Experimental Research on Signal Recognition Algorithm of Wireless Sensor Language

—In the past several decades, much research has been carried out on the wireless sensor network which is widely used in the fields of national defense and national economy within China. The main function of the language sensor is to transfer the voice signal into an electrical signal so as to facilitate the subsequent analysis and processing. Combined with the wireless network signal, it is widely used in banks, shopping malls, examination rooms, prisons, important places of military affairs and other places. This paper makes a detailed introduction of some theoretical knowledge of language recognition and puts forward the recognition algorithm in which the language signal is abstracted by language features. Afterwards, Matlab is used to make simulation in-depth research and classification based on Support Vector Machine. Finally, a large number of samples are collected for the experiment so as to research the effect of weighted feature value and structure of classification on speech recognition rate. The conclusion of the paper provides a basis for further subsequent theoretical study.


INTRODUCTION
Wireless Sensor Network, WSN has the functions of signal collection, storage, processing, and transmission, etc. Many sensor nodes will construct a WSN which includes a variety of sensors including the ones for displaying information in the surrounding area, humidity, the size of the moving object, temperature, electromagnet, noise, earthquake, light intensity, pressure, soil composition, speed and direction. It can also be used in the battlefield, environment, disaster relief, industry and other fields [1].
Language is an important carrier of information. For understanding the real world and obtaining the features in the real world, the information can be acquired by acoustic signal as an alternative to visual observation. Therefore, as the most pervasive physical phenomenon in society, the application of language both in the battlefield and in real life has been highlighted [2]. currently. The acoustic target identification has been widely used in many fields. Currently, the recognition and application of acoustic target are based on the traditional pattern of a recognition algorithm [3]. The key to the technique is to extract some features of the signal of the acoustic target. Furthermore, the basic thought of acoustic target identification is processing the target and then identifying it so as to classify it into the possible category [4]. Research is being conducted on acoustic target in active battlefield environments.
Using wireless sensor networks to achieve language target recognition is a new technology with good application prospects. This paper uses the wireless sensor network platform to identify the language goal [5]. The target categories within the monitoring area can be displayed in real time on computers by building the language hardware platform, combining the recognition algorithm and the development and design of the upper computer software. The system can be used in an unattended military defensive environment. Moreover, this system also has wide application prospects in the civil application field, such as security monitoring, access control systems, etc [6].

II. STRUCTURE OF WIRELESS SENSOR NODES AND PROCESS OF LANGUAGE RECOGNITION
In different applications, the composition of sensor nodes is slightly different. But, the following several modules include the sensor module, wireless communication module, processing module and energy supply module, totalling four major parts, which is shown in Figure.1.  This paper adopts a wireless sensor node based on JN5139 chip. It is a wireless microcontroller with low power consumption and low energy consumption. Its wireless protocol is widely used in ZigBee and its external can be used to stock application codes which will be implemented after entering internal RAM or during the process of operation. The equipment is integrated by hardware MAC, AES encryption acceleratorprocessor, ener-  Figure 2. Voice recognition framework gy-saving and timing for sleep mechanism, as well as a mechanism for guaranteeing safety and code encryption of the program. It has a wide application and it is suitable for safe and reliable wireless applications with low power, utility of the wireless sensor network based on IEEE802.15.4 and ZigBee systems [7]. The main process of acoustic target identification includes language acquisition, pre-processing, storage, feature extraction and recognition algorithm, as shown in Figure 2.

A. Analysis of features
In order to correctly and effectively carry out the classification and recognition of targets, the transfer of the acoustic signal is required so as to obtain the information which can reflect the categories of targets. The process of obtaining the signal features is also the process of feature extraction and feature selection [8]. Therefore, feature extraction is to transform the original language signal so as to obtain the original characteristics which can reflect the essence of the targets through a variety of transformations [9]. In terms of feature extraction for the signals of acoustic targets, the information which can be used to distinguish between a person, a vehicle, a gun and a tank is extracted. Feature extraction methods include the features of energy, amplitude and zero-crossing rate based on time domain, feature of classical power spectrum based on frequency domain, features estimated by the power spectrum estimation of parameter model, the feature based on wavelet packet, etc. The feature of time domain is relatively simple and it can be calculated simply. As the resource of network node for wireless sensor is limited and not able to process very complex algorithm, this paper just compiles the feature of time domain in nodes, including three char-acteristics of short-time energy, short-time average amplitude and zero-crossing rate [10].

1) Time domain waveform of language
In order to observe the time domain of language, this paper puts forward the time domain waveform of language, as shown in Figure 3.

2) Energy distribution of language
In order to observe the energy distribution of language, this paper puts forward the diagram of energy distribution of language [11]. It is able to more intuitively show the  Figure 5. Working sketch energy distribution of each language target with its energy being normalized within [0,1] so that it can be observed that the corresponding energy of different frames with different targets are different, which is conductive for identifying and processing later. This is shown in Figure  4.

