CN113971430A

CN113971430A - Signal detection and model training method, device, equipment and storage medium

Info

Publication number: CN113971430A
Application number: CN202110675785.2A
Authority: CN
Inventors: 李新增; 金婕; 马天鸣; 张嘉
Original assignee: Shanghai University of Engineering Science
Current assignee: Shanghai University of Engineering Science
Priority date: 2021-06-21
Filing date: 2021-06-21
Publication date: 2022-01-25

Abstract

The invention provides a signal detection and model training method, a device, equipment and a storage medium, wherein the method comprises the following steps: demodulating the received signal to obtain corresponding demodulation data; extracting a real part and an imaginary part from the demodulated data to form real data corresponding to the demodulated data; inputting the real data into a target deep learning network model obtained by pre-training to obtain an output result; and obtaining the detection result of the original sending signal according to the output result. The invention carries out channel estimation and equalization by adopting a deep learning network model, adaptively learns the state information of the communication channel, avoids the influence of imaginary part interference and effectively improves the accuracy of signal detection.

Description

Signal detection and model training method, device, equipment and storage medium

Technical Field

The present invention relates to the field of communications technologies, and in particular, to a method, an apparatus, a device, and a storage medium for signal detection and model training.

Background

With the rapid development of communication technology, the communication technology has currently entered the 5G era, and in the 5G era, compared with the OFDM (Orthogonal Frequency Division Multiplexing) technology of LTE (Long Term Evolution), an FBMC (Filter Bank Multi-Carrier) system does not need a cyclic prefix, has low out-of-band leakage, and has high spectrum efficiency, and thus has superior stability, time-Frequency focusing performance, and higher spectrum utilization rate. However, since the FBMC system only satisfies strict orthogonality in the real number domain, there is inherent imaginary interference, and the channel estimation method in the OFDM system cannot be directly used in the FBMC system to detect signals.

For this problem, in the prior art, channel estimation and equalization are usually achieved by using a least square channel estimation, an MMSE equalization algorithm, and a noise estimation method required by MMSE equalization based on a given pilot structure, but the accuracy of a detection result of a transmission signal by the existing channel estimation and equalization method is low.

Disclosure of Invention

The embodiment of the invention provides a signal detection and model training method, a signal detection and model training device and a signal detection and model training storage medium, and aims to solve the problem of low accuracy of a transmission signal detection result in the prior art.

In a first aspect, an embodiment of the present invention provides a signal detection method, including:

demodulating the received signal to obtain corresponding demodulation data;

extracting a real part and an imaginary part from the demodulated data to form real data corresponding to the demodulated data;

inputting the real data into a target deep learning network model obtained by pre-training to obtain an output result;

and obtaining the detection result of the original sending signal according to the output result.

In a second aspect, an embodiment of the present invention provides a method for training a model for signal detection, including:

acquiring a training data set, wherein the training data set comprises training input data and training supervision data;

and training the pre-established deep learning network by adopting the training data set, and finishing training when the loss of the training output result relative to the training supervision data reaches a second preset condition to obtain a trained target deep learning network model.

In a third aspect, an embodiment of the present invention provides a signal detection apparatus, including:

the demodulation module is used for demodulating the received signals to obtain corresponding demodulation data;

the extraction module is used for extracting a real part and an imaginary part from the demodulated data to form real data corresponding to the demodulated data;

the detection module is used for inputting the real data into a target deep learning network model obtained by pre-training to obtain an output result;

and the processing module is used for obtaining the detection result of the original sending signal according to the output result.

In a fourth aspect, an embodiment of the present invention provides an electronic device, including: a memory, a transceiver, and at least one processor;

the processor, the memory and the transceiver are interconnected through a circuit;

the memory stores computer-executable instructions; the transceiver is used for receiving signals sent by a sending end;

the at least one processor executing the computer-executable instructions stored by the memory causes the at least one processor to perform the methods described in the first aspect and various possible designs of the first aspect above, and to perform the methods described in the second aspect and various possible designs of the second aspect above.

In a fifth aspect, embodiments of the present invention provide a computer-readable storage medium, in which computer-executable instructions are stored, and when a processor executes the computer-executable instructions, the method according to the first aspect and various possible designs of the first aspect is implemented, and the method according to the second aspect and various possible designs of the second aspect is implemented.

According to the signal detection and model training method, device, equipment and storage medium provided by the embodiment of the invention, the deep learning network model is adopted to carry out channel estimation and equalization, the communication channel state information is learned in a self-adaptive manner, the influence of imaginary part interference is avoided, and the accuracy of signal detection is effectively improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and those skilled in the art can also obtain other drawings according to the drawings without creative efforts.

Fig. 1 is a schematic flow chart of a signal detection method according to an embodiment of the present invention;

fig. 2 is a schematic block diagram of the overall structure of a communication system according to an embodiment of the present invention;

fig. 3 is a schematic structural diagram of a communication system with a channel classification function according to an embodiment of the present invention;

fig. 4 is a schematic structural diagram of a communication system in a training process according to an embodiment of the present invention;

fig. 5 is a schematic structural diagram of a communication system in a testing or practical application process according to an embodiment of the present invention;

FIG. 6 is a diagram illustrating the comparison between the LSTM network model of short packets and DNN and other conventional methods under the same channel conditions according to an embodiment of the present invention;

FIG. 7 is a diagram illustrating the comparison of the LSTM network model with long packets and codes with DNN and conventional methods under the same channel conditions according to an embodiment of the present invention;

FIG. 8 is a flowchart illustrating a method for training a signal detection model according to an embodiment of the present invention;

fig. 9 is a schematic structural diagram of a signal detection apparatus according to an embodiment of the present invention;

fig. 10 is a schematic diagram illustrating an exemplary structure of a signal detection apparatus according to an embodiment of the present invention;

FIG. 11 is a schematic structural diagram of a model training apparatus for signal detection according to an embodiment of the present invention;

fig. 12 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

With the above figures, certain embodiments of the invention have been illustrated and described in more detail below. The drawings and the description are not intended to limit the scope of the inventive concept in any way, but rather to illustrate it by those skilled in the art with reference to specific embodiments.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

The terms to which the present invention relates will be explained first:

LSTM: long Short-Term Memory Network is a time-cycle Neural Network, a special RNN (Recurrent Neural Network), and compared with a common RNN, LSTM can perform better in a longer sequence.

