Base station interference management method, system and electronic device based on ai-ran architecture

Abstract

The application provides a base station interference management method and system based on an AI-RAN architecture and electronic equipment, the application performs initialization deployment on the AI-RAN, and obtains a normalized multi-dimensional time sequence feature vector set of KPM data and UE measurement data of a serving cell and a neighbor cell; an LSTM-Attention model is used to predict the normalized multi-dimensional time sequence feature vector set to obtain an interference type; an interference decision intelligent agent is constructed based on a degree-deterministic policy gradient (DDPG) reinforcement learning architecture, training and reasoning are performed in a non-real-time base station intelligent controller (Non-RT RIC), and an optimal interference mitigation strategy of the corresponding interference type is output; the optimal interference mitigation strategy is encrypted and issued by using A1, O1 and E2 interfaces, and near real-time execution is completed by a near real-time base station intelligent controller (Near-RT RIC). The application improves the accuracy and response speed of interference identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent network communication technology, and in particular to a base station interference management method, system and electronic device based on AI-RAN architecture. Background Technology

[0002] In modern mobile communication systems, with the continuous increase in the deployment density of 5G networks and the evolution of 6G technology, base stations present a three-dimensional network configuration of macro stations, micro stations, pico stations, indoor distribution systems, and RIS (Radio Router System). The types of interference have expanded from traditional co-channel interference to composite interference such as adjacent channel interference, spurious interference, blocking interference, terminal mutual interference, and multi-slice interference. The overlapping coverage between base stations has led to an increasingly complex wireless environment.

[0003] Traditional centralized network management systems suffer from problems such as prolonged computation time and low data standardization when handling massive amounts of real-time performance metrics across nodes and levels, which limits the accuracy of interference detection. Furthermore, traditional interference management methods often rely on static frequency planning or pre-defined rule models, making it difficult to achieve fine-grained scheduling and precise adaptation of network resources when dealing with sudden interference and dynamic load changes.

[0004] Therefore, there is an urgent need for a method for dynamic interference management of base stations. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a base station interference management method, system, and electronic device based on an AI-RAN architecture, in order to solve the problems of complex co-channel interference caused by increased base station deployment density in large-scale 5G communication networks, as well as the problems of delayed interference identification and low control efficiency in existing technologies.

[0006] In a first aspect, the present invention provides a base station interference management method based on an AI-RAN architecture, comprising:

[0007] Initialize and deploy the AI-RAN system's hardware and software resources;

[0008] KPM data and UE measurement data of the serving cell and neighboring cells are acquired, encapsulated by O1-agent and E2 agent, and transmitted to AI-RAN server through O1 and E2 interfaces respectively. The normalized multi-dimensional time series feature vector set obtained after cleaning and normalization is stored in the time series database InfluxDB.

[0009] The LSTM-Attention model is used to predict the normalized multi-dimensional temporal feature vector set to obtain the interference type and interference source localization results.

[0010] Based on the Degree Deterministic Policy Gradient (DDPG) reinforcement learning architecture, an interference decision agent is constructed, consisting of a state space, an action space, a reward function, and a target network. The agent is trained and inferred in a Non-RT Base Station Intelligent Controller (Non-RT RIC) to output the optimal interference mitigation strategy for the corresponding interference type.

[0011] The optimal interference mitigation strategy is encrypted and distributed using the A1, O1, and E2 interfaces, and executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller).

[0012] Preferably, the initial deployment of AI-RAN system hardware and software resources includes:

[0013] A test environment consisting of domestically produced servers and base station facilities is constructed. The hardware environment of the test environment includes AI-RAN server, network switch, baseband processing unit (BBU), expansion unit (EU), radio frequency unit (RU), and customer terminal equipment (CPE).

[0014] The Non-RT Base Station Intelligent Controller (Non-RT RIC) environment is deployed on the AI-RAN server, and the timing database InfluxDB and visualization tool Grafana are installed. The management plane agent O1-agent is run in the baseband processing unit (BBU), and the near real-time control agent E2-agent is deployed in the radio frequency unit (RU). These agents are responsible for reporting the collected key performance indicators (KPM) and UE measurement data to the Non-RT RIC in real time.

[0015] Preferredly, the key performance indicators (KPMs) of each cell are obtained from the physical nodes and transmitted to the AI-RAN server via the O1 interface. The key performance indicators (KPMs) of each cell include the reference received power (RSRP), received signal strength indication (RSSI), signal-to-interference-plus-noise ratio (SINR), downlink block error rate (BLER), physical resource block (PRB) utilization, uplink and downlink throughput, and modulation and coding scheme (MCS) of the serving cell.

[0016] Preferably, the collected UE measurement data is transmitted to the AI-RAN server via the E2-agent interface; the UE measurement data includes the neighbor cell reference signal received power (RSRP), the neighbor cell signal-to-interference-plus-noise ratio (SINR), the beam-level reception quality and the physical layer real-time interference power, and the real-time occupancy information of the disturbed PRB.

[0017] Preferably, the LSTM-Attention model is used to predict the interference type from the normalized multi-dimensional temporal feature vector set, including:

[0018] Take continuous The input tensor set is constructed from the normalized multi-dimensional temporal feature vectors of each time slice.

[0019] The temporal feature sequence of the input tensor set is extracted iteratively time-by-time using an LSTM network;

[0020] An attention mechanism is introduced to calculate the attention energy value of the hidden state at each time step, and the attention energy value is normalized by Softmax to obtain the attention weight. The hidden states at all time steps are weighted and summed based on the attention weight to obtain the global perturbation feature vector.

[0021] The global perturbation feature vector is input into the fully connected layer and the Sigmoid activation function, and the comprehensive perturbation probability is output.

