated to be satisfactory. Once it is achieved, the final updated weight matrices [] and [] and the bias matrices [] and [] are taken out from the trained neural model and these matrices are used for implementing the testing algorithm on FPGA board which is to be discussed in Section 4.
4. Testing Algorithm: Hardware Implementation
The FPGA logic is a class of digital integrated circuitry, whose internal structure can be configured in parallel computing units and it can be repeatedly reprogrammed after manufacturing. Also, the ability to program parallel computing makes FPGA logic very suitable for implementing the parallel structure of the neural networks. The property of hardware reconfigurability is very attractive for the implementation of neural network, as it allows maximum network flexibility. FPGAs have the advantages to be onsite reconfigurable, allowing the use of circuit in few minutes after the design versus many months for the MPLDs (mask programmable logic devices). The FPGA board is basically characterized as functional unit and storage unit. The functional unit consists of processing operations (e.g., multipliers, adders, and activation functions) and storage unit (e.g., RAM containing synaptic weights, bias values, and input buffers). Also, FPGA programmability authorizes design upgrades in the field with no hardware replacement necessary, an impossibility with ASICs [47]. The Spartan3E family of FPGAs builds on the success of the earlier Spartan3 family by increasing the amount of logic per I/O, significantly reducing the cost per logic cell. These features improve system performance and reduce the cost of configuration. The Spartan3E FPGAs, combined with advanced 90 nm process technology, deliver more functionality and bandwidth [47]. The Spartan3E family is a better alternative to mask programmed ASICs. FPGAs avoid the high initial cost, the lengthy development cycles, and the inherent inflexibility of conventional ASICs. In 1985, Xilinx introduced the world’s first FPGA, “XC2064”, which contained approximately 1000 logic gates. Since then, the gate density of Xilinx FPGAs has increased 500 times in XC3S500E [48]. In the present work, Sparton3E XC3S500E FPGA is used which is having two crystal oscillators “” and “” of frequencies 24.0 MHz and 25.0 MHz, respectively.
In Section 2, the weights [] and [] and the biases [] and [] are optimized during training of the RBF neural network. In this section, these weights and biases have been arranged as per the requirement of testing algorithm. The schematic flow of the testing algorithm to be implemented on FPGA board is shown in Figure 4. The values of weight and bias matrices used in Figure 4 have already been defined in Figure 3. For making it convenient, it has been decided to split the overall implementation into the implementation of inputtohidden layer and the implementation of hiddentooutput layer. There are a total of 45 symmetrical nodes in the hidden layer and single node in the output layer. All the nodes in hidden layer are fired by the radial basis activation function, whereas the output node is fired by the pure linear activation function. The procedure to implement all nodes in the hidden layer is identical and it is described as below.
The input () and output () of hidden node1 in Figure 2(a) are written as As there is no activation function in the output layer, the role of this layer is just to accumulate the results processed in the earlier hidden layer. Hence, the final result obtained at the output node in Figure 2 is also written as It is clear from (1) that 6 multipliers and 6 adders are required to implement this node. As a total of 45 symmetrical nodes are there in the hidden layer, then the number of multipliers and adders required to implement the hidden layer is coming out to be 270 each (). But doing 270 simultaneous multiplications and 270 simultaneous additions is a tedious and timeconsuming job. To reduce this complexity, 6 multipliers and 6 adders are used along with some memorywrite (MW) and memoryread (MR) operations to implement the 45 nodes in the hidden layer. These multipliers and adders process one node at a time and the result corresponding to that node is stored in the memory storage available on the block RAM of the FPGA board using memorywrite (MW) operation. The same procedure is followed for the rest of the hidden nodes. For this purpose, a total of 45 memorywrite (MW) operations are performed for implementing the inputtohidden layer. Again for implementing the hiddentooutput layer defined in (3), the results stored in the memory are retrieved through 45 memoryread (MR) operations. In this work, a VHDL library for an IEEE 32bit floating point number precision is also designed for floating point addition fp_add and floating point multiplication fp_mul. Each node value along with their bias counterpart is further added in the adder circuit. The added values obtained on 45 hidden nodes are then passed in radial basis activation function through lookup tables (LUTs). The results of LUTs are further stored for next layer (i.e., hiddentooutput layer) implementation. The same procedure is adopted for implementing this layer as it is done for implementing the inputtohidden layer. Finally, the required result is obtained at this output node. Another value to the input matrix [] is assigned as per the requirement and it is then passed through the same process to get the corresponding result.
