The longer than twenty-two-year success marching of the IEEE 802.11 communication technology continues in the next years with new standard editions having transfer rate in the multi Gbit/s range. Realistic evaluation of the WiFi controller supervised hot zone service level becomes more and more critical because of the very high number of frames transmitted per unit of time. Online evaluation of the content transmission eficiency on radio channel is afected by several conditions including environment reflection characteristics, multipath influences, movement behaviour of the users and time dependence of the mobile terminals population in the service area. Based on our anterior investigations we found that in special places of the coverage area with WiFi hot zone service high ratio of transmitted frames are temporarily control and management frames even in case of communications with low level of the radio signals. To scan and evaluate IEEE 802.11/n/ac/ax channel usage eficiency we developed a complex scanner and evaluator tool based on neural network stick hardware. The software prototype developed utilize Long-Short Term Memory and Gated Recurrent Unit functions to determine periodically the percent of data frames of the total transmitted radio frames. Constant number of frames and constant time intervals, respectively are applied as two basic approaches of our evaluation methods. Advantages, weaknesses and usability cases in practice of the proposed solutions will be given in the paper.
Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Network, routing, clustering, time series classification.
MSC: 65C60, 60G35, 91A28
Due to the advantageous properties of IEEE 802.11 (WiFi) wireless technology, it has been present with unbroken popularity in the access environment of network infrastructures for over twenty years of its development. Several stable versions (a/b/g/n/ ac/ad/ax) have been used recently for SOHO and other corporate LANs. Apart from its security aspects, each newer standard has been designed to improve communication eficiency and enhance the user experience. There has been relatively little comprehensive research on the complex analysis of frame streams by category, since measurement and processing require mathematical proficiency in addition to technological knowledge. It is also cumbersome to interpret trafifc which is grouped by over seventy subtypes of the three main types of WiFi frames. Access to these has only recently become possible with a public softwarebased protocol analyzer. Using this changed feature, we performed WiFi network measurements, and analysis by trafic type.
The further structure of the paper is as follows: the second section lists some of the related neural network applications used for wireless networks. In the third section, we introduce the most important features of IEEE 802.11 transmission technology. Introducing the neural network stick and using it for fast data processing is discussed in the fourth section. The fifth section deals with the deep learning analysis of IEEE 802.11n/ac/ax trafics. The last section provides a summary of the analysis work described and its possible continuation of the research work.
In recent years, many application of neural networks (NNs) have been used in wireless networks. This is because using NNs allows us to, among other things, map the input data to predictions or classes, e.g. error rate prediction, signal classification, to predict throughput. NNs can be used to model non-linear systems to ensure quality of service (QoS), predict device location, and more. The use of NNs can be divided into two main parts: supervised learning; unsupervised learning. The most common types of supervised NNs used in wireless networks are Multilayer Feed-Forward NNs (ML-FFNNs), and Recurrent NNs (RNNs). The methods described in this paper relate to supervised learning and RNN. ML-FFNNs can learn the relationship between the given inputs and their corresponding outputs through the training process. Using the resulting NN model, it is possible to evaluate outputs for new inputs. RNNs are similar to ML-FFNNs except that they also have a cyclic nature. This allows us to introduce time stamps, which make it possible to model time dependent systems; i.e. the output for the following input example may depends on the previous examples. Use of RNNs in WiFi frame type ratio regression has not been used till now as we know.
Some supervised NN applications in wireless networks are summarized in the
section below. Battiti et al. have created an ML-FFNN-type NN that can predict
the x and y coordinates of a given device location based on RSSI signal intensity
values from three APs [
Altini et al. worked on localization with Bluetooth signals using multiple NNs,
each NN specific to a diferent positional orientation of a user, and each NN is a
MLFFNN [
Baldo and Zorzi proposed usage of ML-FFNN to predict delay, throughput, and
reliability using the SNR and error rate as inputs [
The IEEE 802.11 transmission technology further continues the success story of its development. The most important technological features of the latest three standard versions are summarized in Table 1.