Furthermore, the terms "first", "second", etc. are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. In the description of the following examples, "plurality" means two or more unless specifically limited otherwise.

The following several specific embodiments may be combined with each other, and details of the same or similar concepts or processes may not be repeated in some embodiments. Embodiments of the present invention will be described below with reference to the accompanying drawings.

An embodiment of the present invention provides a signal detection method, which is used for detecting an original transmission signal in an FBMC system. The execution subject of the embodiment is the device signal detection device, and the device may be disposed in an electronic device, which may be a receiver or other computer equipment that can be implemented.

As shown in fig. 1, a schematic flow chart of a signal detection method provided in this embodiment is shown, where the method includes:

step 101, demodulating the received signal to obtain corresponding demodulated data.

Specifically, the received signal is a signal which is transmitted by a transmitting end of the FBMC system through a communication channel and then reaches a receiving end, and in the transmission process of the communication channel, the signal to be transmitted is processed by a multipath channel, a doppler effect, white gaussian noise and the like, so as to finally obtain the received signal of the receiving end; the receiving end demodulates the received signal to obtain the demodulated data of the frequency domain, the demodulated data includes the frequency domain data at each time-frequency grid point, in practical application, the demodulated data can be expressed as a complex matrix, if N represents the number of FBMC symbols, M represents the number of subcarriers, and the demodulated data is an M × N complex matrix.

The demodulation operation for the received signal includes processing of serial-to-parallel conversion, inverse filtering, fast fourier transform, and removal of the initial phase of the received signal.

Step 102, extracting a real part and an imaginary part from the demodulated data to form real data corresponding to the demodulated data.

Specifically, after obtaining demodulated data corresponding to a received signal, real parts and imaginary parts of complex elements in the demodulated data need to be extracted, and the extracted real parts and imaginary parts form corresponding real data, the real data meets the input requirement of a deep learning network model, the deep learning network model is used for further recovering an original transmitted signal, the original transmitted signal refers to an original information bit stream to be transmitted by a transmitting end, the original information bit stream needs to be modulated, the imaginary real parts are taken, pilot frequency is inserted, an initial phase is added, inverse fast fourier transform, filtering, parallel-serial conversion and the like at the transmitting end to obtain a corresponding signal to be transmitted, and the signal to be transmitted is transmitted through a communication channel to obtain a received signal.

Illustratively, the real part and the imaginary part can be arranged in a row to form a vector to input the deep learning network model.

And 103, inputting real number data into a target deep learning network model obtained through pre-training to obtain an output result.

Specifically, the target deep learning network model is a model obtained based on a large amount of training data training and is used for channel estimation and equalization to accurately recover an original transmission signal, the input of the target deep learning network model is real data formed by a real part and an imaginary part of demodulated data, and the output result is parallel data of the recovered original transmission signal.

The target deep learning network model is obtained through pre-training, specifically, a training data set can be obtained, a pre-established deep learning network is trained by the training data set, when the loss of a training output result relative to training supervision data reaches a second preset condition, the training is finished, and the trained target deep learning network model is obtained, wherein the training data set comprises training input data and training supervision data.

In practical application, a training information bit stream generated randomly can be sent to a receiving end through a communication channel after being processed to a certain degree through the sending end, the receiving end obtains training input data by demodulating and extracting a real part and an imaginary part, and the training information bit stream is used as training supervision data; if the transmitting end encodes the training information bit stream, the encoding result can be used as training supervision data at the receiving end; the second preset condition may be set according to actual requirements. Specifically, for example, the training information bit stream may be randomly generated, the pilot is a determined information bit, the training information bit stream is modulated to obtain the second modulation result, and similarly, the pilot needs to be modulated to obtain the first modulation result; the modulation of the training information bit stream and the pilot frequency is not in sequence, and the specific modulation mode can be set according to actual requirements, such as QAM modulation; after modulation, real data to be transmitted can be determined according to a first modulation result of the pilot frequency and a second modulation result of the training information bit stream, both the first modulation result and the second modulation result are complex symbols, and the real data symbols corresponding to the pilot frequency and the real data symbols of the training information bit stream are converted into corresponding real data symbols by extracting a real part and an imaginary part, so that the real data symbols corresponding to the pilot frequency and the real data symbols of the training information bit stream form real data to be transmitted; that is, the real data to be transmitted includes a pilot real part, a pilot imaginary part and training information bit stream real data, the pilot real part and the pilot imaginary part are real parts and imaginary parts of a first modulation result corresponding to the pilot, and the training information bit stream real data is real parts and imaginary parts of a second modulation result; after obtaining real data to be transmitted, adding an initial phase to the real data to be transmitted, aiming at staggering a phase between a real part and an imaginary part of the same complex number and keeping orthogonality, performing inverse fast Fourier transform after adding the initial phase to obtain a second transform result, further filtering the second transform result, specifically, filtering by adopting a prototype filter to obtain a filtering result, performing parallel-serial conversion on the filtering result to obtain a training signal to be transmitted, inputting the training signal to be transmitted to a communication channel for transmission and then reaching a receiving end, obtaining a training received signal by the receiving end, demodulating the training received signal by the receiving end to obtain corresponding training demodulation data, extracting the real part and the imaginary part from the training demodulation data to form training real data as training input data to be input to a deep learning network (such as an LSTM network), and using the coding result of a training information bit stream as training supervision data for supervising the training, and obtaining the trained target deep learning network model after the training is finished.

And 104, obtaining the detection result of the original sending signal according to the output result.

