AI-based 5g mobile terminal signal interference detection system

Abstract

This invention relates to the field of wireless communication interference detection technology, specifically to an AI-based 5G mobile terminal signal interference detection system. The system includes an interference signal modeling module, an interference type clustering module, a dynamic path scheduling module, an interference suppression mapping module, and a baseband retransmission control module. By combining time-frequency analysis and deep learning, this invention can accurately identify diverse interference types. Through joint analysis of instantaneous amplitude and phase transition signal characteristics, it effectively detects non-stationary abrupt change regions, improving the ability to identify novel interference signals. Dynamic path scheduling allocates inference channels based on the complexity of the interference region, enabling real-time signal recovery and reducing communication quality degradation caused by interference. Interference suppression technology and baseband retransmission control mechanisms ensure effective recovery of communication signals in complex interference environments, enhancing the anti-interference capability of the 5G communication system, optimizing communication quality, and improving system stability and robustness.

Description

Technical Field

[0001] This invention relates to the field of wireless communication interference detection technology, and in particular to an AI-based 5G mobile terminal signal interference detection system. Background Technology

[0002] The field of wireless communication interference detection technology involves the identification, analysis, and management of interference sources in wireless signals to ensure the stable operation of communication systems and reliable communication quality. Core aspects of this field include radio spectrum management, interference source type identification, interference signal feature extraction, interference location and tracking, and interference assessment methods. With the rapid development of wireless communication systems, especially 5G networks, spectrum resources are becoming more concentrated, the number of terminal devices is increasing dramatically, and the channel environment is becoming more complex, leading to frequent interference problems, including adjacent channel interference, intermodulation interference, and electromagnetic radiation interference. How to accurately detect interference signals in real time under complex dynamic environments and perform type identification and location determination has become an important research direction in this field.

[0003] Traditional 5G mobile terminal signal interference detection systems refer to systems that detect signal interference encountered by 5G terminals during operation. These systems rely on fixed rules to judge and analyze interference signals. They identify interference through methods such as preset interference feature templates, spectral energy threshold determination methods, and time-frequency domain signal analysis. For example, they might extract spectral features from the received signal by performing a Fourier transform and then match them with known interference templates, or set a signal strength threshold, exceeding which is considered interference. Furthermore, some traditional systems perform statistical analysis based on historical signal records to infer the existence and type of interference. However, these traditional methods, dependent on manually defined interference features, are ill-suited to the complex and ever-changing 5G communication environment and new forms of interference.

[0004] Existing technologies for wireless communication interference detection have certain limitations. Traditional methods rely on fixed rules and preset interference templates for interference identification and analysis. This makes it difficult for the system to cope with the complex and ever-changing 5G communication environment and new forms of interference, and it cannot flexibly adapt to dynamic channel changes. Traditional systems determine interference by performing Fourier transforms on received signals and matching templates, or by basing judgments on set signal strength thresholds. However, this approach has poor adaptability to unknown interference signals and struggles to identify new or unseen interference types. Furthermore, while statistical analysis based on historical signals can infer interference types, it lacks real-time performance and accuracy. Therefore, in practical applications, facing complex interference scenarios and diverse signal environments, traditional methods cannot promptly and accurately determine the source and type of interference, severely impacting communication quality and system stability. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing an AI-based 5G mobile terminal signal interference detection system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: the AI-based 5G mobile terminal signal interference detection system includes: The interference signal modeling module receives the original baseband signal stream from the 5G terminal's radio frequency front end, and jointly analyzes the instantaneous amplitude, phase jump, and spectrum occupancy status of the signal in time and frequency. Based on the location and shape of the non-stationary abrupt change region in the tensor, it generates an initial interference candidate region label set. The interference type clustering module maps each region to the embedding space of the pre-trained neural network based on the initial interference candidate region label set, compares the embedding vector with the distance of the preset interference prototype cluster center, assigns interference category identifiers, and forms a list of signal interference type clustering groups. The dynamic path scheduling module allocates inference channels to the interference area based on the processing complexity level in the clustering list of signal interference types, extracts original baseband signal stream segments for network discrimination, writes the discrimination results into the shared inference result linked list, and generates a dynamic interference identification and processing status linked list. The interference suppression mapping module performs position index matching on the interfered resource units based on the dynamic interference identification and processing state linked list, performs complex domain weighted reconstruction on the interfered symbols in the original baseband signal, and reassembles them into a complete frame structure according to the resource block order, forming the interference suppression reconstruction baseband signal frame detection result.

[0007] As a further aspect of the present invention, the initial interference candidate region marker set includes a time-frequency start index, resource unit span, and amplitude change intensity identifier; the interference type grouping list includes narrowband burst interference group, intermodulation harmonic interference group, adjacent channel leakage interference group, and electromagnetic pulse interference group; the channel-specific buffer includes inference channel ID, interference confidence vector sequence, and buffer fill status flag; and the interference suppression reconstructed baseband signal frame detection result includes resource block index, complex domain correction value array, and interference compensation kernel type encoding.

[0008] As a further aspect of the present invention, the interference signal modeling module includes: The time-frequency joint analysis submodule receives the raw baseband signal stream output from the 5G terminal's radio frequency front end. For each OFDM symbol, the time-domain sampling points obtained by IFFT are processed by a sliding window to calculate the amplitude variance. The phase sequence of the frequency-domain subcarriers in each OFDM symbol is processed by first-order differential processing to extract the coordinates of amplitude change points and phase jump points in the time-frequency grid and generate the original set of anomaly point coordinates. The tensor structure construction submodule performs a two-dimensional mapping of the original anomaly point coordinate set according to the time index and the subcarrier index, and fills it into a three-dimensional tensor structure of a predefined size. The third dimension stores the amplitude change intensity, phase jump variable and spectrum occupancy status flag respectively, forming a multi-dimensional signal state tensor. The candidate region generation submodule scans the distribution of non-zero elements in the third dimension of the multidimensional signal state tensor, uses a connected component labeling algorithm to identify spatially adjacent abnormal point clusters, calculates the time-frequency range and internal feature uniformity of the bounding rectangle for each cluster, and selects clusters that meet the judgment criteria as effective interference candidate regions based on area conditions and internal difference conditions, thus generating an initial interference candidate region label set.