[0022] By setting a preset interference decision threshold The interference is determined. If the overall interference probability is greater than or equal to the interference decision threshold, interference is determined to exist.

[0023] Once interference is detected, the interference type is classified into four types according to the preset multi-condition decision rules: co-channel interference, adjacent channel interference, spurious interference, and blocking interference.

[0024] The interference source is located based on the distribution of indicators and the relationship between neighboring cells, and the disturbed cell, interference source, disturbed user, disturbed physical resource block and interference level are output.

[0025] Preferably, the temporal feature sequence of the input tensor set is extracted iteratively time-by-time using an LSTM network, including:

[0026] For each moment The input tensor, the LSTM network, completes information filtering and memory update through forget gate, input gate, and output gate, and outputs the hidden state at the current time step to represent the time step. The comprehensive characteristics of the wireless environment;

[0027] After traversing the entire time window, a complete time series feature sequence is obtained.

[0028] As a preferred approach, an interference decision-making agent is constructed based on a Deep Deterministic Policy Gradient (DDPG) reinforcement learning architecture, consisting of a state space, action space, reward function, and target network. This agent is then trained and inferred within a Non-Real-Time Base Station Intelligent Controller (Non-RT RIC) to output the optimal interference mitigation strategy for the corresponding interference type. This includes:

[0029] Construct a state space S consisting of the serving cell SINR, RSRP, BLER, PRB utilization, MCS, the RSRP of the strongest interfering neighbor cell, interference type, and interference level;

[0030] Construct a continuous action space A consisting of RU transmit power adjustment, beam offset, disturbed PRB subband avoidance mask, handover trigger threshold adjustment, and cell offset CIO adjustment.

[0031] A reward function is constructed, with SINR improvement and throughput gain as positive rewards and power adjustment magnitude as negative penalties. ;

[0032] The current state space S is input into the Actor network of the Deep Deterministic Policy Gradient (DDPG) reinforcement learning architecture to obtain continuous actions A; the expected reward corresponding to action A is evaluated through the Critic network of the DDPG reinforcement learning architecture; and a reward function is selected. The largest action is taken as the optimal interference mitigation strategy;

[0033] Configuration instructions are generated based on the optimal interference mitigation strategy.

[0034] Preferably, the optimal interference mitigation strategy is encrypted and transmitted using the A1, O1, and E2 interfaces, and executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller); including:

[0035] The optimal interference mitigation strategy is encrypted using the national standard SM4 symmetric encryption.

[0036] The encrypted optimal interference mitigation strategy is encapsulated into a strategy message by the Non-RT RIC and then distributed to the Near-RT RIC via the A1 interface. The Near-RT RIC decrypts the optimal interference mitigation strategy and parses it into executable near-real-time control instructions.

[0037] The Near-RT RIC (Near-RT Intelligent Controller) distributes near-real-time control commands to the centralized unit (CU), distributed unit (DU), and radio frequency unit (RU) via the E2 interface.

[0038] Power and cell-level configurations are distributed via the O1 interface and YANG, a hierarchical data modeling language based on the NETCONF network configuration protocol.

[0039] Secondly, the present invention provides a base station interference management system based on an AI-RAN architecture, comprising:

[0040] The deployment module is used for the initial deployment of the AI-RAN system's hardware and software resources;

[0041] The data acquisition module is used to acquire KPM data and UE measurement data from the serving cell and neighboring cells. After being encapsulated by the O1-agent and E2 agent, the data is transmitted to the AI-RAN server through the O1 and E2 interfaces respectively. After cleaning and normalization, the data is stored in the time-series database InfluxDB.

[0042] The interference prediction module is used to predict the normalized multi-dimensional temporal feature vector set using the LSTM-Attention model to obtain the interference type and interference source localization results.

[0043] The interference mitigation strategy generation module is used to construct an interference decision-making agent based on the degree-deterministic policy gradient (DDPG) reinforcement learning architecture, which consists of a state space, an action space, a reward function, and a target network. The agent is trained and inferred in the Non-RT RIC (Non-Real-Time Base Station Intelligent Controller) and outputs the optimal interference mitigation strategy for the corresponding interference type.

[0044] The strategy distribution module uses the A1, O1, and E2 interfaces to encrypt and distribute the optimal interference mitigation strategy, which is then executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller).

[0045] Thirdly, the present invention provides an electronic device, comprising:

[0046] At least one processor; and

[0047] A memory that is communicatively connected to the at least one processor;

[0048] The memory stores a computer program that can be executed by the at least one processor. When the computer program is executed by the at least one processor, it implements the base station interference management method based on the AI-RAN architecture.

[0049] Fourthly, the present invention provides a computer-readable storage medium storing a computer-executable computer program, which, when executed by a processor, implements a base station interference management method based on an AI-RAN architecture.

[0050] The beneficial technical effects of this invention are as follows:

[0051] 1. This invention uses the O1 interface to collect KPM indicators and the E2 interface to collect near real-time UE measurement data. Combined with LSTM network temporal feature extraction and Attention key feature weighting, it can automatically learn the interference evolution trend and avoid single-point threshold misjudgment. At the same time, through multi-condition joint decision-making, it can accurately distinguish four types of interference: co-channel, adjacent channel, spurious, and blocking, with an interference identification accuracy of no less than 95%.

[0052] 2. This invention adopts a layered control architecture of Non-RT RIC and Near-RT RIC. It completes policy distribution through the A1 interface, achieves near real-time execution at the 10ms-100ms level through the E2 interface, and completes management plane configuration distribution through the O1 interface, realizing the entire process from interference detection to policy execution. At the same time, it adopts coordinated control of power, beam, PRB, handover threshold, and CIO, which improves the interference suppression effect by more than 40%, effectively ensuring user rate and link stability.

