An Online RFID Localization in the Manufacturing Shopfloor
Radio Frequency Identification technology has gained popularity for cheap and easy deployment. In the realm of manufacturing shopfloor, it can be used to track the location of manufacturing objects to achieve better efficiency. The underlying challenge of localization lies in the non-stationary characteristics of manufacturing shopfloor which calls for an adaptive life-long learning strategy in order to arrive at accurate localization results. This paper presents an evolving model based on a novel evolving intelligent system, namely evolving Type-2 Quantum Fuzzy Neural Network (eT2QFNN), which features an interval type-2 quantum fuzzy set with uncertain jump positions. The quantum fuzzy set possesses a graded membership degree which enables better identification of overlaps between classes. The eT2QFNN works fully in the evolving mode where all parameters including the number of rules are automatically adjusted and generated on the fly. The parameter adjustment scenario relies on decoupled extended Kalman filter method. Our numerical study shows that eT2QFNN is able to deliver comparable accuracy compared to state-of-the-art algorithms.
Radio Frequency Identification (RFID) technology has been used to manage objects location in the manufacturing shopfloor. It is more popular than similar technologies for object localization, such as Wireless Sensor Networks (WSN) and WiFi, due to the affordable price and the easy deployment (Ni et al, 2011; Yang et al, 2016).
In the Maintenance, Repair, and Overhaul (MRO) industry, for example, locating the equipments and trolleys manually over the large manufacturing shopfloor area results in time-consuming activities and increases operator work-load. Instead of manually locating the tool-trolleys, RFID localization is utilized to monitor the location real-time. Despite much work and progress in RFID localization technology, it is still challenging problems. The key disadvantage of RFID is that it has low quality signal, which is primarily altered by the complexity and severe noises in the manufacturing shopfloors (Chai et al, 2017).
Generally, an RFID localization system comprises of three components, i.e. RFID tag, RFID reader, and the data processing subsystem. The reader aims to identify the tag ID and obtain the received signal strength (RSS) information from tags. An object’s location can be estimated by observing the RSS. However, the RSS quality in the real-world is very poor, moreover, it keeps changing over time. As an illustration, although RFID tag is used in the static environment, the RSS keeps changing over time. Moreover, a minor change in the surrounding area can greatly fluctuate the RSS. The major factors causing the phenomena are multipath effect and interference. Therefore, obtaining the accurate location relying on RSS information is a hard task. In several works, those challenges are tackled by employing computational techniques, thus the objects location can be achieved accurately (Ni et al, 2011).
There are several techniques which can be utilized to estimate the object’s location. First of all, the distance from an RFID tag to an RFID reader can be calculated via the two-way radar equation for a monostatic transmitter. It can be obtained easily by solving the equation. Another approach, LANDMARC, is proposed for the indoor RFID localization (Ni et al, 2004). It makes us of reference tags, and then it evaluates the RSS similarity between reference tags and object tags. A higher weight will be assigned to the reference tags which posses the similar RSS information to the object tags. In the realm of machine learning, support vector regression (SVR) is implemented for the indoor RFID localization (Chai et al, 2017). It is one-dimensional method and is designed for stationary objects in the small area. Another way to obtain better accuracy is by employing Kalman filter (KF), it has been demonstrated to deal with wavelength ambiguity of the phase measurements (Soltani et al, 2015).
The strategy to estimate the object location via the so-called radar equation is easy to execute. However, the chance to obtain acceptable accuracy is practically impossible due to the severe noises. LANDMARC manage to improve the localization accuracy. Nonetheless, it is difficult to select the reference tag properly in the industrial environments where interference and multipath effect occurred. Improper selection of reference tags can alter the localization accuracy. Similarly, SVR is also designed to address object localization problem in the small area, it even encounters an over-fitting problem. Meanwhile, integrating KF in some works can improve localization accuracy and it also has low computational cost. KF has better robustness and good statistical properties. However, the requirement to calculate the correlation matrix burdens the computation. In addition, the use of KF is also limited by the nonlinearity and non-stationary condition of the real-world (Oentaryo et al, 2014; Chai et al, 2017). The data generated from a non-stationary environment can be regarded as the data stream (Pratama, 2017).