Specifically, after the output result of the target deep learning network model is obtained, the detection result of the original transmission signal may be obtained based on the output result, for example, the output result is subjected to parallel-to-serial conversion to obtain the original transmission signal.

Optionally, before the original information bit stream is modulated by the sending end, the original information bit stream may be further encoded, and then subsequent modulation and other processing are performed, accordingly, at the receiving end, if the output result of the target deep learning network model is an undecoded original sending signal, the output result needs to be decoded to obtain a decoding result, and since the output result is parallel data, in order to obtain a result consistent with the original sending signal, parallel-to-serial conversion needs to be performed to obtain a detection result of the original sending signal, that is, the recovered original sending signal. The decoding mode is adapted to the encoding mode in the processing process of the sending end, and may be specifically set according to actual requirements, for example, a convolutional code mode is adopted during encoding, and a Viterbi decoding algorithm is adopted during decoding.

It can be understood that, for a communication system including an encoding processing process, in a training process of a target deep learning network model, a transmitting end also needs to encode a training information bit stream and then perform subsequent related processing, and a receiving end needs to perform corresponding decoding operation, which is not described in detail.

The signal detection method provided by the embodiment adaptively learns the communication channel state information by adopting the deep learning network model for channel estimation and equalization, avoids the influence of imaginary part interference, and effectively improves the accuracy of signal detection.

In order to make the technical solution of the present invention clearer, the method provided by the above embodiment is further described in an additional embodiment of the present invention.

As a practical way, in order to further improve the accuracy of signal detection, on the basis of the above embodiment, optionally, inputting real number data into a target deep learning network model obtained by training in advance, and obtaining an output result includes: and inputting the real data into the target LSTM network model to obtain an output result.

Specifically, since all bits (bits) of the information bit stream of the communication are correlated and memorized, in order to better utilize the correlation between the bits, the embodiment adopts the trained target LSTM network model as a deep learning network model for channel estimation and equalization, which is used for signal detection, thereby improving the accuracy of the FBMC system and further improving the accuracy of the signal detection.

The target LSTM network model is obtained based on a large amount of training data, and the LSTM network can learn long-term dependence information, so that the correlation among all bits in the information bit stream of communication can be well learned, and the signal detection accuracy is effectively improved.

Further, in an embodiment, in order to better perform channel estimation and equalization, demodulating the received signal to obtain corresponding demodulated data may specifically include:

carrying out serial-to-parallel conversion on the received signals to obtain corresponding parallel signals; inverse filtering is carried out on the parallel signals by adopting an inverse prototype filter to obtain an inverse filtering result; performing fast Fourier transform on the inverse filtering result to obtain a first transform result; and deleting the initial phase of the first conversion result to obtain demodulation data.

Specifically, the received signal is transmitted in serial, and needs to be converted in serial-to-parallel to obtain a parallel signal, the parallel signal is subjected to inverse filtering by using an inverse prototype filter to obtain an inverse filtering result, Fast Fourier Transform (FFT) is further performed on the inverse filtering result to obtain a first transformation result, and an initial phase is deleted from the first transformation result to obtain demodulated data; the initial phase is deleted because the initial phase is added before the sending end performs inverse fast fourier transform, specifically, at the sending end, an original information bit stream to be transmitted is modulated (such as QAM), an imaginary real part is taken, a pilot frequency is inserted, the initial phase is added, Inverse Fast Fourier Transform (IFFT), filtering, parallel-serial conversion and the like to obtain a corresponding signal to be transmitted, in practical application, the original information bit stream can be further encoded before modulation, the signal to be transmitted reaches a receiving end through communication channel transmission, and the receiving end receives a received signal; specifically, the real part and the imaginary part of the complex number may be divided into two pure real numbers (which may be referred to as a real part symbol and an imaginary part symbol, respectively) through an OQAM (Offset Quadrature Amplitude Modulation) Modulation preprocessing, so as to separately process the real part and the imaginary part of the complex signal, and set a time interval as a symbol period T/2, and the FBMC system uses an OQAM Modulation technique to ensure orthogonality between carriers, where a specific Modulation principle is the prior art and is not described herein again.

Illustratively, as shown in fig. 2, the functional block diagram of the overall structure of the communication system provided for this embodiment is shown, where Real () and Imag () represent taking Real part and taking imaginary part, j ^ (n +2m) and j ^ (n +2m +1) represent adding initial phase, j ^ (n +2m) and j ^ (n +2m +1) represent deleting initial phase, and routing represents the Round function process, i.e. the result of Rounding according to the specified decimal place number, converting each decimal value of the target LSTM network model output result into 0 or 1; it should be noted that, in this example, the target LSTM network model outputs a decimal between a group (0,1), which indicates the probability of whether each bit of transmission is 1 or 0, and obtains a recovered un-decoded original transmission signal after Rounding, and then obtains the original transmission signal after Decoding (Decoding) and parallel-to-serial conversion.

It can be understood that in practical applications, the sounding can also be used as a part of the target LSTM network model, and the output result of the target LSTM network model is an un-decoded parallel original transmission signal, and the original transmission signal can be obtained through decoding and parallel-serial conversion.

In another embodiment, to further improve the accuracy of the signal detection result, obtaining the detection result of the original transmission signal according to the output result includes: and decoding and parallel-serial conversion are carried out on the output result by adopting a Viterbi decoding algorithm to obtain the detection result of the original transmission signal.

Specifically, before the original information bit stream is modulated by the transmitting end, the original information bit stream is encoded to obtain an encoding result, and then subsequent modulation, imaginary real part taking, pilot frequency inserting, initial phase adding, inverse fast fourier transform, filtering, parallel-to-serial conversion and other processing are performed. The decoding mode is adapted to the encoding mode in the processing process of the sending end, the convolutional code mode is adopted during encoding, and the Viterbi decoding algorithm is adopted during decoding.

