Applying RFID and NLP for efficient warehouse picking

RFID Indoor positioning technology of the system

RFID technology refers to the use of wireless radio frequency signals as a carrier for information transmission, and the exchange of information between readers and tags to achieve applications. RFID is a type of information exchange that utilizes wireless communication technology. Its characteristics are non-contact, fast, and reliable. RFID systems have the advantages of faster reading and writing speed, longer working time, and easier reading and writing. They can also recognize multiple objects and have broad development space (Li et al., 2020). The RFID basic structure is shown below.

From Fig. 1, the RFID system includes a reader, a label, and a computer control network. Among them, the reader is a device in the RFID system used to read and write information about RFID tags that communicate with it. A tag is a device on an identified object in an RFID system. When a tag communicates with a reader within a certain range, the tag processes the received signal and replies with data about the tag itself and the object it represents. The computer control network is a core component of the RFID system, used to control and manage the operation of the RFID system. These three parts work together to form a complete RFID system (Zhang, Kane & Wang, 2019). The reader reads the data of the label, the computer control network processes and manages it, and takes corresponding measures according to the needs. This system architecture provides enterprises and organizations with more efficient, accurate, and intelligent logistics management and control capabilities.

OPEN IN VIEWER

Fig. 1

The Basic Structure of RFID System

The RFID indoor positioning is a technique that utilizes the RFID technology to determine the indoor location by leveraging the interaction between tags and readers within RFID systems. RFID indoor positioning requires the deployment of RFID tags, which can be attached to objects or fixed in specific locations. The position of the tags can be selected and set according to positioning needs and scene characteristics. Then, the reader-tag interaction is carried out, followed by information processing and calculation. After receiving the information returned by the tag, the reader will transmit this information to the computer control network for subsequent processing. Computer control networks can process and calculate the received signals through algorithms and technologies to achieve indoor positioning. Then, positioning calculations can be carried out to calculate the distance and orientation of the tag relative to the reader by calculating the signal transmission time and strength between the tag and multiple readers. Finally, the indoor location calculation can be used to calculate the tag’s indoor location based on the relative distance and orientation information between the tag and the reader (Zan, 2022) (Yike, Pengyuan & Yonghong, 2019).

NLP-based term speech recognition

NLP-based term speech recognition is an innovative technology that combines the capabilities of speech recognition and NLP, providing a new way to understand and process speech information. Speech recognition technology has achieved tremendous success, and the NLP-based term speech recognition has been widely applied in different fields. It can improve user experience, improve operational efficiency and accuracy, and has broad development prospects.

This study mainly focuses on speech recognition of terms, so it mainly focuses on word segmentation and semantic analysis. The core of dictionary-based word segmentation is the dictionary and rules. A dictionary is like a dictionary, which contains indexes and body text. The difference lies in its query process. It saves words or combinations of words that appear in the segmented corpus in the dictionary and then matches them through query mechanisms to complete word segmentation, such as the maximum matching algorithm and minimum word segmentation algorithm. Among them, in the general field, the maximum matching segmentation algorithm has a 99% accuracy and the widest application range (Malek & Bohac, 2020).

The maximum matching segmentation algorithm is a basic segmentation method used in Chinese NLP, mainly divided into three types: forward, backward, and bidirectional. Forward maximum matching and backward maximum matching are two commonly used word segmentation algorithms in Chinese NLP. Their main difference lies in scanning and matching from the beginning and end of the text to be processed. The error rate of forward and reverse maximum matching segmentation algorithms is as low as 0.6%, but the latter is greatly limited by the size of the vocabulary. Therefore, the research will utilize the forward maximum matching segmentation algorithm to form a segmentation algorithm that is more suitable for term speech recognition. Before that, it is necessary to first build a domain dictionary, which will use two methods: extraction tools and manual direct addition. The specific process is depicted below.

From Fig. 2, the collected terminology corpus is first preprocessed, then extracted by a vocabulary extraction tool, and finally filtered manually to construct a domain dictionary. Term extraction software can be used to automatically extract terminology corpus. The vocabulary is initialized to facilitate queries on words and word attributes. Arrays dictionary [ ] and hash [ ] are defined. The former mainly stores all words, while the latter mainly stores words with consistent key values. Firstly, the first word w and length L of the word in the vocabulary are obtained, the hash function is used to calculate the w corresponding hash value, and the value is obtained. The expression is shown in Equation (1).

v = w + L

(1)

OPEN IN VIEWER

Fig. 2

The construction process of domain dictionaries

In Equation (1), v represents the value. Assuming that key = v %n in the hash function, due to the small size of the vocabulary, n = 300 is assigned to make the data distribution more uniform (Chen, 2021). Then the key values of the words are combined and stored in the corresponding hash[key], following this process until all the words are stored in the hash[]. Then, the forward maximum matching algorithm is used for word segmentation, part of speech tagging, and stopping word filtering, as shown below.