Evolving intelligent system (EIS) is an innovation in the field of computational intelligent to deal with data stream (Angelov and Zhou, 2006; Lughofer, 2008). EIS has an open structure, it implies that it can starts the learning processes from scratch or zero rule base. Its rules are automatically formed according to the data stream information. EIS adapts online learning scenario, it conducts the training process in a single-pass mode (Pratama et al, 2016b). EIS can either adjusts the network parameters or generates fuzzy rule without retraining process. Hence, it is capable to deal with the severe noises and the systems dynamics (Pratama et al, 2014b; Lughofer et al, 2015; Pratama et al, 2016b). In several works, the gradient descent (GD) is utilized to adjust the EIS parameters. However, it is vulnerable to noises due to its sensitivity (Oentaryo et al, 2014). In the premise part, the Gaussian membership function (GMF) is usually employed to capture the input features of EIS. The drawback of GMF is its inadequacy to detect the overlaps between classes. Several works are employed quantum membership function (QMF) to tackle the problem (Purushothaman and Karayiannis, 1997; Chen et al, 2008; Lin and Chen, 2006). Nevertheless, it is type-1 QMF which is lack of robustness to deal with uncertainties in real-world data streams (Pratama et al, 2016b).
This research proposes an EIS, namely evolving Type-2 Quantum Fuzzy Neural Network (eT2QFNN). The eT2QFNN adopts online learning mechanism, it processes the incoming data one-by-one and the data is discarded after being learned. Thus, eT2QFNN has high efficiency in terms of computational and memory cost. The eT2QFNN is encompassed by two learning policies, i.e. the rule growing mechanism and parameters adjustment. The first mechanism enables eT2QFNN to start the learning process from zero rule base. It can automatically add the rule on demands. Before a new rule is added to the network, it is evaluated by a proposed formulation, namely modified Generalized Type-2 Datum Significance (mGT2DS). The second mechanism performs parameters adjustment whenever a new rule is not formed. It aims to keep the network adapted to the current data stream. This mechanism is accomplished by decoupled extended Kalman filter (DEKF). It is worth noting that the DEKF algorithm performs localized parameter adjustment, i.e. each rule can be adjusted independently (Oentaryo et al, 2014). In this research, the adjustment is only undertaken on a winning rule, i.e. a rule with the highest contribution. Therefore, DEKF is more efficient than extended Kalman filter (EKF), and yet it still preserves the EKF performance. On the premise part, the interval type-2 QMF (IT2QMF) is proposed to approximate the desired output. It worth noting that it is a universal function approximator which has been demonstrated by several researchers(Purushothaman and Karayiannis, 1997; Lin and Chen, 2006; Chen et al, 2008). Moreover, it is proficient to form a graded class partition, such that the overlaps between classes can be identified (Chen et al, 2008).
The major contributions of this research are summarized as follows: 1) This research proposes the IT2QMF with uncertain jump position. It is the extended version of QMF. That is, IT2QMF has the interval type-2 capability in terms of incorporating data stream uncertainties. 2) The eT2QFNN is equipped with rule growing mechanism. It can generate its rule automatically in the single-pass learning mode, if a condition is satisfied. The proposed mGT2DS method is employed as the evaluation criterion. 3) The network parameter adjustment relies on DEKF. The mathematical formulation is derived specifically for eT2QFNN architecture. 4) The effectiveness of eT2QFNN has been experimentally validated using real-world RFID localization data.
The remainder of this book chapter is organized as follows. Section II presents the RFID localization system. The proposed type-2 quantum fuzzy membership function and the eT2QFNN network architecture are presented in section III. Section IV presents the learning policies of eT2QFNN and the DEKF for parameter adjustment are discussed. Section V provides empirical studies and comparisons to state-of-the-art algorithms to evaluate the efficacy of eT2QFNN. Finally, section VI concludes this book chapter.
