CN112968749B - WCDMA cell searching method based on machine learning - Google Patents

Abstract

The invention provides a WCDMA cell search method based on machine learning, which is characterized in that during a training stage, a WCDMA downlink signal without an auxiliary synchronous code is generated to manufacture a training data set, so that the dependence relationship between a primary scrambling code and the auxiliary synchronous code is eliminated; in the testing stage, on the basis of the signal after time slot synchronization, corresponding to-be-identified prediction data is manufactured according to 15 positions possibly appearing in the frame header to carry out primary scrambling code identification, a primary scrambling code classification result and corresponding probability of the group of 15 to-be-identified prediction data are obtained, and finally, the frame header position and the primary scrambling code used by the signal are obtained by searching the maximum value of the 15 prediction probabilities. The invention can achieve the rapid frame synchronization and the classification and identification of 512 scramblers in a machine learning mode without knowing the secondary synchronization codes and the pattern table corresponding to the primary scramblers, thereby more flexibly finishing the cell search.

Description

WCDMA cell searching method based on machine learning

Technical Field

The invention relates to a WCDMA communication technology, in particular to a WCDMA signal scrambling code identification technology.

Background

A WCDMA frame is 38400 chip chips divided into 15 slots, each slot containing 2560 chips. The position of a certain time slot can be determined in the primary synchronization process, and then the position of the head of a radio frame is determined. The secondary synchronization process may determine the position of the frame header, i.e., timeslot number 0, and the scrambling code group number. The downlink primary scrambling codes are 512, and each group of 8 is divided into 64 groups. Each cell is allocated only one primary scrambling code.

The structural design of the secondary synchronization channel S-SCH channel is shown in FIG. 1. The structure of the S-SCH channel is very similar to that of the P-SCH channel, which is a primary synchronization channel, and the first 256chips of a slot are used to store secondary synchronization codes (SSC codes). The secondary synchronization codes in the S-SCH channel are 16 in total, and 15 slots in one physical frame data are regularly distributed according to the protocol, as shown in fig. 2.

In the process of signal transmission, modulated secondary synchronization codes occupying 15 time slots are repeatedly transmitted in each wireless frame, 16 secondary synchronization codes are specified in a protocol, the length of the secondary synchronization codes is 256chips, 64 distribution modes are extracted through permutation and combination, each distribution mode can be called as a pattern, 64 patterns are provided, the following table shows that the scrambling code group number is 64 groups, and each group uses 8 secondary synchronization codes:

TABLE 1 corresponding table of WCDMA secondary synchronization code sequence and scrambling code group number

Secondary synchronization code (SSC code) allocation: the cyclic shifts of the 64 spare SCH code sequences are unique, that is, any non-0 less than 15 cyclic shift of the 64 code sequences is not equal to the cyclic shifts of the other 64 sequences, and any non-0 less than 15 cyclic shift is not equal to its own other less than 15 cyclic shift.

The secondary synchronization code pattern and the scrambling code group number are in one-to-one correspondence, so the frame synchronization process is to determine the starting position of the frame header and the scrambling code group by detecting the sequence order of the secondary synchronization code in the frame.

The secondary synchronization code is generated as follows:

two vectors b and z are defined first:

b＝(1,1,1,1,1,1,-1,-1,-1,1,-1,1,-1,1,1,-1)

z＝(b,b,b,-b,b,b,-b,-b,b,-b,b,-b,-b,-b,-b,-b)

and then generating a Hadamard matrix of 256 orders:

H₀＝(1)

p represents the matrix order, 256-order Hardman matrix is numbered in rows from the top, and the nth row sequence is h_nK is 0,1,2,.., 255, and the sequence h is repeated_nThe ith symbol of the sum sequence z is denoted h_n(i) And z (i), i ═ 0,1,2,.., 255, the mth SSC code C_ssc,k,k＝0,1,2,...,15：

C_ssc,k＝(1+j)*(h_n(0)*z(0),h_n(1)*z(1),...,h_n(255)*z(255))

Wherein m 16 k, k 0,1_nI.e. row m of the Hadamard matrix. The above is the generation method of 16 secondary synchronization codes.

The WCDMA communication technology has been developed, and a series of standardized protocols are established for the WCDMA technology by the 3GPP organization. When receiving WCDMA downlink signals, the terminal firstly carries out cell search, and the cell search is divided into three steps:

the first step is time slot synchronization, and data after time slot synchronization is obtained through main synchronization code correlation; first, the slot synchronization can be performed by a primary synchronization code (PSC code), and the start position of the slot is found.

The second step is frame synchronization, finding patterns corresponding to continuous 15 time slot data through 16 secondary synchronization codes (SSC codes) specified by a protocol, and obtaining the initial position and the scrambling code group number of a frame header in a table look-up mode; determining the scrambling code group number, namely determining the scrambling code group where the primary scrambling code is located;

and thirdly, finding out the correct primary scrambling code from the 8 primary scrambling codes in one scrambling code group through the scrambling code group number, and finishing the cell search.

