Method and system for detecting active state of base station access device

Abstract

The present invention discloses a method and system for detecting the active state of a base station access device, including: constructing and training a preamble detection neural network to extract first channel state information from the preamble set and covariance information of the received signal; constructing and training a data detection neural network to extract second channel state information from the data signal, preamble set and first channel state information in the received signal; a spatial expansion unit converts the second channel state information into an active state estimation result based on a preset conversion threshold; the base station aggregates the results of all spatial expansion units and outputs a final active state detection result. The present invention improves the accuracy of channel estimation and active state estimation; combines deep learning and covariance detection to solve the defects of traditional compressed sensing methods, better utilizes short non-orthogonal preambles, improves the detection performance under the application of short preamble sequences, and performs active detection and channel estimation without a complex scheduling process.

Description

Technical Field

[0001] The present invention relates to the technical field of wireless communication networks, and particularly to a method and system for detecting the active state of a base station access device. Background Art

[0002] With the rapid development of the power Internet of Things and wireless communication, the power Internet of Things needs to support the simultaneous access and communication of a large number of devices. These devices usually carry very small data packets and require extreme coverage. In an unlicensed communication mode, a device can directly transmit a data signal after a preamble signal without the authorization of a base station, which leads to a high collision risk and the problem that the number of potential active users is difficult to predict. Traditional compressive sensing methods have performance limitations in the joint detection of user activity and channel state. Especially when the number of users exceeds the length of the preamble sequence, they cannot meet the access requirements of a large number of devices, and the performance decays severely when the coverage area is large. In addition, the detection performance of compressive sensing methods decreases significantly when processing short preamble sequences.

[0003] In the "Data Transmission Method, Device, Electronic Device and Storage Medium" disclosed in the Chinese patent literature, with the publication number CN115333709A and the publication date of November 11, 2022, the method includes: determining a first compressive sensing model and a second compressive sensing model based on target signals sent by at least one active terminal; the target signals include a preamble and valid data; determining the channel state information corresponding to each active terminal and each active terminal based on the first compressive sensing model; and determining the valid data corresponding to each active terminal based on each active terminal, the channel state information corresponding to each active terminal, and the second compressive sensing model. This technology realizes the one-step synchronous transmission of the preamble and valid data through the target signals sent by the active terminal including the preamble and valid data, and determines the valid data transmitted by each active terminal based on the first compressive sensing model and the second compressive sensing model, completing the transmission of the valid data while completing the transmission of the preamble, so that the transmission delay and signaling overhead are greatly reduced. However, it does not solve the performance limitations of compressive sensing methods in the joint detection of user activity and channel state, cannot meet the access requirements of a large number of devices, and the performance decays severely when the coverage area is large. Summary of the Invention

[0004] The present invention is to overcome the problems that when using the compressive sensing method to detect the active state of access devices in the prior art, there are performance limitations, cannot meet the access requirements of a large number of devices, and the performance decays severely when the coverage area is large, and provides a method and system for detecting the active state of a base station access device.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for detecting the active state of a base station access device, comprising:

[0007] Construct and train a preamble detection neural network, and extract first channel state information from the preamble set and covariance information of the received signal;

[0008] Construct and train a data detection neural network, and extract second channel state information from the data signal, preamble set and first channel state information in the received signal;

[0009] The spatial expansion unit converts the second channel state information into an active state estimation result based on a preset conversion threshold;

[0010] The base station aggregates the active state estimation results output by all spatial expansion units and outputs a final active state detection result.

[0011] In the present invention, after the spatial expansion unit receives the signals in its respective regions, the sample covariance of the received signal is calculated; the processed signal covariance information and preamble set are used as inputs and fed into the preamble detection neural network (PDNN) to extract the first channel state information; the output first channel state information, the data segment signal of the received signal, and the preamble set are used as inputs to the detection neural network (DDMM), and the detection neural network outputs more accurate second channel state information; the second channel state information is processed using a preset conversion threshold to output an active state estimation result; finally, the spatial expansion unit transmits the trained model parameters to the base station, and the base station integrates them into the global model to obtain the final global prediction detection result; the present invention improves the accuracy of channel estimation and active state estimation when the device uses short preamble sequences; by combining deep learning and covariance detection methods to better solve the problems of traditional compressive sensing methods, better utilize short non-orthogonal preambles, improve the detection performance under short preamble sequence applications, and perform user activity detection and channel estimation without a complex scheduling process.