For the abovementioned procedure, the testing algorithm of RBF neural network is coded by means of VLSI hardware description language (VHDL) required for implementation on FPGA board “XC3S500E.” The generalized RBF neural network is implemented using Xilinx Embedded Development Kit (EDK) version 9.2i, simulated with Mentor Graphics ModelSim PE 10.1, and further downloaded and tested on the FPGA board. The different attributes available on the FPGA board and their use for this implementation are given in Table 5. The screenshot of the setup used during this implementation is shown in Figure 5.
Some basic operations, data representation, multiplications, additions, and activation functions implementation, used in present work are also discussed below.
4.1. Data Representation
Neural Networks, in general, work with floatingpoint numbers but working with floatingpoint numbers in FPGAbased digital hardware is a difficult problem. As a solution to make it easier and further to improve its performance, it is required to convert the floatingpoint numbers to binary ones. The data patterns, synaptic weights, and bias values are encoded in binary representation. Each value of weights and biases is encoded into 32bit binary equivalent in which the MSB (most significant bit) is employed as signed bit (0 for plus and 1 for minus) and rest 31 bits represented as magnitude values for weights and biases, whereas the input pattern is represented by 32bit binary values as there is no negative sign with the input pattern.
4.2. Multiplications and Additions
FPGAbased digital processors perform a variety of informationprocessing tasks. The basic functions encountered in a processing are the various arithmetic operations and, particularly, multiplication and addition of two binary strings are, no doubt, the basic arithmetic building blocks of the present work. The multiplication is done using lookup tables (LUTs) approach. To realize these LUTs, configurable logic blocks (CLBs) have been programmed.
The 32bit input is multiplied by a 32bit signed weight to form a 64bit signed product that is further accumulated with a 32bit signed bias value into a 64bit signed sum. Here, 2’s complement representation is implemented to handle the multiplication and addition of negative and/or positive numbers. To save the memory space, this 64bit signed sum is scaled down to 40bit value. This 24bit scaling is done after running a software program of neural network in a computer system using the same input data of that of hardware implementation. The 40 bits are the minimum number of bits that can be retained without deteriorating the accuracy ofthe sum. Multiplying a 32bit input by a 32bit signed weight produces a 64bit signed product that is further accumulated with a 32bit signed bias value into a 64bit signed sum. The 32 × 32 bit multiplication is broken into four 8 × 8 bit multiplications with four addition processes. The most significant partial product (8 × 8) of each of multiplicand and multiplier is shifted by eight bits before adding it to the least significant partial product.
4.3. Implementation of Activation Functions
The most challenging job in reconfigurable hardware (design or implementation) is the handling of activation functions. Multivariate Gaussian nonlinear function given in (4) is used for the hidden neurons. This Gaussian function is not suitable for FPGA implementation as it consists of an infinite terms in the exponential series. Thus, it needs some truncation as done in (5) as follows: where represents the output of th neuron in the input layer and represents the output of th neuron in the hidden layer.
5. Computed Results and Comparison
The computed seven different parameters of RMSAs, CMSAs, and TMSAs are given in Table 6, whereas their reference counterparts are mentioned in Tables 1, 2, and 3, respectively. For RMSAs, the computed results are compared with their measured counterparts mentioned in Table 1. For CMSAs, the computed results are compared with their measured counterparts mentioned in Table 2 and, for TMSAs, the computed results are compared with their measured counterparts mentioned in Table 3. It is clear from this comparison that most of the calculated points are in very good convergence and only few are off.