IEEE 802.11n HT (High
2,4 | 5 GHz OFDM; DSSS/CCK 20 | 40 MHz 600 Mb/s indoor: 70 m | outdoor: 250 m 2 Tx | 3 Rx antennae Yes
IEEE 802.11ac
5 GHz OFDM; DSSS/CCK 20 | 40 | 80 | 160 MHz 6,93 Gb/s
The IEEE 802.11 standard is basically a WiFi communication technology that operates in the bottom two layers of the OSI reference model. In our case, we will describe the key elements of the physical layer of the 802.11n/ac protocol, as this technology was available for us to measure wireless trafic. The task of the physical layer is to appear the bit sequence in its original, error-free state on the receiving node where the bit sequence was transmitted on the medium between the two communicating devices. In our case, the medium in question is the electromagnetic ifeld.
In order to avoid eavesdropping and signal interference, and to increase fault
tolerance, several transmission modes have been developed for data transmission
in the electromagnetic field. 802.11n uses HT-OFDM (Higher Throughput
Orthogonal Frequency Division Multiplexing) technology [
Basically, the task of the media access and control (MAC - Medium Access
Control) sub-layer is to reliably transmit frames to be transmitted even in noisy
environments. Since the transmission medium is not at all as reliable as, say,
on an optical line, “Best efort” transmission would not have been an expedient
mechanism. Therefore, for the sake of reliability, the receiver always acknowledges
the received frames. To conclude, the simplest frame transfer event looks like the
following. The source station sends the desired data with a collision prevention
strategy (CSMA/CA) then the receiving station acknowledges these. If the
acknowledgment is not received, the sender retransmits the bit sequence previously
transmitted by itself [
The FC (Frame Control) field contains the protocol version, the main and
subtype of the frame, the distribution system bits, the fragment bit, which indicates
whether the current frame is the last fragment of a longer frame. This field contains
the send-retry bit, the bit to proclaim energy conservation, the bit to indicate more
data. The latter indicates that at least one more frame is currently bufered at the
base station for the mobile station. Other fields are the WEP encryption bit and
the request defining bit, which indicates whether the given data frame is needed
for a well-defined service [
The Duration/ID field has a duration length or an identifier function. In the ifrst case, it specifies the time slice as much time the sending device needs to use the radio wave channel for transmitting the frame. In the second case, the mobile station identifies the receipt of its frame bufered by the base station.
The four address fields (Addr.1, . . . , Addr.4) contain the physical addresses of the destination and source, as well as the recipient and sender addresses. Source addresses are unique, and destination addresses can be multicast/broadcast addresses. The Sequence Control field contains the sequence number and the fragment number. These can help you deal with communication disorder caused by the possible occurrence of duplicate frames.
The Data section contains the payload bytes. If the frame contains bytes of
data type, this field occupies a significant part of the size of the radio frame [
FCS (Frame Check Sequence) enables error checking of the ordinary IEEE 802 standard. The receiver compares the received code number with the code value calculated from the frame and, if they are the same, records the receipt of the frame as error-free delivery. The possible discrepancy gives the error rate. At a large Hamming distance, the frame is damaged and therefore discarded, but can repair a one-bit error.
With this communication technology, it is not enough to talk about just one kind of frame. The “Type” field in the FC record allows you to distinguish three frame types: i) Management frames; ii) Control frames; iii) Data frames. There is also a S“ubtype” field, which further subdivides the types into subtypes.
00 01 01 01 10 11
Data Reserved
From the viewpoint of the paper one of the dominantly important subtypes of management frames is the Beacon frame. Beacon frames are theoretically sent by base stations at periodic intervals, but in practice this period may sometimes vary slightly due to the base station’s occupancy due to the transmission of the current frame.