[0009] As a further aspect of the present invention, the interference type clustering module includes: The embedding vector mapping submodule reads the multidimensional signal state tensor block corresponding to each region in the initial interference candidate region label set, inputs it into the pre-trained convolutional autoencoder, and the encoder part outputs a fixed-dimensional embedding vector to establish a mapping relationship from region to embedding feature. The prototype distance calculation submodule loads the interference prototype cluster center vectors pre-trained by unsupervised clustering according to the mapping relationship from the region to the embedded features. It calculates the Euclidean distance between each embedded vector and the center of the interference prototype cluster, calculates the perturbation distance feature value, selects the interference type corresponding to the cluster center with the smallest distance, and generates a set of interference region category identifiers. The group list generation submodule merges the area numbers of the same category according to the interference area category identifier set to form an interference type group set, and outputs a signal interference type cluster group list after sorting according to the preset category priority.

[0010] As a further aspect of the present invention, the dynamic path scheduling module includes: The channel resource configuration submodule calculates the load weight corresponding to the interference in the clustering list of signal interference types. The load weight includes four types: narrowband burst, intermodulation harmonics, adjacent channel leakage, and electromagnetic pulse. Independent neural network inference channels are allocated to each channel. Each channel is bound to dedicated weight parameters and activation function configuration to generate a multi-type interference inference channel configuration list. The signal segment extraction submodule extracts complex signal samples of corresponding resource units from the original baseband signal stream based on the time-frequency start index and span of each type of interference region in the initial interference candidate region marker set. The samples are then sent to the matching lightweight convolutional neural network branch according to the channel ID in the multi-type interference inference channel configuration list to calculate the amplitude difference feature value of the complex samples and form a channel signal sample set to be processed. The cache link management submodule monitors the number of data fillings of the channel signal sample set to be processed in the dedicated cache area of ​​the channel. When the number of items to be processed in any cache area reaches the preset capacity threshold, the interference confidence vector sequence output by the channel is packaged and written into the globally shared state chain, and the cache clearing flag of the channel is updated to generate a dynamic interference identification and processing state chain list.

[0011] As a further aspect of the present invention, the interference suppression mapping module includes: The resource unit index matching submodule parses the time-frequency start index of each record in the dynamic interference identification and processing state chain list, converts it into the resource block number and subcarrier offset in the 5G physical layer resource grid, locates the position of the interfered symbol in the frame structure, and generates the location information of the interfered resource unit. The compensation kernel calling submodule combines the location information of the interfered resource unit, selects the corresponding complex domain compensation kernel function, which is optimized for key interference types during the offline training phase, and applies directional attenuation and phase rotation to the real and imaginary parts of the interfered symbols in the original baseband signal, respectively, to generate a corrected signal sample processed by the compensation kernel. The signal reconstruction submodule applies the corrected signal samples processed by the compensation kernel to the complex samples at the corresponding positions in the original baseband signal, and splices the correction blocks in the order of resource blocks to form the interference suppression reconstructed baseband signal frame detection result.

[0012] As a further aspect of the present invention, the complex domain compensation kernel function refers to extracting the corresponding real part samples and imaginary part samples according to the location information of the disturbed resource unit, applying an amplitude attenuation coefficient threshold to the real part samples, applying a phase rotation angle threshold to the imaginary part samples, and forming a complex domain compensation parameter set by combining the amplitude attenuation coefficient threshold and the phase rotation angle threshold to obtain the complex domain compensation kernel function.

[0013] As a further aspect of the present invention, the system also includes a baseband re-demodulation control module: The baseband re-demodulation control module reads the detection result of the interference suppression and reconstruction baseband signal frame, writes the corrected signal block into the demodulation buffer of the 5G terminal baseband processor, triggers the physical layer re-demodulation operation within the current transmission time interval, replaces the original failed soft information bits, performs signal recovery under interference environment, and generates 5G terminal anti-interference communication recovery dataset. The 5G terminal anti-interference communication recovery dataset includes re-demodulated soft bitstreams and physical layer HARQ feedback states.

[0014] As a further aspect of the present invention, the baseband re-demodulation control module includes: The demodulation buffer writing submodule writes the corrected complex samples of each resource block in the detection result of the interference suppression and reconstruction baseband signal frame into the corresponding address of the demodulation buffer of the 5G terminal baseband processor according to the physical layer mapping rules, so as to obtain the corrected signal of the demodulation buffer. Based on the correction signal of the demodulation buffer, the re-demodulation trigger submodule sends a re-demodulation command to the physical layer control unit after the demodulation phase of the current transmission time interval ends, forcing the re-execution of channel estimation, equalization and soft demodulation operations on the corrected resource blocks to obtain the re-demodulated soft bit stream; The output result encapsulation submodule collects the re-demodulated soft bitstream, combines the HARQ process status and error detection results, encapsulates and outputs the 5G terminal anti-interference communication recovery dataset.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention combines time-frequency analysis and deep learning, employing a dynamic interference detection and suppression strategy to achieve efficient and real-time interference identification and location determination in complex environments. It accurately identifies diverse interference types and effectively detects non-stationary abrupt change regions through joint analysis of signal features such as instantaneous amplitude and phase jumps in time and frequency, thus improving the ability to identify novel interference signals and avoiding the limitations of relying on manually defined templates. Dynamic path scheduling allocates inference channels based on the complexity of the interference region, making the processing more flexible and efficient, and enabling real-time signal recovery operations. This effectively reduces communication quality degradation caused by interference. Interference suppression technology and baseband retransmission control mechanisms ensure effective recovery of communication signals in complex interference environments, enhancing the anti-interference capability of the 5G communication system, optimizing communication quality, and improving system stability and robustness. Attached Figure Description

