5G PDCCH (Physical Downlink Control Channel) blind detection method

Abstract

The invention provides a 5G PDCCH (Physical Downlink Control Channel) blind detection method, which belongs to the technical field of channel blind detection, and is characterized in that an information filtering module is constructed, and the information filtering module is arranged behind a CA-SCL decoder and is in communication connection with the CA-SCL decoder; the feature extraction module extracts feature information from each candidate path output by CA-SCL decoding of the CA-SCL decoder to form a feature vector; the BP neural network discrimination module adopts a back propagation algorithm to train an input feature vector so as to output a decoding credibility score; and the credibility judgment module comprehensively utilizes BP neural network output, decoding characteristics and training prior distribution information to carry out posterior probability calculation on each decoding candidate, and judges whether the decoding candidate is a real and effective transmission code word according to the posterior probability calculation. An information classification function is cascaded on the basis of an existing 5G communication data transmission link framework, so that the false alarm rate in the PDCCH blind detection process is reduced, and the Polar code decoding performance is improved.

Description

Technical Field

[0001] This invention belongs to the field of channel blind detection technology and also relates to the field of wireless communication technology. Specifically, it relates to a polar code decoding method, and more particularly to a PDCCH channel blind detection method under the fifth generation mobile communication system (5G NR). Background Technology

[0002] 5G communication systems use polar codes as the coding scheme for control channels. In 5G New Radio (NR), downlink control information carries information about channel resource allocation, transmission format, and Hybrid Automatic Repeat reQuest (HARQ). The base station encodes the downlink control information into polar codes and transmits it to the user equipment (UE) via the Physical Downlink Control Channel (PDCCH).

[0003] The downlink control information encoding process of 5G NR is as follows: Figure 1 As shown, in 5G NR, a 24-bit Cyclic Redundancy Check (CRC) is concatenated with the downlink control information sequence. This concatenated sequence is scrambled using the Radio Network Temporary Identity (RNTI) of the target user equipment and encoded as a polar codeword. Multiple code blocks belonging to different user equipments are placed in different valid positions, and all valid positions constitute the search space for blind decoding. Each position in the search space is called a physical downlink control channel candidate, and the code length of each candidate is determined by the format of the physical downlink control channel. To effectively utilize channel resources, at the transmitting end, the base station dynamically selects the downlink control information format and the physical downlink control channel position to transmit the downlink control information.

[0004] At the receiving end, the user equipment does not know whether downlink control information has been sent, nor does it know the length, location, or other format of the downlink control information. Therefore, it needs to try decoding the polar code throughout the entire search space to identify the target downlink control information.

[0005] As mentioned in the prior art with patent publication number "CN118399979B", when a terminal device receives control information, it often needs to decode Polar codes without fully knowing whether the information exists; this is called blind detection. During this process, the decoder may mistake noise or erroneous candidate information for valid Polar code messages, thus generating a "false alarm." The probability that the system incorrectly identifies invalid information as a valid Polar code block is called the false alarm rate. It is an important indicator for measuring the decoder's misjudgment behavior in the absence of information input, affecting system resource scheduling and energy efficiency.

[0006] During the user equipment decoding process, it is necessary to decode physical downlink control channel candidates in all possible downlink control information formats to search for the target downlink control information, such as... Figure 2 As shown. Based on the 3rd Generation Partnership Project (3GPP) NR protocol, and according to the processing method for transmitting link control information, the following is designed: Figure 2 The PDCCH receive data processing procedure is shown below. First, based on the relevant parameters and RNTI type provided by the higher layers, it is determined which search spaces the current UE needs to parse and the corresponding Downlink Control Information (DCI) format. Then, resource mapping is performed according to the relevant parameters configured by the higher layers to obtain the data resources in all control resource sets (CORESET) belonging to the UE. Next, resource element group bundles (REG bundles) are de-interleaved according to the interleaving type of the transmitter to obtain the order of control channel elements (CCEs) before mapping to the REG bundle, and the demodulation reference signal (DM-RS) is removed. Then, the transmitted signal of the transmitter is estimated using channel estimation and received signal estimation. Finally, blind detection is performed, and the AL is determined according to different search space types, so that parsing starts from the beginning of the CCE in the corresponding candidate set. The parsing process includes, for example,... Figure 1 The demodulation and other processing steps are shown in the red box. Blind detection performance is crucial for data parsing at the PDCCH receiver, and among the blind detection modules, Polar decoding has the highest complexity.