In the encoding process of the convolutional code, the input information bits are encoded in groups, the encoding output bits of each code group are not only related to the information bits of the group, but also related to the information bits of other groups at the previous moment, similarly, in the decoding process of the convolutional code, decoding information is obtained from the group received at the current moment, and related information is extracted from the related groups, and as the correlation of each group is fully utilized in the encoding process of the convolutional code, the convolutional code has quite good performance gain; in a communication system, when a signal is transmitted in a channel, due to the fact that the self condition of the channel is not ideal, the signal is affected by a plurality of noise interferences, and therefore an error code is generated, the embodiment of the invention carries out convolutional coding on the transmitted information bit stream to generate a convolutional code, and the convolutional code is transmitted after being modulated and the like, so that the error rate can be effectively reduced, and the accuracy of a signal detection result is improved.

In practical application, if a vehicle-mounted communication system enters a tunnel from an open environment, scenes of communication channels are completely different, so that the communication system can perform scene switching in as short a time as possible.

As another implementable manner, to improve the universality of the communication system, the method may further include:

acquiring characteristic data of a target channel of a current transmission signal; inputting the characteristic data of the target channel into the trained channel classification model to obtain the scene type of the target channel; acquiring corresponding target LSTM parameters according to the scene type of the target channel; the network model based on the target LSTM parameters is taken as the target LSTM network model.

Specifically, for channels of different scenes, different LSTM parameters can be obtained by training and stored, so that the original transmission signal can be accurately detected for transmission of different channels. In practical application, the feature data of the target channel of the current transmission signal can be acquired, the feature data of the target channel is the channel parameter of the target channel, the scene type of the target channel is identified based on the trained channel classification model, so that the LSTM parameter (referred to as target LSTM parameter) of the corresponding scene type can be acquired and transmitted to the corresponding network to form an LSTM network model as the target LSTM network model, and the original transmission signal is further acquired based on the output result of the target LSTM network model.

The channel classification model can be set according to actual requirements.

Further, in order to improve the classification accuracy, the embodiment of the invention adopts a deep neural network model as a channel classification model.

The deep neural network model comprises an input layer, a hidden layer and an output layer, the number of layers of the hidden layer can be set according to actual requirements, the layers are all connected, the number of neurons in each layer is set according to the actual requirements, the number of neurons in the output layer is the number of scene types, for example, 14 scene types exist, and the number of neurons in the output layer is 14.

Exemplary communication coverage propagation scenarios are: indoor wireless communication (such as indoor office, residential), indoor-to-outdoor wireless communication, urban micro-cells, bad urban macro-cells, suburban macro-cells, outdoor-to-indoor wireless communication, fixed feeders, urban macro-cells, rural mobile networks, etc., the number Of maximum delays and/or multipaths for different scene types is different, as shown in table 1 below, where los (Line Of signal) and nlos (not Line Of signal) refer to Line-Of-sight transmission and non-Line-Of-sight transmission Of wireless signals, respectively; these scenes can therefore be categorized into a certain number of scene types.

TABLE 1

Illustratively, the scenes are divided into 14 scene types, and the channel classification model includes an input layer, a first hidden layer, a second hidden layer and an output layer, and the number of neurons in each layer is, for example, 32, 128, 64 and 14, respectively.

Illustratively, as shown in fig. 3, a schematic structural diagram of a communication system with a channel classification function provided for this embodiment includes a channel classification module (i.e., a gray portion in the figure), where the Channal classification module is a deep neural network model and is used for identifying a scene type of a channel, LSTM1-LSTM14 represents LSTM parameters corresponding to 14 scene types, the LSTM parameters are obtained by pre-training, and after the scene type of the channel is identified, the corresponding LSTM parameters are obtained from the 14 LSTM parameters and transmitted to an LSTM network, so as to form a target LSTM network model for estimation and equalization processing of a currently transmitted signal, so as to recover an original transmitted signal.

In one embodiment, to ensure the accuracy of the classification result of the channel classification model, the method further includes:

acquiring channel training characteristic data and corresponding label data; and training the pre-established deep neural network based on the channel training characteristic data and the corresponding label data, and ending the training if the loss reaches a first preset condition to obtain a deep neural network model.

Specifically, in order to enable the channel classification model to accurately classify the channel, the channel classification model needs to be obtained through training of a large amount of training data in advance, specifically, channel training feature data of each scene type and corresponding label data can be obtained, the channel training feature data comprise channel parameter information of various scene types, the label data are actual scene types of the channel, the channel training feature data are input into a pre-established deep neural network according to a preset rule, loss is calculated by adopting a preset loss function based on a prediction result output by the network and the corresponding label data, and when the loss reaches a first preset condition, the training can be ended to obtain the trained channel classification model; the preset loss function may adopt any implementable loss function according to an actual requirement, such as a cross entropy loss function, and may be specifically set according to the actual requirement, which is not limited in this embodiment.

Illustratively, the total sample data includes 400 txt files, 300 of which can be used as training data and 100 as test data; the channel training characteristic data comprises 300 txt files, each txt file comprises a matrix with the size of 10000 × 32, namely 10000 rows and 32 numbers in each row, during training, the 32 numbers in each row in each file can be used as one-time input, each row is sequentially input into the deep neural network, for the 32 numbers input each time, the deep neural network calculates an activation value layer by layer from the first layer, a prediction result is finally output, then the prediction result can be subjected to Boolean function calculation, different vectors are obtained by combining different forms of 0 and 1, each vector represents one classification, the prediction result is compared with the label data, loss is calculated, if the loss does not meet a first preset condition, training is continued, if the loss meets the first preset condition, the training is ended, and a trained channel classification model is obtained; the first preset condition may be set according to actual requirements.

As another practical way, in order to ensure the channel estimation and equalization effect of the deep learning network model, training optimization needs to be performed in advance, that is, the method further includes:

acquiring a training data set, wherein the training data set comprises training input data and training supervision data; and training the pre-established deep learning network by adopting a training data set, and finishing training when the loss of a training output result relative to the training supervision data reaches a second preset condition to obtain a trained target deep learning network model.