[0016] Figure 1 This is a system flowchart of the present invention; Figure 2 This is a flowchart of the interference signal modeling module in this invention; Figure 3 This is a flowchart of the interference type clustering module in this invention; Figure 4 This is a flowchart of the dynamic path scheduling module in this invention; Figure 5 This is a flowchart of the interference suppression mapping module in this invention; Figure 6 This is a flowchart of the baseband re-demodulation control module in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0019] Please see Figure 1 The AI-based 5G mobile terminal signal interference detection system includes: The interference signal modeling module receives the original baseband signal stream from the 5G terminal's radio frequency front end, and jointly analyzes the instantaneous amplitude, phase jump, and spectrum occupancy status of the signal in time and frequency. Based on the location and shape of the non-stationary abrupt change region in the tensor, it generates an initial interference candidate region label set. Tensors are constructed by fusing data from multiple sensors in different dimensions such as the time domain and frequency domain. The interference type clustering module maps each region to the embedding space of the pre-trained neural network based on the initial interference candidate region label set, compares the embedding vector with the distance of the preset interference prototype cluster center, assigns interference category labels, and forms a list of signal interference type clustering groups. The dynamic path scheduling module allocates inference channels to the interference area based on the processing complexity level in the clustering list of signal interference types, extracts original baseband signal stream segments for network discrimination, writes the discrimination results into the shared inference result linked list, and generates a dynamic interference identification and processing status linked list. The interference suppression mapping module is based on the dynamic interference identification and processing state linked list. It performs position index matching on the interfered resource units, performs complex domain weighted reconstruction on the interfered symbols in the original baseband signal, and reassembles them into a complete frame structure according to the resource block order, forming the interference suppression reconstruction baseband signal frame detection result. The baseband re-demodulation control module reads the detection results of the interference suppression and reconstruction baseband signal frame, writes the corrected signal block into the demodulation buffer of the 5G terminal baseband processor, triggers the physical layer re-demodulation operation within the current transmission time interval, replaces the original failed soft information bits, completes the signal recovery under interference environment, and generates the 5G terminal anti-interference communication recovery dataset.

[0020] The initial interference candidate region label set includes time-frequency start index, resource unit span, and amplitude change intensity identifier. The interference type group list includes narrowband burst interference group, intermodulation harmonic interference group, adjacent channel leakage interference group, and electromagnetic pulse interference group. The channel-specific buffer includes inference channel ID, interference confidence vector sequence, and buffer fill status flag. The interference suppression reconstructed baseband signal frame detection results include resource block index, complex domain correction value array, and interference compensation kernel type encoding. The 5G terminal anti-interference communication recovery dataset includes re-demodulated soft bit stream and physical layer HARQ feedback status.

[0021] Please see Figure 2 The interference signal modeling module includes: The time-frequency joint analysis submodule receives the raw baseband signal stream output from the 5G terminal's radio frequency front end. For each OFDM symbol, the time-domain sampling points obtained by IFFT are processed by a sliding window to calculate the amplitude variance. The phase sequence of the frequency-domain subcarriers in each OFDM symbol is processed by first-order differential processing to extract the coordinates of amplitude change points and phase jump points in the time-frequency grid and generate the original set of anomaly point coordinates. The module receives the raw baseband signal stream output from the 5G terminal's RF front-end. This signal stream consists of continuous complex I / Q samples. With a 5G NR bandwidth of 100MHz and a subcarrier spacing of 15kHz, the baseband signal sampling rate is 61.44MSPS. For each OFDM symbol, the module performs a sliding window amplitude variance calculation on the 2048 time-domain sampling points obtained through IFFT. The sliding window has 64 points and a step size of 16 points. The module calculates the power variance of the 64 complex sampling points within the window. The calculated variance is compared to an amplitude mutation variance threshold of 0.05. This threshold is determined by the 99.9th percentile of the amplitude variance of thousands of OFDM symbols in a normal communication environment. A variance exceeding 0.05 is considered an amplitude mutation. The module also performs first-order differential processing on the phase sequence of 3300 subcarriers in the frequency domain, calculating the phase difference between adjacent subcarriers and limiting it to... The absolute value of the phase difference and the phase jump threshold For comparison, the threshold references the 3GPP TS 38.104 EVM tolerance requirements, and the absolute value of the phase difference exceeds... To determine phase jumps, the module records the coordinates of amplitude abrupt change points and phase jump points in the time-frequency grid, generating the original set of anomaly point coordinates.

[0022] The tensor structure construction submodule performs a two-dimensional mapping of the original anomaly point coordinate set according to the time index and the subcarrier index, and fills it into a three-dimensional tensor structure of predefined size. The third dimension stores the amplitude change intensity, phase jump variable and spectrum occupancy status flag respectively, forming a multi-dimensional signal state tensor. The original set of anomaly point coordinates is received, such as coordinate pairs (1, 200) and (1, 201). The coordinates are then mapped in two dimensions according to the time index and the subcarrier index to construct a three-dimensional tensor of size 140×3300×3. The tensor size is determined based on the 5GNR physical layer frame structure and the 100MHz maximum bandwidth configuration. Corresponding to 140 OFDM symbols in 10 time slots within a 10ms frame. Corresponding to 3300 subcarriers with a 100MHz bandwidth, the module will analyze each anomaly point in the original anomaly point coordinate set. The data is mapped to the corresponding position in the tensor, and the third dimension is filled. The third dimension contains three channels. Channel 1 stores the amplitude change intensity. If an amplitude change occurs in the time domain, the calculated amplitude variance value is stored; otherwise, it is filled with 0. Channel 2 stores the phase jump variable. If a phase jump occurs in the frequency domain subcarrier, the absolute value of the first-order phase difference is stored; otherwise, it is filled with 0. Channel 3 stores the spectrum occupancy status flag. The occupancy flag for normal 5G signals is 1; otherwise, it is 0. For example, the channel corresponding to tensor position T is filled with [0.08, 0, 1], and the channel corresponding to tensor position T is filled with [0, 0.8, 1]. The module forms a multi-dimensional signal state tensor.