0.2 RFID Localization System
RFID localization technology has three major components, i.e. RFID tags, RFID readers, and data processing subsystem. There are two types of RFID tags, i.e. the active and passive tag. The active RFID tag is battery-powered and has its own transmitter. It is capable of sending out beacon message, i.e. tag ID and RSS information, actively at specified time window. The transmitted signal can be read up to radius. In contrast, passive tag does not have independent power source. It exploits the reader power signal, thus it cannot actively send the beacon message in a fixed period of time. The signal can only be read around radius from the reader, which is very small compared to the manufacturing shopfloor. After that, the transmitted signal is read by the RFID reader. And then it is propagated to the data processing subsystem where the localization algorithm is executed (Ni et al, 2011). The configuration of RFID localization is illustrated in the Fig. 0.1.
The RSS information can be utilized to estimate the object location. It can be achieved by solving the two-way radar equation for a monostatic transmitter as per (0.1). The variables are explained as follows. , , , , and are the reader signal power, the antenna gain, carrier wavelength, tag radar cross-section, and the distance between reader and tag, respectively. However, due to the occurrence of multipath effect and interference, the RSS information becomes unreliable and keeps changing over time. Consequently, the satisfying result cannot be obtained (Ni et al, 2011; Chai et al, 2017). Instead of employing (0.1) to locate the RFID, this research utilizes eT2QFNN to process the RSS information to obtain the precise object location.
One of many objectives of this research is to demonstrate the eT2QFNN to deal with the RFID localization problem. The RSS informations of reference tags are utilized to train the network. The reference tags are tags placed at several known and static positions. The eT2QFNN can estimate the new observed tags location according to its RSS. It worth noting that eT2QFNN learning processes are achieved in the evolving mode, it keeps the network parameters and structure adapted to the current data stream. This benefits the network to deal with multipath effect and interference occurred in the manufacturing shopflor.
Now suppose there are reference tags which are deployed in locations, then the RSS measurement vector at th time-step can be expressed as . Afterward, the network outputs for RFID localization problem can be formulated into a multiclass classification problem. As an illustration, if there exist reference tags deployed in the shopfloor it will indicates that the number of classes is equal to 4.
0.3 eT2QFNN Architecture
This section presents the network architecture of eT2QFNN. The network architecture, as illustrated in the Fig. 0.2, consists of a five-layer, multi-input-single-output (MISO) network structure. It is systematized into input features, outputs nodes and -term nodes for each input feature. The rule premise is compiled of IT2QMF and is expressed as follows:
where and are the th input feature and the regression output of the th class in the th rule, respectively. denotes the set of upper and lower linguistic term of IT2QMF, is the extended input and is the set of upper and lower consequent weight parameters which are defined as .
The membership function applied in this study is different from the typical QMF and GMF. The QMF concept is extended into interval type-2 membership function with uncertain jump position. Thus, the network can identify overlaps between classes and capable to deal with the data stream uncertainties. The IT2QMF output of the th rule for the th input feature is given in (0.3)
where , , and are mean of th input feature in th rule, slope factor, and number of grades, respectively. is the set of uncertain jump position, it is defined as . The upper and lower jump position is expressed as and . It is defined that , thus the execution of (0.3) leads to interval type-2 inference scheme which produces a footprint of uncertainties (Pratama et al, 2016b), it can be clearly seen in the Fig. 0.3. The eT2QFNN operation in each layer is presented in the following passages.
Input layer This layer performs no computation. The data stream is directly propagated to the next layer. The input at th observation is defined by . And the output of the th node is given as follows:
Rule layer This layer functions to combine the membership degree of th rule denoted as , which is known as spatial firing strength. This can be achieved by employing product T-norm of IT2QMF, as per (0.7) and (0.8). The set of upper and lower firing strengths are expressed as and , respectively.
Output processing layer The calculation of the two endpoints output, i.e. and , is conducted here. These variables represent the lower and upper crisp output of the th class, respectively. The design factor are employed to convert the interval type-2 variable to type-1 variable, this is known as the type reduction procedure. This requires less iterative steps compare to the Karnik Mendel (KM) type reduction procedure (Pratama et al, 2016a). The design factor will govern the proportion of upper and lower IT2QMF and it is defined such that . The design factor is adjusted using DEKF, thus the proportion of upper and lower outputs keeps adapting to the data streams uncertainties. The lower and upper outputs are given as:
where , are the design factors of all classes, while , and express the upper and lower consequent weight parameters of the th rule for the th class. In addition, is the extended input vector. For example, has input features , then the extended input vector is . The entry 1 is included to incorporate the intercept of the rule consequent and to prevent the untypical gradient (Pratama et al, 2016b).