[0007] CORESET defines the frequency domain range and number of symbols occupied within the corresponding BWP; SearchSpaces defines the timing and period for initiating blind detection within the corresponding CORESET; PDCCH set defines the set of PDCCHs for each aggregation level (AL) within the corresponding SearchSpaces; CCE represents the actual time-frequency resource to be detected, and blind detection of possible DCIs is performed at the corresponding CCE. A PDCCH can consist of 1, 2, 4, 8, or 16 CCEs, with the specific value determined by the DCI payload size and the required coding rate; the number of CCEs constituting a PDCCH is called the aggregation level. The base station can adjust the aggregation level of the PDCCH according to the actual wireless channel conditions.

[0008] Polar code decoding employs the CRC-Aided Successive Cancellation List (CA-SCL) decoding scheme. The CA-SCL decoding algorithm is currently the most popular form of SCL decoder, using CRC checksums to select the decoding result instead of relying on PM (Proof-of-Stake) decisions. The encoding and decoding process of the CA-SCL algorithm is as follows: Figure 3 As shown. Wherein, the source bits Add a CRC of length r, and transmit it through the channel, such as... Figure 4 The image shows L paths obtained after decoding using the SCL algorithm:

[0009] (1) If no path passes the CRC check, the decoding fails and the algorithm ends; (2) If only one path passes the CRC check, that path is determined to be the decoding result; (3) If multiple paths pass the CRC check, the path with the smallest path metric (PM, PathMetrics) is determined to be the decoding result, as shown in the following formula:

[0010] in,

[0011] For each physical downlink control channel candidate, the user equipment (UE) iterates through the downlink control information format to decode the polar code and descrambles the CRC based on the UE's radio network identifier. The CRC is used to identify whether the candidate carries the target downlink control information. If the CRC is successful, the candidate is determined to carry the target downlink control information; otherwise, it is determined not to. Figure 5 As shown.

[0012] A crucial parameter for evaluating the performance of a coding scheme is the false alarm rate (FRA). The FRA is the ratio of the number of blocks that the receiver believes it has successfully decoded, when in fact they differ from the blocks transmitted by the transmitter, or the transmitter never transmitted a block at all, to the total number of decoding attempts. Similarly, the FRA is an important indicator of the coding performance of the PDCCH channel. However, currently, there is no way to reduce the FRA during PDCCH blind detection, leading to poor Polar code decoding performance. Summary of the Invention

[0013] To address the shortcomings of existing technologies, this invention provides a 5G PDCCH channel blind detection method. This method cascades an information classification function onto the existing 5G communication data transmission link framework, thereby reducing the false alarm rate during PDCCH blind detection and improving Polar code decoding performance.

[0014] The present invention employs the following technical solution.

[0015] A 5G PDCCH channel blind detection method includes:

[0016] Step 1: Construct an information filtering module, which is placed after the CA-SCL decoder and communicates with it;

[0017] Step 2: The feature extraction module extracts feature information from each candidate path in the CA-SCL decoder output to form a feature vector;

[0018] Step 3: Back Propagation (BP) The neural network discrimination module uses the back propagation algorithm to train the input feature vector, thereby outputting a decoding confidence score;

[0019] Step 4: The credibility decision module comprehensively utilizes the output of the BP neural network, decoding features, and training prior distribution information to calculate the posterior probability of each decoding candidate and determine whether it is a genuine and valid transmission codeword.

[0020] Furthermore, in step 1, the information filtering module includes a feature extraction module, a BP neural network discrimination module, and a credibility decision module that are sequentially connected by communication.

[0021] Furthermore, in step 2, the extracted feature information includes the following types of feature information:

[0022] Log-Likelihood Ratio (LLR): mean, maximum, and variance;

[0023] The consistency ratio between frozen bits and decoded output;

[0024] The difference in path metrics between multiple path candidates;

[0025] Hamming distance between the best and second-best paths;

[0026] Check if the CRC check passes.

[0027] Furthermore, in step 2, the extracted feature information is standardized to form a feature vector y∈R, which is then used as the input to the BP neural network.