Beacon frames advertise basic information about the base station to other
mobile terminals that want to use this WiFi service. From the viewpoint of the paper
less important but noteworthy subtypes are the Authentication and
Deauthentication frames, which serve to authenticate terminals. Like the Beacon frame, Probe
Request and Response advertise base station information, such as supported data
transmission rates [
The control frame also has subtypes. Some of these frames are used to prevent
collision. The dialogue with the control frames of subtypes RTS (Request to Send)
and CTS (Clear to Send) determines whether the entity wishing to send can start
sending the data frame. The potential sending terminal requests transmission
permission with RTS control frame from the base station. If it receives a CTS control
frame response, it can begin transmitting its data frame. The ACK (Acknowledge)
subtype acknowledges the received data frame to the sender [
The subtypes of the data frame difer only in those that also contain smaller
control information for the transmitted data, such as an ACK (Acknowledgment)
acknowledgment, which confirms the successful transmission of the previous data
frame [
The Intel Neural Compute Stick 2 (Intel NCS2) is a plug-and-play device that looks like a standard USB drive designed to execute inferences using neural network and input data in a quickly and energy eficiently way. It is also suitable for real-time inference. It requires no cloud infrastructure and can be used with lowpowered devices such as the Raspberry Pi 3 and other ARM host devices. It can be connected to USB 3.0 Type-A compatible connectors. Inside the device is an Intel® Movidius ™ Myriad ™ X Vision Processing Unit (VPU) that contains 16 processing cores and a dedicated deep neural network hardware accelerator. Multiple NCS 2s can be used to infer at the same time, and can even be combined with CPU and GPU. NCS 2 is primarily used to deploy vision-oriented solutions using convolutional neural networks (CNNs).
The general workflow as shown in the Figure 2. is the following: First, the raw data is preprocessed, resulting in the training set, validation set and test set. These sets are used to train the model. The model is trained using frameworks or formats supported by Intel OpenVINO software, and this process should be done on the development hardware, not on the stick.
Briefly, OpenVINO is a cross-platform (Windows, Ubuntu, macOS, etc.) software that enables deep learning inferences, supports various heterogeneous execution across an Intel Integrated Graphics, Intel FPGA, NCS 2, Intel CPU, etc., supports well-known deep learning frameworks and formats (TensorFlow, MXNet, Cafee, ONNX, etc.) and includes optimized calls for computer vision libraries such as OpenCV, OpenCL and OpenVX. Model Optimizer and Inference Engine API are part of OpenVINO. Once the model is complete, it must be translated with the Model Optimezer into the so-called Intermediate Representation (IR) format. This format consists of two files, .xml describes the network topology, and .bin contains the network weights and bias values. Finally, the user application uses IR and infers new data both using the Inference Engine API, which is executed on Intel NCS 2, and outputs the result.
Some real-time applications need to process large amounts of data in a short amount of time, which in our case means that our application must infer for hundreds of incoming frames in less than 1 second. This requires a powerful computing device, such as the Intel NCS 2, which is optimized for neural network inferences. Experiments have already been made by us to use this device for enhancement of RNN calculations, but issues have been detected while running the Inference Engine. Further in-depth research and software development is in progress to eliminate these compatibility problems concerning neural network topologies supported explicitly by the Openvino.
There were captured WiFi frame streams at a four floor building with 26 base
stations managed by a common controller. The measurements were executed step by
step in 56 physical dispersed in space in diferent time moments. Measured values
were: Frame Length (B), Frame Type/Subtype, Frame Arrival Time (sec) and
Received Signal Strength Indicator - RSSI (dBm). Duration of one measurement was
in the scale of ten seconds. Sampling of the WiFi frames was made with Wireshark
protocol analyser software running on a notebook. It was activated deep mode of
the Npcap v0.9984 driver to capture IEEE 802.11 format of all the frames instead
of the basic mode showing just IEEE 802.3 frames. In this way all the management,
control and data type frames transmitted on the wireless channel were captured
by the software. Based on the investigation results described in papers [
As an example frame length of diferent types captured at sampling point no. 23 can be seen in Figure 3. In this case majority of the frames were Beacon management (34.5%) and Request-to-Send control (17.69%) frames. It should be noted that the IEEE 802.11n/ac/ax communication mechanism use intensive management and control frame transmissions to supervise nodes and deliver data type frames between the wireless nodes. Intensity of the unidentified type frames was 0.61% which is relatively small but not negligible and are sent by new smart phones without access point responses. We enrolled such frames in a separated type named Other.