From Fig. 3, after initialization, all words with the same initial word and word count are stored in the same linked list. During the segmentation process, when encountering scenarios with no logged-in words or single-character words, the algorithm efficiently filters out stop words or identifies keywords by traversing the dictionary only once. This approach reduces the number of searches and significantly decreases the algorithm’s running time (Dujmešić, Bajor & Rožić, 2018).

OPEN IN VIEWER

Fig. 3

Forward maximum matching word segmentation algorithm process

After completing the segmentation, semantic analysis can be performed. Semantic templates can abstract database terminology into a model and divide it into two types: conceptual and relational, with sentences represented by conceptual relational models (Zhong et al., 2022). The study will utilize concepts, their attributes, and interrelationships to form a semantic template and calculate the probability parameters of language statistical models based on statistical semantic templates to determine the likelihood of language components appearing in sentences.

The semantic template is composed of 2 tuples 〈M, N〉, with M and N representing keywords and keyword weight sets respectively (Hou et al., 2020). In the implementation process of semantic analysis based on the semantic template, it is necessary to first input segmentation text, then extract vocabulary, complete the matching of the semantic template, and finally obtain the results and output them. Semantic analysis based on semantic templates utilizes keyword matching to understand the semantics of sentences. The algorithm first stores some basic semantic templates with N values, each corresponding to a semantic. Then, it extracts vocabulary from the segmented text and matches them one by one with the template. If the matching is successful, the corresponding semantics will be output. Except for the keyword part, other parts have no impact on semantics, so they will not be considered.

Intelligent warehouse picking based on LANDMARC algorithm and speech recognition

RFID-based indoor positioning algorithms are mainly divided into range-based and non-range-based positioning algorithms. The latter is currently the most popular indoor positioning method on the market due to its small impact on multipath effects (Peng, Zhao, & Jiang, 2019). The LANDMARC algorithm is a relatively mature algorithm among non-ranging positioning algorithms. It is based on the Received Signal Strength Indicator (RSSI) of wireless signals and achieves the localization of moving objects in indoor environments by establishing an association model between RSSI and location. The LANDMARC structure is shown below.

In Fig. 4, the tag to be located can be connected to an RFID reader through a reference tag. In the LANDMARC algorithm, it is assumed that there is n RFID readers, m reference tags, and u tags to be tested (Legowo & Wijaya, 2022). When positioning, the reader first sends out an RF signal, and within the RF range, the tag returns the RF information identifying itself, and the reader obtains the strength value. Assuming the obtained signal strength vector of the tag to be tested is S → = ( S 1 , S 2 , . . . , S n ) , S _i , i ∈ (1, n) is the RSSI of the tag to be tested on the second reader. Assuming that the signal strength vector corresponding to the reference label is θ → = ( θ 1 , θ 2 , … , θ n ) , θ _i , i ∈ (1, n) is the RSSI of the reference label on the second reader. As for all the labels p to be tested, p ∈ (1, u), and the expression is shown in Equation (2).

E j = ∑ i = 1 n ( θ i - S i ) 2 , j ∈ ( 1 , m )

(2)

OPEN IN VIEWER

Fig. 4

Schematic diagram of LANDMARC structure

In Equation (2), E _j is used to describe the Euclidean distance between the reference label and the label to be tested. The smaller the value, the closer the signal strength of the two labels is, that is, the closer the actual distance is. Among the m reference labels, E → = ( E 1 , E 2 , . . . , E m ) , all the tested labels are sorted in ascending order based on the Euclidean distance obtained from all guessed labels, resulting in the smallest first k reference label of E _j . The position of the label to be tested is calculated using the k reference label coordinates. The coordinates (x, y) of the label to be tested can be obtained using Equation (3).

( x , y ) = ∑ i = 1 k w i ( x i , y i )

(3)

In Equation (3), w _i is used to describe the coordinate weight value of the i nearest neighbor label, and the weight value of the k nearest neighbor label is 1. Each reference label must be assigned values according to a certain pattern. From the relationship between signal strength and transmission distance, the closer the distance is, the greater its reference weight. The calculation is shown in Equation (4).

w l = 1 E l 2 ∑ i = 1 k ( 1 E l 2 )

(4)

In Equation (4), when E _j is the smallest, its weight is the highest. Next, the positioning error distance e is located, which means the algorithm estimates the geometric distance between the coordinates (x, y) and the time coordinates (x₀, y₀) of the tag to be tested, as shown in Equation (5).

e = ( x - x 0 ) 2 + ( y - y 0 ) 2

(5)

The LANDMARC algorithm is influenced by various factors when using k nearest neighbors, but k size plays a crucial role in the LANDMARC positioning accuracy. Scholars have found that the cumulative error of the tested label is the most ideal (Pier & Weidinger, 2022). In the LANDMARC algorithm, a reference label is added, which does not require distance measurement in the environment, thus achieving higher positioning accuracy. Typically, when the location of the items to be picked is known, the k-nearest neighbor labels of the items are utilized as a reference to identify the operator nearest to the items. This approach helps minimize positioning errors resulting from selecting the wrong neighbor label in LANDMARC.