[0023] The candidate region generation submodule scans the distribution of non-zero elements in the third dimension of the multidimensional signal state tensor, uses a connected component labeling algorithm to identify spatially adjacent anomalous point clusters, calculates the time-frequency range and internal feature uniformity of the bounding rectangle for each cluster, and selects clusters that meet the judgment criteria as effective interference candidate regions based on area conditions and internal difference conditions, thus generating an initial interference candidate region label set. The module receives a multidimensional signal state tensor, for example, with dimensions of 140×3300×3. It scans the distribution of non-zero elements in the third dimension of the tensor, checking for non-zero values ​​in channels 1 and 2. If a non-zero value is found, the time-frequency location is marked as an outlier. Subsequently, a connected component labeling algorithm is used to identify spatially adjacent outlier clusters. Outliers connected within eight neighborhoods are grouped into discrete outlier clusters. For each cluster, the module calculates the time-frequency range of its minimum bounding rectangle, for example, covering OFDM symbols. arrive , subcarrier arrive The uniformity of features within the cluster is calculated, and the average value of the amplitude abrupt change intensity and phase jump variable of all outliers within the cluster is taken, and their respective standard deviations are calculated. All of them passed The criteria are used to select clusters that meet the standards as effective interference candidate regions based on area and internal variability conditions. The area condition requires that the cluster's time-frequency region coverage area be greater than or equal to the minimum area threshold of 10 time-frequency resource units. The threshold is determined by narrowband interference time-frequency characteristic analysis. The internal variability condition requires that the uniformity of the internal features of the cluster be lower than the maximum variability threshold of 0.3. The threshold is determined by statistical analysis of various interference samples. For example, a cluster with an area of ​​18 and a uniformity of 0.25 is determined as an effective interference candidate region, and an initial interference candidate region label set is generated.

[0024] Please see Figure 3 The interference type clustering module includes: The embedding vector mapping submodule reads the multidimensional signal state tensor block corresponding to each region in the initial interference candidate region label set, inputs it into the pre-trained convolutional autoencoder, and the encoder part outputs a fixed-dimensional embedding vector to establish the mapping relationship between the region and the embedding feature. Read the initial interference candidate region marker set. The marker set records the time-frequency range of region 1 as (OFDM symbols 5-7, subcarriers 100-105). Based on the marker set, accurately extract the sub-tensor block corresponding to each candidate region from the multidimensional signal state tensor. For example, extract the tensor... The subtensor of size 3×6×3 is used as input. The extracted multidimensional signal state subtensor block is then fed into a pre-trained convolutional autoencoder (HAE). This HAE uses a U-Net architecture and consists of convolutional layers, pooling layers, and transposed convolutional layers. It is trained offline using a large amount of data containing known interference types and normal noise. The training objective is to minimize the reconstruction error between the input and output. The encoder performs multiple convolutional and pooling operations, outputting a fixed-dimensional 128-dimensional embedding vector. For example, a 3×6×3 subtensor block processed by the encoder outputs a floating-point vector of length 128. The module establishes a mapping relationship between regions and embedded features.

[0025] The prototype distance calculation submodule loads the pre-trained perturbation prototype cluster center vectors, based on the mapping relationship from region to embedded features. For each embedded vector, it calculates the Euclidean distance to the perturbation prototype cluster center using the formula: ; Calculate the disturbance distance feature value, select the disturbance type corresponding to the cluster center with the smallest distance, and generate a set of disturbance region category identifiers; in, Represents the perturbation distance characteristic value. The embedding vector is represented in the th... Values ​​in each dimension Represents the corresponding cluster center at the th Values ​​in each dimension The embedding vector is represented in the th... The mean across each dimension Represents the normalization adjustment constant. Indicates the total number of dimensions of the embedding vector; The mapping relationship between the receiving region and the embedded features, for example, the embedding vector corresponding to a candidate region. Simultaneously, it loads the cluster center vectors of interference prototypes, pre-trained through unsupervised clustering (e.g., K-means algorithm). The cluster centers represent typical features of different interference types. For example, the system pre-defines four main interference types: narrowband burst interference, intermodulation harmonic interference, adjacent channel leakage interference, and electromagnetic pulse interference. Each interference type corresponds to a 128-dimensional cluster center vector. These vectors are obtained through clustering training on the embedding vectors of a large number of labeled interference samples, reflecting the core features of each interference type. For example, the cluster center vector for narrowband burst interference... Intermodulation harmonic interference cluster center vector The module processes each received embedding vector. Calculate its perturbation distance eigenvalues ​​relative to the centers of all perturbed prototype clusters. Using formula ,in, The perturbation distance feature value measures the similarity between the embedding vector and the cluster center. The embedding vector is represented in the th... Values ​​in each dimension Represents the corresponding cluster center at the th Values ​​in each dimension The embedding vector is represented in the th... The mean across each dimension Represents the normalization adjustment constant. This represents the total number of dimensions of the embedded vectors. In this embodiment, This formula introduces This section perturbs the Euclidean distance, allowing the distance calculation to consider not only the absolute difference between the embedding vector and the cluster center, but also the degree of deviation of the embedding vector itself in each dimension. That is, the greater the deviation from its own mean, the greater the contribution of the difference in that dimension to the total distance, thus enhancing the sensitivity to anomalies or unique features. For example, assuming the embedding vector... Its mean Normalization adjustment constant a cluster center ; calculate Dimensional contributions: , ; but Item for The final distance is obtained by summing all 128 dimensions and then taking the square root. For example, the distance to the center of the narrowband burst interference cluster can be calculated. Distance from the center of the intermodulation harmonic interference cluster Distance from the center of the adjacent channel leakage interference cluster Distance from the center of the electromagnetic pulse interference cluster The interference type corresponding to the cluster center with the smallest distance is selected as the interference type to which the embedding vector belongs. In this example, It is the minimum value, so the region corresponding to the embedding vector is determined to be the adjacent channel leakage interference type. By performing this calculation and selection process for each embedding vector, the module generates a set of interference region category identifiers, which contains a unique identifier for each interference candidate region and the interference type it is classified as. Table 1: Center vectors of some interference prototype clusters and example embedding vector data; As shown in Table 1, the example embedding vectors Its mean Normalization adjustment constant ; Assuming the narrowband burst cluster center vector Calculate the perturbation distance from the center of the narrowband burst cluster. : for : ; for : ; for : ; Assuming that summing all 128 dimensions yields ; but ; The results show that, by calculating the perturbation distance, the distance between the embedding vector and the center of the narrowband burst interference cluster is 0.7. After comparing the distances with other cluster centers, the final interference type can be determined. The advantage of the formula is that by introducing the deviation of the embedding vector from its mean to weight the Euclidean distance, the sensitivity to anomalous features is enhanced, which makes it possible to more accurately identify subtle but key feature differences in similarity measures, thereby improving the accuracy of interference type classification.

