Neural Networks Exploration of Neural Network-based Multidimensional Spectral Inference Methods Neural Networks

With the development of data center architecture and the application of new technologies such as Vxlan and RDMA, traditional SNMP and other operation and maintenance methods are no longer able to meet the current IDC operation and maintenance needs. The larger and more complex cloud data centers and high-performance computing clusters have put forward higher requirements for operation and maintenance. How to find more refined and intelligent operation and maintenance methods to achieve more efficient and timely monitoring and maintenance of the network has become a major challenge faced by data center network operation and maintenance.

Abstract: This paper briefly introduces the current development status of spectrum inference and proposes a multi slot spectrum inference method - Long Short Term Memory Network (LSTM) inference method. Based on it, combined with neural network algorithms, it extends and derives the time-frequency two-dimensional spectrum inference method and the spatiotemporal three-dimensional spectrum inference method. It can achieve spectrum inference and prediction in limited and discontinuous spatiotemporal multi-dimensional spectrum monitoring data, providing support for the generation of electromagnetic spectrum situation and spectrum control in the battlefield.

Under modern battlefield conditions, not all combat platforms or frequency systems are necessarily equipped with professional spectrum monitoring equipment, but rather rely on the spectrum sensing capabilities of the combat platforms or frequency systems themselves. Therefore, these spectrum monitoring data are very sparse or even scarce in multiple dimensions such as time domain, spatial domain, and frequency domain. It is necessary to collect and fuse data from limited and discontinuous spatiotemporal multi-dimensional distributed spectrum monitoring, and perform spectrum inference and prediction on these data to form situational information of complex electromagnetic environment on the battlefield, and to form predictive analysis of available spectrum resources, in order to provide a basis for battlefield spectrum control.

Spectrum inference technology is an important foundation of cognitive communication, which can provide equipment systems with channel occupancy information in unknown areas, future times, or unfamiliar spectra through methods such as pattern mining and machine learning. It greatly improves the survival and adaptability of equipment in complex electromagnetic environments, and its significance in modern warfare cannot be ignored.

Spectrum inference is based on historical known spectrum data to explore the inherent correlation or regularity of the data, in order to obtain the data state of unknown wireless spectra in the future. Spectrum inference is also a double-edged sword. On the one hand, it can predict the channel state for a certain period of time in the future through historical data, provide throughput for unauthorized users, and shorten the energy consumption of perception time in adaptive spectrum sensing; On the other hand, it will inevitably lead to false alarms and missed detections in predictions [2-3].

The existing spectrum inference techniques mainly focus on the one-dimensional time dimension, while the inference techniques in the frequency and spatial dimensions are rarely involved. G Ding et al. [4] verified the predictability of spectrum states. For temporal spectrum inference, Hidden Markov Model (HMM) inference method and Neural Network (NN) method have good inference accuracy. However, most research on these two methods is limited to inferring the occupancy of the next time slot, which still has a significant gap with real-world application scenarios. In addition, the accuracy of the Hidden Markov Model inference method [5] is limited by the order of the transformation matrix used. The more complex the scenario, the higher the order required for the model, and the more difficult it is to describe the hidden state rules. Moreover, the complexity of the model operation will increase exponentially. Therefore, the HMM model is not suitable for complex spectral inference scenarios. Relatively speaking, NN algorithm [6] can simulate the nonlinear transformation relationships existing in application scenarios through the action of neurons, and obtain the weight values between nodes in the network through offline learning. For more complex electromagnetic spectrum environments, it can increase the number of hidden layers [7] and neurons in each layer [8] of the neural network. Therefore, NN algorithm is more suitable for multi-dimensional complex electromagnetic spectrum inference scenarios compared to HMM algorithm.

Given the correlation characteristics of spectra in both spatial and frequency dimensions, NN algorithms can be used for one-dimensional spatial and one-dimensional frequency dimension spectrum inference, respectively. This study proposes an inference algorithm, LSTM (Long Short Term Memory), that is suitable for multi time slot spectrum in the time dimension, based on the advantages and disadvantages of NN algorithm. Based on this, two spectrum inference methods are designed for time-frequency two-dimensional spectrum inference and spatiotemporal frequency three-dimensional spectrum inference, respectively. The latter can be fully transplanted into the application scenario of grid structure. At present, multidimensional spectrum inference technology is still in the preliminary exploration and research stage, and has not yet formed a relatively complete theoretical system, nor does it lack practical application scenario verification.

Long Short Term Memory (LSTM) inference method

Related studies have shown that channel states have a high degree of autocorrelation, and with different time shifts, the correlation shows a gradually decreasing trend. The LSTM inference method has significant advantages in processing time or spatial sequence data [9]. The LSTM network unit structure is shown in Figure 1:

Figure 1 LSTM Network Unit Structure

The Long Short Term Memory (LSTM) network overcomes the drawback of RNN (Recurrent Neural Networks) networks that cannot converge quickly during training [10]. It introduces three gate units (input gate, forget gate, output gate) to reasonably control the relationship between historical and current information, as shown in Figure 1. The main formulas involved in LSTM are as follows:

it=σ(Wxixt+Whiht-1+Wcict-1+bi)                      (1)

ft=σ(Wxfxt+Whfht-1+Wcfct-1+bf)                      (2)

ct=ft⊙ct-1+it⊙tanh(Wxcixt+Whcht-1+bc)                      (3)

ot=σ(Wxoxt+Whoht-1+Wcoct+bo)                      (4)

ht=ot⊙tanh(ct)                                  (5)

Among them, σ is a logical S-shaped function, i, f, o, c are the corresponding input gates, forget gates, output gates, and memory cell activation vectors, ⊙ represents the multiplication of corresponding elements between vectors, W (..) Representing the corresponding weight matrices, where the matrices from cells to gates are diagonal matrices, while the remaining weight matrices are non diagonal matrices.