In LANDMARC, assuming there are n readers, m reference tags, and u tags to be located, the signal strength matrix of u tags to be located concerning n readers is obtained as shown in Equation (6).

S = [ S 11 ⋯ S 1 u ⋮ ⋱ ⋮ S n 1 ⋯ S nu ]

(6)

Then four reference labels are obtained on rectangular fixed points surrounding the goods to be picked, and their signal strength matrix expressions for n readers are shown in Equation (7).

S = [ θ 1 a θ 1 b θ 1 c θ 1 d ⋮ ⋮ ⋮ ⋮ θ na θ nb θ nc θ nd ]

(7)

When the goods to be picked are within the box area enclosed by the four reference labels ③, ④, ⑦, and ⑧, as shown in Fig. 5.

In Fig. 5, if the goods to be picked are located near reference labels 3, 4, 7, and 8, then reference labels 3, 4, 7, and 8 can be used as neighbor labels. Equation (8) can be used to calculate the E _j values of these four neighbor labels and the test label, which is the Euclidean distance.

E j = ∑ i = 1 n ( θ i - S i ) 2 , j = 3 , 4 , 7 , 8

(8)

OPEN IN VIEWER

Fig. 5

Schematic diagram of indoor positioning layout for voice warehouse sorting

Then, based on the distance relationship between the reference label and the goods to be picked, the E _j values of each label to be picked are calculated, as shown in Equation (9).

M j = ∑ k j · E j , j = 3 , 4 , 7 , 8

(9)

The calculation of k _j in Equation (9) can be obtained by Equation (10).

k j = 7 - d j 7 n 7 - d 3 7 n + 7 - d 4 7 n + 7 - d 7 7 n + 7 - d 8 7 n , j = 3 , 4 , 7 , 8

(10)

In Equation (10), is the corresponding distance. The smallest label is selected, which is the operator closest to the goods to be picked. Then, using speech recognition, staff can use voice commands to interact with the system, such as indicating the location, quantity, or special requirements for picking goods. The system will convert voice commands into text and perform corresponding processing and execution. This speech recognition technology can reduce manual input and search time in warehouse operations and improve operation speed and accuracy. It can also improve the work experience of staff and reduce fatigue and errors. By integrating voice technology with warehouse picking systems, more intelligent and efficient warehouse management and operation can be achieved. In addition, it can also adapt to the needs of diversity and flexibility, providing solutions for different warehousing environments and work scenarios.

Warehouse picking performance analysis

Firstly, the forward maximum matching segmentation algorithm performance was verified. This algorithm, combined with the hash function, was beneficial for reducing search times and running time, and improving the efficiency of the algorithm. Due to the overall small size of the corpus, this study did not separate the training and testing sets. To maximize the training corpus, cross-validation was used for testing. That was to randomly divide the 500-sentence sample into 50 copies, each with 10 sentence samples, and the total number of Chinese characters included was close (Ali, Elrouby & Haddad, 2022). During the testing process, 20 copies were selected for testing, and the remaining 30 copies for training. Two different methods were used to segment sentences, namely the forward maximum matching segmentation algorithm and the traditional segmentation algorithm. The two algorithms’ running time and processing speed were compared in Fig. 6.

From Fig. 6, after combining the hash function, the overall running time of the forward maximum matching segmentation algorithm was reduced by 50% compared to traditional segmentation algorithms. When there were fewer stop words, the algorithm ran relatively less time. The emergence of stop words had a greater impact on the running time of traditional algorithms, as it required traversal of the word library, which consumed more time. After combining with the hash function, the forward maximum matching word segmentation algorithm only needed to use one operation to remove the stop word when it appeared, greatly reducing the impact of the stop word on the algorithm’s operation time (Gajšek et al., 2022). Meanwhile, the overall processing speed of the forward maximum matching segmentation algorithm was significantly faster than traditional segmentation algorithms.

OPEN IN VIEWER

Fig. 6

Running time of word segmentation algorithm using forward maximum matching

Then, the speech recognition performance based on NLP algorithms was verified and compared with traditional speech recognition methods. The accuracy, recall, and F1 values of the two methods were compared, as shown in Fig. 7.