The commercialized WCDMA technology protocol has strict rules on which chips are used in each link and in what manner to obtain information, and the information of the signal can be extracted by processing according to the method specified by the protocol. The process of cell search is relatively complex, and the secondary synchronization code patterns specified by the protocol are required to be corresponding to the primary scrambling codes for processing, and if the pattern table is changed or the secondary synchronization codes are changed, the subsequent processing cannot be performed.

Disclosure of Invention

The technical problem to be solved by the invention is to provide a method for identifying and classifying 512 kinds of established scrambling codes in a machine learning mode and quickly completing cell search under the condition of unknown pattern table, in order to solve the problem of excessively depending on the scrambling code group pattern table to carry out frame synchronization and scrambling code identification in a WCDMA communication system.

The technical scheme adopted by the invention for solving the technical problems is that the WCDMA cell searching method based on machine learning comprises the following steps:

1) training:

1-1) generating training data: generating WCDMA downlink frame synchronization signals without adding an auxiliary synchronization channel, after IQ separation of the frame synchronization signals is completed, printing corresponding scrambling code numbers on IQ two paths of signals as labels to complete the manufacturing scrambling code numbers of a training set, wherein the manufacturing scrambling code numbers are 512 and respectively correspond to 512 types of primary scrambling codes;

1-2) inputting training data into the set up machine learning network model for training, and storing the trained network model; the output layer of the machine learning network model has 512 classification outputs;

2) the testing steps are as follows:

2-1) generating data to be predicted: in a WCDMA downlink, carrying out one-frame length interception on a time slot data length by once sliding a time slot synchronized WCDMA signal, and making a group of 15 data to be predicted by 15 times of sliding; 15 data to be predicted correspond to 15 frame header positions respectively;

2-2) inputting a group of 15 data to be predicted into the trained network model to obtain the recognition results of the 15 data to be predicted, and obtaining the correct frame head position and the main scrambling code by searching the maximum probability value in the recognition results to complete the cell search.

In the training stage, the invention generates WCDMA downlink signals without secondary synchronization codes to manufacture a training data set, so as to get rid of the dependence relationship between primary scrambling codes and secondary synchronization codes; in the testing stage, on the basis of a signal after time slot synchronization, corresponding to-be-identified prediction data is manufactured according to 15 positions possibly appearing in a frame header to carry out primary scrambling code identification, a primary scrambling code classification result and corresponding probability of the group of 15 to-be-identified prediction data are obtained, and finally, the frame header position and a primary scrambling code used by the signal are obtained by searching 15 maximum prediction probabilities; the cyclic neural network RNN is utilized to extract the characteristics of the primary scrambling codes in the WCDMA signals for the training data or the prediction data to be identified, so that the classification training or classification identification of the primary scrambling codes is carried out, and the purpose of completing cell search is achieved.

The invention has the advantages that the quick frame synchronization and the classification and identification of 512 scrambling codes can be realized in a machine learning mode without knowing the secondary synchronization codes and the pattern table corresponding to the primary scrambling codes, thereby more flexibly finishing the cell search.

Drawings

FIG. 1 is a schematic diagram of an S-SCH channel structure;

FIG. 2 is a diagram illustrating a secondary synchronization code corresponding to a timeslot;

FIG. 3 is a flow chart of an embodiment;

FIG. 4 is a schematic diagram of a network structure of machine learning according to an embodiment;

FIG. 5 is a diagram illustrating WCDMA downlink baseband complex signal generation;

FIG. 6 is a data generation diagram of data to be predicted.

Detailed Description

The specific flow of cell search of the embodiment is shown in fig. 3:

1) network construction

Because the structure of WCDMA downlink signals in each time slot is basically consistent, a frame signal is scrambled by using a GOLD code truncation sequence of 38400, the GOLD code is a pseudo-random sequence generated by a certain rule, the corresponding characteristics are expected to be trained and learned in a neural network mode, and the RNN network family is suitable for learning sequence data and has excellent effect in the natural language processing field and time sequence processing, so that the RNN neuron is adopted for building the network.

The network structure is shown in fig. 4, and includes an input layer InputLayer (Simple _ rnn _1_ input), a hidden layer SimpleRNN (Simple _ rnn _1) built by a recurrent neural network layer, a fully-connected layer sense (sense _1) and an output layer Activation (Activation _ 1); the data form (batch size, number of channels, data length) none indicates that the default value is used. Input layer input data input: (None, 38400), input layer output data output: (None, 38400); input data of the recurrent neural network layer: (None, 38400), output data output of recurrent neural network layer: (None, 256); input of full connection layer input: (None, 256), full connection layer output data output: (None, 512); output layer input data input: (None, 512), output layer output data output: (None, 512).