[0053] 3. This invention constructs a multi-objective reward function using DDPG reinforcement learning, incorporating SINR gain, throughput gain, power cost, and BLER reduction. This minimizes power loss and coverage contraction while suppressing interference, avoiding coverage holes caused by blindly reducing power. Through PRB subband avoidance, beamforming offset, and joint optimization of handover and CIO, global resource collaborative scheduling in the time-frequency, power, spatial, and mobility domains is achieved, increasing PRB utilization by 15%-30% and reducing handover failure rate by 50%.

[0054] 4. This invention adopts a modular deployment approach of KPM-rApp, ID-rApp, and IM-rApp, without relying on specific hardware or proprietary protocols. It is built on domestically produced AI-RAN servers, baseband, and radio frequency hardware, and uses the national cryptographic SM4 algorithm to encrypt the transmission of interface commands throughout the process to prevent signaling tampering, replay attacks, and unauthorized access; it meets the high reliability requirements of URLLC, industrial internet, power, rail transit, and other scenarios. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating the method of Embodiment 1 of the present invention;

[0056] Figure 2 This is a system hardware architecture diagram in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of the software architecture in an embodiment of the present invention. Detailed Implementation

[0058] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0059] like Figure 1 As shown, this embodiment provides a base station interference management method based on an AI-RAN architecture, including:

[0060] S1. Initialize and deploy the AI-RAN system's hardware and software resources;

[0061] S2. Obtain KPM data and UE measurement data of the serving cell and neighboring cells. After being encapsulated by O1-agent and E2-agent, the data is transmitted to the AI-RAN server through the O1 and E2 interfaces respectively. After cleaning and normalization, the data is stored in the time series database InfluxDB.

[0062] S3. The LSTM-Attention model is used to predict the normalized multi-dimensional temporal feature vector set to obtain the interference type and interference source localization results.

[0063] S4. Based on the degree-deterministic policy gradient (DDPG) reinforcement learning architecture, an interference decision agent is constructed, consisting of a state space, an action space, a reward function, and a target network. The agent is trained and inferred in the Non-RT RIC (Non-Real-Time Base Station Intelligent Controller) to output the optimal interference mitigation strategy for the corresponding interference type.

[0064] S5. The optimal interference mitigation strategy is encrypted and sent through the A1, O1, and E2 interfaces, and is executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller).

[0065] In this embodiment, step S1 involves the initial deployment of the AI-RAN system's hardware and software resources, including:

[0066] A test environment consisting of domestically produced servers and base station facilities is constructed. The hardware environment of the test environment includes AI-RAN server, network switch, baseband processing unit (BBU), expansion unit (EU), radio frequency unit (RU), and customer terminal equipment (CPE).

[0067] like Figure 3 As shown, a non-real-time base station intelligent controller (Non-RT RIC) environment is deployed on the AI-RAN server, and the time-series database InfluxDB and the visualization tool Grafana are installed.

[0068] The management plane agent O1-agent runs in the baseband processing unit (BBU), and the near real-time control agent E2-agent is deployed in the radio frequency unit (RU). These agents are responsible for reporting the collected key performance indicators (KPM) and UE measurement data to the non-real-time base station intelligent controller (Non-RT RIC) in real time.

[0069] In this embodiment, in step S2, KPM data and UE measurement data of the serving cell and neighboring cells are acquired, encapsulated by O1-agent and E2-agent, and transmitted to the AI-RAN server through the O1 and E2 interfaces respectively. After cleaning and normalization, the data is stored in the time-series database InfluxDB; including:

[0070] The key performance indicators (KPMs) of each cell are obtained from the physical nodes and transmitted to the AI-RAN server via the O1 interface after being encapsulated by the O1-agent. The key performance indicators (KPMs) of each cell include the reference received power (RSRP), received signal strength indication (RSSI), signal-to-interference-plus-noise ratio (SINR), downlink block error rate (BLER), physical resource block (PRB) utilization, uplink and downlink throughput, and modulation and coding scheme (MCS) of the serving cell.

[0071] The collected UE measurement data is encapsulated by the E2-agent and transmitted to the AI-RAN server via the E2 interface. The UE measurement data includes the neighbor cell reference signal received power (RSRP), the neighbor cell signal-to-interference-plus-noise ratio (SINR), the beam-level reception quality and the physical layer real-time interference power, and the real-time occupancy information of the disturbed PRB.

[0072] The key performance indicator (KPM) and UE measurement data are timestamped to obtain an aligned multi-dimensional temporal feature vector set.

[0073] The aligned multi-dimensional temporal feature vector set is cleaned of outliers to remove abnormal data that exceed the preset range, resulting in a cleaned multi-dimensional temporal feature vector set.

[0074] The cleaned multi-dimensional time series feature vector set is mapped to the interval [0, 1] using min-max normalization to obtain the normalized multi-dimensional time series feature vector set, which is then stored in the time series database InfluxDB.

[0075] In this embodiment, step S3 involves using an LSTM-Attention model to predict the normalized multi-dimensional temporal feature vector set to obtain the interference type; including:

[0076] S31, Take Continuity The input tensor set is constructed from the normalized multi-dimensional temporal feature vectors of each time slice. ;

[0077] S32. Iteratively extract the input tensor set using an LSTM network at each time step. The temporal dependency features are as follows:

[0078] For each moment Input tensor The LSTM network described above completes information filtering and memory updates through forget gates, input gates, and output gates, and outputs the current hidden state. Used to characterize time The comprehensive characteristics of the wireless environment; namely:

[0079] ;

[0080] In the formula, The hidden state at the above time.