Output layer The crisp network output of the th class is the sum of and as per (0.11). Furthermore, if the network structure of eT2QFNN is utilized to deal with multiclass classification, the multi-model (MM) classifier can be employed to obtain the final classification decision. The MM classifier splits the multiclass classification problem into binary sub-problems, then MISO eT2QFNN is built accordingly. The final class decision is the index number of the highest output, as per (0.12).
0.4 eT2QFNN Learning Policy
The online learning mechanism of eT2QFNN consists of two scenarios, i.e. the rule growing and the parameter adjustment which is executed in every iteration. The eT2QFNN starts it learning process with an empty rule base and keeps updating its parameters and network structure as the observation data comes in. The proposed learning scenario is presented in the Algorithm 1, while subsections 4.1 and 4.2 further explain the learning scenarios.
0.4.1 Rule growing mechanism
The eT2QFNN is capable of automatically evolving its fuzzy rule on demands using the proposed mGT2DQ method. First of all, it is achieved by forming a hypothetical rule from a newly seen sample. The initialization of hypothetical rule parameters is presented in the sub-subsection 0.4.2. Before it is added to the network, it is required to evaluate its significance. The significance of the th rule is defined as an of weighted by the input density function as follows (Huang et al, 2005):
From the (0.13), it is obvious that the input density greatly contributes to . In practical, it is hard to be calculate a priori because the data distribution is unknown. Huang et al (2003) and Huang et al (2004) calculated (0.13) analytically with the assumption of being uniformly distributed. However, Zhang et al (2004) demonstrated that it leads to performance degradation for complex . To overcome this problem, Bortman and Aladjem (2009) proposed Gaussian mixture model (GMM) to approximate the complicated data stream density. The mathematical formulation of GMM is given as:
where is the Gaussian function of variable as per (0.15), with the mean vector , variance matrix , denotes the number of mixture model, and represent the mixing coefficients ().
In the next step, the estimated significance of th rule is calculated. Vuković and Miljković (2013) derived the mathematical formulation to obtain by substituting (0.14) to (0.13) and then solving the closed form analytical solution, it yields to the following result:
where is the vector of GMM mixing coefficients, denotes as the positive definite weighting matrix which is expressed as , and is given as:
where is the mean vector of th rule defined as . And the GMM parameters , , and can be calculated by exploiting pre-recorded data. This technique is feasible and easy to implement because the pre-recorded input data is most likely to be stored especially in the era of data stream. The number of pre-recorded data is somewhat smaller than the training data, it is denoted as . It is not problem-specific and it can be set to a fixed value (Pratama, 2017). In this research, it is set as 50 for simplicity.
The method (0.16) could not, however, be applied directly to estimate the eT2QFNN rule significance, because eT2QFNN utilizes IT2QMF instead of Gaussian membership function (GMF). The key idea to overcome this problem is by approximating IT2QMF using interval type-2 Gaussian Membership Function (IT2GMF). The mathematical formulation of this approach can be written as follows:
The mean of IT2GMF is defined to equal the mean of IT2QMF, i.e. . And the width of upper and lower IT2GMF are obtained by taking the minimum value of as per (0.19). By selecting these criteria, the whole area of IT2GMF will be located inside the area of IT2QMF. As illustrated in the Fig. 0.4, both upper and lower area of IT2GMF covers the appropriate area of IT2QMF. Therefore, this approach can provide a good approximation of IT2QMF.
Afterward, the rule significance of eT2QFNN can be estimated by executing (0.16) with the design factor as per (0.20). In (0.21) and (0.22), and are denoted as the upper and lower consequent weight parameters of all classes, respectively. The variance matrices and are formed of and as per (0.24), while and are given in (0.25). The hypothetical rule will be added to the network as a new rule if it posses statistical contribution over existing rules. The mathematical formulation of rule growing criterion is given in (0.23), where the constant is defined as the vigilance parameter and in this research it is fixed at 0.65 for simplicity.