[0026] The group list generation submodule merges the area numbers of the same category according to the interference area category identifier set to form an interference type group set, and outputs a signal interference type cluster group list after sorting according to the preset category priority. The module receives a set of interference region category identifiers, such as "Region A: Adjacent channel leakage, Region B: Narrowband burst, Region C: Adjacent channel leakage, Region D: Intermodulation harmonic interference". It iterates through the identifier set, creating a list for each interference type and adding all region identifiers of that type to the corresponding list. For example, Region A and Region C are merged into the "Adjacent channel leakage interference" type, forming an interference type grouping set with the structure {Adjacent channel leakage interference: [Region A, Region C], Narrowband burst interference: [Region B], Intermodulation harmonic interference: [Region D]}. Subsequently, the module groups the regions according to a preset category priority. The sets are sorted, and the priority is set based on the severity of the impact of different interference types on 5G communication performance. For example, electromagnetic pulse > narrowband burst > intermodulation harmonic interference > adjacent channel leakage. The priority setting refers to the 3GPP standard's assessment report on the impact of various interferences on bit error rate, throughput and latency, as well as the actual network operation and repair difficulty. The module sorts the interference type group sets in descending order according to this priority and outputs a list of signal interference type clusters. For example, the list is: [Narrowband burst interference: [Region B], intermodulation harmonic interference: [Region D], adjacent channel leakage interference: [Region A, Region C]].

[0027] Please see Figure 4 The dynamic path scheduling module includes: The channel resource configuration submodule calculates the load weight corresponding to the interference in the clustering list based on the type of signal interference. The load weight includes four types of interference: narrowband burst, intermodulation harmonics, adjacent channel leakage, and electromagnetic pulse. Independent neural network inference channels are allocated to each channel. Each channel is bound to dedicated weight parameters and activation function configuration, generating a multi-type interference inference channel configuration list. A clustering list of received signal interference types is generated, such as narrowband burst interference, intermodulation harmonic interference, and adjacent channel leakage interference. The computational load weight for each type of interference is determined based on the computational resources and latency required for neural network model inference. For example, narrowband burst interference has a load weight of 0.2, intermodulation harmonic interference has a load weight of 0.5, and electromagnetic pulse interference has a load weight of 0.9. These weights are obtained by testing model inference on a high-performance computing platform, recording CPU / GPU utilization, RAM usage, and average inference time, and normalizing these values ​​to an interval. Computational resources on the hardware accelerator are dynamically allocated based on the load weights. Subsequently, dedicated weight parameters and activation function configurations are bound to each independent neural network inference channel. For example, the narrowband burst channel is configured with a ReLU activation function and a set of narrowband feature-optimized convolutional kernel weight parameters, while the intermodulation harmonic channel is configured with a LeakyReLU activation function and a set of harmonic feature-optimized convolutional kernel weight parameters. These configurations are optimized for their respective interference type datasets during offline training, generating a multi-type interference inference channel configuration list.

[0028] The signal segment extraction submodule extracts complex signal samples corresponding to resource units from the original baseband signal stream based on the time-frequency start index and span of each type of interference region in the initial interference candidate region label set. These samples are then fed into a matching lightweight convolutional neural network branch according to the channel ID in the multi-type interference inference channel configuration list, using the following formula: ; Calculate the complex sample amplitude difference eigenvalues ​​to form the channel signal sample set to be processed; in, Represents the characteristic value of the amplitude difference of complex samples. The representative is in the sample sequence number The amplitude of the complex signal sample at the location, The representative is in the sample sequence number The reference value for the amplitude of the complex signal sample at the location. The upper limit of the sample number to which the summation terminates; The module receives an initial set of interference candidate region markers. For example, the marker set indicates that a certain region is a narrowband burst interference region, with its time-frequency start index being OFDM symbol 10, subcarrier 500, and a span of 2 OFDM symbols and 10 subcarriers. Based on this information, the module accurately extracts complex signal samples corresponding to the time-frequency resource units from the original baseband signal stream. For example, for the aforementioned region, all complex I / Q samples corresponding to OFDM symbols 10 to 11 and subcarriers 500 to 509 will be extracted from the original baseband signal stream. These samples form a sample array with a size of [missing information]. A complex matrix, where The number of time-domain sampling points corresponding to each subcarrier is determined. Subsequently, according to the channel ID in the multi-type interference inference channel configuration list, the intercepted complex signal samples are fed into the matching lightweight convolutional neural network branch. For example, samples of narrowband burst interference are fed into their corresponding neural network branch for further analysis. The module uses the formula... Calculate the complex sample amplitude difference eigenvalues, where, This represents the amplitude difference eigenvalue of complex samples, measuring the degree of amplitude difference between the actual signal sample and the reference signal sample. The representative is in the sample sequence number The amplitude of the complex signal sample at the location, i.e. , The representative is in the sample sequence number The amplitude reference value of the complex signal sample at the location, for example, the average amplitude value obtained through long-term channel state assessment or reference signal transmitted by the base station (CSI-RS). The upper limit of the sample index where the summation terminates, i.e., the total number of sampling points in this time-frequency region, for example... The core of this formula lies in quantifying the degree of amplitude deviation of the signal within the interference region by calculating the square root of the absolute value of the ratio of the sum of the differences between the actual signal amplitude and the reference amplitude to the sum of the reference amplitudes. For example, suppose a segment of the signal contains 10 complex samples with amplitudes of... The corresponding reference range is ; but ; ; but ; The results indicate a slight deviation between the signal amplitude and the reference amplitude, with an amplitude difference feature value of 0.079. This feature value will serve as one of the inputs to the neural network, helping the model to more accurately determine the presence and type of interference. In this way, a sample set of signals to be processed for each channel is formed, containing the original complex signal samples of each interference candidate region and the calculated amplitude difference feature value. The advantage of this formula is that it provides a standardized measure of amplitude deviation. By normalizing the amplitude difference relative to the reference amplitude, the feature values ​​are comparable across different signal power levels, thereby enhancing the robustness of interference feature extraction.