Because it exists correlation between the intensity of diferent types of the frames we grouped subtypes of the frames in four: management (MNG), control (CTR), data (DAT) and other (OTH). We split each ten second measurement in sequences having total duration of one second. Each sequence belongs to the event series of frame arrivals during one second sampling time interval. The sequence has three predictor variables: v1: Frame Length (B), v2: Frame Interarrival Time (sec) and v3: Received Signal Strength Indicator - RSSI (dBm). The interarrival time is the time diference between sampling times of two consecutive frames. The number of events is given by the number of frames, n.
The sequence is characterized by the four metrics: relative amount of p1: MNG, p2: CTR, p3: DAT and p4: OTH bytes transmitted on the radio channel during the sequence time interval. We have p1 + p2 + p3 + p4 = 100% in reality.
In this way we got N = 591 sequences, each having three predictor variables: v1,
v2, v3 and four score metrics (ratios): p1, p2, p3, p4. It is obvious that the sum of
these four metrics gives 100% for each sequence. The length of predictor variables
is ni, i = 1, 2, . . . , N. Because the number of frames during the sequence sampling
time varied in function of the WiFi trafic intensity, length n i is not constant and
takes values in the range of [
Parameters of the tensor data analysed with diferent recurrent neural networks are given in Table 4. For each neural network learning it was used Adaptive Moment Estimation (Adam) optimizer with GradientDecayFactor = 0.9, SquaredGradientDecayFactor = 0.9990 and InitialLearnRate = 0.01. Topology of these networks are including Long-Short Term Memory (LSTM) or bidirectional LSTM (BiLSTM) layers (see Figure 6 and 7). For faster convergence two fully connected (FC) layers were used. The first layer has 512 number of classes and the second layer just four because we are predicting scores p1, p2, p3 and p4 of the test dataset. Number of hiden units in LSTM or BiLSTM layers was set to be 100 or 200.
Parameter RNN1 RNN Type LSTM SmiizneiBatch- 50 maxEpochs 360 LSTM numHidde- 100 nUnits FCClassesnum- 512 tLiemaern[isnegc] 348 100 180 100 512 107 300 180 100 512 84 RNN2 LSTM
RNN3 LSTM
RNN4 BiLSTM
RNN5 BiLSTM
RNN6 BiLSTM
RNN7 BiLSTM
RNN8
BiLSTM 50 360 200 512 137 100 180 100 512 154 100 180 200 512 154 100 360 100 512 271 50 360 100 512 378 stated that for the sampled WiFi channel sequences BiLSTM has better performance than LSTM. Having information about the past and the future sequences of the trafic in WiFi channel helps to train better the neural network. To have best performance of the RNN needs at least 200 hidden neurons of the recurrent layer. Computation capacity has impact on the training time. To decrease the computation time dedicated hardware tool, like compute stick is required.
Best neural network to predict the IEEE 802.11n/ac channel usage eficiency was found to be RNN5. Prediction with 0.70 correlation is based on frame size, interarrival time of the frames and the received signal strength intensity.
Coef. of correlation R: MNG R: CTR R: DAT R: OTH
RNN1
RNN2
RNN3
RNN4
RNN5
RNN6
RNN7
RNN8
In this paper we analysed IEEE 802.11n/ac/ax technology behaviour based on neural network modelling. Based on the special features of the wireless medium access communication mechanism there were captured all the management, control and data frames transmitted on the radio channel. It is settled that majority of the wireless bandwidth is used for management and control byte transmissions and just 50% or less of the channel capacity is allocated to transmit data bytes. The proposed neural network contains recurrent layers and the learning process is supervised. There were compared usability of LSTM and BiLSTM layers for prediction of the WiFi channel usage eficiency. It was found that BiLSTM has better performance to prognoses transmitted byte ratios belonging to management, control and data classes. Possible continuation of this research may be oriented to evaluate other topologies of the deep neural network.
Acknowledgements. This work was supported by the construction EFOP-3.6.3VEKOP-16-2017-00002. The project was supported by the European Union, coifnanced by the European Social Fund. and EFOP-3.6.3-VEKOP-16-2017-00002, respectively. This paper was supported also by the FIKP-20428-3/2018/FEKUTS TRAT project of the University of Debrecen, Hungary and by the QoS-HPC-IoT Laboratory.