The network inputs 38400 chips of one frame length, according to the characteristics of the WCDMA downlink signal: the maximum spreading length of each channel is 256, one time slot is 2560 chips, and meanwhile, in order to reduce the complexity of the network, the hidden layer is constructed by using 256 RNN neurons, and finally, in order to perform identification and prediction of scrambling codes, 512 output neurons Softmax are required to be classified, so that the 256 RNN neurons can keep respective states and are fully connected with the 512 output neurons. The network model will have 512 outputs, each representing the probability of using the corresponding primary scrambling code for the data, and the output with the highest probability in the whole network is used as the primary scrambling code identification result. Since the overall network goal is class prediction, the output layer uses softmax as the activation function, while choosing cross entropy as the loss function. The whole network adopts an Adam optimizer for a more appropriate learning rate, and meanwhile, in order to reduce the risk of overfitting, a regularization parameter droupout is set to be 0.2 in a hidden layer.

2) Training phase

As shown in fig. 5, the WCDMA downlink baseband complex signal generation process is: pilot channel, broadcast channel, paging channel, forward access signal, special service channel, paging indication data, access capture indication pass through their respective physical channels, then are modulated and coded, spread spectrum and gain generated, and then are accumulated; the accumulated channel information is scrambled by a primary scrambling code and superposed with a synchronous channel, the conventional synchronous channel SCH superposes a secondary synchronous code SSC on the primary synchronous code PSC, and the synchronous channel SCH is superposed with a frame synchronous signal scrambled by the primary scrambling code and then generates an analog signal through filtering modulation to be transmitted to a wireless channel.

In order to better adapt to the change of a pattern table and the change of secondary synchronous codes to finish the identification of primary scrambling codes, the training data uses WCDMA downlink baseband complex signals without the secondary synchronous codes, namely as shown in figure 5, the secondary synchronous code SSC part in a dotted line is removed, and IQ two-path signals obtained by only using primary synchronous codes PSC to generate synchronous channels SCH and overlapping channel information scrambled by the primary scrambling codes and then filtering and modulating are marked with corresponding scrambling code numbers (0-511) as labels.

The tag is one-hot coded to correspond to 512 sorted outputs of the network.

Inputting training data into the constructed network model, setting corresponding parameters for training, and then storing the trained network model.

3) Testing phase

The test data is data after time slot synchronization.

Because the characteristics learned by the network are extracted by training by using the data after frame synchronization, the data after frame synchronization must be input into a trained model to accurately identify the primary scrambling code used by the signal, and after time slot synchronization, the occurrence positions of the frame headers in the data of one frame length are 15 possible. In order to better reduce the data amount and the recognition error rate, the WCDMA signal after slot synchronization is slid one time by a slot data length (2560 chips) to be intercepted with a length of 38400 chips, and a group of 15 data to be predicted is made by 15 times of sliding, as shown in fig. 6.

Inputting data [15,1,38400] to be predicted, which is generated according to test data, into a trained network model to obtain recognition results with matrixes [15,1,512], searching the position of the maximum value of the recognition result of each data to obtain the recognized scrambling code number and probability of the recognition result, thereby obtaining a primary scrambling code classification result formed by 2 rows of data, wherein one row in the classification result represents the scrambling code numbers of 15 data recognition predictions, and the other row represents the corresponding prediction probability. Since the network model is trained with frame-synchronized data lines, the probability of correspondence of the frame-synchronized data prediction result should be very high, and the probability of correspondence of the recognition prediction result of 14 data is low because correct features cannot be extracted. By means of the difference, the position of the maximum value of the column of data representing the probability is obtained, the position of the frame head and the primary scrambling code used by the signal can be obtained, and cell search is completed.

Claims (3)

1. The WCDMA cell searching method based on machine learning comprises the following steps:

1) training:

1-1) generating training data: generating WCDMA downlink frame synchronization signals without adding an auxiliary synchronization channel, after completing in-phase and quadrature IQ separation of the frame synchronization signals, marking corresponding scrambling code numbers on IQ two-path signals as labels to complete the manufacturing of training data; 512 scrambling code numbers respectively correspond to 512 types of primary scrambling codes;

1-2) inputting training data into the set up machine learning network model for training, and storing the trained network model; the output layer of the machine learning network model has 512 classification outputs;

2) the testing steps are as follows:

2-1) generating data to be predicted: in a WCDMA downlink, carrying out one-frame length interception on a time slot data length by once sliding a time slot synchronized WCDMA signal, and making a group of 15 data to be predicted by 15 times of sliding; 15 data to be predicted correspond to 15 frame header positions respectively;

2-2) inputting a group of 15 data to be predicted into the trained network model to obtain the recognition results of the 15 data to be predicted, and obtaining the correct frame head position and the main scrambling code by searching the maximum probability value in the recognition results to complete the cell search.

2. The method of claim 1, wherein the label is processed using one-hot encoding.

3. The method of claim 1, in which the machine learning network model is a Recurrent Neural Network (RNN).

Images