[0029] The cache link management submodule monitors the number of data fillings of the channel signal sample set to be processed in the dedicated cache area of ​​the channel. When the number of items to be processed in any cache area reaches the preset capacity threshold, the interference confidence vector sequence output by the channel is packaged and written into the globally shared state chain, and the cache clearing flag of the channel is updated to generate a dynamic interference identification and processing state chain list. The monitoring channel's dedicated buffer contains a large number of sample data entries. Each lightweight convolutional neural network branch has its own independent buffer for temporarily storing signal samples to be processed. For example, the buffer capacity for the narrowband burst interference channel is 100 samples. The number of currently filled, unprocessed entries in each buffer is continuously monitored, for example, every 1 millisecond. When the number of unprocessed entries in any buffer reaches a preset capacity threshold, for example, 100 for the narrowband burst interference channel, the processing flow is triggered. This involves packaging all interference confidence vector sequences that have been inferred to that channel, for example, into... The values ​​[0.95, 0.02, 0.01, 0.02] represent a 95% probability of narrowband burst interference. The packaged confidence vector sequence, along with the corresponding time-frequency region information and classification results, is written into a globally shared dynamic interference identification and processing status linked list. For example, the linked list records "Interference ID: X, Type: Narrowband Burst, Confidence: 0.95, Time-Frequency Range: (Symbol 10-11, Subcarrier 500-509)". The linked list adopts a doubly linked list structure. The channel cache clear flag is updated to "cleared" to generate the dynamic interference identification and processing status linked list.

[0030] Please see Figure 5 The interference suppression mapping module includes: The resource unit index matching submodule parses the time-frequency start index of each record in the dynamic interference identification and processing state chain list, converts it into the resource block number and subcarrier offset in the 5G physical layer resource grid, locates the position of the interfered symbol in the frame structure, and generates the location information of the interfered resource unit. The module parses the time-frequency start index of each record in the dynamic interference identification and processing state chain, such as "OFDM symbol 10, subcarrier 500". First, it extracts the OFDM symbol index and subcarrier index from the record, for example, "OFDM symbol 10" and "subcarrier 500". Then, according to the 5GNR physical layer resource grid mapping rules, the module converts the time-frequency index into a 5G physical layer resource block number and subcarrier offset. The 5GNR resource grid defines a resource block as consisting of 12 subcarriers, with a 30kHz subcarrier spacing of 27. There are 5 resource blocks. The subcarrier index 500 corresponds to the resource block number [500 / 12]=41. The subcarrier offset is 500÷12=41 remainder 8. The OFDM symbol index is combined with the resource block number to locate the precise position of the interfered symbol in the 5G frame structure. For example, it can be located as "the 10th OFDM symbol, the 41st resource block, and the 8th subcarrier". The conversion process follows the resource grid mapping standard defined in 3GPP TS38.211 to generate the location information of the interfered resource unit. The information is stored in the form of structured data.

[0031] The compensation kernel calling submodule combines the location information of the interfered resource unit and selects the corresponding complex domain compensation kernel function. The function is optimized for key interference types during the offline training stage. Directional attenuation and phase rotation are applied to the real and imaginary parts of the interfered symbols in the original baseband signal, respectively, to generate a corrected signal sample processed by the compensation kernel. The complex domain compensation kernel function refers to extracting the corresponding real and imaginary samples based on the location information of the disturbed resource units, applying an amplitude attenuation coefficient threshold to the real samples, and applying a phase rotation angle threshold to the imaginary samples. The complex domain compensation kernel function is obtained by combining the amplitude attenuation coefficient threshold and the phase rotation angle threshold to form a set of complex domain compensation parameters. Combining the location information of the interfered resource units with the original baseband signal stream, the interfered complex symbol samples are accurately extracted from the original baseband signal based on the location information. For example, the complex sample on the 8th subcarrier of the 41st resource block within the 10th OFDM symbol is extracted. Subsequently, based on the classification results of the interference type clustering module, the corresponding complex domain compensation kernel function is selected. The compensation kernel function is optimized for key interference types during the offline training phase. Each kernel function corresponds to a set of preset compensation parameters. The specific execution process is as follows: extract the corresponding real part sample I and imaginary part sample Q; apply an amplitude attenuation coefficient threshold of 0.6 to the real part sample I, with the threshold verified through narrowband burst interference experiments; when I is greater than 0.7, I is multiplied by 0.6 for attenuation; apply a phase rotation angle threshold to the imaginary part sample Q. The radians and threshold are adjusted through intermodulation harmonic interference experiments. If the phase corresponding to Q exceeds... If the range is determined, a rotation operation is performed on Q. A set of complex-domain compensation parameters is formed by combining the amplitude attenuation coefficient threshold and the phase rotation angle threshold, resulting in a complex-domain compensation kernel function applicable to specific interference types, such as narrowband burst interference. , If the original sample is The real part becomes With the imaginary part unchanged, the corrected signal sample is obtained. The results show that by selecting a compensation kernel function that matches the type of interference and applying its specific parameters, the real and imaginary parts of the interfered signal sample can be precisely adjusted, thereby effectively suppressing the interference effect.

[0032] The signal reconstruction submodule applies the correction signal samples processed by the compensation kernel to the complex samples at the corresponding positions in the original baseband signal, and splices the correction blocks in the order of resource blocks to form the interference suppression reconstructed baseband signal frame detection result; Receive the corrected signal sample processed by the compensation kernel, such as the corrected sample. The process involves precisely applying the correction signal sample to the corresponding complex sample in the original baseband signal stream, based on the time-frequency position information of the correction signal sample, such as the 10th OFDM symbol, the 41st resource block, or the 8th subcarrier. Then, using replace This enables signal correction in the affected area, such as the original signal. quilt After replacement, the module performs splicing operations on the corrected resource blocks according to the order of the resource blocks. For example, the sample data of resource blocks 0 to 274 in OFDM symbol 10 are arranged sequentially to ensure their time-frequency continuity and integrity. The splicing process follows the resource block arrangement rules of the 5GNR physical layer frame structure, thereby reconstructing a complete baseband signal frame and forming an interference suppression reconstructed baseband signal frame detection result. This result is a complete and corrected baseband signal frame.