OPEN IN VIEWER

Fig. 7

Speech recognition performance based on NLP algorithms

In Fig. 7, the accuracy of traditional combined speech recognition methods was between 60% –70% . However, due to its loose distribution and a higher concentration within smaller ranges, a large amount of the corpus goes unrecognized. Speech recognition based on NLP algorithms had a more centralized distribution and the addition of domain word banks, greatly reducing the number of unrecognized corpus and improving accuracy, and the accuracy was concentrated between 90% –95% . The accuracy, recall, and F1 values of traditional combined speech recognition methods were 66.72% and 61, respectively. 23%, 63%. 45%, the accuracy, recall, and F1 values of speech recognition using NLP were 90.15%, 92.32%, and 91.67%, respectively. Compared with traditional methods, the accuracy of speech recognition using NLP improved by 23.43% . Speech recognition based on NLP algorithms was more suitable for application in speech recognition, which ensured the accuracy of the application effect. This indicated the efficiency of NLP technology in processing and understanding speech data. The NLP method effectively reduced the number of unrecognized corpora, and improved accuracy, recall, and F1 value through the use of domain lexicons and a more refined algorithm design. This improvement meant that speech recognition became more reliable and better captured and understood user voice commands, especially in specific fields. This progress was of great significance for improving user experience and developing more efficient voice interaction systems in specific fields.

Finally, an analysis was conducted on the CPU usage during the warehouse picking process using the LANDMARC algorithm and speech recognition based on RFID indoor positioning technology, and it was compared with conventional warehouse picking. The results are shown in Fig. 8.

From Fig. 8, there was little difference in CPU usage between the LANDMARC algorithm and speech recognition warehouse picking process based on RFID indoor positioning technology and conventional warehouse picking. The research method can ensure improved accuracy and efficiency while not occupying too much CPU memory, indicating good overall performance.

OPEN IN VIEWER

Fig. 8

CPU usage of warehouse picking operation based on RFID indoor positioning technology and speech recognition

Warehouse picking simulation analysis

This study conducted a simulation analysis on the LANDMARC algorithm and speech recognition warehouse picking based on RFID indoor positioning technology. Firstly, the simulation environment was set up below.

OPEN IN VIEWER

Table 1

Environment for simulation experiments

Project	Name	Function/Types
Hardware devices	RFID reader/writer	Support indoor positioning
	RFID tags	Attached to items that
		require positioning
	Microphone	Speech recognition
	Speaker	Voice indication
Software environment	Install RFID indoor positioning	LANDMARC algorithm
		software
	Install speech recognition	Speech Recognition Based
	software	on NLP Algorithms
Warehouse layout design	Arrange shelves, cargo storage areas, and picking areas, install RFID antennas to cover the entire positioning range of the warehouse, affix RFID tags on each shelf, and assign a unique RFID tag to each storage unit.

In the warehouse layout, it was necessary to arrange shelves, cargo storage areas, and picking areas, install RFID antennas to cover the entire positioning range of the warehouse, affix RFID tags on each shelf, and assign a unique RFID tag to each storage unit. The specific simulation design of the logistics warehouse picking layout is shown in Fig. 9.

OPEN IN VIEWER

Fig. 9

Simulation Design of Logistics Warehouse Picking Layout

From Fig. 9, within an 8 m×8 m area, there were 4 readers, 20 reference tags, 7 test tags (operators), and 10 items to be picked. The distance between adjacent reference tags in each row was 2 m, and there was 1.5 m between the two rows. Subsequently, simulations were conducted on the research method and traditional method for warehouse picking, and the results are shown in Fig. 10.

OPEN IN VIEWER

Fig. 10

Simulation Design of Logistics Warehouse Picking Layout

Figures 10 (a) and 10 (b) show the strict positioning accuracy and effective positioning accuracy. Compared with traditional methods, the research method’s accuracy was greatly improved. For example, the strict positioning accuracy of research methods and traditional methods for picking goods 1 were 0.57 and 0.38, respectively, and the effective positioning accuracy was 0.80 and 0.70, respectively. In addition, in experiments, when operators made mistakes, they were often located closest to the position of the goods to be picked. Therefore, the positioning accuracy of the system can satisfy picking and positioning. In the case of strict positioning, the goods to be picked with relatively low positioning accuracy had a significant improvement in effective positioning, and the final effective positioning accuracy of the research was relatively high or the same. The research method ensured high precision and effective positioning accuracy, thereby ensuring the efficiency of warehouse sorting. In terms of effective positioning, both methods had high accuracy, proving the applicability and reliability of these technologies under general conditions. This is very useful in ordinary warehouse operations and daily item picking and selection.