A comparison between the present method results and previously computed neural networks results using software implementations [5–11] is given in Table 7, which shows that, in the neural models [5, 10–12], the total absolute error for resonance frequency of RMSAs is calculated as 751.0 MHz, 750.0 MHz, 203.6 MHz, and 557.1 MHz, whereas, in the present model, it is only 55.2 MHz. In case of physical dimensions (both widths and lengths) of rectangular MSAs, the model [12] is having the total absolute error of 0.0653 cm and 0.0490 cm, whereas, in the present work, these are calculated as 0.0179 cm and 0.0146 cm, respectively. In case of resonance frequency of circular MSAs, the proposed method is having the total absolute error of 24 MHz, whereas, in the models [9–11], it is calculated as 92 MHz, 116 MHz, and 462 MHz, respectively. In case of radius of circular MSAs, the present model is having the total absolute error of 0.0280 cm, whereas there is no neural model proposed for computing the radius of circular MSAs in the open literature [5–12]. For resonance frequency of equilateral triangular MSAs, the models [6, 9–11] are having total absolute error as 27 MHz, 220 MHz, 272 MHz, and 23.00 MHz, respectively, whereas, in the present method, it is only 19.5 MHz. In calculating the side length of equilateral triangular MSAs, the error in testing patterns of the present method is only 0.0022 cm, whereas, in the model [8], it is calculated as 0.0213 cm.
The processing time of the hardware implementation is measured by interfacing the FPGA board with TLA 601 logic analyzer, which is a 34channel Tektronix logic analyzer, having external display with 2 GHz timing, 100 MHz State, and 64 K depth options for up to 256 K depth and/or 200 MHz State [47]. The screen shot of the logic analyzer (TLA601) and FPGA board (XC3S500E) is shown in Figure 4. This measured processing time is in the confirmation to the calculated processing time from the datasheet of FPGA board [48, 49]. The testing algorithm is also implemented in MATLAB software on a computer system for making a comparison of processing time with its hardware counterpart. The time elapsed in software implementation is calculated using MATLAB syntax “cputime” as mentioned in Algorithm 1.

By doing this procedure, it has been observed that the average time elapsed in software implementation of the testing algorithm on computer system is coming out to be 65 ms, whereas, in its hardware counterpart, it was measured as 320.301 ns.
6. Conclusions
In this paper, the authors have proposed a novel approach for prototyping the neural network models on FPGA platform, which has not been attempted earlier for analyzing and/or designing the microstrip antennas (MSAs). Secondly, the approach is capable of designing the circular MSAs as there is no reported neural model available in the open literature [1–12]. Thirdly, the prototype is capable of computing seven different parameters of three different MSAs, simultaneously, as no neural model has been suggested in the literature [1–12] for computing more than three parameters, simultaneously. Finally, at the same time, a common model has produced more encouraging results for all seven different cases.
The processing time for the approach has been calculated in software implementation and measured in hardware implementation. The average time elapsed in software implementation is coming out to be 65 ms, whereas, in its hardware counterpart, it has been measured as 320.301 ns. Thus, in the present case, the processing of hardware implementation is faster than its software counterparts of the same problem.
In general, the hardware implementation of seven different parameters may require seven different hardware modules, whereas, in the present work, only one hardware module is fulfilling the requirement of seven different hardware chips. The present approach can be considered low cost in this sense.
This study provides a roadmap for implementing neural networks on reconfigurable hardware for microwave applications. As the approach is comparatively faster, having low manufacturing cost, it can be used for creating neuralnetworkbasedFPGA simulators for microwave applications. Also, the proposed approach can be generalized for any number of computing parameters of the microstrip antennas. But, as the dimensionality of the parameters to be computed or the dimensionality of the input patterns increases, the structural configuration such as number of nodes in the hidden layer or sometimes the number of hidden layers also increases. Thus, this results in a considerable increase in the reconfiguration time and finally the training time. It may further require the large memory space if it is implemented on hardware platform. Thus, there is a tradeoff between the computing parameters of the microstrip antennas and available environment for the implementing the hardware neural networks.
Conflict of Interests
The authors declare that there is no conflict of interests regarding the publication of this paper.