[0081] After traversing the entire time window, the complete temporal feature sequence is obtained. ;

[0082] S33. Introduce an attention mechanism for the hidden state at each time step. Calculate attention energy value And on attention energy value Attention weights are obtained by performing Softmax normalization. Based on attention weights The global perturbation feature vector is obtained by weighted summation of the hidden states at all time points. ;Right now:

[0083] ;

[0084] ;

[0085] ;

[0086] In the formula, , These are the weight matrix and bias term of the attention module, respectively; It is a value vector; It is the hyperbolic tangent function;

[0087] S34. Transfer the global interference feature vector Input a fully connected layer and a sigmoid activation function, output the combined perturbation probability. ,Right now:

[0088]

[0089] In the formula, It is the sigmoid activation function; , These are the weight matrix and bias term of the fully connected layer, respectively;

[0090] S35, using a preset interference decision threshold Interference is assessed to determine whether the CPE has entered the interference zone; the overall interference probability is considered. Greater than or equal to the interference decision threshold Interference was detected.

[0091] S36. Upon determining the presence of interference, the interference type is categorized into four types based on preset multi-condition decision rules: co-channel interference, adjacent-channel interference, spurious interference, and blocking interference; as detailed below:

[0092] If all of the following conditions are met, it is determined to be co-channel interference:

[0093] The signal-to-interference-plus-noise ratio (SINR) of the serving cell is less than the SINR threshold.

[0094] The received power of the strongest neighbor cell reference signal is greater than the neighbor cell interference intensity threshold.

[0095] Physical resource block utilization exceeds the high congestion PRB utilization threshold;

[0096] The absolute difference between the received reference signal power of the serving cell and the received reference signal power of the strongest neighboring cell is less than the threshold value of the difference between the RSRP of the serving cell and the neighboring cell.

[0097] If all of the following conditions are met, it is determined to be adjacent channel interference:

[0098] Adjacent channel received signal strength refers to a signal strength greater than the adjacent channel RSSI interference threshold;

[0099] The increase in background noise exceeds the background noise increase threshold.

[0100] The RSRP of neighboring cells at the same frequency is less than the weak signal threshold of neighboring cells at the same frequency.

[0101] The downlink block error rate is greater than the block error rate interference threshold.

[0102] If all of the following conditions are met, it is determined to be spurious interference;

[0103] The absolute value of the system noise floor is greater than the noise floor anomaly threshold;

[0104] All neighboring cell reference signal received power (RSRP) is less than the neighboring cell weak signal threshold;

[0105] The average block error rate across the entire bandwidth is greater than the block error rate interference threshold.

[0106] The average signal-to-interference-plus-noise ratio (SIR) across the entire bandwidth is less than the SIR interference threshold.

[0107] If all of the following conditions are met, it is determined to be a blocking interference:

[0108] The noise floor increases dramatically by a much greater margin than the noise floor threshold for blocking interference.

[0109] The serving cell's RSRP is greater than the spurious strong signal threshold;

[0110] The block error rate is close to 100%;

[0111] The uplink and downlink throughput of the cell is close to 0.

[0112] S37. Based on the distribution of indicators and the relationship between neighboring cells, locate the interference source and output the disturbed cell, interference source, disturbed user, disturbed physical resource block and interference level.

[0113] In this embodiment, step S4 involves constructing an interference decision agent based on the Deep Deterministic Policy Gradient (DDPG) reinforcement learning architecture, consisting of a state space, action space, reward function, and target network. This agent is then trained and inferred within a Non-Real-Time Base Station Intelligent Controller (Non-RT RIC) to output the optimal interference mitigation strategy for the corresponding interference type. This includes:

[0114] S41) Construct a state space S consisting of the serving cell SINR, RSRP, BLER, PRB utilization, MCS, the RSRP of the strongest interfering neighbor cell, interference type, and interference level; that is:

[0115]

[0116] In the formula, To serve the current signal-to-interference-plus-noise ratio of the community; Reference signal received power for the serving cell; This represents the current downlink block error rate; For physical resource block utilization; For modulation and coding strategies; This represents the received power of the reference signal from the strongest interfering neighboring cell. Interference type; Interference level.

[0117] S42) Construct a continuous action space A consisting of RU transmit power adjustment, beam offset, disturbed PRB subband avoidance mask, handover trigger threshold adjustment, and cell offset CIO adjustment, i.e.:

[0118] ;

[0119] In the formula, This is the RU transmit power adjustment amount; This refers to the beam tilt angle or horizontal offset adjustment amount. For circumvention masking of the disturbed PRB subband; Adjust the trigger threshold amount for switching; Adjust the CIO for the cell offset.

[0120] S43) Construct a reward function with SINR improvement and throughput gain as positive rewards and power adjustment magnitude as negative penalties. ;Right now:

[0121] ;

[0122] In the formula, This is an increase in SINR. For throughput gain; This represents the decrease in the block error rate. These are the weighting coefficients for the corresponding items.

[0123] S44) Input the current state space S into the Actor network of the Deep Deterministic Policy Gradient (DDPG) reinforcement learning architecture to obtain continuous actions A; and evaluate the expected reward corresponding to action A through the Critic network of the Deep Deterministic Policy Gradient (DDPG) reinforcement learning architecture. Choose the reward function The largest action is taken as the optimal interference mitigation strategy;

[0124] S45) Generate configuration instructions based on the optimal interference mitigation strategy;

[0125] use Adjusting the RU transmit power, in this embodiment, if there is user coverage No user coverage .