[0012] Preferably, the converting the second channel state information into an active state estimation result based on a preset conversion threshold includes:

[0013] Compare the square of the modulus value of the second channel state information with the conversion threshold. If it is greater than the conversion threshold, the active state estimation result is the active state, otherwise it is the inactive state.

[0014] Preferably, the preamble detection neural network includes T preamble neural network layers, and each preamble neural network layer receives the channel estimation result of the previous preamble neural network layer, the preamble set of the received signal, and the covariance information for estimation update, and outputs a new channel estimation result.

[0015] Preferably, each preamble neural network layer in the preamble detection neural network includes a number of first hidden layers; iterative calculations are performed:

[0016] The output of each first hidden layer is the product of the output of the previous first hidden layer and the weight of the current layer, added with the bias vector of the current layer, and then multiplied by the activation function of the current layer; in the first first hidden layer, the input of the corresponding preamble neural network layer is used as the output of the previous first hidden layer.

[0017] Preferably, when training the preamble detection neural network, the square of the Euclidean norm is calculated based on the difference between the channel estimation result and the true result of each preamble neural network layer, and then weighted summation is performed as the first loss function; the square of the Euclidean norm is calculated based on the difference between the channel estimation result and the true result of the preamble detection neural network and summed as the second loss function;

[0018] First, training is performed to reduce the value of the first loss function until convergence, and then training is performed to minimize the second loss function.

[0019] Preferably, the data detection neural network includes a number of second hidden layers;

[0020] A preset layer number threshold is set. For the second hidden layer with a layer number less than the layer number threshold, its input is the output of the previous second hidden layer and the input of the data detection neural network; for the second hidden layer with a layer number greater than or equal to the layer number threshold, its input is the output of the previous second hidden layer.

[0021] Preferably, the data processing process of the second hidden layer includes: for any second hidden layer, its output is the product of the input of the second hidden layer and the corresponding weight, added with the corresponding bias vector, and then multiplied by the corresponding activation function.

[0022] Preferably, when training the data detection neural network, training is performed to minimize the value of the third loss function;

[0023] The third loss function is the sum of the squares of the Euclidean norms calculated based on the difference between the channel estimation result and the true result of the data detection neural network.

[0024] Preferably, a global detection model is set in the base station. After averaging the activity state estimation results from the spatial expansion unit, the weight and bias parameters of the fully connected layer are processed, and finally batch normalization operation is performed to obtain the final activity state detection result.

[0025] An activity state detection system for a base station access device includes a base station and a number of spatial expansion units within the range of the base station;

[0026] The spatial expansion unit is provided with a preamble detection neural network, a data detection neural network, and a conversion threshold. After receiving the signals within its own coverage area, it outputs the estimated results of the device activity status in the corresponding area.

[0027] The base station is provided with a global detection model, which aggregates the estimated results of the activity status output by all spatial expansion units and outputs the final activity status detection result.

[0028] The present invention has the following beneficial effects: The base station efficiently utilizes the covariance structure of the signals through a neural network to detect the activity status. Detecting based on the covariance structure of the signals overcomes the problems existing in the compressed sensing technology, such as the inability to meet the access requirements of a large number of devices and the serious performance degradation when the coverage area is large, improves the performance in the small data packet transmission scenario, and at the same time applies the neural network of deep learning. The trained deep learning model not only performs well under the channel conditions during training, but also has good generalization ability under different channel statistical conditions; after offline training, it can be deployed in real time at the base station site to extract the useful features of the received signals and is suitable for application in various actual environments; by adopting the method combining covariance activity detection and deep learning neural network, the detection efficiency and accuracy of activity status detection in the background of large-scale access are improved, and the base station can maximize the efficiency of activity status and channel estimation when a large number of devices are accessing. Description of the Drawings

[0029] Figure 1 It is a flowchart of the method for detecting the activity status of the devices accessed by the base station in the present invention.