IV. SUPPORT VECTOR MACHINE ALGORITHM
The support vector machine algorithm is a method for the study of machines, and it is endowed with inherent risk of experience and minimum scope of confidence. The basic principle of the support vector machine is to determine an optimal hyper plane which will make two kinds of targets have the maximum classification interval [12]. Optimal classification hyper plane with the maximum structural interval is regarded as the decision function. Its structure of working principle for classification is shown Figure 5.
In terms of its calculation principle, there are two types of samples which can be divided [13].
are the category markers of the targets. In order to determine the plane which will distinguish the two categories, the hyper plane is recorded as (!*x)+b=0. Under the rule of minimization of empirical risk, there are numerous such kinds of hyper planes, but in terms of the training sample, some hyper planes have better effect of classification than others. The core idea of support vector machine is that the larger the classification interval is, the better the promotional ability will be [14][15].
In order to distinguish the target effectively, it must meet: The above two equations can be integrated into a normalizing condition the distance of each sample to hyper plane is: In this way, 2 1 2 ! will be the minimum.
In the optimum classification plane, in order to make all the samples classify correctly with the maximum classification interval, the conditions in equation (6) should be met. Therefore, in terms of determining the optimum solution, it can be regarded as a simple optimization problem.
That is to say: At the same time, the constraint condition in the above normalization , when y=1, it is classified as one category, when y=-1, it is categorized as another category.
the problem of quadratic programming problem of minimization is actually the problem of minimization in the original constraint condition.
Its partial derivation is calculated and its value set as 0; that is: After calculating the above equation, the results will be substituted into the original lagrange function and the result is: If the function W(") is the maximum The equation:  can be obtained and the final optimal classification function obtained is: In the equation, b is the bias, according to 3-15, thus yielding:

V. STRUCTURE AND COMPUTATIONAL PRINCIPLE OF MULTI-CLASS CLASSIFIER
The support vector machine is dominated by two categories of classification. If the multi-class classification is to be carried out, the multi-class classifier must be established [16]. This paper carries out research on four multiclass classifiers of one-on-one, one-on-many, two fork tree-SVM multiple classification method and decisiondirected acyclic graph respectively. In the following content, the algorithms of these classifiers are introduced.

1) One-on-one method
The basic principle of one-on-one structure is that in n categories of samples, a classifier is created for every two samples. Afterwards, these classifiers are trained respectively with the testing sample being input after the training [17]. Every sample will be tested in these classifiers and each classifier will obtain an identification result. All of the identification results will be voted on and the samples with the most votes are the category they belong to.

2) One-on-many method
The principle of one-on-many algorithm is that in n categories of samples, n classifiers are created. When creating the classifier of the first class of samples, its training samples are regarded as one category which is marked as 1 while the other training samples are regarded as another category which will be marked as -1. When testing the samples, the voting method is adopted. The testing sample will be input into n classifiers and then n classification results will be obtained. The results will be voted on afterwards, and the samples with the most votes are the category they belong to.

3) Two fork tree-SVM multiple classification method
The principle of two fork tree-SVM multiple classification method is that there are a training period and a testing period [18]. When training, it starts from the leaf node and then moves towards the root node. Each time, a category will be formed until it reaches the root node. The testing starts from the root node, which is carried out by adopting preorder traversal two fork tree until it reaches the leaf node. The category of the leaf node is in the category of the testing sample. Figure 6 shows the structural diagram of the two fork tree.

4) Decision-directed acyclic graphs
The decision-directed acyclic graph must create n(n-1)/2 classifiers. For a classification problem of n category, there are n (n-1)/2 nodes [19]. There are two categories for each classifier which is distributed in the structure of n layers. On the top layer, only one node is included, also referred to as the root node, while in the second layer there are two nodes. In ith layer, there are i nodes. In these nodes, the ith node in jth layer pints to the ith node and (i+1)th node in j+1 layer. In the training process of the DDAG algorithm, only the sub-classifier is trained. The error rate can be reduced by maximizing the interval of the two-valued classifier. The structure of the DDAG classifier is different from the three classifier. In the general decision of the three methods, when some node is mistaken, the error will last until the end of the training. As there is redundancy in the DDAG structure, the ratio of classification path of the same kind must be the same. For the three common decisions, its calculation is simple and highly efficient. Figure 7 shows the structural diagram of the decision-directed acyclic graph.  (1). A sensor node for collection of language and a coordinator node; (2). A language box used to simulate the language of guns and tanks; (3). A laptop; 2 The experimental environment of this paper is on a road outdoors with low noise. This paper does not consider background noise.
3 Experimental content The experiment is carried out with a single node. On the side of the road, a node for collecting language and a coordinator node are placed. The language of a person and vehicles are collected in real life while the language of tanks and guns are simulated by the language box. It collects the language characteristics of the four targets which are person, vehicle, guns and tanks whose language will be transmitted to the sensor node. The sensor node will extract and send the characteristic quantity of language to the coordinator after processing it. The coordinator will then send it to a serial assistant passing through the serial port. The four targets will collect 200 samples respectively including 100 training samples and 100 identification samples, after which the following research on parameters is carried out.