[0126] according to The interference direction beam is compressed to reduce the coverage overlap area;

[0127] according to Skip the disturbed subband and reallocate resources;

[0128] according to Lower the switching threshold and speed up user migration;

[0129] according to Increase the offset of neighboring cells to guide handover and traffic diversion.

[0130] In this embodiment, in step S5, the optimal interference mitigation strategy is encrypted and sent using the A1, O1, and E2 interfaces, and is executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller); specifically as follows:

[0131] S51. Encrypt the optimal interference mitigation strategy using the national standard SM4 symmetric encryption, including:

[0132] The optimal interference mitigation strategy is encapsulated in plaintext using the cell identifier, timestamp, and device ID as the encryption initialization vector (IV).

[0133] The SM4 algorithm is used to encrypt the plaintext encapsulation using the optimal interference mitigation strategy, generating an encrypted instruction string, which is then decrypted by the receiving end using a preset key.

[0134] S52. After encapsulating the encrypted optimal interference mitigation strategy into a policy message through the Non-RT RIC (Non-Real-Time Base Station Intelligent Controller), the message is distributed to the Near-RT RIC (Near-Real-Time Base Station Intelligent Controller) via the A1 interface.

[0135] The Near-RT RIC optimal interference mitigation strategy of the aforementioned near-real-time base station intelligent controller is decrypted and parsed into executable near-real-time control commands;

[0136] S53, the Near-RT RIC (Near-Real-Time Base Station Intelligent Controller) distributes near-real-time control commands to the centralized unit CU, distributed unit DU, and radio frequency unit RU via the E2 interface; wherein:

[0137] The beamforming bias command is sent to the RF unit RU to complete the beam angle or gain adjustment.

[0138] The PRB subband avoidance instruction is sent to the distributed unit (DU) to complete the shielding of the disturbed resource block and resource reallocation;

[0139] The handover trigger threshold adjustment and cell offset CIO adjustment commands are sent to the central unit (CU) to update mobility parameters;

[0140] S54. Power and cell-level configurations are distributed via the O1 interface and the YANG hierarchical data modeling language based on the NETCONF network configuration protocol, as detailed below:

[0141] Encapsulate the RU transmit power adjustment amount as a YANG model configuration item;

[0142] The YANG model configuration items are sent to the baseband processing unit (BBU) via the O1 interface and forwarded to the radio frequency unit (RU). The RU performs power adjustment and returns the configuration result. The O1-agent reports the configuration status in real time.

[0143] In some embodiments, such as Figure 2 As shown, RU- and RU-2 represent radio frequency units of two adjacent cells. When CPE- or CPE-2 moves to the overlapping coverage area between the two cells, it will receive radio signals from both radio frequency units simultaneously. The signal from the non-serving cell becomes inter-cell interference for the CPE.

[0144] Example 2

[0145] This embodiment provides a base station interference management system based on an AI-RAN architecture, including:

[0146] The deployment module is used for the initial deployment of the AI-RAN system's hardware and software resources;

[0147] The data acquisition module is used to acquire KPM data and UE measurement data of the serving cell and neighboring cells respectively. After being encapsulated by O1-agent and E2 agent, the data is transmitted to the AI-RAN server through O1 and E2 interfaces respectively. After cleaning and normalization, the data is stored in the time series database InfluxDB.

[0148] The interference prediction module is used to predict the normalized multi-dimensional temporal feature vector set using the LSTM-Attention model to obtain the interference type and interference source localization results.

[0149] The interference mitigation strategy generation module is used to construct an interference decision-making agent based on the degree-deterministic policy gradient (DDPG) reinforcement learning architecture, which consists of a state space, an action space, a reward function, and a target network. The agent is trained and inferred in the Non-RT RIC (Non-Real-Time Base Station Intelligent Controller) and outputs the optimal interference mitigation strategy for the corresponding interference type.

[0150] The strategy distribution module uses the A1, O1, and E2 interfaces to encrypt and distribute the optimal interference mitigation strategy, which is then executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller).

[0151] In this embodiment, the deployment module performs initial deployment of the AI-RAN system's hardware and software resources, including:

[0152] A test environment consisting of domestically produced servers and base station facilities is constructed. The hardware environment of the test environment includes AI-RAN server, network switch, baseband processing unit (BBU), expansion unit (EU), radio frequency unit (RU), and customer terminal equipment (CPE).

[0153] The Non-RT Base Station Intelligent Controller (Non-RT RIC) environment is deployed on the AI-RAN server, and the timing database InfluxDB and visualization tool Grafana are installed. The management plane agent O1-agent is run in the baseband processing unit (BBU), and the near real-time control agent E2-agent is deployed in the radio frequency unit (RU). These agents are responsible for reporting the collected key performance indicators (KPM) and UE measurement data to the Non-RT RIC in real time.

[0154] In this embodiment, the data acquisition module obtains the key performance indicators (KPM) of each cell from the physical node and transmits them to the AI-RAN server through the O1 interface. The key performance indicators (KPM) of each cell include the reference signal received power (RSRP), received signal strength indication (RSSI), signal-to-interference-plus-noise ratio (SINR), downlink block error rate (BLER), physical resource block (PRB) utilization, uplink and downlink throughput, and modulation and coding scheme (MCS) of the serving cell.

[0155] In this embodiment, the data acquisition module transmits the acquired UE measurement data to the AI-RAN server via the E2 interface; the UE measurement data includes the neighbor cell reference signal received power (RSRP), the neighbor cell signal-to-interference-plus-noise ratio (SINR), the beam-level reception quality and the physical layer real-time interference power, and the real-time occupancy information of the disturbed PRB.