[0028] Furthermore, in step 3, the topology of the BP neural network includes three layers of feedforward network, namely the input layer, the hidden layer and the output layer; each layer of neurons is fully connected only to the neurons of the adjacent layers, there are no connections between neurons within the same layer, and there are no feedback connections between neurons of different layers, thus forming a feedforward neural network system with a hierarchical structure.

[0029] Furthermore, in step 3, the specific structure of the BP neural network is as follows:

[0030] Input layer: dimension d, corresponding to the length of the extracted feature vector;

[0031] Hidden layer 1: 64 neurons, activated by Rectified Linear Unit (ReLU);

[0032] Hidden layer 2: 32 neurons, ReLU activated;

[0033] Output layer: The output layer consists of 1 neuron with the activation function being the Sigmoid function, used to output the decoding confidence score s∈[0,1].

[0034] Furthermore, in step 3, the training methods for the BP neural network include:

[0035] Employ the cross-entropy loss function;

[0036] The dataset was generated by simulation and includes both correct and incorrect candidate samples.

[0037] Supervised learning methods are used, with the label being whether the candidate is a real transmitted codeword;

[0038] The BP neural network outputs a decoding confidence score, which indicates the probability that the candidate result is a real data codeword.

[0039] Furthermore, in step 3, the training process of the BP neural network includes:

[0040] If no path in the CA-SCL decoder's path list passes CRC, the received signal is determined not to be the target downlink control information. Conversely, if a candidate path passes CRC, the path with the lowest metric value is selected as the decoder's output. This output is then re-encoded into a polar codeword c1. After rate adaptation, the N-window polar codeword c1 yields a transmission codeword d1 of length M. d1 is modulated to obtain the valid transmit signal x1. Based on the received signal and the reconstructed transmit signal x1, the squared Euclidean distance γ is calculated using the following formula:

[0041]

[0042] in, Represents |y-x1| 2 , Represents ‖y-0‖ 2 y i Let y represent the i-th type of received signal, and x1 represent the feature vector. i This indicates the i-th type of transmitted signal.

[0043] Furthermore, in step 4, the inputs to the credibility decision module include:

[0044] The feature vector x corresponding to the candidate codeword is generated by the feature extraction module and includes path metric, LLR statistics, CRC check result and sorting position.

[0045] BP neural network prediction s: the confidence score output by the BP neural network discriminator, s∈[0,1]; prior probability P(H1): the probability of the true codeword appearing;

[0046] The output of the credibility decision module includes:

[0047] The posterior probability P(H1 / x) is calculated based on the Bayesian decision model: the confidence level of the candidate path as a true and valid codeword.

[0048] The decision was to determine whether to retain the path as the final decoding result based on the credibility of the actual valid codeword.

[0049] Furthermore, in step 4, the Bayesian decision model includes:

[0050] Let H1: candidate codewords be real data codewords; H0: candidate codewords be error paths;

[0051] Based on the information observation x, the objective is to calculate the posterior probability:

[0052]

[0053] in:

[0054] P(x / H1): The characteristic distribution under the condition that H1 is true;

[0055] P(H1): The prior probability of H1, which can be set as a fixed value or dynamically estimated;

[0056] P(x): The total probability of all cases, expanded by the following formula:

[0057] P(x) = P(x / H1)·P(H1) + P(x / H0)·P(H0), where P(x / H0) represents the feature distribution under the condition that H0 is true; P(H0) represents the prior probability of H0, which can be set as a fixed value or dynamically estimated.

[0058] The decision is made jointly with the CRC check result, forming a dual verification mechanism:

[0059] If the CRC fails but the network is trusted, the path can be retained as the final decoding result.

[0060] If the CRC passes but the network determines the path to be of low trust, the path can be rejected as the final decoding result.