[0030] Figure 2 It is a schematic diagram of the activity detection channel estimation neural network in the present invention. Detailed Embodiments

[0031] The following further describes the present invention in combination with the drawings and specific embodiments.

[0032] Traditional compressed sensing methods have performance limitations in the joint detection of user activity and channel status, especially when the number of users exceeds the length of the preamble sequence. In addition, the performance of compressed sensing methods degrades significantly when dealing with short preamble sequences. To solve these problems, the present invention proposes a covariance detection method, the advantage of which is that it can utilize the covariance structure of the signals to detect user activity. This method does not rely on orthogonal pilot sequences, so it can work under non-orthogonal pilot sequences and allows more users to access the network simultaneously. In addition, this method improves the detection performance by utilizing the macro diversity gain provided by large-scale antenna arrays, especially showing better performance compared to the co-located deployment architecture when the coverage area is large. At the same time, in response to the defects of the compressed sensing algorithm, an activity detection channel estimation neural network UAD-CE-NN is designed to improve the detection accuracy.

[0033] As shown in Figure 1 , a method for detecting the active state of a base station access device includes:

[0034] Construct and train a preamble detection neural network to extract the first channel state information from the preamble set and covariance information of the received signal;

[0035] Construct and train a data detection neural network to extract the second channel state information from the data signal, preamble set, and the first channel state information in the received signal;

[0036] The spatial expansion unit converts the second channel state information into an active state estimation result based on a preset conversion threshold;

[0037] The base station aggregates the active state estimation results output by all spatial expansion units and outputs the final active state detection result.

[0038] It should be noted that in the present invention, after the spatial expansion unit receives the signals in its respective regions, the sample covariance of the received signal is calculated; the processed signal covariance information and the preamble set are used as inputs and sent to the preamble detection neural network (PDNN) to extract the first channel state information; the output first channel state information, the data segment signal of the received signal, and the preamble set are used as inputs to the detection neural network (DDMM), and the detection neural network outputs more accurate second channel state information; the second channel state information is processed using a preset conversion threshold to output an active state estimation result; finally, the spatial expansion unit transmits the trained model parameters to the base station, and the base station integrates them into the global model to obtain the final global prediction detection result; the present invention improves the accuracy of channel estimation and active state estimation when the device uses short preamble sequences; by combining deep learning and covariance detection methods, it better solves the problems of traditional compressed sensing methods, makes better use of short non-orthogonal preambles, improves the detection performance under short preamble sequence applications, and performs user activity detection and channel estimation without a complex scheduling process.

[0039] It should be noted that the base station efficiently utilizes the covariance structure of signals through a neural network to detect the activity status. Detecting based on the covariance structure of signals overcomes the problems existing in compressive sensing technology, such as the inability to meet the access requirements of a large number of devices and the serious performance degradation when the coverage area is large, improves the performance in the small data packet transmission scenario, and at the same time applies the neural network of deep learning. The trained deep learning model not only performs well under the channel conditions during training, but also has good generalization ability under different channel statistical conditions; after offline training, the activity detection channel estimation neural network UAD-CE-NN can be deployed in real time at the base station site to extract useful features of the received signals and is suitable for application in various actual environments; by adopting the method of combining covariance activity detection and deep learning neural network, the detection efficiency and accuracy of activity status detection in the context of large-scale access are improved, and the base station can maximize the efficiency of activity status and channel estimation when a large number of devices are accessing.

[0040] As a specific embodiment, converting the second channel state information into an activity status estimation result based on a preset conversion threshold includes:

[0041] Compare the square of the modulus value of the second channel state information with the conversion threshold. If it is greater than the conversion threshold, the activity status estimation result is the active state; otherwise, it is the inactive state.