[0156] In this embodiment, the interference prediction module uses the LSTM-Attention model to predict the interference type from the normalized multi-dimensional temporal feature vector set; specifically as follows:

[0157] Take continuous The input tensor set is constructed from the normalized multi-dimensional temporal feature vectors of each time slice. .

[0158] Iterative extraction of the input tensor set using an LSTM network at each time step The temporal dependency features are as follows:

[0159] For each moment Input tensor The LSTM network described above completes information filtering and memory updates through forget gates, input gates, and output gates, and outputs the hidden state at the current time. Used to characterize time The comprehensive characteristics of the wireless environment; namely:

[0160] ;

[0161] In the formula, The hidden state at the above time.

[0162] After traversing the entire time window, the complete temporal feature sequence is obtained. .

[0163] Introducing an attention mechanism for the hidden state at each time step Calculate attention energy value And on attention energy value Attention weights are obtained by performing Softmax normalization. Based on attention weights The global perturbation feature vector is obtained by weighted summation of the hidden states at all time points. ;Right now:

[0164] ;

[0165] ;

[0166] ;

[0167] In the formula, , These are the weight matrix and bias term of the attention module, respectively; It is a value vector; It is the hyperbolic tangent function.

[0168] global interference feature vector Input a fully connected layer and a sigmoid activation function, output the combined perturbation probability. ,Right now:

[0169]

[0170] In the formula, It is the sigmoid activation function; , These are the weight matrix and bias term of the fully connected layer, respectively.

[0171] By setting a preset interference decision threshold To determine the interference, consider the overall interference probability. Greater than or equal to the interference decision threshold Interference was detected.

[0172] Once interference is detected, it is categorized into four types based on preset multi-condition decision rules: co-channel interference, adjacent-channel interference, spurious interference, and blocking interference; as detailed below:

[0173] If all of the following conditions are met, it is determined to be co-channel interference:

[0174] The signal-to-interference-plus-noise ratio (SINR) of the serving cell is less than the SINR threshold.

[0175] The received power of the strongest neighbor cell reference signal is greater than the neighbor cell interference intensity threshold.

[0176] The physical resource block utilization rate is greater than the high congestion PRB utilization threshold.

[0177] The absolute difference between the received reference signal power of the serving cell and the received reference signal power of the strongest neighboring cell is less than the threshold value of the difference between the RSRP of the serving cell and the neighboring cell.

[0178] If all of the following conditions are met, it is determined to be adjacent channel interference:

[0179] Adjacent channel received signal strength refers to a signal strength greater than the adjacent channel RSSI interference threshold;

[0180] The increase in background noise exceeds the background noise increase threshold.

[0181] The RSRP of neighboring cells at the same frequency is less than the weak signal threshold of neighboring cells at the same frequency.

[0182] The downlink block error rate is greater than the block error rate interference threshold.

[0183] If all of the following conditions are met, it is determined to be spurious interference;

[0184] The absolute value of the system noise floor is greater than the noise floor anomaly threshold;

[0185] All neighboring cell reference signal received power (RSRP) is less than the neighboring cell weak signal threshold;

[0186] The average block error rate across the entire bandwidth is greater than the block error rate interference threshold.

[0187] The average signal-to-interference-plus-noise ratio (SIR) across the entire bandwidth is less than the SIR interference threshold.

[0188] If all of the following conditions are met, it is determined to be a blocking interference:

[0189] The noise floor increases dramatically by a much greater margin than the noise floor threshold for blocking interference.

[0190] The serving cell's RSRP is greater than the spurious strong signal threshold;

[0191] The block error rate is close to 100%;

[0192] The uplink and downlink throughput of the cell is close to 0.

[0193] The interference source is located based on the distribution of indicators and the relationship between neighboring cells, and the disturbed cell, interference source, disturbed user, disturbed physical resource block and interference level are output.

[0194] In this embodiment, the interference mitigation strategy generation module constructs an interference decision agent based on the Deep Deterministic Policy Gradient (DDPG) reinforcement learning architecture, consisting of a state space, action space, reward function, and target network. This agent is trained and inferred within a Non-Real-Time Base Station Intelligent Controller (Non-RT RIC) to output the optimal interference mitigation strategy for the corresponding interference type. This includes:

[0195] Construct a state space S consisting of the serving cell's SINR, RSRP, BLER, PRB utilization, MCS, the RSRP of the strongest interfering neighbor cell, interference type, and interference level; that is:

[0196]

[0197] In the formula, To serve the current signal-to-interference-plus-noise ratio of the community; Reference signal received power for the serving cell; This represents the current downlink block error rate; For physical resource block utilization; For modulation and coding strategies; This represents the received power of the reference signal from the strongest interfering neighboring cell. Interference type; Interference level.

[0198] Construct a continuous action space A consisting of RU transmit power adjustment, beam offset, disturbed PRB subband avoidance mask, handover trigger threshold adjustment, and cell offset CIO adjustment, namely:

[0199] ;

[0200] In the formula, This is the RU transmit power adjustment amount; This refers to the beam tilt angle or horizontal offset adjustment amount. For circumvention masking of the disturbed PRB subband; Adjust the trigger threshold amount for switching; Adjust the CIO for the cell offset.

[0201] A reward function is constructed, with SINR improvement and throughput gain as positive rewards and power adjustment magnitude as negative penalties. ;Right now:

[0202] ;

[0203] In the formula, This is an increase in SINR. For throughput gain; This represents the decrease in the block error rate. These are the weighting coefficients for the corresponding items.

[0204] The current state space S is input into the Actor network of the Deep Deterministic Policy Gradient (DDPG) reinforcement learning architecture to obtain continuous actions A; and the expected reward corresponding to action A is evaluated through the Critic network of the Deep Deterministic Policy Gradient (DDPG) reinforcement learning architecture. Choose the reward function The largest action is taken as the optimal interference mitigation strategy;

[0205] Generate configuration instructions based on the optimal interference mitigation strategy;

[0206] use Adjusting the RU transmit power, in this embodiment, if there is user coverage No user coverage .

