A NOMA signal detection method based on deep learning for non-stationary industrial Internet

Abstract

The present invention belongs to the technical field of multi-user signal processing in wireless communication systems, and specifically relates to a NOMA signal detection method based on deep learning for non-stationary industrial Internet; considering the upper-layer retransmission mechanism, an LSTM (long short-term memory network) is designed to fully utilize the time-related characteristics of the active terminal set to estimate the active terminal set, and combining the DNN (deep neural network) and SIC (serial interference cancellation) signal detection structure, a hierarchical DNN structure is proposed to effectively improve the accuracy of multi-user superposition signal detection in an overloaded NOMA system.

Description

Technical Field

[0001] The present invention belongs to the technical field of multi-user signal processing in wireless communication systems, and specifically relates to a NOMA signal detection method based on deep learning for non-stationary industrial Internet. Background Art

[0002] The Industrial Internet is a product of 5G mobile communications empowering traditional manufacturing. Mobile communication networks and technologies are the foundation of the Industrial Internet. The Industrial Internet has the characteristics of a large number of nodes and low data packet load, and there are critical instruction-type services and emergency burst services that require high reliability and low latency. In this case, multiple access technology plays a key role in effectively supporting large-scale connections and low-latency communications.

[0003] Traditional multiple access technology based on centralized control first performs signaling interactions such as scheduling authorization before starting data communication. The signaling overhead is approximately proportional to the number of terminals. Due to the large signaling overhead and long signaling interaction time, multiple access technology based on centralized control is no longer suitable for the industrial Internet. In view of this, 3GPP proposed scheduling-free access for short packet high reliability and low latency communication. In addition, non-orthogonal multiple access (NOMA) can eliminate collision problems and achieve system overload through technologies such as multi-user detection. The combination of scheduling-free and NOMA will effectively reduce signaling overhead and improve access capabilities, providing the possibility of supporting large connections, high reliability and low latency. In scheduling-free NOMA, time domain and / or frequency domain resource blocks are divided into non-orthogonal sub-blocks, and all potential terminals in the system share these resources. Active terminals can freely access these resource blocks without waiting for any scheduling authorization. This greatly reduces the overhead of control signaling. Therefore, scheduling-free NOMA has a natural advantage in supporting the industrial Internet. However, since scheduling-free NOMA omits the scheduling authorization process, the set of active terminals and their channel information are unknown, which brings new challenges to the design of NOMA receiving signal processing methods and the design of NOMA transmission service quality assurance methods.

[0004] In response to the multi-user detection problem of scheduling-free NOMA, the current research based on deterministic reasoning such as compressed sensing and convex optimization proposes a joint design of active terminal detection, channel estimation of active terminals, and multi-user detection methods. The prior art provides a multi-task variational autoencoder based on deep learning (DL), and a method for jointly optimizing constellation point distribution and multi-user detection based on variational reasoning. There is also a joint detection method for active users and multi-user signals based on a generative neural network framework. The existing NOMA multi-user detection has evolved from deterministic reasoning to machine learning methods, solving the multi-user detection problem in independent random scenarios, but it is difficult to cope with highly random and strongly correlated industrial Internet communication scenarios. More importantly, when the industrial Internet generates emergency burst services, it will bring about a series of chain reactions, causing the number of active terminals in the network and the service characteristics carried by the terminals to mutate, and the number of active terminals will produce non-stationary points. The accuracy of active user estimation and channel estimation will directly affect the accuracy of the multi-user signal detection method of the NOMA system, and thus affect the reliability of the system.

[0005] The existing scheduling-free NOMA technology can only cope with the detection problems of independent random scenarios, but cannot cope with the detection problems of highly random, strongly correlated, and non-stationary communication scenarios of the Industrial Internet.

[0006] In summary, in the industrial Internet, a single terminal service has the characteristics of sporadic arrival, and the arrival between different terminals has spatial correlation and non-stationarity. Wireless channels and scheduling-free transmission aggravate the randomness of the system, and retransmission causes the terminal activation to have temporal correlation. The space-time correlation and non-stationarity of the service will lead to the space-time correlation and non-stationarity of the active terminal set. How to give full play to the model-free characteristics of machine learning and establish the relationship between the received signal and the active terminal set and the transmitted symbol in a highly random, strongly correlated and non-stationary environment is a core technical problem that needs to be solved in this field. Summary of the invention

[0007] In order to overcome the above problems, the present invention provides a NOMA signal detection method based on deep learning for the non-stationary industrial Internet, and studies multi-user signal detection based on DNN (Deep Neural Networks), mainly including non-stationary data processing, LSTM, the design of DNN network structure, and the design of low-complexity DNN training sequence; it can provide a transmission scheme for multi-user NOMA system with security and unauthorized communication, propose a non-stationary data processing method, and an active user estimation method based on LSTM (Long Short-Term Memory Network), and feed back the results.