[0061] The beneficial effects of the present invention are as follows, compared with the prior art:

[0062] This invention constructs an information filtering module, which is placed after and communicatively connected to the CA-SCL decoder; a feature extraction module extracts feature information from each candidate path of the CA-SCL decoding output of the CA-SCL decoder to form a feature vector; a BP neural network discrimination module trains the input feature vector using the backpropagation algorithm to output a decoding credibility score; and a credibility decision module comprehensively utilizes the BP neural network output, decoding features, and training prior distribution information to calculate the posterior probability of each decoding candidate and determine whether it is a genuine and valid transmission codeword. By cascading an information classification function onto the existing 5G communication data transmission link framework, the false alarm rate in the PDCCH blind detection process is reduced, and the Polar code decoding performance is improved. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the downlink control information encoding process in existing 5G NR technology;

[0064] Figure 2 This is a schematic diagram of the data processing procedure at the PDCCH receiver in existing technology;

[0065] Figure 3 This is a schematic diagram of the encoding and decoding process of CA-SCL in the existing technology;

[0066] Figure 4 This is a schematic diagram of the SCL decoding tree in the prior art;

[0067] Figure 5 This is a flowchart of 5G NR CA-CRC decoding in existing technology;

[0068] Figure 6 This is a schematic diagram of the decoding functional structure in this invention;

[0069] Figure 7 This is a schematic diagram of the functional structure of the information filtering module in this invention;

[0070] Figure 8 This is a flowchart of the training process of the BP neural network in this invention;

[0071] Figure 9 This is a structural diagram of the BP neural network model in this invention;

[0072] Figure 10 This is a flowchart of 5G NR CA-CRC decoding in this invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0074] The 5G PDCCH channel blind detection method of the present invention includes:

[0075] Considering that the Physical Downlink Control Channel (PDCCH) carrying Downlink Control Information (DCI) requires multiple blind detections, and these blind detections might detect DCI from other users, or even when the base station hasn't transmitted such DCI at all (this can still happen even with CRC), there are two false alarm scenarios for PDCCH blind detection: one is that the user equipment (UE) identifies the target DCI when no DCI is transmitted; the other is that the UE identifies another user's DCI as its own target DCI. The False Alarm (FAR) metric is crucial for PDCCH. The target FAR for 5G NR is the same as for 4G systems, which is below 2. -16 =1.52*10 -5 To reduce the false alarm rate in blind detection, this invention proposes a blind detection method using cascaded neural networks.

[0076] In 5G communication systems, the CRC (Corrective Graft Detector) must both detect errors during blind decoding to identify whether the received signal carries the target downlink control information of the user equipment, and correct transmission errors by checking multiple candidate paths in the SCL (Search Channel Logical Path) decoding. This dual working mode of the CRC negatively impacts the FAR (Failure-Average Recognition) performance of blind decoding.

[0077] To reduce the false alarm rate during PDCCH blind detection, this invention cascades an information classification function onto the existing 5G communication data transmission link framework. Two innovations are proposed in this invention:

[0078] 1. An information filtering module for reducing false alarm rate was cascaded on the existing 5G communication PDCCH transmission link;

[0079] 2. A neural network model is used as the information filtering module.

[0080] 5G NR uses a 24-bit CRC for blind decoding error detection and SCL decoding error correction, which is an increase compared to LTE's 16-bit CRC. This additional CR overhead causes a significant rate loss, especially for the short code long polar code used in the 5G control channel. This rate loss ultimately manifests as BLER loss in CA-SCL decoding. Regarding the two types of false alarms, the second type occurs less frequently due to the RITI process in downlink DCI; this invention primarily targets the first type of false alarm.

[0081] The technical solution proposed in this invention is as follows: Figure 6 As shown, further detection is performed on the existing CA-SCL decoding results to distinguish false alarm information.

[0082] As mentioned above, by adding an information filtering module to enhance the decoding algorithm, the CA-SCL decoding results are further inspected to filter out interference information that could cause false alarms, thereby reducing the false alarm rate. Figure 6 As shown, the information filtering module is placed after the CA-SCL decoder, and its functional structure is as follows: Figure 7 As shown, the information filtering module consists of a feature extraction module, a BP neural network discrimination module, and a credibility decision module. The specific method of the 5G PDCCH channel blind detection method is described below:

[0083] Step 1: Construct an information filtering module, which is placed after the CA-SCL decoder and communicates with it;

[0084] In a preferred but non-limiting embodiment of the present invention, in step 1, the information filtering module includes a feature extraction module, a BP neural network discrimination module, and a credibility determination module that are sequentially connected in communication.