[0033] Please see Figure 6 The baseband re-demodulation control module includes: The demodulation buffer writing submodule writes the corrected complex samples of each resource block in the interference suppression and reconstructed baseband signal frame detection results into the corresponding address of the demodulation buffer of the 5G terminal baseband processor according to the physical layer mapping rules, so as to obtain the corrected signal of the demodulation buffer. The system receives the detection result of the reconstructed baseband signal frame after interference suppression. This result is a complete and corrected baseband signal frame. The corrected complex samples of each resource block in the reconstructed baseband signal frame are then precisely written to the corresponding address in the demodulation buffer of the 5G terminal's baseband processor according to the 5G physical layer mapping rules. Specifically, the process involves first parsing the reconstructed baseband signal frame, identifying all resource blocks and their internal complex samples. For example, for 275 resource blocks within an OFDM symbol, each block is processed individually, and the resource elements are mapped to the demodulation buffer according to the resource elements defined in the 5G_NR physical layer specification. The mapping rule for the demodulation buffer is to calculate the target storage address of the complex samples in each resource block in the demodulation buffer. For example, the complex samples corresponding to the 10th OFDM symbol, the 41st resource block, and the 8th subcarrier are mapped to a specific memory address A0X3C in the demodulation buffer. Then, the corrected complex samples, such as 0.48+j0.3, are written to the calculated address A0X3C in the demodulation buffer. All corrected resource block samples will be written to their predetermined positions in the demodulation buffer to obtain the corrected signal of the demodulation buffer.

[0034] Based on the correction signal in the demodulation buffer, the re-demodulation trigger submodule sends a re-demodulation command to the physical layer control unit after the demodulation phase of the current transmission time interval ends, forcing the re-execution of channel estimation, equalization and soft demodulation operations on the corrected resource blocks to obtain the re-demodulated soft bit stream; Based on the state of the corrected signal in the demodulation buffer, after the demodulation phase of the current transmission time interval ends, a re-demodulation command is sent to the 5G terminal physical layer control unit. For example, if a flag indicating that the demodulation buffer needs to be updated and re-demodulated is detected, a re-demodulation command with a specific identifier is sent to the physical layer control unit. The command forces the physical layer control unit to re-execute channel estimation, equalization, and soft demodulation operations on the corrected resource blocks in the demodulation buffer. Specifically, the physical layer control unit first calls the MMSE estimator to re-evaluate the channel state information of the current time-frequency resource block based on the corrected signal and the reference signal to obtain an accurate channel response. Then, using the updated channel estimation results, the ZF or MMSE equalizer is executed to eliminate fading and multipath effects caused by the channel and restore the original amplitude and phase of the signal. Finally, the ML or LLR demodulator is executed on the equalized signal to perform soft demodulation operations, mapping complex symbols into soft bits with reliability information to obtain the re-demodulated soft bit stream.

[0035] The output result encapsulation submodule collects the re-demodulated soft bit stream, combines the HARQ process status and bit error detection results, encapsulates and outputs the 5G terminal anti-interference communication recovery dataset; The module collects re-demodulated soft bitstreams, such as a set of soft bit sequences containing high reliability information. Simultaneously, it combines this with Hybrid Automatic Repeat Request (HARQ) for final data encapsulation. The specific execution process is as follows: First, all re-demodulated soft bitstreams are collected. Then, the current HARQ process status is queried. The HARQ process status records the data block transmission history. For example, if the current data block is the third retransmission and has been successfully decoded after re-demodulation, the HARQ process status will be updated from "Waiting for retransmission" to "Successful." Simultaneously, a Cyclic Redundancy Check (CRC) code is executed to verify the re-demodulated bitstream to determine if there are any transmission errors. If the CRC check passes, the data block is error-free. The collected re-demodulated soft bitstreams, the updated HARQ process status, and the error detection results are encapsulated to form a structured dataset. For example, the dataset contains {"Data Block ID": 123, "Content": [Soft Bit Sequence], "HARQ Status": "Successfully Decoded", "CRC Result": "Passed", "Delay": "X microseconds"}. Finally, the module outputs a 5G terminal anti-interference communication recovery dataset.

[0036] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An AI-based 5G mobile terminal signal interference detection system, characterized in that, The system includes: The interference signal modeling module receives the original baseband signal stream from the 5G terminal's radio frequency front end, and jointly analyzes the instantaneous amplitude, phase jump, and spectrum occupancy status of the signal in time and frequency. Based on the location and shape of the non-stationary abrupt change region in the tensor, it generates an initial interference candidate region label set. The interference type clustering module maps each region to the embedding space of the pre-trained neural network based on the initial interference candidate region label set, compares the embedding vector with the distance of the preset interference prototype cluster center, assigns interference category identifiers, and forms a list of signal interference type clustering groups. The dynamic path scheduling module allocates inference channels to the interference area based on the processing complexity level in the clustering list of signal interference types, extracts original baseband signal stream segments for network discrimination, writes the discrimination results into the shared inference result linked list, and generates a dynamic interference identification and processing status linked list. The interference suppression mapping module performs position index matching on the interfered resource units based on the dynamic interference identification and processing state linked list, performs complex domain weighted reconstruction on the interfered symbols in the original baseband signal, and reassembles them into a complete frame structure according to the resource block order, forming the interference suppression reconstruction baseband signal frame detection result.

2. The AI-based 5G mobile terminal signal interference detection system according to claim 1, characterized in that, The initial interference candidate region marker set includes a time-frequency start index, resource unit span, and amplitude change intensity identifier. The interference type grouping list includes narrowband burst interference group, intermodulation harmonic interference group, adjacent channel leakage interference group, and electromagnetic pulse interference group. The channel-specific buffer includes inference channel ID, interference confidence vector sequence, and buffer fill status flag. The interference suppression and reconstruction baseband signal frame detection result includes resource block index, complex domain correction value array, and interference compensation kernel type encoding.