[0008] A NOMA signal detection method based on deep learning for non-stationary industrial Internet, including the following contents:

[0009] Step 1: Use the maximum likelihood estimation method to detect non-stationary points and estimate the duration of the stationary process; specifically:

[0010] Decompose the received signal y(n) with W points in the discrete time domain to obtain M basic mode components c 1 ,c 2 ,…,c M And the residual component r, the received signal y(n) is:

[0011]

[0012] Where j = 1, 2, ..., M; n is the discrete time domain value of the received signal, n = 1, 2, ..., W;

[0013] Performing a Hilbert transform on each basic mode component, y(n) is written as a real number a j (n) and plural The product is in the form of:

[0014]

[0015] The stationarity of the received signal is defined as:

[0016]

[0017] The Hilbert spectrum H of the received signal y(n) is:

[0018]

[0019] b j is the frequency of the jth fundamental mode component. When the signal frequency is equal to ω j When b j =1; when the signal frequency is not equal to ω j When b j =0,ω j For b j The relevant frequency domain is a real number;

[0020] The average boundary spectrum B of the received signal y(n) is:

[0021]

[0022] The stationarity DS(ω) can quantitatively detect the stationarity of data: for a stationary process, the Hilbert spectrum H of y(n) does not change with time, and DS(ω) = 0; if DS(ω) is not zero, it is a non-stationary time point, and as DS(ω) increases, the non-stationarity of the signal increases; the duration of DS(ω) = 0 is recorded, that is, the duration of the stationary process, and the time period with the most occurrences in the duration of the stationary process is selected as the predicted duration of the next stationary process, so the time when the next non-stationary point will appear can be predicted based on the duration of the stationary process;

[0023] Step 2: construct an LSTM unit series network consisting of L LSTM units;

[0024] Step 3: 1) When the received signal y(n) is in a stable period, the received signal is used as the input of the LSTM unit series network, and the weighted output of the output gates of L LSTM units is used as the overall output of the LSTM unit series network. The output is the state of each terminal, so the state Ω of the kth terminal is k Then we have:

[0025]

[0026] Where l = 1.2.3…L, w l is the weight coefficient of the output gate of the lth LSTM unit, is the output vector of the output gate of the lth LSTM unit in the LSTM unit series network of the sending signal of the kth terminal;

[0027] For Ω k , if Ω k The value of is greater than 0.5, then Ω k =1, that is, terminal k is active; if Ω k The value of is less than 0.5, then Ω k =0, that is, terminal k is inactive, so the active terminal state set is obtained;

[0028] 2) When the received signal y(n) is at a non-stationary point, the LSTM unit series network established in step 2 is retrained, and then the active terminal state set is obtained according to step 1);

[0029] Step 4: Demodulate the input signal using the SIC signal detection method based on deep neural network:

[0030] Set all active terminal states obtained in step 3 And the received signal y(n) is input into the DNN signal detector, the DNN signal detector demodulates the input information and outputs the demodulated signal.

[0031] The DNN signal detector includes an input layer, a plurality of hidden layers and an output layer, and the number of hidden layers is equal to the number of terminals; The input signal is sent to the input layer of the DNN signal detector, passes through all the hidden layers in sequence, obtains the demodulated signal, and is output through the output layer.

[0032] In the step 2, an LSTM unit is composed of three parts: a forget gate, an input gate and an output gate; wherein:

[0033] The forget gate is expressed as follows:

[0034] f t =σ(W f ×[h t-1 ,y(n)]+b f )

[0035] Where: h t-1 is the output of the previous LSTM unit output gate, y(n) is the received signal, W f is the forget gate weight, b f is the forget gate bias, σ is the Sigmoid activation function;

[0036] The input gate is represented as follows:

[0037] i t =σ(W i *[h t-1 ,y(n)]+b i )

[0038] Where: W i is the input gate weight, b i Bias for input gate;

[0039] The output gate is represented as follows:

[0040] o t =σ(W o *[h t-1 ,y(n)]+b o )

[0041] Where: W o is the output gate weight, b o Bias for the output gate.