Specifically, a training information bit stream generated randomly can be sent to a receiving end through a communication channel after being processed to a certain degree by the sending end, the receiving end obtains training input data by demodulating and extracting a real part and an imaginary part, and the training information bit stream is used as training supervision data; if the transmitting end encodes the training information bit stream, the encoding result can be used as training supervision data at the receiving end; the second preset condition may be set according to actual requirements.

Optionally, since the transmission content is unknown in a real communication system, the training process may also determine whether the training process is finished by the accuracy of the recovered pilot, that is, the training supervision data is the pilot data.

Exemplarily, as shown in fig. 4, a schematic diagram of a communication system structure of a training process provided for this embodiment is provided, and as shown in fig. 5, a schematic diagram of a communication system structure of a testing or practical application process provided for this embodiment is provided; wherein, the deep learning network includes 4 layers of LSTM blocks (LSTM blocks) and one layer of full connection layer (FC), the number of LSTM units (LSTM cells) included in the LSTM blocks can be set according to actual requirements, in this example, the number of LSTM units included in each LSTM Block is 384/700/256/64, the number of neurons included in the full connection layer is 16, and a loss calculation module (calcute Cost) extracts recovered pilot frequency partial data according to the output of the deep learning network, where the output of the full connection layer is a group of decimal between (0 and 1), the loss calculation module can convert each decimal into 0 or 1 according to a certain rule, such as converting through a sigmoid function, and then comparing the conversion result with the corresponding pilot frequency label value to determine whether the conversion result is equal to the pilot frequency label value, if equal, indicating that the pilot frequency of the digit is recovered correctly, otherwise, recovering errors, and sending the average error rate of pilot frequency recovery calculated according to the average error rate as loss cost to a Training network module (Training Net), wherein the Training Net receives the cost calculated by the loss calculation module, judges whether the cost value meets a second preset condition, and if so, takes the current network parameters obtained from the deep learning network as well-trained LSTM parameters, thereby forming and storing a well-trained LSTM network model; if the cost does not meet the second preset condition, the fact that further Training is needed is indicated, then the Training Net updates the current network parameters according to preset rules and transmits the new network parameters to the deep learning network, optimization Training is continued, and the like is performed until the cost value meets the second preset condition.

In practical applications, the specific number of layers of the LSTM network model may be set according to practical requirements, and is not limited to the above-mentioned 4 layers of LSTM blocks and one layer of fully-connected layer, and may also be 1 layer, 2 layers or 3 layers of LSTM blocks.

The deep learning network model is obtained by training according to the training method and is stored correspondingly to each scene type, and then the corresponding deep learning network model can be selected as the target deep learning network model for channel estimation and equalization according to the scene type identified based on the channel classification model, so that the universality of the communication system is effectively improved.

Further, to obtain a training data set of the deep learning network model, the method may further include:

acquiring a training information bit stream and a corresponding pilot frequency; coding the training information bit stream to obtain a coding result; modulating the pilot frequency to obtain a first modulation result, and modulating a coding result to obtain a second modulation result; determining real data to be sent according to the first modulation result and the second modulation result, wherein the real data to be sent comprises a pilot real part, a pilot imaginary part and training information bit stream real data; adding an initial phase to real data to be sent, and performing inverse fast Fourier transform to obtain a second transform result; filtering the second transformation result to obtain a filtering result; performing parallel-serial conversion on the filtering result to obtain a training signal to be sent; transmitting a training signal to be transmitted through a communication channel to obtain a training received signal; demodulating the training received signal to obtain corresponding training demodulation data; extracting a real part and an imaginary part of the training demodulation data to form training real data corresponding to the training demodulation data; and taking real training data as training input data and taking an encoding result as training supervision data.

Specifically, the training information bit stream may be randomly generated, the pilot frequency is a determined information bit, and in order to reduce the error rate, the training information bit stream is encoded to obtain an encoding result, for example, the training information bit stream is encoded by using a convolutional coding method to obtain an encoding result, where the encoding result is a convolutional code; after coding, the coding result can be modulated to obtain a second modulation result, and similarly, the pilot frequency also needs to be modulated to obtain a first modulation result; the coding result and the pilot frequency are modulated in no sequence, and the specific modulation mode can be set according to actual requirements, for example, QAM modulation is adopted, and specifically, 4QAM modulation, 16QAM modulation, 64QAM modulation, 256QAM modulation and the like can be adopted; after modulation, real data to be transmitted can be determined according to a first modulation result of the pilot frequency and a second modulation result of the coding result, both the first modulation result and the second modulation result are complex symbols, and the real data symbols corresponding to the pilot frequency and the real data symbols of the training information bit stream need to be converted into corresponding real data symbols by extracting a real part and an imaginary part, so that the real data symbols corresponding to the pilot frequency and the real data symbols of the training information bit stream form real data to be transmitted; that is, the real data to be transmitted includes a pilot real part, a pilot imaginary part, and training information bit stream real data, where the pilot real part and the pilot imaginary part are real parts and imaginary parts of a first modulation result corresponding to the pilot, and the training information bit stream real data is real parts and imaginary parts of a second modulation result.

Illustratively, the formed real data to be transmitted is an mxn data block on a frequency-domain time coordinate, the first two columns are a pilot real part and a pilot imaginary part, respectively, and the last column is training information bit stream real data.

After real data to be sent is obtained, adding an initial phase to the real data to be sent in order to stagger a phase between a real part and an imaginary part of the same complex number and keep the real part and the imaginary part orthogonal, performing inverse fast Fourier transform after adding the initial phase to obtain a second transformation result, and further filtering the second transformation result, specifically, filtering with prototype filter to obtain filtering result, the filtering result is converted in parallel and serial to obtain a training signal to be sent, the training signal to be sent is input into a communication channel to be transmitted and then reaches a receiving end, the receiving end obtains a training received signal, the receiving end demodulates the training received signal to obtain corresponding training demodulation data, extracting a real part and an imaginary part of training demodulation data to form training real data serving as training input data to be input into a deep learning network (such as an LSTM network), and using a coding result of a training information bit stream as training supervision data to supervise the end of training; the specific processing of the training received signal at the receiving end is consistent with the processing procedure of the received signal, and details are not repeated.