3. The AI-based 5G mobile terminal signal interference detection system according to claim 1, characterized in that, The interference signal modeling module includes: The time-frequency joint analysis submodule receives the raw baseband signal stream output from the 5G terminal's radio frequency front end. For each OFDM symbol, the time-domain sampling points obtained by IFFT are processed by a sliding window to calculate the amplitude variance. The phase sequence of the frequency-domain subcarriers in each OFDM symbol is processed by first-order differential processing to extract the coordinates of amplitude change points and phase jump points in the time-frequency grid and generate the original set of anomaly point coordinates. The tensor structure construction submodule performs a two-dimensional mapping of the original anomaly point coordinate set according to the time index and the subcarrier index, and fills it into a three-dimensional tensor structure of a predefined size. The third dimension stores the amplitude change intensity, phase jump variable and spectrum occupancy status flag respectively, forming a multi-dimensional signal state tensor. The candidate region generation submodule scans the distribution of non-zero elements in the third dimension of the multidimensional signal state tensor, uses a connected component labeling algorithm to identify spatially adjacent abnormal point clusters, calculates the time-frequency range and internal feature uniformity of the bounding rectangle for each cluster, and selects clusters that meet the judgment criteria as effective interference candidate regions based on area conditions and internal difference conditions, thus generating an initial interference candidate region label set.

4. The AI-based 5G mobile terminal signal interference detection system according to claim 3, characterized in that, The interference type clustering module includes: The embedding vector mapping submodule reads the multidimensional signal state tensor block corresponding to each region in the initial interference candidate region label set, inputs it into the pre-trained convolutional autoencoder, and the encoder part outputs a fixed-dimensional embedding vector to establish a mapping relationship from region to embedding feature. The prototype distance calculation submodule loads the interference prototype cluster center vectors pre-trained by unsupervised clustering according to the mapping relationship from the region to the embedded features. It calculates the Euclidean distance between each embedded vector and the center of the interference prototype cluster, calculates the perturbation distance feature value, selects the interference type corresponding to the cluster center with the smallest distance, and generates a set of interference region category identifiers. The group list generation submodule merges the area numbers of the same category according to the interference area category identifier set to form an interference type group set, and outputs a signal interference type cluster group list after sorting according to the preset category priority.

5. The AI-based 5G mobile terminal signal interference detection system according to claim 4, characterized in that, The dynamic path scheduling module includes: The channel resource configuration submodule calculates the load weight corresponding to the interference in the clustering list of signal interference types. The load weight includes four types: narrowband burst, intermodulation harmonics, adjacent channel leakage, and electromagnetic pulse. Independent neural network inference channels are allocated to each channel. Each channel is bound to dedicated weight parameters and activation function configuration to generate a multi-type interference inference channel configuration list. The signal segment extraction submodule extracts complex signal samples of corresponding resource units from the original baseband signal stream based on the time-frequency start index and span of each type of interference region in the initial interference candidate region marker set. The samples are then sent to the matching lightweight convolutional neural network branch according to the channel ID in the multi-type interference inference channel configuration list to calculate the amplitude difference feature value of the complex samples and form a channel signal sample set to be processed. The cache link management submodule monitors the number of data fillings of the channel signal sample set to be processed in the dedicated cache area of ​​the channel. When the number of items to be processed in any cache area reaches the preset capacity threshold, the interference confidence vector sequence output by the channel is packaged and written into the globally shared state chain, and the cache clearing flag of the channel is updated to generate a dynamic interference identification and processing state chain list.

6. The AI-based 5G mobile terminal signal interference detection system according to claim 5, characterized in that, The interference suppression mapping module includes: The resource unit index matching submodule parses the time-frequency start index of each record in the dynamic interference identification and processing state chain list, converts it into the resource block number and subcarrier offset in the 5G physical layer resource grid, locates the position of the interfered symbol in the frame structure, and generates the location information of the interfered resource unit. The compensation kernel calling submodule combines the location information of the interfered resource unit, selects the corresponding complex domain compensation kernel function, which is optimized for key interference types during the offline training phase, and applies directional attenuation and phase rotation to the real and imaginary parts of the interfered symbols in the original baseband signal, respectively, to generate a corrected signal sample processed by the compensation kernel. The signal reconstruction submodule applies the corrected signal samples processed by the compensation kernel to the complex samples at the corresponding positions in the original baseband signal, and splices the correction blocks in the order of resource blocks to form the interference suppression reconstructed baseband signal frame detection result.

7. The AI-based 5G mobile terminal signal interference detection system according to claim 6, characterized in that, The complex domain compensation kernel function refers to extracting the corresponding real and imaginary samples based on the location information of the disturbed resource units, applying an amplitude attenuation coefficient threshold to the real samples, and applying a phase rotation angle threshold to the imaginary samples. The complex domain compensation kernel function is obtained by combining the amplitude attenuation coefficient threshold and the phase rotation angle threshold to form a set of complex domain compensation parameters.

8. The AI-based 5G mobile terminal signal interference detection system according to claim 1, characterized in that, The system also includes a baseband re-demodulation control module: The baseband re-demodulation control module reads the detection result of the interference suppression and reconstruction baseband signal frame, writes the corrected signal block into the demodulation buffer of the 5G terminal baseband processor, triggers the physical layer re-demodulation operation within the current transmission time interval, replaces the original failed soft information bits, performs signal recovery under interference environment, and generates 5G terminal anti-interference communication recovery dataset. The 5G terminal anti-interference communication recovery dataset includes re-demodulated soft bitstreams and physical layer HARQ feedback states.

9. The AI-based 5G mobile terminal signal interference detection system according to claim 8, characterized in that, The baseband re-demodulation control module includes: The demodulation buffer writing submodule writes the corrected complex samples of each resource block in the detection result of the interference suppression and reconstruction baseband signal frame into the corresponding address of the demodulation buffer of the 5G terminal baseband processor according to the physical layer mapping rules, so as to obtain the corrected signal of the demodulation buffer. Based on the correction signal of the demodulation buffer, the re-demodulation trigger submodule sends a re-demodulation command to the physical layer control unit after the demodulation phase of the current transmission time interval ends, forcing the re-execution of channel estimation, equalization and soft demodulation operations on the corrected resource blocks to obtain the re-demodulated soft bit stream; The output result encapsulation submodule collects the re-demodulated soft bitstream, combines the HARQ process status and error detection results, encapsulates and outputs the 5G terminal anti-interference communication recovery dataset.

Images