[0042] It should be noted that the activity detection channel estimation neural network UAD-CE-NN of the present invention mainly includes a preamble detection neural network and a data detection neural network located in the spatial expansion unit, and a global detection model located in the base station. According to the signals received in real time, the preamble detection neural network and the data detection neural network output the corresponding second channel state information, and then convert the second channel state information into an activity status estimation result and transmit it to the base station, and the global detection model fuses the results of each spatial expansion unit to obtain the final activity status estimation result.

[0043] Specifically, when converting the second channel state information into an activity status estimation result, by selecting an appropriate conversion threshold as an offset and estimating with a shrinkage function with an offset:

[0044]

[0045] where is the estimation of the activity status by the data detection neural network, is the conversion threshold. Specifically:

[0046]

[0047] where Represents the square of its modulus value. When the result is 1, the device is in an active state, and when the result is 0, the device is in an inactive state.

[0048] As a specific embodiment, as Figure 2 shown, the preamble detection neural network includes T preamble neural network layers. Each preamble neural network layer receives the channel estimation result of the previous preamble neural network layer, as well as the preamble set and covariance information of the received signal for estimation update, and outputs a new channel estimation result.

[0049] It should be noted that in the spatial expansion unit SEU, training is performed on the preamble detection neural network PDNN to generate training data, which includes the preamble of the received signal in the corresponding area of the spatial expansion unit and the covariance information of the signal. According to the defined loss function, the neural network model is optimized and trained so that the preamble detection neural network PDNN can learn how to extract the first channel state information from the preamble set and covariance information. The entire preamble detection neural network PDNN is divided into T layers and arranged in sequence. Each preamble detection neural network PDNN layer is actually a specific neural network, and the updated estimate of the channel is obtained by using the received preamble set and the output of the previous PDNN layer. Finally, the result output by the last PDNN layer can be obtained as the first channel state information.

[0050] Furthermore, each preamble neural network layer in the preamble detection neural network includes several first hidden layers;

[0051] Iterative calculations are performed: the output of each first hidden layer is the product of the output of the previous first hidden layer and the weight of the current layer, added with the bias vector of the current layer, and then multiplied by the activation function of the current layer; in the first first hidden layer, the input of the corresponding preamble neural network layer is used as the output of the previous first hidden layer.

[0052] It should be noted that in a neural network, each input feature is multiplied by an element in the weight vector, and then these products are summed. The role of the weight vector is to determine the linear combination method of the input signal when passing through this layer. By adjusting the weights, the neural network can learn the features and patterns of the input data. The bias vector is a constant vector that is added to the linear combination result of the weights and the input signal. The role of the bias is to provide an offset for the output of the neuron, so that the output of the neuron not only depends on the strength of the input signal, but can also be adjusted independently of the input signal to a certain extent, enabling it to better fit the data. For active users, their channel state information can be observed in the received signal, while the channel information of inactive users is not available.

[0053] Further, when training the preamble detection neural network, the square of the Euclidean norm is calculated based on the difference between the channel estimation result and the true result of each preamble neural network layer, and then weighted summation is performed as the first loss function; the square of the Euclidean norm is calculated based on the difference between the channel estimation result and the true result of the preamble detection neural network and summed as the second loss function;

[0054] First, train with the value of the first loss function reduced until convergence, and then train with the second loss function minimized.

[0055] It should be noted that the process of first reducing the first loss function to convergence and then minimizing the second loss function is to improve the estimation accuracy and convergence speed of the neural network; the first loss function evaluates the overall accuracy of the channel estimation of all preamble neural network layers PDNN. By continuously reducing the result of the first loss function, all neural network layers in the PDNN preamble neural network are collectively trained; when the first training step converges, the second training step aims to fine-tune the channel estimation by minimizing the second loss function.

[0056] As a specific embodiment, as Figure 2 shown, the data detection neural network includes several second hidden layers;

[0057] A preset layer number threshold is set. For the second hidden layer with a layer number less than the layer number threshold, its input is the output of the previous second hidden layer and the input of the data detection neural network; for the second hidden layer with a layer number greater than or equal to the layer number threshold, its input is the output of the previous second hidden layer.