[0042] The training process of the LSTM unit series network in step 2 is as follows:

[0043] There is a known set of terminal states There are several groups of sending signals corresponding to each terminal in the terminal state set. The known sending signals of each terminal are used as the input of the LSTM unit series network, and the terminal state set is used as the output of the LSTM unit series network. The average error is used as the loss function. The weighted output O of the output gates of L LSTM units p As the overall output of the LSTM unit series network, the average error Where K is the number of terminals, and the number of terminals is equal to the number of LSTM units L, p = 1.2.3...K, then the weight matrix of the LSTM unit series network is Θ A , update the weight matrix Θ according to the Adam algorithm A get Repeat the above process and use the Adam algorithm to update the weight matrix until Θ A and The training process of the LSTM unit series network is completed.

[0044] The training process of the DNN signal detector in step 4 is as follows:

[0045] Let several groups of known received signals Y = {s 1 ,s 2 ,…,s M} and the set of active terminal states of the transmitted signal are used as the input of the DNN signal detector, and the set of received signals corresponding to each active terminal state is U={u 1 ,u 2 ,…,u M} is used as the output of the DNN signal detector. The DNN signal detector is trained and the average error is used as the loss function. The loss function Repeat the above process, use the gradient descent method to update the weight matrix, take the weight matrix with the smallest loss function as the optimal weight matrix, and obtain the optimal weight matrix for the DNN signal detector At this point the DNN signal detector training is complete.

[0046] Beneficial effects of the present invention:

[0047] 1. This invention is the first to use mutation detection theory to detect non-stationary time points in the industrial Internet. At non-stationary time points, the LSTM-based active terminal detection module is restarted to improve the accuracy of active terminal estimation and achieve highly reliable unlicensed transmission.

[0048] 2. The present invention fully considers the impact of random arrival of services, HARQ and other retransmission mechanisms on active terminals in the NOMA system, and adopts a long short-term memory network to establish a nonlinear mapping relationship between received signals and active user sets and transmitted symbols, thereby achieving high-accuracy multi-user signal detection and improving the reliability of the system.

[0049] 3. This invention is the first to propose a machine learning-based scheduling-free NOMA technology for the Industrial Internet, which alleviates the contradiction between high reliability, low latency and large connections, and supports the implementation of the Industrial Internet.

[0050] 4. The results of the present invention are applicable to industrial Internet of Things wireless communication systems. In particular, for industrial Internet applications involving automatic control, the research results of the present invention can also assist in guiding the setting of service arrival parameters such as the sampling frequency of terminal services and the sending frequency of control / interaction commands, and provide technical guidance for the integrated design of control and communication. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a schematic diagram of the LSTM unit series network in the present invention.

[0052] Figure 2 This is a schematic diagram of the LSTM unit structure in the present invention. DETAILED DESCRIPTION

[0053] Example 1

[0054] A NOMA signal detection method based on deep learning for non-stationary industrial Internet, including the following contents:

[0055] Step 1: The change in the number of active terminals affects the stability of the received signal. That is to say, the sudden change in the number of active terminals will cause the stability of the signal to change, making the duration of the signal stability uncertain. Therefore, it is necessary to predict the non-stationary point and estimate the duration of the stable process. A non-stationary time point detection module is designed to detect the non-stationary point. The module uses the maximum likelihood estimation method to detect the non-stationary point and estimate the duration of the stable process. Specifically:

[0056] Decompose the received signal y(n) with W points in the discrete time domain to obtain M basic mode components c 1 ,c 2 ,…,c M And the residual component r, the received signal y(n) is:

[0057]

[0058] Where j = 1, 2, ..., M; n is the discrete time domain value of the received signal, n = 1, 2, ..., W;

[0059] Performing a Hilbert transform on each basic mode component, y(n) is written as a real number a j (n) and plural The product is in the form of:

[0060]

[0061] The stationarity of the received signal is defined as:

[0062]

[0063] The Hilbert spectrum H of the received signal y(n) is:

[0064]

[0065] b j is the frequency of the jth fundamental mode component. When the signal frequency is equal to ω j When b j =1; when the signal frequency is not equal to ω j When b j =0,ω j For b j The relevant frequency domain is a real number;

[0066] The average boundary spectrum B of the received signal y(n) is:

[0067]

[0068] The stationarity DS(ω) can quantitatively detect the stationarity of data: for a stationary process, the Hilbert spectrum H of y(n) does not change with time, and DS(ω) = 0; if DS(ω) is not zero, it is a non-stationary time point, and as DS(ω) increases, the non-stationarity of the signal increases; the duration of DS(ω) = 0 is recorded, that is, the duration of the stationary process, and the time period with the most occurrences in the duration of the stationary process is selected as the predicted duration of the next stationary process, so the time when the next non-stationary point will appear can be predicted based on the duration of the stationary process;

[0069] Step 2: construct a LSTM unit series network consisting of L LSTM units, such as Figure 1 and Figure 2 As shown;

[0070] Step 3, 1) when the received signal y(n) is in a stable time period, the received signal is used as the input of the LSTM unit series network, and the weighted output of the output gates of L LSTM units is used as the overall output of the LSTM unit series network, the output is the state of each terminal (the output gate of each LSTM unit has an output, and the overall output of the LSTM unit series network is obtained by weighting these outputs, that is, the input to the LSTM unit series network is the received signal, and each LSTM unit predicts the terminal state of the sending terminal corresponding to each signal in the received signal, and then the terminal state obtained after each signal is processed by L LSTM units is weighted to be Ω k), so for the state Ω of the kth terminal k Then we have:

[0071]

[0072] Where l = 1.2.3…L, w l is the weight coefficient of the output gate of the lth LSTM unit, o l k is the output vector of the output gate of the lth LSTM unit in the LSTM unit series network of the sending signal of the kth terminal;

[0073] For Ω k , if Ω k The value of is greater than 0.5, then Ω k =1, that is, terminal k is active; if Ω k The value of is less than 0.5, then Ω k =0, that is, terminal k is inactive, so the active terminal state set is obtained;

[0074] 2) When the received signal y(n) is at a non-stationary point, the LSTM unit series network established in step 2 is retrained, and then the active terminal state set is obtained according to step 1);

[0075] Step 4: Demodulate the input signal using the SIC signal detection method based on deep neural network:

[0076] Set all active terminal states obtained in step 3 And the received signal y(n) is input into the DNN signal detector, the DNN signal detector demodulates the input information and outputs the demodulated signal.

[0077] The DNN signal detector includes an input layer, a plurality of hidden layers and an output layer, and the number of hidden layers is equal to the number of terminals; The input signal is input into the input layer of the DNN signal detector, passes through all the hidden layers in sequence, obtains the demodulated signal, and outputs it through the output layer. The DNN signal detector finds out which active terminal each signal is sent from from the received signal, that is, restores the mixed received signals and finds out which terminal each signal is sent from. That is, the demodulated signal obtained by the DNN signal detector corresponds all the signals in the received signal to the active terminals that actually sent them.

[0078] In the step 2, an LSTM unit is composed of three parts: a forget gate, an input gate and an output gate; wherein:

[0079] The forget gate is expressed as follows:

[0080] f t =σ(Wf ×[h t-1 ,y(n)]+b f )

[0081] Where: h t-1 is the output of the previous LSTM unit output gate, y(n) is the received signal, W f is the forget gate weight, b f is the forget gate bias, σ is the Sigmoid activation function;

[0082] The input gate is represented as follows:

[0083] i t =σ(W i *[h t-1 ,y(n)]+b i )

[0084] Where: W i is the input gate weight, b i Bias for input gate;

[0085] The output gate is represented as follows:

[0086] o t =σ(W o *[h t-1 ,y(n)]+b o )

[0087] Where: W o is the output gate weight, b o Bias for the output gate.

[0088] The training process of the LSTM unit series network in step 2 is as follows:

[0089] There is a known set of terminal states There are several groups of sending signals corresponding to each terminal in the terminal state set. The known sending signals of each terminal are used as the input of the LSTM unit series network, and the terminal state set is used as the output of the LSTM unit series network. The average error is used as the loss function. The weighted output O of the output gates of L LSTM units p As the overall output of the LSTM unit series network, the average error Where K is the number of terminals, and the number of terminals is equal to the number of LSTM units L, p = 1.2.3...K, then the weight matrix of the LSTM unit series network is Θ A , update the weight matrix Θ according to the Adam algorithm A get Repeat the above process and use the Adam algorithm to update the weight matrix until Θ A and The training process of the LSTM unit series network is completed.