[0085] Step 2: The feature extraction module extracts the following feature information from each candidate path of the CA-SCL decoder output to form a feature vector;

[0086] In a preferred but non-limiting embodiment of the present invention, the extracted feature information in step 2 includes the following types of feature information:

[0087] LLR (log-likelihood ratio) mean, maximum, and variance;

[0088] The ratio of frozen bits to decoded output consistency (i.e., the proportion of frozen bits that were successfully recovered);

[0089] Path Metric Gap (PMG) between multiple path candidates;

[0090] Hamming distance between the best and second-best paths;

[0091] Whether the CRC check passes (can be used as auxiliary input).

[0092] In a preferred but non-limiting embodiment of the present invention, in step 2, the extracted feature information is standardized to form a feature vector y∈R, which is used as the input of the BP neural network. That is, the above feature information is standardized (such as normalized) and then input into the next module.

[0093] Step 3: The BP neural network discrimination module uses the back propagation (BP) algorithm to train the input feature vector, thereby outputting a decoding confidence score;

[0094] In preferred but non-limiting embodiments of the present invention, such as Figure 9 As shown, in step 3, the topology of the BP neural network includes three layers of feedforward network, namely the input layer, the hidden layer and the output layer. Its characteristics are: each layer of neurons is fully connected only to the neurons of the adjacent layers, there is no connection between neurons in the same layer, and there is no feedback connection between neurons in different layers, thus forming a feedforward neural network system with a hierarchical structure.

[0095] This module is a multi-layer feedforward neural network (FNN) trained using the back propagation algorithm. The corresponding example structure is as follows:

[0096] In a preferred but non-limiting embodiment of the present invention, the specific structure of the BP neural network in step 3 is as follows:

[0097] Input layer: dimension d, corresponding to the length of the extracted feature vector (e.g., d = 5);

[0098] Hidden layer 1: 64 neurons, ReLU activated;

[0099] Hidden layer 2: 32 neurons, ReLU activated;

[0100] Output layer: The output layer consists of 1 neuron with the activation function being the Sigmoid function, used to output the decoding confidence score s∈[0,1].

[0101] In a preferred but non-limiting embodiment of the present invention, the training method of the BP neural network in step 3 includes:

[0102] Employ the cross-entropy loss function;

[0103] The dataset was generated by simulation and includes both correct and incorrect candidate samples.

[0104] Supervised learning methods are used, with the label being whether the candidate is a real transmitted codeword;

[0105] The BP neural network outputs a "confidence score", which is the decoding confidence score representing the probability that the candidate result is a real data codeword.

[0106] This BP neural network has a credibility scoring function. To implement a neural network with credibility scoring capabilities, the BP neural network is trained using training samples. The training process is as follows: Figure 8 As shown.

[0107] In a preferred but non-limiting embodiment of the present invention, step 3, the training process of the BP neural network includes:

[0108] If no path in the CA-SCL decoder's path list passes CRC, the received signal is determined not to be the target downlink control information. Conversely, if a candidate path passes CRC, the path with the lowest metric value is selected as the decoder's output. This output is then re-encoded into a polar codeword c1. After rate adaptation, the N-window polar codeword c1 yields a transmission codeword d1 of length M. d1 is modulated to obtain the valid transmit signal x1. Based on the received signal and the reconstructed transmit signal x1, the squared Euclidean distance γ can be calculated using the following formula:

[0109]

[0110] in, Represents |y-x1| 2 , Represents ‖y-0‖ 2 y i Let y represent the i-th type of received signal (feature information), and let x1 represent the feature vector. iThis represents the i-th type of transmitted signal. The N-window refers to the original code length of the polar code (i.e., the length of the information block before encoding).

[0111] Step 4: The credibility decision module is based on a credibility decision mechanism. This module comprehensively utilizes information such as the BP neural network output, decoding features, and training prior distribution to calculate the posterior probability of each decoding candidate.

[0112] Based on this, a judgment is made as to whether the transmitted codeword is genuine and valid.

[0113] In a preferred but non-limiting embodiment of the present invention, in step 4, the inputs to the confidence determination module include:

[0114] The feature vector x corresponding to the candidate codeword is generated by the feature extraction module and includes path metric (PM), LLR statistics, CRC check results and sorting position, etc.