[0058] It should be noted that the data detection neural network DDNN is constructed and trained after the preamble detection neural network PDNN is set. It takes the preamble set of the received signal, the data signal, and the first channel state information output by the preamble detection neural network as inputs, and trains it to learn and extract more accurate second channel state information. For the data detection neural network, a specific layer number threshold is set based on its several second hidden layers, so as to distinguish two different hidden layer data processing methods. In the first few layer number threshold second hidden layers, the selection of input data is expanded, so as to ensure data integrity during the training process and detection process. In the subsequent second hidden layers, the focus is on training the output result of the previous layer, which can improve the convergence speed.

[0059] Further, the data processing process of the second hidden layer includes: for any second hidden layer, its output is obtained by multiplying the input of this second hidden layer by the corresponding weight, adding the corresponding bias vector, and then multiplying by the corresponding activation function.

[0060] It should be noted that the data processing method in the second hidden layer of the data detection neural network is the same as that in the first hidden layer of the preamble detection neural network, which is to perform iterative calculations on a number of hidden layers arranged in sequence. The difference lies in that the weights, bias vectors, activation functions, and input data for calculation in each hidden layer are different. After continuous iterative calculations, the output result of the last second hidden layer is used as the second channel state information output by the data detection neural network.

[0061] Furthermore, when training the data detection neural network, it is trained to minimize the value of the third loss function;

[0062] The third loss function is to calculate the sum of the squares of the Euclidean norms of the differences between the channel estimation results and the true results of the data detection neural network.

[0063] It should be noted that the training of the data detection neural network DDNN is similar to that of the preamble detection neural network PDNN. The overall channel evaluation accuracy of the data detection neural network is evaluated through the third loss function, and all neural networks in the data detection neural network DDNN are trained by continuously reducing the result of the third loss function. Since the data detection neural network only includes a number of second hidden layers, which is different from the preamble neural network that includes T preamble neural network layers and each layer includes a number of first hidden layers, only one third loss function is required for neural network training.

[0064] As a specific embodiment, a global detection model is set in the base station. After averaging the active state estimation results from the spatial expansion unit, the weights and bias parameters of the fully connected layer are processed, and finally, a batch normalization operation is performed to obtain the final active state detection result.

[0065] It should be noted that the weights and bias parameters of the fully connected layer are trainable parameters in the global detection model. After training the preamble neural network PDNN and the data detection neural network DDNN, the second channel state information output by the data detection neural network is used as the training set to input into the global detection model of the base station for training, and finally, the required weights and bias parameters are obtained to complete the model configuration in the base station.

[0066] In addition to the method for detecting the active state of the base station access device, the present invention also discloses a system for detecting the active state of the base station access device, including a base station and a number of spatial expansion units within the range of the base station;

[0067] The spatial expansion unit is provided with a preamble detection neural network, a data detection neural network, and a conversion threshold. After receiving the signals within its coverage area, it outputs the device active state estimation results in the corresponding area;

[0068] The base station is provided with a global detection model, which aggregates the active state estimation results output by all spatial expansion units and outputs the final active state detection result.

[0069] It should be noted that the present invention considers configuring several spatial expansion units around the base station. The range of access devices covered by the base station includes the sum of the ranges covered by each spatial expansion unit. Therefore, the specific connection state is that the base station is connected to each spatial expansion unit, and each spatial expansion unit is connected to the access devices within its coverage area. The spatial expansion unit performs local device active state estimation within its coverage area, and then each spatial expansion unit transmits its respective estimation results to the base station for aggregation. The base station performs data processing to obtain the final active state detection result. By adopting the method of combining covariance activity detection and deep learning neural network, the detection efficiency and accuracy of active state detection in the context of large-scale access are improved; at the same time, the base station can maximize the efficiency of active state and channel estimation when a large number of devices are accessing.