In practical application, the pilot frequency may also be encoded and then subjected to subsequent processing, which may be specifically set according to actual requirements, and this embodiment is not limited.

For example, the learning rate of the LSTM network may be set to 0.001, the optimization function may adopt Adam optimization algorithm, and may also adopt rmsprop (root Mean Square prop) optimization algorithm, and the output layer may adopt sigmoid function as the activation function.

In practical applications, an FBMC baseband equivalent transmission signal (i.e., a signal to be transmitted at a transmitting end) can be represented as s (t):

wherein N represents the number of FBMC symbols, M represents the number of subcarriers, a_m,nIs the real-valued data symbol (i.e., complex symbol), g, transmitted on the mth subcarrier in the nth FBMC symbol_m,n(t) represents the odd function at time-frequency grid point coordinates (m, n), which can be obtained by the following time-frequency transformation:

wherein v is₀Indicating the spacing, τ, between subcarriers₀Representing the time offset, τ, between the real and imaginary parts of adjacent symbols₀v₀1/2; initial phase

Further, the received signal at the receiving end can be represented as r (t):

wherein h is_m,nIs the channel value at the time-frequency grid point coordinate (m, n),

the convolution operation is represented by w (t), gaussian white noise is represented by w (t), and the signal transmission and reception principle of the FBMC communication system is the prior art and is not described herein again.

In an exemplary embodiment, the beneficial effects of the present invention are explained in conjunction with the results of simulation experiments.

In the example, the simulation condition adopts a new wireless system channel model 'WINNER II channel models' standard, the system bandwidth is 2.5GHz, the number of subcarriers is 64, a Hermite prototype filter is adopted, the overlap coefficient is 4, and an OQAM modulation mode is adopted. In the simulation, a (2,1,3) convolutional code is adopted, the code rate is 1/2, and a polynomial coefficient C is generated₁＝(1,1,1)，C₂Viterbi decoding ═ 1,0, 1; fig. 6 is a schematic diagram showing a comparison result between the LSTM network model of the short packet and the DNN and other conventional methods under the same channel condition provided in this embodiment, and fig. 7 is a schematic diagram showing a comparison result between the LSTM network model of the long packet and the coding provided in this embodiment and the DNN and conventional methods under the same channel condition; in the figure, the abscissa represents the Signal-to-Noise Ratio (SNR) in dB, and the ordinate represents the bit error rate (BitError Ratio); pilot LS (OFDM) is an LS (least squares) channel estimation method under the existing OFDM system, and the calculation method is that pilot frequency of a transmitting end is H_TThe pilot frequency received by the receiving end is H_RThen the estimated impact value of the channel is:

pilot Deep Learning (OFDM) is a channel estimation method adopting a five-layer DNN network structure under the existing OFDM system, and the number of neurons in each layer is: 256,500,250,120,16, five layers in total; pilot MMSE (OFDM) is an MMSE (minimum mean square) channel estimation method under the existing OFDM system; pilot mapping (FBMC) is an LS (least square) channel estimation method under the existing FBMC system, an overlap coefficient O is 4, and a prototype filter is a Hermite filter; pilot DL-CE (FBMC) is a channel estimation method of a five-layer DNN network structure under the existing FBMC system: the number of neurons in each layer is: 384,700,300,100,16,five layers in total, wherein the overlapping coefficient O is 4, and the prototype filter is a Hermite filter; pilot LSTM (FBMC) is a simulation result of the channel estimation method of the short packet LSTM network under the FBMC system of the present invention, each packet length is 100 × 384, and the five-layer LSTM network structure: the number of neurons in each layer is: 384,700,256,100,16, five layers in total, wherein the overlap coefficient O is 4, and the prototype filter is a Hermite filter; pilot LSTM long (FBMC) is a long packet simulation result under the FBMC system of the present invention, each packet is 800 × 384, and a five-layer LSTM network structure: the packet length is 800 × 384, and the number of neurons in each layer is: 384,700,256,100,16, five layers in total; the overlap coefficient O is 4, and the prototype filter is a Hermite filter; the pilot LSTM long decode (FBMC) is the long packet simulation result of the convolutional coding Viterbi decoding under the FBMC system of the invention, each packet is 800X 384, and the five-layer LSTM network structure: the number of neurons in each layer is: 384,700,256,100,16, five layers in total; the overlap coefficient O is 4, and the prototype filter is a Hermite filter;

obviously, as can be seen from fig. 6 and 7, under the same BER, the signal-to-noise ratio of the LSTM is 3.74dB to 5.27dB lower than DNN, the short packet LSTM is 0.67dB to 2.14dB lower than the long packet LSTM, and the long packet coded LSTM is 0.71dB to 2.94dB lower than the long packet LSTM, and in conclusion, the LSTM-based signal detection method of the present invention has a better effect compared with the existing method.

It should be noted that the respective implementable modes in the embodiment may be implemented individually, or may be implemented in combination in any combination without conflict, and the present invention is not limited thereto.

The signal detection method provided by the embodiment performs channel estimation and equalization through the LSTM network model, thereby effectively improving the detection accuracy; the error rate is effectively reduced and the detection accuracy is further improved by encoding the original sending signal; and various channels can be identified through the channel classification model, and a corresponding LSTM network model is provided for each channel to carry out channel estimation and equalization, so that the universality of the communication system is improved.

Another embodiment of the present invention provides a model training method for signal detection, which is used to train and obtain the deep learning network model required by the above embodiments.

As shown in fig. 8, a schematic flow chart of the model training method for signal detection provided in this embodiment is provided, where the method specifically includes:

step 301, a training data set is obtained, wherein the training data set comprises training input data and training supervision data.

And 302, training the pre-established deep learning network by adopting a training data set, and finishing training when the loss of a training output result relative to the training supervision data reaches a second preset condition to obtain a trained target deep learning network model.