[0115] BP neural network prediction value s: that is, the confidence score output by the BP neural network discriminator, s∈[0,1];

[0116] Prior probability P(x / H1): The probability of the actual codeword appearing, which can generally be set to 1 / L or estimated from actual data, where L is a preset positive integer;

[0117] The output of the credibility decision module includes:

[0118] The posterior probability P(H1|x) is calculated based on the Bayesian decision model: the confidence level of the candidate path as a true and valid codeword.

[0119] The decision is based on the credibility of the actual valid codewords to determine whether to retain the path as the final decoding result. If the credibility is higher than a preset threshold, the network is considered trustworthy; otherwise, the network is considered low-credibility.

[0120] In a preferred but non-limiting embodiment of the present invention, in step 4, the Bayesian decision model includes:

[0121] Let H1: candidate codewords be real data codewords; H0: candidate codewords be error paths;

[0122] Based on the information observation x, the objective is to calculate the posterior probability:

[0123]

[0124] in:

[0125] P(x / H1): The characteristic distribution (likelihood) given that H1 is true;

[0126] P(H1): The prior probability of H1, which can be set as a fixed value or dynamically estimated;

[0127] P(x): The total probability of all cases, expanded by the following formula:

[0128] P(x) = P(x / H1)·P(H1) + P(x / H0)·P(H0), where P(x / H0) represents the feature distribution (likelihood) given that H0 is true; P(H0) represents the prior probability of H0, which can be set as a fixed value or dynamically estimated.

[0129] The decision is made jointly with the CRC check result, forming a dual verification mechanism:

[0130] If the CRC fails but the network is trusted, the path can be retained as the final decoding result.

[0131] If the CRC passes but the network determines the path to be of low trust, the path can be rejected as the final decoding result.

[0132] The decoding workflow proposed in this invention is as follows: Figure 10 As shown:

[0133] After receiving the PDCCH, multipath candidate CA-SCL decoding is performed.

[0134] Each path output is checked using a CRC check;

[0135] Simultaneously, features related to the decoding path are extracted and fed into the information filtering module;

[0136] The feasibility reliability s is calculated using a BP neural network.

[0137] The final trusted path is parsed by the downstream feasible decision module, while other paths are discarded.

[0138] The training process of the BP model is as follows: Figure 9 As shown, one feature in the training set is the Euclidean distance of the reconstructed codeword, and another feature is the signal-to-noise ratio of the transmission channel. These two features serve as inputs to the training model. The output of the BP neural network is either 0 or 1, where 1 indicates a false alarm and 0 indicates information. The trained BP neural network model is then applied to CA-SCL decoding to further determine whether the decoded codeword is a false alarm, thereby reducing the false alarm rate.

[0139] The beneficial effects of the present invention are as follows, compared with the prior art:

[0140] This invention constructs an information filtering module, which is placed after and communicatively connected to the CA-SCL decoder; a feature extraction module extracts feature information from each candidate path of the CA-SCL decoding output of the CA-SCL decoder to form a feature vector; a BP neural network discrimination module trains the input feature vector using the backpropagation algorithm to output a decoding credibility score; and a credibility decision module comprehensively utilizes the BP neural network output, decoding features, and training prior distribution information to calculate the posterior probability of each decoding candidate and determine whether it is a genuine and valid transmission codeword. By cascading an information classification function onto the existing 5G communication data transmission link framework, the false alarm rate in the PDCCH blind detection process is reduced, and the polar code decoding performance is improved.

[0141] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention without departing from the spirit and scope of the present invention. Any modifications or equivalent substitutions should be covered within the scope of protection of the claims of the present invention.

Claims

1. A 5G PDCCH channel blind detection method, characterized in that, include: Step 1: Construct an information filtering module, which is placed after the CA-SCL decoder and communicates with it; Step 2: The feature extraction module extracts feature information from each candidate path in the CA-SCL decoder output to form a feature vector; Step 3: The BP neural network discrimination module uses the backpropagation algorithm to train the input feature vector, thereby outputting a decoding credibility score; Step 4: The credibility decision module comprehensively utilizes the output of the BP neural network, decoding features, and training prior distribution information to calculate the posterior probability of each decoding candidate and determine whether it is a genuine and valid transmission codeword.

2. The 5G PDCCH channel blind detection method according to claim 1, characterized in that, In step 1, the information filtering module includes a feature extraction module, a BP neural network discrimination module, and a credibility decision module that are sequentially connected by communication.