[0070] The following takes a certain industrial park as an example to illustrate the active state detection method of the base station access device of the present invention. A large number of power Internet of Things devices are deployed in the park, such as smart meters, environmental monitoring sensors, industrial control systems, etc. These devices need to frequently transmit and communicate data across multiple networks. At the same time, the devices and systems in the park need to perform user activity detection and channel estimation without a complex scheduling process to ensure the efficiency and reliability of communication. The specific parameters in the industrial park are as follows:

[0071]

[0072] Performing active state detection in this embodiment includes the following steps.

[0073] Step 1: The base station assigns a pilot sequence to each device:

[0074]

[0075] where represents the pilot sequence of the k-th user, represents pi, represents the exponential expression with the natural constant e as the base, Let \(L\) denote the length of the pilot sequence, \(j\) denote the imaginary unit, and \(c\) be a constant used to control the phase interval in the pilot sequence. The pilot sequence, as a known reference signal, helps the receiving end estimate the characteristics of the wireless channel, including information such as channel gain and phase. Training is carried out for the PDNN model in the spatial expansion unit SEU to generate training data, which includes the preamble set of the regional received signal and the covariance information of the signal. According to the defined loss function, the deep neural network model is optimized and trained to enable the preamble detection neural network PDNN to learn how to extract the first channel state information from the preamble set and covariance information.

[0076] The preamble set is the set of preamble signals of the received signal; the covariance information of the received signal can be expressed as:

[0077]

[0078] where \(Y\) is the received signal, denotes the conjugate transpose operation of a vector or matrix, and \(K\) is the number of antennas equipped at the base station.

[0079] The entire preamble detection neural network PDNN is divided into T layers, arranged in sequence. Each PDNN layer is actually a specific neural network that uses the received preamble sequence and the output of the previous PDNN layer to obtain an updated estimate of the channel. Finally, the final estimation result output by the last PDNN layer can be obtained.

[0080] To clearly illustrate the processing process of PDNN, take the \(t\)-th layer in PDNN as an example. To ensure that PDNN gets rid of overfitting, vanishing gradients, or falling into local optimal points, the input of the \((t + 1)\)-th PDNN layer is set to

[0081]

[0082] where represents the estimated channel state information of the \(k\)-th antenna in the \(t\)-th PDNN layer, and \(K\) is the number of antennas. The specific input is the vector of covariance information, the estimated channel state information output by the previous PDNN layer, and the product of the preamble set \(S\) and the estimated channel state information output by the previous PDNN layer. As a special case, there is no previous PDNN layer for the first PDNN layer, and at this time , where represents the complex conjugate transpose, represents converting the matrix into a column vector. represents the channel estimation result of the \(k\)-th antenna in the \((t - 1)\)-th PDNN layer.

[0083] The calculation process in the \(t\)-th PDNN layer is expressed as:

[0084]

[0085] where and are the weight and bias vectors in the th first hidden layer of the th PDNN layer, and there are first hidden layers;

[0086] The weight vector represents the connection strength between the input signal and the neurons in this layer. In a neural network, each input feature is multiplied by an element in the weight vector, and then these products are summed. The role of the weight vector is to determine the linear combination method of the input signal when passing through this layer. By adjusting the weights, the network can learn the features and patterns of the input data. The bias vector is a constant vector that is added to the result of the linear combination of the weights and the input signal. The role of the bias is to provide an offset for the output of the neuron, so that the output of the neuron not only depends on the strength of the input signal, but can also be adjusted independently of the input signal to a certain extent, enabling it to better fit the data. The activation function is used to process the output of each layer, enabling the network to extract useful features from the received signals and perform non-linear transformations. By selecting an appropriate activation function, PDNN can perform user activity detection and channel estimation more effectively.

[0087] Thus, the channel estimation result of the th antenna at

[0088]

[0089] where is the weight of the mean square error (MSE) of the th PDNN layer, represents the channel estimation result of the th antenna at the is the number of antennas of the base station. The overall accuracy of the channel estimation of all PDNN units is evaluated according to the first loss function. By continuously reducing the first loss function, all neural networks of the PDNN are collectively trained. When the first training step converges, the second training step aims to fine-tune the channel estimation by minimizing the second loss function:

[0090]

[0091] where represents the channel estimation result of the k-th antenna in the preamble detection neural network.