The specific operations of the above steps have been described in detail in the foregoing embodiments, and are not described herein again.

Further, the acquiring of the training data set may specifically include:

acquiring a training information bit stream and a corresponding pilot frequency; coding the training information bit stream to obtain a coding result; modulating the pilot frequency to obtain a first modulation result, and modulating a coding result to obtain a second modulation result; determining real data to be sent according to the first modulation result and the second modulation result, wherein the real data to be sent comprises a pilot real part, a pilot imaginary part and training information bit stream real data; adding an initial phase to real data to be sent, and performing inverse fast Fourier transform to obtain a second transform result; filtering the second transformation result to obtain a filtering result; performing parallel-serial conversion on the filtering result to obtain a training signal to be sent; transmitting a training signal to be transmitted through a communication channel to obtain a training received signal; demodulating the training received signal to obtain corresponding training demodulation data; extracting a real part and an imaginary part of the training demodulation data to form training real data corresponding to the training demodulation data; and taking real training data as training input data and taking an encoding result as training supervision data. For specific operations, refer to the above embodiments, which are not described herein again.

In one embodiment, the method may further comprise:

acquiring channel training characteristic data and corresponding label data; and training the pre-established deep neural network based on the channel training characteristic data and the corresponding label data, and if the loss reaches a first preset condition, finishing the training to obtain a trained deep neural network model for detecting the channel scene type.

Yet another embodiment of the present invention provides a signal detection apparatus for performing the method of the above embodiment.

As shown in fig. 9, a schematic structural diagram of the signal detection apparatus provided in this embodiment is shown. The device 30 comprises: demodulation module 31, extraction module 32, detection module 33 and processing module 34.

The demodulation module is used for demodulating the received signals to obtain corresponding demodulation data; the extraction module is used for extracting a real part and an imaginary part from the demodulation data to form real data corresponding to the demodulation data; the detection module is used for inputting real data into a target deep learning network model obtained by pre-training to obtain an output result; and the processing module is used for obtaining the detection result of the original sending signal according to the output result.

Specifically, the demodulation module receives a received signal sent by another module (for example, the receiving module), demodulates the received signal to obtain corresponding demodulation data, and sends the demodulation data to the extraction module, the extraction module extracts a real part and an imaginary part from the demodulation data to form real data corresponding to the demodulation data, and sends the real data to the detection module, the detection module inputs the real data to a target deep learning network model obtained by pre-training, obtains an output result, and sends the output result to the processing module, and the processing module obtains a detection result of an original sent signal according to the output result.

The specific manner in which each module performs the operation has been described in detail in the embodiment of the method, and the same technical effect can be achieved, and will not be described in detail herein.

In order to make the device of the present invention clearer, the device provided by the above embodiment is further described in an additional embodiment of the present invention.

As an implementable manner, in order to further improve the accuracy of the detection result, on the basis of the foregoing embodiment, optionally, the detection module is specifically configured to input real data to the target LSTM network model, and obtain an output result.

Further, the demodulation module is specifically configured to:

In practical applications, the demodulation module may further include a serial-to-parallel conversion sub-module, an inverse filtering sub-module, a fast fourier transform sub-module, and a deletion sub-module.

The serial-parallel conversion sub-module is used for performing serial-parallel conversion on the received signals to obtain corresponding parallel signals; the inverse filtering submodule is used for performing inverse filtering on the parallel signals by adopting an inverse prototype filter to obtain an inverse filtering result; the fast Fourier transform submodule is used for carrying out fast Fourier transform on the inverse filtering result to obtain a first transform result; and the deleting submodule is used for deleting the initial phase of the first conversion result to obtain the demodulation data.

As another implementable manner, in order to further improve the accuracy of the detection result, optionally, the processing module is specifically configured to:

and decoding and parallel-serial conversion are carried out on the output result by adopting a Viterbi decoding algorithm to obtain the detection result of the original transmission signal.

As shown in fig. 10, an exemplary structure diagram of the signal detection apparatus provided in this embodiment is shown.

As another practical way, in order to improve the universality of the communication system, the apparatus further includes a first obtaining module 35, a classifying module 36 and a determining module 37.

The first acquisition module is used for acquiring the characteristic data of a target channel of a current transmission signal; the classification module is used for inputting the characteristic data of the target channel into the trained channel classification model to obtain the scene type of the target channel; and the determining module is used for acquiring corresponding target LSTM parameters according to the scene type of the target channel and taking the network model based on the target LSTM parameters as a target LSTM network model.

Specifically, the first obtaining module may obtain feature data of the target channel from a configuration file of the target channel and send the feature data to the classifying module, the classifying module inputs the feature data of the target channel into the trained channel classification model to obtain a scene type of the target channel, and sends the scene type to the determining module, the determining module obtains a corresponding target LSTM parameter according to the scene type of the target channel, and transmits the target LSTM parameter to the LSTM network to form an LSTM network model as the target LSTM network model.

Further, in order to improve the accuracy of the classification result, the channel classification model is a deep neural network model.

Furthermore, in order to ensure the accuracy of the classification result, the acquisition module is further configured to acquire channel training feature data and corresponding label data; and the classification module is further used for training the pre-established deep neural network based on the channel training characteristic data and the corresponding label data, and if the loss reaches a first preset condition, ending the training to obtain a deep neural network model.

As another practicable way, the apparatus further includes a second acquisition module 38 and a second training module 39.

The second acquisition module is used for acquiring a training data set, and the training data set comprises training input data and training supervision data; and the second training module is used for training the pre-established deep learning network by adopting the training data set, and when the loss of the training output result relative to the training supervision data reaches a second preset condition, finishing the training to obtain a trained target deep learning network model.

In one embodiment, for the transmission of the channel corresponding to the target deep learning network model, the target deep learning network model trained by the second training module may be sent to the detection module, and directly used for the detection of the original transmission signal.