[0090] The training process of the DNN signal detector in step 4 is as follows:

[0091] Let several groups of known received signals Y = {s 1 ,s 2 ,…,s M} and the set of active terminal states of the transmitted signal are used as the input of the DNN signal detector, and the set of received signals corresponding to each active terminal state is U={u 1 ,u 2 ,…,u M} is used as the output of the DNN signal detector. The DNN signal detector is trained and the average error is used as the loss function. The loss function Repeat the above process, use the gradient descent method to update the weight matrix, take the weight matrix with the smallest loss function as the optimal weight matrix, and obtain the optimal weight matrix for the DNN signal detector At this point the DNN signal detector training is complete.

[0092] Example 2

[0093] A NOMA signal detection method based on deep learning for non-stationary industrial Internet, including the following contents:

[0094] For the Industrial Internet, in non-emergency situations, the terminal's services are mainly command-type services and status detection feedback services. The arrival process of these services is a stable random process. However, in emergency situations such as sending alarms, burst services will be generated, the number of active terminals in the network, and the service characteristics carried by the terminals will suddenly change, and the aggregation arrival process and even the terminal arrival process will produce non-stationary points. According to the above analysis, the Industrial Internet terminal service belongs to the situation of burst non-stationary but piecewise stable. Based on wavelet decomposition, the present invention designs a non-stationary time point detection module, and uses the cumulative sum test method and the maximum likelihood estimation method to detect non-stationary points and estimate the duration. At the non-stationary point, retrain the LSTM network.

[0095] In the industrial Internet uplink scheduling-free NOMA system, there is a base station and K terminals. Both the base station and the terminal are equipped with a single RF antenna. The upper-layer application data packets arrive at the terminal MAC layer cache randomly. When the terminal cache is not empty, it will transmit data packets to the base station with a certain probability.

[0096] The terminals share N resource blocks, and the signal y(n) received by the base station (each active terminal sends a signal, and the received signal of the base station is a mixture of the signals sent by all active terminals) is expressed as:

[0097]

[0098] Among them, P k represents the transmission power of terminal k. If terminal k transmits a data packet to the base station, the transmission power P k >0, otherwise P k =0; H k represents the channel coefficient vector of terminal k, c k represents the spreading sequence of terminal k, x k represents the complex signal symbol modulated by terminal k, and v represents the noise vector; for the unlicensed NOMA system that combines power domain and code domain, when terminal k transmits, it will randomly select the transmission power P from the candidate power set. k ; Let Z k represents the product of the channel coefficient vector of terminal k and the spreading sequence, and equation (1) is rewritten as:

[0099]

[0100] The problem of estimating the active terminal set can be abstracted as follows:

[0101]

[0102] in, is the optimal weight matrix of LSTM; f A Represents the relationship between active terminals, input signals, and the optimal weight matrix.

[0103] The multi-user signal detection problem is abstracted as:

[0104]

[0105] where f D Represents the mapping relationship between the DNN input and weights and the received signal, active terminals and the optimal weight matrix; is the set of active terminals, is the optimal weight matrix of DNN.

[0106] In the scheduling-free NOMA system, on the one hand, the impact of the MAC layer retransmission mechanism on reliability cannot be ignored. By controlling the time interval and number of retransmissions, the research results show that a balance can be achieved between latency and reliability QoS; on the other hand, under the HARQ retransmission mechanism, that is, the terminal will receive ACK feedback from the base station after successful transmission, and the evolution of the active terminal set over time is closely related to the retransmission mechanism. The potential time correlation of the evolution process provides the possibility of estimating the active terminal set in the current time slot from historical information. Therefore, an active terminal estimation algorithm based on LSTM is proposed.

[0107] Step 1: Due to the non-stationary factors caused by the arrival of services and changes in wireless channels in the industrial Internet, such as the sudden increase in the number of communication terminals caused by the surge in services, or the significant changes in the state of the channel, the non-stationary factors will lead to the deterioration of delay and reliability performance. In order to resist the negative impact of non-stationary factors, a non-stationary time point detection module is designed to detect non-stationary points and reduce the impact of the above non-stationary factors. The module uses the maximum likelihood estimation method to detect non-stationary points and estimate the duration; specifically:

[0108] Based on the local characteristics of the signal, the received signal y(n) with a length of W is decomposed to obtain M basic mode components c 1 ,c 2 ,…,c M And the residual component r, where the received signal y(n) is:

[0109]

[0110] Where j = 1, 2, ..., M; n = 1, 2, ..., W, is a discrete time series;

[0111] Performing a Hilbert transform on each basic mode component, y(n) is written as a real number a j (n) and plural The product is in the form of:

[0112]

[0113] Among them, a is a real number, i is an imaginary number indicating the sign, ω j (n) is the phase of the fundamental mode component;

[0114] Stationarity is defined as:

[0115]

[0116] The Hilbert spectrum H of the received signal y(n) is expressed as:

[0117]

[0118] ω j is the frequency of the jth fundamental mode component, ω is the signal frequency, and i is the imaginary number representing the sign;

[0119] The average boundary spectrum B of the received signal y(n) is:

[0120]

[0121] The stationarity DS(ω) can quantitatively detect the stationarity of data: for a stationary process, the Hilbert spectrum H of y(n) does not change with time and only contains horizontal contour lines. At this time, DS(ω)=0; if DS(ω) is not zero, it is a non-stationary time point. As DS(ω) increases, the non-stationarity of the signal increases; record the duration of the stationary process when DS(ω)=0, and predict the time of the next non-stationary point according to the maximum likelihood principle, that is, select the time period with the most occurrences in the duration of the stationary process as the predicted next stationary duration, so the time point when the next non-stationary point will appear can be predicted according to the stationary duration;

[0122] Step 2: construct an LSTM unit series network consisting of L LSTM units connected in series (a series network is a network in which multiple LSTM units are connected end to end in sequence to form a straight line), and the number of LSTM units in the network is determined by the maximum number of retransmissions of a data packet (the maximum number of retransmissions of a data packet is determined by the protocol and is a known constant for the present invention);

[0123] An LSTM unit consists of three parts: forget gate, input gate and output gate; among them:

[0124] The forget gate formula is as follows:

[0125] f t =σ(W f ×[h t-1 ,y(n)]+b f )

[0126] where h t-1 is the output of the previous LSTM unit output gate, y(n) is the received signal, W f is the forget gate weight, b f is the forget gate bias, σ is the Sigmoid activation function;

[0127] The input gate formula is as follows:

[0128] i t =σ(W i *[h t-1 ,y(n)]+b i )

[0129] Where: W i is the input gate weight, b i Bias for input gate;

[0130] The cell state formula is as follows:

[0131]

[0132] The output gate formula is as follows:

[0133] o t =σ(W o *[h t-1 ,y(n)]+b o )

[0134] Where: W o is the output gate weight, b o Bias for the output gate;

[0135] The LSTM unit structure is as follows Figure 2 As shown:

[0136] Where: h t is the output of the current LSTM unit, σ is the Sigmoid activation function, f t is the forget gate of the current LSTM unit, i t is the input gate of the current LSTM unit, o t is the output gate of the current LSTM unit, is the new data of the current LSTM unit, C t is the cell state of the current LSTM unit;

[0137] Step 3: 1) When the received signal y(n) is in a stable period, the signals (received signals) sent by each terminal in step 1 are used as the input of the LSTM unit series network, and the weighted outputs of the output gates of the L LSTM units are used as the overall output of the LSTM unit series network. The state Ω of the kth terminal is k (The terminal state is estimated by weighting these LSTM unit gates) then:

[0138]

[0139] Where l = 1.2.3…L, is the weight coefficient of the output gate of the lth LSTM unit of the kth terminal, The signal sent by the kth terminal is the output of the output gate of the lth LSTM unit in the LSTM unit series network;

[0140] For Ω k , if Ω k The value of is greater than 0.5, then Ω k =1, that is, terminal k is active; if Ω k The value of is less than 0.5, then Ω k =0, that is, terminal k is inactive, so the active terminal set is obtained;

[0141] 2) When the received signal y(n) is at a non-stationary point (the non-stationary point is caused by a sudden change in the number of terminals at a certain time point), the LSTM unit series network established in step 2 is retrained to improve accuracy, and then the active terminal set is obtained according to step 1);

[0142] Step 4: Demodulate the input signal using the SIC (Long Short-Term Memory Network) signal detection method based on a deep neural network (DNN):

[0143] All active terminals obtained in step 3 The signal sent by the active terminal, i.e., the received signal y(n), is input into the DNN signal detector, which demodulates the input information and outputs the demodulated signal sent by the active terminal. According to the above, all terminal states constitute the terminal state set If the kth terminal state is active, then The corresponding terminal state Ω k is 1; if the kth terminal state is inactive, then The corresponding terminal state Ω k is 0. The received signal y(n) is the signal {y 1 ,y 2 ,…,y L}composition.