[0207] according to The interference direction beam is compressed to reduce the coverage overlap area;

[0208] according to Skip the disturbed subband and reallocate resources;

[0209] according to Lower the switching threshold and speed up user migration;

[0210] according to Increase the offset of neighboring cells to guide handover and traffic diversion.

[0211] In this embodiment, the strategy distribution module uses the A1, O1, and E2 interfaces to encrypt and distribute the optimal interference mitigation strategy, which is then executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller); specifically as follows:

[0212] The optimal interference mitigation strategy is encrypted using the national standard SM4 symmetric encryption, including:

[0213] The optimal interference mitigation strategy is encapsulated in plaintext using the cell identifier, timestamp, and device ID as the encryption initialization vector (IV).

[0214] The SM4 algorithm is used to encrypt the plaintext encapsulation using the optimal interference mitigation strategy, generating an encrypted instruction string, which is then decrypted by the receiving end using a preset key.

[0215] After the encrypted optimal interference mitigation strategy is encapsulated into a policy message by the Non-RT RIC (Non-Real-Time Base Station Intelligent Controller), it is distributed to the Near-RT RIC (Near-Real-Time Base Station Intelligent Controller) through the A1 interface.

[0216] The Near-RT RIC optimal interference mitigation strategy of the aforementioned near-real-time base station intelligent controller is decrypted and parsed into executable near-real-time control commands;

[0217] The Near-RT RIC (Near-Real-Time Base Station Intelligent Controller) distributes near-real-time control commands to the centralized unit (CU), distributed unit (DU), and radio frequency unit (RU) via the E2 interface; wherein:

[0218] The beamforming bias command is sent to the RF unit RU to complete the beam angle or gain adjustment.

[0219] The PRB subband avoidance instruction is sent to the distributed unit (DU) to complete the shielding of the disturbed resource block and resource reallocation;

[0220] The handover trigger threshold adjustment and cell offset CIO adjustment commands are sent to the central unit (CU) to update mobility parameters;

[0221] Power and cell-level configurations are distributed via the O1 interface and the YANG hierarchical data modeling language based on the NETCONF network configuration protocol, as detailed below:

[0222] Encapsulate the RU transmit power adjustment amount as a YANG model configuration item;

[0223] The YANG model configuration items are sent to the baseband processing unit (BBU) via the O1 interface and forwarded to the radio frequency unit (RU). The RU performs power adjustment and returns the configuration result. The O1-agent reports the configuration status in real time.

[0224] Example 3

[0225] This embodiment provides an electronic device, including:

[0226] At least one processor; and

[0227] A memory that is communicatively connected to the at least one processor;

[0228] The memory stores a computer program that can be executed by the at least one processor. When the computer program is executed by the at least one processor, it implements the base station interference management method based on AI-RAN architecture as described in Embodiment 1.

[0229] Example 4

[0230] This embodiment provides a computer-readable storage medium storing a computer-executable computer program. When the computer program is executed by a processor, it implements the base station interference management method based on AI-RAN architecture described in Embodiment 1.

[0231] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A method for base station interference management based on AI-RAN architecture, characterized in that, include: Initialize and deploy the AI-RAN system's hardware and software resources; KPM data and UE measurement data of the serving cell and neighboring cells are acquired, encapsulated by O1-agent and E2 agent, and transmitted to AI-RAN server through O1 and E2 interfaces respectively. The normalized multi-dimensional time series feature vector set obtained after cleaning and normalization is stored in the time series database InfluxDB. The LSTM-Attention model is used to predict the normalized multi-dimensional temporal feature vector set to obtain the interference type and interference source localization results. Based on the Degree Deterministic Policy Gradient (DDPG) reinforcement learning architecture, an interference decision agent is constructed, consisting of a state space, an action space, a reward function, and a target network. The agent is trained and inferred in a Non-RT Base Station Intelligent Controller (Non-RT RIC) to output the optimal interference mitigation strategy for the corresponding interference type. The optimal interference mitigation strategy is encrypted and sent using the A1, O1, and E2 interfaces, and is executed in near real-time by the Near-Real-Time Base Station Intelligent Controller (Near-RTRIC). 2.The AI-RAN architecture based base station interference management method of claim 1, wherein: Initial deployment of AI-RAN system hardware and software resources, including: A test environment consisting of domestically produced servers and base station facilities is constructed. The hardware environment of the test environment includes AI-RAN server, network switch, baseband processing unit (BBU), expansion unit (EU), radio frequency unit (RU), and customer terminal equipment (CPE). By deploying a non-real-time base station intelligent controller (Non-RT RIC) environment on the AI-RAN server and installing the timing database InfluxDB; running the management plane agent O1-agent in the baseband processing unit (BBU) and deploying the near real-time control agent E2-agent in the radio frequency unit (RU), the system is responsible for reporting the collected key performance indicators (KPM) and UE measurement data to the non-real-time base station intelligent controller (Non-RT RIC) in real time. 3.The AI-RAN architecture based base station interference management method of claim 2, wherein: The key performance indicators (KPMs) of each cell are obtained from the physical nodes and transmitted to the AI-RAN server via the O1 interface. The key performance indicators (KPMs) of each cell include the reference received power (RSRP), received signal strength indication (RSSI), signal-to-interference-plus-noise ratio (SINR), downlink block error rate (BLER), physical resource block (PRB) utilization, uplink and downlink throughput, and modulation and coding scheme (MCS) of the serving cell. The collected UE measurement data is transmitted to the AI-RAN server via the E2-agent interface. The UE measurement data includes the neighbor cell reference signal received power (RSRP), the neighbor cell signal-to-interference-plus-noise ratio (SINR), the beam-level reception quality and the physical layer real-time interference power, and the real-time occupancy information of the disturbed PRB. 4.The AI-RAN architecture based base station interference management method of claim 1, wherein: The LSTM-Attention model is used to predict the interference type by analyzing the normalized multi-dimensional temporal feature vector set. include: taking consecutive normalized multi-dimensional temporal feature vectors of consecutive time slices to construct an input tensor set; The temporal feature sequence of the input tensor set is extracted iteratively time-by-time using an LSTM network; An attention mechanism is introduced to calculate the attention energy value of the hidden state at each time step, and the attention energy value is normalized by Softmax to obtain the attention weight. The hidden states at all time steps are weighted and summed based on the attention weight to obtain the global perturbation feature vector. The global perturbation feature vector is input into the fully connected layer and the Sigmoid activation function, and the comprehensive perturbation probability is output. Interference is determined by a preset interference decision threshold. If the overall interference probability is greater than or equal to the interference decision threshold, interference is determined to exist. Once interference is detected, the interference type is classified into four types according to the preset multi-condition decision rules: co-channel interference, adjacent channel interference, spurious interference, and blocking interference. The interference source is located based on the distribution of indicators and the relationship between neighboring cells, and the disturbed cell, interference source, disturbed user, disturbed physical resource block and interference level are output.