3. The 5G PDCCH channel blind detection method according to claim 2, characterized in that, In step 2, the extracted feature information includes the following types of feature information: LLR mean, maximum value and variance; The consistency ratio between frozen bits and decoded output; The difference in path metrics between multiple path candidates; Hamming distance between the best and second-best paths; Check if the CRC check passes.

4. The 5G PDCCH channel blind detection method according to claim 3, characterized in that, In step 2, the extracted feature information is standardized to form a feature vector y∈R, which is then used as the input to the BP neural network.

5. The 5G PDCCH channel blind detection method according to claim 4, characterized in that, In step 3, the topology of the BP neural network includes three layers of feedforward network, namely the input layer, the hidden layer and the output layer; each layer of neurons is fully connected only to the neurons of the adjacent layers, there are no connections between neurons within the same layer, and there are no feedback connections between neurons of different layers, thus forming a feedforward neural network system with a hierarchical structure.

6. The 5G PDCCH channel blind detection method according to claim 5, characterized in that, In step 3, the specific structure of the BP neural network is as follows: Input layer: dimension d, corresponding to the length of the extracted feature vector; Hidden layer 1: 64 neurons, ReLU activated; Hidden layer 2: 32 neurons, ReLU activated; Output layer: The output layer consists of 1 neuron with the activation function being the Sigmoid function, used to output the decoding confidence score s∈[0,1].

7. The 5G PDCCH channel blind detection method according to claim 6, characterized in that, In step 3, the training methods for the BP neural network include: Employ the cross-entropy loss function; The dataset was generated by simulation and includes both correct and incorrect candidate samples. Supervised learning methods are used, with the label being whether the candidate is a real transmitted codeword; The BP neural network outputs a decoding confidence score, which indicates the probability that the candidate result is a real data codeword.

8. The 5G PDCCH channel blind detection method according to claim 7, characterized in that, In step 3, the training process of the BP neural network includes: If no path in the CA-SCL decoder's path list passes CRC, the received signal is determined not to be the target downlink control information. Conversely, if a candidate path passes CRC, the path with the lowest metric value is selected as the decoder's output. This output is then re-encoded into a polar codeword c1. After rate adaptation, the N-window polar codeword c1 yields a transmission codeword d1 of length M. d1 is modulated to obtain the valid transmit signal x1. Based on the received signal and the reconstructed transmit signal x1, the squared Euclidean distance γ is calculated using the following formula: in, Represents |y-x1| 2 , Represents ‖y-0‖ 2 y i Let y represent the i-th type of received signal, and x1 represent the feature vector. i This indicates the i-th type of transmitted signal.

9. The 5G PDCCH channel blind detection method according to claim 8, characterized in that, In step 4, the inputs to the credibility decision module include: The feature vector x corresponding to the candidate codeword is generated by the feature extraction module and includes path metric, LLR statistics, CRC check result and sorting position. BP neural network prediction s: the confidence score output by the BP neural network discriminator, s∈[0,1]; prior probability P(H1): the probability of the true codeword appearing; The output of the credibility decision module includes: The posterior probability P(H1 / x) is calculated based on the Bayesian decision model: the confidence level of the candidate path as a true and valid codeword. The decision was to determine whether to retain the path as the final decoding result based on the credibility of the actual valid codeword.

10. The 5G PDCCH channel blind detection method according to claim 9, characterized in that, In step 4, the Bayesian decision model includes: Let H1: candidate codewords be real data codewords; H0: candidate codewords be error paths; Based on the information observation x, the objective is to calculate the posterior probability: in: P(x / H1): The characteristic distribution under the condition that H1 is true; P(H1): The prior probability of H1, which can be set as a fixed value or dynamically estimated; P(x): The total probability of all cases, expanded by the following formula: P(x) = P(x / H1)·P(H1) + P(x / H0)·P(H0), where P(x / H0) represents the feature distribution under the condition that H0 is true; P(H0) represents the prior probability of H0, which can be set as a fixed value or dynamically estimated. The decision is made jointly with the CRC check result, forming a dual verification mechanism: If the CRC fails but the network is trusted, the path can be retained as the final decoding result. If the CRC passes but the network determines the path to be of low trust, the path can be rejected as the final decoding result.

Images