[0144] The DNN signal detector includes an input layer, several hidden layers and an output layer. The number of hidden layers is equal to the number of terminals. The input layer is used to input data. The value of the hth node of the zth hidden layer is expressed as: Where D is the number of nodes in the h-1th hidden layer, v dh is the weight, is the input of the dth node in the h-1th layer, and the output layer is used to output the result; The input signal is input into the input layer of the DNN signal detector, passes through all the hidden layers in sequence, and obtains the demodulated signals sent by all active terminals. The demodulated signals are output through the output layer.

[0145] The problem abstracted by the DNN signal detector is:

[0146]

[0147] in is the output of the DNN signal detector, y(n) is the received signal, is the terminal state set consisting of all terminal states, is the optimal weight matrix for detecting signals in the DNN signal detector, f DRepresents the mapping relationship between the output of the DNN signal detector, the received signal, the active terminal, and the optimal weight matrix;

[0148] The training process of the LSTM cell cascade network in the second step is as follows:

[0149] To optimize the performance of the LSTM cascade network, Adam optimization is used to train the LSTM cell cascade network to obtain the optimal weight matrix inside the LSTM cascade network

[0150] There is a known set of terminal states And the received signals corresponding to each terminal state. The received signals of each known terminal state are used as the input of the LSTM cell cascade network. The output of the LSTM cell cascade network is the active terminal to be obtained. The mean squared error is used as the loss function, and the weighted output O of the output gates of L LSTM cells p Is used as the overall output of the LSTM cell cascade network, and the mean squared error Among them, K is the number of terminals, p = 1, 2, 3... K. At this time, the weight matrix of the LSTM cell cascade network is Θ A , and the weight matrix Θ is updated according to the Adam algorithm A To obtain Repeat the Adam algorithm to update the weight matrix until Θ A And Are equal, and the training process of the LSTM cell cascade network ends.

[0151] The role of the LSTM cascade network is to detect active terminals. The input is a set of signals, and the output is the active terminals that send signals.

[0152] The training process of the DNN signal detector in the fourth step is as follows:

[0153] Among them, during the training phase of the DNN signal detector, the weight matrix of the DNN signal detector is Θ D , and the loss function is J(Θ D ). The weight matrix is updated using the gradient descent method, as shown in Equation (7):

[0154]

[0155] Among them, q represents the q-th update using the gradient descent method;

[0156] A number of known received signal sets Y = {s 1 , s 2 , …, s M} are used as the input of the DNN signal detector, then the output of the DNN signal detector is the demodulated signal set U = {u 1 , u2 ,…,u M}; For the training of DNN signal detector, the average error is used as the loss function, and the loss function The weight matrix is ​​updated using the gradient descent method, and the weight matrix with the smallest loss function is taken as the optimal weight matrix to obtain the optimal weight matrix for the DNN signal detector. At this point the DNN signal detector training is complete.

[0157] The training of DNN is to input data into DNN to compare the difference between DNN's prediction and reality, adjust DNN's internal matrix according to the loss function, and repeat the process of minimizing the loss function.