[0092] After training is completed, the covariance information of the received signal and the preamble set are input in actual applications, and finally the first channel state information of the final output is obtained from the preamble detection neural network .

[0093] Step 2: The data detection neural network DDNN is trained after setting the PDNN. The DDNN uses the received signal and the first channel estimation information output by the PDNN as the training set to train the data detection neural network DDNN to learn to extract the channel state information. The input of the data detection neural network is defined as:

[0094]

[0095] where is the signal of the m-th data segment received by the k-th antenna (there are a total of M data segments in the signal), and the square brackets indicate the stacking of all M data segment signals.

[0096] There are several second hidden layers in the DDNN. In the l-th second hidden layer, the input is defined as

[0097]

[0098] where is the output of the (l - 1)-th second hidden layer, is a specific hidden layer index in the DDNN, corresponding to the layer number threshold, used to distinguish two different data processing methods.

[0099] The output of the l-th second hidden layer is calculated by the following formula:

[0100]

[0101] where represents the weight matrix of the l-th second hidden layer, is the activation function used in the l-th second hidden layer, is the bias vector of the l-th second hidden layer. The bias is a parameter learned in the network and is used to control the output of the activation function. Then, and are both regarded as the input to the next hidden layer until the last second hidden layer of the DDNN, i.e., the output of the last second hidden layer, is used as the channel estimation result .

[0102] For training the DDNN model, similar to the PDNN, the third loss function of the DDNN is defined as:

[0103]

[0104] The overall accuracy of the channel estimation of the DDNN is evaluated according to the third loss function. By continuously reducing the value of the third loss function, all the neural networks of the DDNN are collectively trained.

[0105] After training is completed, similar to the PDNN, the data processing process in the data detection neural network DDNN can be represented by a mapping function:

[0106]

[0107] where is the data processing process in the DDNN, m is the index of the signal received by the m-th antenna in the base station, Y = { } is the set of received signals of the preamble and M data segments, and in addition to the received signals, it also includes the preamble set and the first channel state information output by the PDNN as inputs. In the DDNN, the information of the data segments can be jointly utilized with the preamble set to improve the output result of the PDNN. Finally, more accurate second channel state information is output.

[0108] Step 3: According to the second channel state information obtained by the DDNN, select an appropriate conversion threshold to convert the channel state output result into a device activity result. The activity state is estimated by the following shrinking function with an offset as follows:

[0109]

[0110] where is the estimation of the activity state by the data detection neural network, is the conversion threshold. Specifically:

[0111]

[0112] where Represents the square of its modulus value. When the result is 1, the device is in an active state, and when the result is 0, the device is in an inactive state.

[0113] The training phases proposed by PDNN and DDNN are carried out offline. After the offline training converges, the entire neural network can be deployed in the Spatial Expansion Unit (SEU) to extract useful features of the received signal.

[0114] Step 4: After the local training of all Spatial Expansion Units (SEUs) is completed, the output model parameters are transmitted to the base station, and the base station conducts the training of the global model:

[0115]

[0116] where is the active state detection result output by the global model, is the active state estimation result output by the Spatial Expansion Unit (SEU), BN represents the Batch Normalization layer, Average represents the averaging operation on the local estimation results of all SEUs, W and b respectively represent the weights and bias parameters of the fully connected layer, which are the trainable parameters of the global model on the base station. After the training is completed, the global model collects the output results of all Spatial Expansion Units (SEUs) and outputs the final active state detection result.

[0117] The method of the present invention reduces the active state detection error rate and improves the active state detection accuracy. The following are the detection error probabilities (detecting the active state as an inactive state) of different detection algorithms when the signal-to-noise ratio is 20 dB:

[0118]

[0119] And the detection error probabilities of different detection algorithms when the signal-to-noise ratio is 10 dB:

[0120]

[0121] It is found through comparison that the error probabilities of the traditional approximate message passing method and the orthogonal matching pursuit method in different signal-to-noise ratio environments are higher than those of the neural network detection method of the present invention.