In another embodiment, for a communication system with multiple scene types, the target deep learning network model or the model parameters trained by the second training module may be sent to the determining module, and the deep learning network models corresponding to the various scene types may be trained, and the model parameters may be sent to the determining module, so that the determining module may obtain the deep learning network model corresponding to the scene type as the target deep learning network model for signal detection under different channel scene types.

In another embodiment of the present invention, a model training apparatus for signal detection may be further provided, which is used for training a deep learning network model and/or a channel classification model.

As shown in fig. 11, a schematic structural diagram of a model training apparatus for signal detection provided in this embodiment includes a third obtaining module 41 and a third training module 42.

The third acquisition module is used for acquiring a training data set, wherein the training data set comprises training input data and training supervision data; and the third training module is used for training the pre-established deep learning network by adopting the training data set, and when the loss of the training output result relative to the training supervision data reaches a second preset condition, finishing the training to obtain a trained target deep learning network model.

Specifically, the third obtaining module may obtain a training data set from a preset storage area and send the training data set to the third training module, the third training module trains the pre-established deep learning network by using the training data set, and when a loss of a training output result relative to the training supervision data reaches a second preset condition, the training is ended, and a trained target deep learning network model is obtained.

In an embodiment, the third training module is specifically configured to:

In practical application, the third training module may be further divided into a plurality of sub-modules, the specific division manner may be set according to actual requirements, and this embodiment is not limited.

In one embodiment, the third obtaining module is further configured to obtain channel training feature data and corresponding label data; and the third training module is further used for training the pre-established deep neural network based on the channel training characteristic data and the corresponding label data, and if the loss reaches a first preset condition, ending the training to obtain a trained deep neural network model as a channel classification model for detecting the scene type of the channel.

Still another embodiment of the present invention provides an electronic device, configured to perform the method provided by the foregoing embodiment. The electronic device may be a receiver or other implementable computer device.

As shown in fig. 12, is a schematic structural diagram of the electronic device provided in this embodiment. The electronic device 50 includes: memory 51, transceiver 52, and at least one processor 53.

The processor, the memory and the transceiver are interconnected through a circuit; the memory stores computer-executable instructions; the transceiver is used for receiving the signal sent by the receiving end; the at least one processor executes computer-executable instructions stored by the memory to cause the at least one processor to perform a method as provided by any of the embodiments above.

Specifically, the transceiver receives a signal sent by the receiving end and sends the signal to the processor, and the processor reads and executes the computer execution instruction stored in the memory, so as to implement the method provided in any of the above embodiments.

Optionally, the transceiver may further receive training data, a channel profile, and other data input by a user, and may be specifically set according to actual requirements.

The electronic device provided by the invention can be applied to any scene based on an FBMC communication system, particularly to receiving end channel estimation and equalization.

It should be noted that the electronic device of this embodiment can implement the method provided in any of the above embodiments, and can achieve the same technical effect, which is not described herein again.

Yet another embodiment of the present invention provides a computer-readable storage medium, in which computer-executable instructions are stored, and when the processor executes the computer-executable instructions, the method provided in any one of the above embodiments is implemented.

It should be noted that the computer-readable storage medium of this embodiment can implement the method provided in any of the above embodiments, and can achieve the same technical effects, which are not described herein again.

In the embodiments provided in the present invention, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the units is only one logical division, and other divisions may be realized in practice, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.

Other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention following, in general, the principles of the invention and including such departures from the present disclosure as come within known or customary practice within the art to which the invention pertains. It is intended that the specification and examples be considered as exemplary only, with a true scope and spirit of the invention being indicated by the following claims.

It will be understood that the invention is not limited to the precise arrangements described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the invention is limited only by the appended claims.

Claims

1. A method of signal detection, comprising:

demodulating the received signal to obtain corresponding demodulation data;

2. The method of claim 1, wherein demodulating the received signal to obtain corresponding demodulated data comprises:

performing serial-to-parallel conversion on the received signals to obtain corresponding parallel signals;

inverse filtering is carried out on the parallel signals by adopting an inverse prototype filter, and an inverse filtering result is obtained;

performing fast Fourier transform on the inverse filtering result to obtain a first transform result;

and deleting the initial phase of the first transformation result to obtain the demodulation data.

3. The method of claim 1, wherein obtaining the detection result of the original transmission signal according to the output result comprises:

and decoding and parallel-serial conversion are carried out on the output result by adopting a Viterbi decoding algorithm to obtain a detection result of the original transmission signal.

4. The method of claim 1, wherein the target deep learning network model is a target LSTM network model;

the method further comprises the following steps:

acquiring characteristic data of a target channel of a current transmission signal;

inputting the characteristic data of the target channel into a trained channel classification model to obtain the scene type of the target channel, wherein the channel classification model is a deep neural network model;

acquiring corresponding target LSTM parameters according to the scene type of the target channel;

and taking the network model based on the target LSTM parameters as the target LSTM network model.

5. The method of claim 4, further comprising:

acquiring channel training characteristic data and corresponding label data;

and training a pre-established deep neural network based on the channel training characteristic data and the corresponding label data, and ending the training if the loss reaches a first preset condition to obtain the deep neural network model.

6. A method for model training for signal detection, comprising:

7. A signal detection device, comprising:

8. The apparatus of claim 7, wherein the target deep learning network model is a target LSTM network model;

the device further comprises:

the first acquisition module is used for acquiring the characteristic data of a target channel of a current transmission signal;

the classification module is used for inputting the characteristic data of the target channel into a trained channel classification model to obtain the scene type of the target channel, and the channel classification model is a deep neural network model;

and the determining module is used for acquiring corresponding target LSTM parameters according to the scene type of the target channel and taking a network model based on the target LSTM parameters as the target LSTM network model.

9. An electronic device, comprising: a memory, a transceiver, and at least one processor;

the at least one processor executing the computer-executable instructions stored by the memory causes the at least one processor to perform the method of any one of claims 1-6.

10. A computer-readable storage medium having computer-executable instructions stored thereon which, when executed by a processor, implement the method of any one of claims 1-6.