The LSTM network structure that matches the one-dimensional time series spectrum inference is shown in Figure 2. The spectrum inference method based on LSTM network can flexibly adjust the inference network structure, and can switch between applications that infer different time slot lengths. At the same time, another advantage of this method is that ordinary inference algorithms (such as HMM or NN) can only make inferences about the availability of shared channels (i.e. 0-1 values), while the LSTM method can also infer the energy level of the channel, greatly increasing the flexibility of cognitive devices in designing access decision criteria during channel access, further improving the efficiency of channel access, and reducing the perceived energy consumption level of cognitive devices.

Figure 2 Schematic diagram of LSTM network structure for one-dimensional time series inference

Three time frequency two-dimensional spectrum inference method (2D hybrid network)

The combination algorithm based on multiple network algorithms can achieve more satisfactory processing or inference results by matching the strengths and advantages of different networks. Therefore, the application prospects of hybrid neural networks in the field of spectral inference are enormous. Multiple research analyses have shown that the usage of spectrum is also highly correlated in the frequency domain. Here, an algorithm combining LSTM network and neural network is proposed for time-frequency two-dimensional spectrum inference. The network relationship in the algorithm is shown in Figure 3:

Figure 3 Relationship diagram of time-frequency two-dimensional spectrum inference network

This hybrid network is mainly constructed based on the correlation of time-frequency spectrum. LSTM expresses the potential impact relationship between historical perceptual data and future inference data through a forget gate. When no neural network is added, the process of spectral inference realizes multiple parallel time series inference functions, and there is no correlation between them. After the combination of neural network and LSTM network, the hybrid network can be used to infer the occupancy of one or more unknown channels, as well as the usage of the next time slot (or multiple time slots) in the frequency domain occupied by the entire channel.

he parameter structure of the hybrid neural network used for time-frequency domain spectrum inference is shown in Figure 4. Figure 4 only describes the spectrum inference process at time slot t. From Figure 4, it can be seen that the LSTM structures of different channels are independent and parallel. Based on this, the outputs of each channel obtained at time slot t are used as input values and transmitted to the MLP inference engine. After processing by the inference engine, the occupancy of each channel in the next time slot (t+1) or multiple future time slots can be obtained by continuing to input. Due to the relatively clear connection between LSTM and neural networks, the backward propagation process of the hybrid network during training is essentially equivalent to the sum of the original two training methods. Therefore, this hybrid network does not significantly increase the complexity of training. After certain targeted training operations, the forward weight matrices and bias values converge to stable values, which can then be used for subsequent two-dimensional spectral inference processes.

Figure 4 Schematic diagram of parameter relationship in time-frequency two-dimensional spectrum inference network

4. Spatiotemporal frequency three-dimensional spectrum inference method (3D hybrid network)

On the basis of LSTM algorithm and time-frequency spectrum inference algorithm, a spatiotemporal frequency spectrum inference algorithm suitable for grid structure is proposed. In the spectrum inference of grid structure, given the spectrum sensing information of multiple different spatial nodes in the time and frequency dimensions, it is necessary to calculate and infer the spectrum occupancy (or power level) of unknown spatial positions. A inference algorithm network structure that can meet this requirement is proposed, as shown in Figure 5:

Figure 5 Schematic diagram of spatiotemporal frequency three-dimensional spectrum inference network

In this network architecture, LSTM based networks are used for the time dimension, and neural networks are used for the frequency and spatial dimensions. Among them, the unknown nodes are located in space S, and the spectrum usage of spatial nodes such as (S-1) and (S+1) is known information. During the training process of the network, historical information from multiple spaces (not limited to the three shown in the figure) is used as input information, and the output can be adjusted according to actual needs. If only the future spectrum usage of space S is needed, it can be used as the output. Meanwhile, the weight relationships can be appropriately reduced to match the actual impact relationships, while reducing the computational pressure during the weight reverse iteration process.

In this 3D neural network, the traditional neural network structure can be appropriately improved or directly replaced with structures such as Convolutional Neural Networks (CNN) that have better inference performance. It can be seen that the 3D neural network combines the relative relationship of spatial spectra and can adjust the connection relationship in the network to be applicable to multi-dimensional spectral data of time, space, and frequency. Specifically, the three-dimensional hybrid neural network is capable of inferring the spectrum occupancy of point (channel slot point), line (single channel or single slot or different spatial slot points), plane (multi-channel slot or multi-channel space or multi slot space), and volume (channel slot space) structures.

Comparison of Spectrum Reasoning Techniques

The comparison of parameters and performance between existing spectrum inference algorithms and currently proposed spectrum inference algorithms is shown in Table 1:

The traditional Hidden Markov Model algorithm is not suitable for complex scenarios; Neural network algorithms are suitable for complex scenarios, but their training is slow; The long short-term memory network reasoning method is suitable for sequence relationships in complex scenarios and can be trained quickly; The time-frequency two-dimensional spectrum inference method is suitable for two-dimensional relational scenarios, and the training speed can be improved by combining it with genetic algorithms; The spatiotemporal frequency three-dimensional spectrum inference method is suitable for three-dimensional relational scenes, and the training speed can be improved by combining it with genetic algorithms. The network structure can also be further optimized.

  (Source: Mobile Communications, February 2018 article translated from C114 Communication Network)