^{*}

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (http://creativecommons.org/licenses/by/3.0/).

Recently, the range of available Radio Frequency Identification (RFID) tags has been widened to include smart RFID tags which can monitor their varying surroundings. One of the most important factors for better performance of smart RFID system is accurate measurement from various sensors. In the multi-sensing environment, some noisy signals are obtained because of the changing surroundings. We propose in this paper an improved Kalman filter method to reduce noise and obtain correct data. Performance of Kalman filter is determined by a measurement and system noise covariance which are usually called the R and Q variables in the Kalman filter algorithm. Choosing a correct R and Q variable is one of the most important design factors for better performance of the Kalman filter. For this reason, we proposed an improved Kalman filter to advance an ability of noise reduction of the Kalman filter. The measurement noise covariance was only considered because the system architecture is simple and can be adjusted by the neural network. With this method, more accurate data can be obtained with smart RFID tags. In a simulation the proposed improved Kalman filter has 40.1%, 60.4% and 87.5% less Mean Squared Error (MSE) than the conventional Kalman filter method for a temperature sensor, humidity sensor and oxygen sensor, respectively. The performance of the proposed method was also verified with some experiments.

In the field of Radio Frequency Identification (RFID) technology, a tremendous variety of novel RFID sensor tags has emerged. The RFID sensor tags also known as smart RFID tags are able to measure and compute data from the environmental such as temperature, humidity, oxygen concentration, pressure, tampering, shock,

Multi-sensors such as resistive, capacitive and inductive type sensors can be combined into smart RFID tags. These diverse sensor data can represent the freshness and vitality of living organism. In a multi-sensing RFID tag environment, the tags can obtain data of multi-sensors using both just one port or several ports. In a multi-sensor system that correlates noises caused by multiple sensors, accuracy of sensor data is one of the most important factors to evaluate the monitoring system. The Kalman filter has been widely applied to solve the noise problem of measurement systems [

When a system designer organizes a multi sensor system, each sensor may not work properly because of some factors that disturb their work. Those factors are noise and interference, which are caused by the measurement system in a multi-sensing environment. The multi-sensing RFID system used was composed of a EVB90129 by Melexis Microeletronic Systems and a sensor board combined with a temperature sensor, humidity sensor and oxygen sensor as shown in

The measurement data of the multi-sensing environment has more noise and disturbances than the single measurement as shown in

The Kalman filter requires that all the plant dynamics and noise processes be known exactly and when the noise processes are zero it means there is white noise. If the theoretical behavior of a filter and its actual behavior are not matched, a divergence problem will occur. When the error covariance is computed from the actual error in the measurement, satisfactory results are obtained without divergence. Noise covariance in the Kalman filter acts as in the role of controlling the bandwidth and modulates the Kalman gain. Abnormal choice of noise covariance is one of the most important factors which make Kalman filters diverge. The purpose of the proposed method is the estimation of the noise covariance by using neural networks to prevent divergence of Kalman filter.

The Kalman Filter is an algorithm which makes optimal use of imprecise data in a linear system with noises to continuously update the best estimate of the system’s current state. Kalman filter theory is based on a state-space approach in which a state equation models the dynamics of the signal generation process and an observation equation models the noisy and distorted observation signal.

The random variables _{k}_{k}

The matrix _{k}

In practice, Q represents the process noise covariance and R is the measurement noise covariance. In deriving the Kalman filter formulation, we begin with the goal of finding an equation that computes an a posteriori state estimate as a linear combination of an

The time update equations are responsible for projecting forward (in time) the current state and error covariance estimates to obtain the

We define _{k/k−1} to be our _{k}_{k/k} to be our _{k}_{k}_{k/k−1} is a presumption of the pre-measurement value. _{k/k} can be expressed as in

The Kalman gain, _{k}_{k/k}e_{k/k}^{T}_{k}

Derivation of the equation of estimate error covariance to compute the Kalman gain, _{k}

The above expected value is equivalent to minimizing the trace of the _{k}_{k}

The _{k/k}

The

The complete operation of the Kalman filter is described in

The measurement equation from the system state equation is shown in

When we develop the _{k}

The discrepancy between the predicted measurement and the actual measurement

A neural network can obtain the optimal noise covariance estimate _{k}_{k}_{k}

In the proposed algorithm, the structure of the neural network is a multi-layer neural network model. We used the error back propagation algorithm (EBA) [

In _{k}_{k}

In this case, _{k}_{0} is the output vector of the input-layer in the neural network, and _{1} is the output vector of a hidden-layer neuron. _{1} is a weight-sum vector, and _{1} is the bias vector of the hidden-layer output vector. _{2} indicates the output vector of the output-layer in the neural network. _{k}_{1} is a nonlinear function of the hidden-layer neurons. The hidden-layer function is a log-sigmoid function and could be expressed as in

In this case, n is pure linear input value. The original network utilized multiple layers of weight-sum units of the type _{1} = _{1}(_{1}_{0} + _{1}), where f was a sigmoid function or logistic function such as used in logistic regression. Training was done by a form of stochastic gradient descent. The use of the chain rule of differentiation in deriving the appropriate parameter updates results in an algorithm that seems to “back propagate errors”, hence the nomenclature. However it is essentially a form of gradient descent. Determining the optimal parameters in a model of this type is not trivial, and local numerical optimization methods such as gradient descent can be sensitive to initialization because of the presence of local minima of the training criterion:

The term _{2} is the sensitivity units of the output layer neuron, and the _{2} is a linear output vector of the output layer neurons. The term _{k}_{k}_{2}(_{2}) is a derivative of the nonlinear function of the output layer neuron. _{1} is a sensitivity unit of the input layer neuron, and _{1}(_{1}) is a derivative of the input layer neuron. Lastly, weighted value and bias are updated by the steepest descent rule. It could be formulated as

_{m}_{m}_{m}_{m}

In order to evaluate the effectiveness of the proposed method, and using ther measurement system represented in

Since the measurement system described in

To compare the results of the common Kalman method and the improved Kalman filter, in this section simulations were examined with an assumed measurement noise covariance. The Kalman filter can provide optimal solutions if the system model is correctly defined and the noise statistics for the measurement and system are completely known [

The R value of the Kalman filter influences the weight that the filter applies between the existing process information and the latest measurements. Mismatch in any of them may result in the filter being suboptimal or even cause divergence, as shown in

The measurement noise covariance R is a most significant factor when designing a Kalman filter. As shown in

In order to evaluate the performance of the proposed method, we also conducted some simulations of a Kalman filter with computed measurement noise covariance using a neural network in this section. A complex and large neural network performs better calculations, but its calculation time is long. In the simulation of the proposed method, to decrease the calculation time the structure of the neural network is composed of a simple three-layer feed forward neural network with one hidden layer which is the most widely spread architecture type. The input layer, hidden layer and output layer had 3, 5 and 3 neurons, respectively. The activation function of the hidden nodes is chosen to be a sigmoid type function and the output nodes are linear. The design factor of a neural network can change the performance of the system and can be determined through trial and error. In this simulation, the learning rate and initial bias of neural network were set at 0.01 and 1, respectively. A random value between −1 and 1 is employed as the initial weight which was computed by a trial and error method. The input vector of the neural network is determined by the three previous measurement values. Variances of 0.01625124 at the temperature sensor, 0.00166088 at the humidity sensor and 0.00165041 at the oxygen sensor were used as target values of each neural network, respectively. The measurement noise variance of each sensor was calculated from several measurement data which was varied by the sensing environment.

As shown in

The MSE between measured data and

The experiments were conducted with an EVB90129 connected with a TMS320F28X EVM with a DSPF2812 microprocessor to evaluate the proposed method in the implemented RFID sensor tag. The temperature, humidity and oxygen sensors are connected to the ADC port on the EVM board. A Serial Peripheral Interface (SPI) was used to interlock between EVB90129 and the EVM. The data from the multi-sensing environment is obtained from the organized module. In the EVM, the sensor data was filtered by the improved Kalman filter implemented on the EVM. The filtered data is transmitted to the EVB90129 and sent to a RFID reader.

The sensor data from the designed module was sent into the EVM. The improved Kalman filter which has simple neural network structure to evaluate the measurement noise covariance was realized in the EVM. Previous measurement data was an input of the neural network and the target value is the same as the simulation conditions.

In order to obtain accurate sensor data in a multi-sensing environment and prevent divergence of the Kalman filter caused by disturbances of the measurement environment, we have proposed an improved Kalman filter which can estimate its measurement noise covariance using a neural network. The improved Kalman filter is realized with a neural network to estimate measurement noise covariance for preventing divergence of the Kalman filter and reduction of the measurement noise. The target value of the neural network was computed from a large number of measurement data in a multi-sensing environment, and the input is the previous measurement data. The proposed method was applied to reduce measurement noise and prevent divergence with some simulations and experiments. In the simulations of multi-sensing environments, the Kalman filter method and its divergence condition was compared to the improved Kalman filter which was proposed in this paper. From several simulation results and experimental results, the performance of the improved Kalman filter is good as excellent as those of the Kalman filter. The MSE of the improved Kalman filter were 2.4633 × 10^{−4} mV^{2}, 3.7143 × 10^{−4} mV^{2} and 0.00105 mV^{2} lower than previous method with temperature sensor, humidity sensor and oxygen sensor, respectively, under the simulation conditions. The experimental results show that the MSE of the improved Kalman filter were 1.6589 × 10^{−4} mV^{2}, 2.5203 × 10^{−4} mV^{2} and 0.0.00102 mV^{2} lower than those of the common Kalman filter with a temperature sensor, humidity sensor and oxygen sensor, respectively.

This research was supported by the Agriculture Research Center program of the Ministry for Food, Agriculture, Forestry and Fisheries, Korea.

Measurement system configuration based on EVB 90129.

Compare sensor data between the single (solid line) and multi-sensing environment (dotted line).

The operation of the Kalman filter.

Neural network for evaluating measurement noise covariance.

Simulation results of common Kalman filter.

Simulation results of the Kalman filter with divergence condition.

Simulation results of the improved Kalman filter.

Configuration of the system for experiments.

Experimental results of the improved Kalman filter.

MSE between measured data and the

Common Kalman method | 6.1445 × 10^{−4} (mV^{2}) |
6.1453 × 10^{−4} (mV^{2}) |
0.0012 (mV^{2}) |

Kalman method in divergence condition | 0.0013 (mV^{2}) |
0.0011 (mV^{2}) |
0.0063 (mV^{2}) |

Improved Kalman method | 3.6812 × 10^{−4} (mV^{2}) |
2.4310 × 10^{−4} (mV^{2}) |
0.00015 (mV^{2}) |

MSE between measured data and the

4.4856 × 10^{−4} (mV^{2}) |
3.6250 × 10^{−4} (mV^{2}) |
0.00018 (mV^{2}) |