0.4.2 Parameter adjustment
This phase comprises of two alternative strategies. The first strategy is carried out to form a hypothetical rule. It will be added into the network structure if the condition in (0.23) is satisfied. This strategy is called the fuzzy rule initialization. And the second mechanism is executed whenever (0.23) is not satisfied. It is aimed to adjust the network parameters according to the current data stream. This is called the winning rule update. These strategies are elaborated in the following sub-subsections.
Fuzzy rule initialization
The rule growing mechanism first of all is conducted by forming a hypothetical rule according to the current data stream. The data at th time-step is assigned as the new mean of IT2QMF, as per (0.26). And then, the new jump position is achieved via distance-based formulation inspired by (Lin and Chen, 2006), as per (0.27). In this research, however, it is modified such that the new distance is obtained utilizing the mixed mean of GMM , as per (0.28). Thanks to the GMM features which is able to approximate the mean and variance of very complex input. For this reason, instead of using to calculate the jump position of the first rule, the eT2QFNN utilizes the diagonal entries of the mixed variance matrix, as per (0.30). The constant is introduced to create the footprint of uncertainty. In this study, it is set for simplicity.
In the next stage, the consequent weight parameters of hypothetical rule are determined. It is equal to the consequent weight of the winning rule as per (0.31). The key idea behind this strategy is to acquire the knowledge of winning rule in terms of representing the current data stream (Oentaryo et al, 2014). The way to select the winning rule is presented in sub-subsection (0.4.2). Finally, if the hypothetical rule passes the evaluation criterion in (0.23), it is added as the new rule () and its covariance matrix is initialized via (0.32).
In contrast, a consideration is required to adjust the covariance matrices of other rules, because the new rule formation corrupts those matrices. This phenomenon has been investigated in SPLAFIS (Oentaryo et al, 2014), the research revealed that those matrices need to be readjusted. The proper readjustment technique is achieved by multiplication of those matrices and as per (0.33). This strategy is signified to take into account the contribution that a new rule would have if it existed from the first iteration. It, therefore, will decrease the corruption effect (Oentaryo et al, 2014).
Winning rule update
The hypothetical rule would not be added to the network structure if it failed the evaluation in (0.23). To maintain the eT2QFNN performance, the network parameters is required to be adjusted according to the information provided by the current data stream. In this research, the adjustment is only undertaken on the winning rule which is defined as a rule having the highest average of the spatial firing strength. The mathematical formulation is given in (0.34). It worth noting that spatial firing strength represents the degree to which the rule antecedent part is satisfied. The rule having higher firing strength possesses higher correlation to the current data stream (Pratama et al, 2014b), therefore it deserves to be adjusted.
The DEKF is employed to adjust the network parameters of winning rule. It maintains local learning property of EIS because it can adjust parameters locally (Oentaryo et al, 2014). In this research, the local parameters are classified by rule, i.e. the parameters in the same rule are grouped together. This leads to the formation of block-diagonal covariance matrix as per (0.36). There is only one block covariance matrix updated in each time-step, i.e. . The localized adjustment property of DEKF enhances the algorithm efficiency in terms of computational complexity and memory requirements, moreover it still maintains the same robustness as the EKF (Puskorius and Feldkamp, 1994).
The mathematical formulations of DEKF algorithm are given in (0.37)-(0.39). The designation of each parameter in the equation is as follows. and are the Kalman gain matrix and covariance matrix, respectively. The covariance matrix represents the interaction between each pair of the parameters in the network. is the parameter vector of at th iteration, it consists of all the network parameters which is about to be adjusted. It is expressed as , which is respectively given in (0.40)-(0.46). The length of is equal to . The Jacobian matrix , presented in (0.47), contains the output gradient with respect to the network parameters, and it is arranged into -by- matrix. The gradient vectors are specified in (0.48) and is calculated using (0.49)-(0.52). The output and target vectors are defined as and . It is utilized to calculate the error vector in (0.39). The last parameters, , is a learning rate parameter (Puskorius and Feldkamp, 1994). This completes the second strategy to maintain the network adapted to the current data stream.