B. The effect of weighted results of different eigenvalues
on recognition rate of speech 1) Weighting of short-time average amplitude Kernel function selects Ker="rbf" with the parameter of P1=2 and C=1000. The classifier of the support vector machine is the structure of a two fork tree. When other parameters remain unchanged and only the weight of the short-time average amplitude is changed, the change for recognition rate is shown in Table 1.
The results of the above experiment show that within a certain scope, when the weight of the short-time average amplitude is increased, the recognition rate is also increased. Furthermore, when it reaches a certain parameter, it decreases accordingly. When the short-time average amplitude is at about 800,000, the average recognition rate of the four targets reaches the maximum value.
2) Weight of zero-crossing rate The kernel function selects Ker="rbf" and the parameters are P1=2 and C=1000. The classifier of the support vector machine is the structure of two fork tree. When other parameters remain unchanged and only the weight of zero-crossing rate is changed, the change of recognition rate is shown in table 2.
The experiment shows that within a certain scope, when the weight of zero-crossing rate is increased, the recognition rate is also increased. Furthermore, when it reaches a certain value, it decreases accordingly. Although the situation varies, the overall trend is the same. When zerocrossing rate is at about 3,000,000, the average recognition rate of the four targets reaches the maximum value.

3) Weighting of short-time energy
The kernel function selects Ker="rbf" and the parameters are P1=2 and C=1000. The classifier of the support vector machine is the structure of a two fork tree. When other parameters remain unchanged and only the weight of energy is changed, the situation of change for recognition rate is shown in Table 3.
The above experiment shows that if only the weight of energy is changed, the recognition rate remains the unchanged.
4) The weighting of short-time average amplitude and zero-crossing rate are at the same time.
The kernel function selects Ker="rbf" and the parameter P1=2 and C=1000. The classifier of the support vector machine is the structure of a two fork tree. When other parameters remain unchanged and only the weight of energy is changed, the change of recognition rate is shown in Table 4.
The experiment shows that when the weight of amplitude is 5,000,000 and the weight of zero-crossing rate is 2,000,000, the average recognition rate of the four targets reaches the maximum of 89%.
5) In weighting of average short-term amplitude, the zero-crossing rate and energy weight are considered at the same time.
The kernel function selects Ker="rbf" and the parameters are P1=2 and C=1000. The classifier of the support vector machine is the structure of the two fork tree. When other parameters remain unchanged and only the weight of energy is changed, the change for recognition rate is shown in Table 5.
The above experiment shows that although there is no effect in the improvement of recognition rate if just the energy is weighted, the situation in which three characteristics are weighted at the same time has a better effect than weighing for just one or two characteristics. The overall effect can be heightened by 0.25%.

VII. CONCLUSION
A wireless sensor network is widely used in various fields. This paper builds a platform of a recognition system for an acoustic target based on a wireless sensor network by combining the wireless sensor network with language recognition. It conducts research on the language recognition of a person, a vehicle, guns and tanks. These four categories are targets, and the extraction methods and realization in sensor nodes, as well as recognition of the algorithm of support vector machine and a variety of structures of multi-class classifier. Afterwards, the joint identification algorithm for multiple sensor nodes is studied and then the experimental test of a single node is carried out resulting in weighting of different characteristic values on identification rate being analyzed. The following conclusions are obtained.
1. A platform of recognition system for acoustic target based on wireless sensor network is established which includes compiling the programs of language collection, characteristics extraction and the interface design of the upper computer. Finally, the category of the target region can be displayed in real time for observation.
2 Research on joint decision algorithm of multiple nodes includes the analysis of the advantages and disadvantages of voting judgment and energy judgment... Furthermore, based on the two algorithms, the improved voting judgment and energy judgment are put forward which can be used more widely theoretically. PAPER EXPERIMENTAL RESEARCH ON SIGNAL RECOGNITION ALGORITHM OF WIRELESS SENSOR LANGUAGE 3 Research on identification of sensor node includes the recognition algorithm of support vector machine being studied with the four structures of multi-class classifiers which include the structure of one-on-one, one-on-many, and the structure of two fork tree multi-class classifications and the structure of decision-directed acyclic graph being researched. Further experiments were also conducted. After the experiment and research on the effect of each parameter on identification rate, it shows that the average identification rate of the structure of decision-directed acyclic graph is better comparative to its identification rate being up to 93.5%.