[0122] The above embodiments are further elaborations and explanations of the present invention for easy understanding, and are not any limitations on the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for detecting the activity status of a base station access device, characterized in that: include: Constructing and training a preamble detection neural network to extract first channel state information from a preamble set and covariance information of a received signal; The preamble detection neural network includes T preamble neural network layers, each preamble neural network layer receives the channel estimation result of the previous preamble neural network layer and the preamble set and covariance information of the received signal to perform estimation update, and outputs a new channel estimation result; Constructing and training a data detection neural network to extract second channel state information from a data signal and a preamble set in a received signal and the first channel state information; The spatial expansion unit converts the second channel state information into an activity state estimation result based on a preset conversion threshold; The base station aggregates the activity state estimation results output by all spatial extension units and outputs a final activity state detection result.

2. The method for detecting the activity status of a base station access device according to claim 1, characterized in that: The converting the second channel state information into an activity state estimation result based on a preset conversion threshold comprises: The square of the modulus value of the second channel state information is compared with the conversion threshold. If the modulus value is greater than the conversion threshold, the activity state estimation result is an active state, otherwise it is an inactive state.

3. The method for detecting the activity status of a base station access device according to claim 1, characterized in that: Each preamble neural network layer in the preamble detection neural network includes a plurality of first hidden layers; performing iterative calculation: The output of each first hidden layer is the output of the previous first hidden layer multiplied by the weight of the current layer, plus the bias vector of the current layer, and then multiplied by the activation function of the current layer; in the first first hidden layer, the input of the corresponding preamble neural network layer is used as the output of the previous first hidden layer.

4. The method for detecting the activity status of a base station access device according to claim 1, characterized in that: When training the preamble detection neural network, the square of the Euclidean norm is calculated based on the difference between the channel estimation result and the true result of each preamble neural network layer, and then the weighted sum is performed as the first loss function; the square of the Euclidean norm is calculated based on the difference between the channel estimation result and the true result of the preamble detection neural network, and the sum is used as the second loss function; First, train to reduce the value of the first loss function until convergence, and then train to minimize the second loss function.

5. A method for detecting the activity status of a base station access device according to claim 1 or 2, characterized in that: The data detection neural network includes a plurality of second hidden layers; A layer number threshold is preset. For the second hidden layer whose number of layers is less than the layer number threshold, its input is the output of the previous second hidden layer and the input of the data detection neural network; for the second hidden layer whose number of layers is greater than or equal to the layer number threshold, its input is the output of the previous second hidden layer.

6. The method for detecting the activity status of a base station access device according to claim 5, characterized in that: The data processing process of the second hidden layer includes: for any second hidden layer, its output is, after the input of the second hidden layer is multiplied by the corresponding weight and added with the corresponding bias vector, it is multiplied by the corresponding activation function.

7. The method for detecting the activity status of a base station access device according to claim 5, characterized in that: When training the data detection neural network, the training is performed by minimizing the value of the third loss function; The third loss function is to calculate the square of the Euclidean norm of the difference between the channel estimation result of the data detection neural network and the true result and sum them up.

8. A method for detecting the activity status of a base station access device according to claim 1 or 2 or 3 or 4 or 6 or 7, characterized in that: The base station is provided with a global detection model, which averages the activity state estimation results from the spatial extension units, processes the weights and bias parameters of the fully connected layer, and finally performs a batch normalization operation to obtain the final activity state detection result.

9. An activity status detection system for a base station access device, applicable to the activity status detection method according to any one of claims 1 to 8, characterized in that: It includes a base station and several spatial extension units within the base station; The spatial expansion unit is provided with a preamble detection neural network, a data detection neural network and a conversion threshold, and after receiving a signal within its own coverage area, outputs an estimation result of the device activity state in the corresponding area; The base station is provided with a global detection model, which aggregates the activity state estimation results output by all the spatial expansion units and outputs a final activity state detection result.

Images