5. The AI-RAN architecture based base station interference management method of claim 4, wherein: The temporal feature sequence of the input tensor set is extracted iteratively time-by-time using an LSTM network, including: For each time of the input tensor, the LSTM network completes information screening and memory update through the forget gate, input gate and output gate, and outputs the hidden state at the current time, which is used to represent the comprehensive features of the wireless environment at time . After traversing the entire time window, a complete time series feature sequence is obtained. 6.The AI-RAN architecture based base station interference management method of claim 1, wherein: A deep deterministic policy gradient (DDPG) reinforcement learning architecture is used to construct an interference decision agent consisting of a state space, action space, reward function, and target network. This agent is trained and inferred within a non-real-time base station intelligent controller (Non-RT RIC) to output the optimal interference mitigation strategy for the corresponding interference type. This includes: Construct a state space S consisting of the serving cell SINR, RSRP, BLER, PRB utilization, MCS, the RSRP of the strongest interfering neighbor cell, interference type, and interference level; Construct a continuous action space A consisting of RU transmit power adjustment, beam offset, disturbed PRB subband avoidance mask, handover trigger threshold adjustment, and cell offset CIO adjustment. A reward function is constructed with SINR promotion amount and throughput gain as positive rewards and power adjustment amplitude as negative punishment ; inputting the current state space S into an Actor network of a deep deterministic policy gradient (DDPG) reinforcement learning architecture to obtain a continuous action A; evaluating an expected return corresponding to the action A through a Critic network of the deep deterministic policy gradient (DDPG) reinforcement learning architecture; and selecting an action that maximizes a return function as an optimal interference mitigation strategy. Configuration instructions are generated based on the optimal interference mitigation strategy.

7. The AI-RAN architecture based base station interference management method of claim 6, wherein: The optimal interference mitigation strategy is encrypted and sent through the A1, O1, and E2 interfaces, and is executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller). include: The optimal interference mitigation strategy is encrypted using the national standard SM4 symmetric encryption. The encrypted optimal interference mitigation strategy is encapsulated into a strategy message by the Non-RT RIC and then distributed to the Near-RT RIC via the A1 interface. The Near-RT RIC decrypts the optimal interference mitigation strategy and parses it into executable near-real-time control instructions. The Near-RT RIC (Near-RT Intelligent Controller) distributes near-real-time control commands to the centralized unit (CU), distributed unit (DU), and radio frequency unit (RU) via the E2 interface. Power and cell-level configurations are distributed via the O1 interface and YANG, a hierarchical data modeling language based on the NETCONF network configuration protocol.

8. A base station interference management system based on AI-RAN architecture, characterized by, include: The deployment module is used for the initial deployment of the AI-RAN system's hardware and software resources; The data acquisition module is used to acquire KPM data and UE measurement data of the serving cell and neighboring cells. After being encapsulated by O1-agent and E2agent, the data is transmitted to the AI-RAN server through the O1 and E2 interfaces respectively. The normalized multi-dimensional time series feature vector set obtained after cleaning and normalization is stored in the time series database InfluxDB. The interference prediction module is used to predict the normalized multi-dimensional temporal feature vector set using the LSTM-Attention model to obtain the interference type and interference source localization results. The interference mitigation strategy generation module is used to construct an interference decision-making agent based on the degree-deterministic policy gradient (DDPG) reinforcement learning architecture, which consists of a state space, an action space, a reward function, and a target network. The agent is trained and inferred in the Non-RT RIC (Non-Real-Time Base Station Intelligent Controller) and outputs the optimal interference mitigation strategy for the corresponding interference type. The strategy distribution module uses the A1, O1, and E2 interfaces to encrypt and distribute the optimal interference mitigation strategy, which is then executed in near real-time by the Near-RT RIC (Near-RT Intelligent Controller).

9. An electronic device, comprising: At least one processor; as well as A memory that is communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor, wherein the computer program, when executed by the at least one processor, implements the base station interference management method based on the AI-RAN architecture as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program executable by a computer, characterized in that: When the computer program is executed by the processor, it implements the base station interference management method based on the AI-RAN architecture as described in any one of claims 1-7.

Images