Claims

1. A NOMA signal detection method based on deep learning for non-stationary industrial Internet. Features It includes the following: Step 1: Use the maximum likelihood estimation method to detect non-stationary points and estimate the duration of the stationary process; specifically: Decompose the received signal y(n) with W points in the discrete time domain to obtain M basic mode components c 1 ,c 2 ,…,c M And the residual component r, the received signal y(n) is: Where j = 1, 2, ..., M; n is the discrete time domain value of the received signal, n = 1, 2, ..., W; Performing a Hilbert transform on each basic mode component, y(n) is written as a real number a j (n) and plural The product is in the form of: The stationarity of the received signal is defined as: The Hilbert spectrum H of the received signal y(n) is: b j is the frequency of the j-th basic mode component. When the signal frequency equals ω j , b j = 1; when the signal frequency is not equal to ω j , b j = 0, where ω j is a real number in the frequency domain related to b j . The average boundary spectrum B of the received signal y(n) is: The stationarity DS(ω) can quantitatively detect the stationarity of data: for a stationary process, the Hilbert spectrum H of y(n) does not change with time, and DS(ω) = 0; if DS(ω) is not zero, it is a non-stationary time point, and as DS(ω) increases, the non-stationarity of the signal increases; the duration of DS(ω) = 0 is recorded, that is, the duration of the stationary process, and the time period with the most occurrences in the duration of the stationary process is selected as the predicted duration of the next stationary process, so the time when the next non-stationary point will appear can be predicted based on the duration of the stationary process; Step 2: construct an LSTM unit series network consisting of L LSTM units; Step 3: 1) When the received signal y(n) is in a stable period, the received signal is used as the input of the LSTM unit series network, and the weighted output of the output gates of L LSTM units is used as the overall output of the LSTM unit series network. The output is the state of each terminal, so the state Ω of the kth terminal is k Then we have: Where l = 1.2.3…L, w l is the weight coefficient of the output gate of the lth LSTM unit, is the output vector of the output gate of the lth LSTM unit in the LSTM unit series network of the sending signal of the kth terminal; For Ω k , if Ω k The value of is greater than 0.5, then Ω k =1, that is, terminal k is active; if Ω k The value of is less than 0.5, then Ω k =0, that is, terminal k is inactive, so the active terminal state set is obtained; 2) When the received signal y(n) is at a non-stationary point, the LSTM unit series network established in step 2 is retrained, and then the active terminal state set is obtained according to step 1); Step 4: Demodulate the input signal using the SIC signal detection method based on deep neural network: Set all active terminal states obtained in step 3 And the received signal y(n) is input into the DNN signal detector, the DNN signal detector demodulates the input information and outputs the demodulated signal.

2. According to the NOMA signal detection method based on deep learning for non-stationary industrial Internet according to claim 1, Features The DNN signal detector includes an input layer, a plurality of hidden layers and an output layer, and the number of hidden layers is equal to the number of terminals; The input signal is sent to the input layer of the DNN signal detector, passes through all the hidden layers in sequence, obtains the demodulated signal, and is output through the output layer.

3. According to the NOMA signal detection method based on deep learning for non-stationary industrial Internet according to claim 1, Features In the step 2, an LSTM unit is composed of three parts: a forget gate, an input gate and an output gate; wherein: The forget gate is expressed as follows: f t =σ(W f ×[h t-1 ,y(n)]+b f ) Where: h t-1 is the output of the previous LSTM unit output gate, y(n) is the received signal, W f is the forget gate weight, b f is the forget gate bias, σ is the Sigmoid activation function; The input gate is represented as follows: i t =σ(W i *[h t-1 ,y(n)]+b i ) Where: W i is the input gate weight, b i Bias for input gate; The output gate is represented as follows: the t =σ(W o *[h t-1 ,y(n)]+b o ) Where: W o is the output gate weight, b o Bias for the output gate.

4. According to the NOMA signal detection method based on deep learning for non-stationary industrial Internet according to claim 3, Features The training process of the LSTM unit series network in step 2 is as follows: There is a known set of terminal states There are several groups of sending signals corresponding to each terminal in the terminal state set. The known sending signals of each terminal are used as the input of the LSTM unit series network, and the terminal state set is used as the output of the LSTM unit series network. The average error is used as the loss function. The weighted output O of the output gates of L LSTM units p As the overall output of the LSTM unit series network, the average error Where K is the number of terminals, and the number of terminals is equal to the number of LSTM units L, p = 1.2.3...K, then the weight matrix of the LSTM unit series network is Θ A , update the weight matrix Θ according to the Adam algorithm A get Repeat the above process and use the Adam algorithm to update the weight matrix until Θ A and The training process of the LSTM unit series network is completed.

5. According to a NOMA signal detection method based on deep learning for non-stationary industrial Internet according to claim 2, Features The training process of the DNN signal detector in step 4 is as follows: Let several groups of known received signals Y = {s 1 ,s 2 ,…,s M } and the set of active terminal states of the transmitted signal are used as the input of the DNN signal detector, and the set of received signals corresponding to each active terminal state is U={u 1 ,u 2 ,…,u M } is used as the output of the DNN signal detector. The DNN signal detector is trained and the average error is used as the loss function. The loss function Repeat the above process, use the gradient descent method to update the weight matrix, take the weight matrix with the smallest loss function as the optimal weight matrix, and obtain the optimal weight matrix for the DNN signal detector At this point the DNN signal detector training is complete.

Images