Control device for a wireless access network

Abstract

To optimize allocation of functional blocks related to AI / ML to Near-RT RIC and Non-RT RIC in a control device for a radio access network.SOLUTION: A non-real-time control unit (Non-RT RIC) and a near-real-time control unit (Near-RT RIC) are hierarchized, and an AI / ML training function (11, 12, 16) for generating a learning model on the basis of data collected from an O-RAN base station device 10, and a retraining function (14, 15, 19, 20) for retraining the learning model when the occurrence of a concept drift is detected on the basis of inference performance data are allocated to the non-real-time control unit. An AI / ML inference function (13, 17, 18) for controlling a radio access network on the basis of a result inferred by applying the latest data to the learning model, and transmitting the inference performance data to the retraining function is allocated to the near-real-time control unit.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a control device for a radio access network, and more particularly to a control device for a radio access network having a function of re-learning a learning model generated by learning data collected from the radio access network.

Background Art

[0002] In a radio access network (RAN), conventionally, the functions of a base station that were integrated have been divided into a CU (Centralized Unit) that performs session processing, a DU (Distributed Unit) that performs baseband processing, and an RU (Radio Unit) that performs radio processing, and a specification study for opening the interface specifications between the units is being promoted by the O-RAN Alliance.

[0003] In a Beyond 5G system, it is expected to further expand performance such as throughput, communication delay, and the number of connections, and provide a variety of services (e.g., robot control, connected cars, AR / VR, etc.), and AI (Artificial Intelligence) / ML (Machine Learning) has attracted attention as a key technology for realizing these.

[0004] In Non-Patent Documents 1 and 2, in order to maximize network performance among limited network resources in the RAN, the application of AI / ML to various applications such as beamforming control, radio resource allocation, traffic prediction, and base station function placement has been studied.

[0005] Non-Patent Document 3 discloses a technique of generating a learning model by performing learning based on data collected from the RAN, performing inference using the data collected from the RAN and the learning model, and controlling the RAN according to the inference result.

[0006] However, due to the passage of time and environmental changes, the characteristics of the data used in inference may change from the data during learning (concept drift), and the inference performance of the model may decrease.

[0007] To address such technical problems, the inventors of the present invention proposed and filed a patent application for an AI system that accumulates and monitors data related to AI / ML learning and inference from an O-RAN base station device, detects concept drift, and performs relearning (Patent Document 1).

[0008] FIG. 5 is a functional block diagram showing a conventional configuration of an AI system that detects concept drift and performs relearning.

[0009] The data collection unit 11 repeatedly collects the latest data from the O-RAN base station device 10, provides the collected latest data (collected data) to the AI / ML learning unit 12 and the AI / ML inference unit 13, and stores it in the data storage unit 14. The collected data stored in the data storage unit 14 is managed by the AI / ML database 15. The AI / ML learning unit 12 learns the collected data and generates a learning model for controlling the O-RAN base station device 10.

[0010] The AI / ML model management unit 16 manages the learning models generated by the AI / ML learning unit 12 in the past. The AI / ML inference unit 13 performs inference based on the collected data newly collected by the data collection unit 11 and the learning model, and outputs the inference result to the control unit 17 and the inference performance measurement unit 18. The control unit 17 controls the O-RAN base station device 10 based on the inference result.

[0011] The inference performance measurement unit 18 determines the inference performance based on the latest data collected after the control unit 17 controls the O-RAN base station device 10 based on the inference result and the inference result, and stores the inference performance data indicating the determined inference performance in the AI / ML database 15.

[0012] The concept drift detection unit 19 periodically acquires at least one of the collected data and the inference performance data from the AI / ML database 15, and determines whether concept drift has occurred. When detecting the occurrence of concept drift, the concept drift detection unit 19 instructs the relearning control unit 20 to generate (relearn) a new learning model. The relearning control unit 20 provides the AI / ML learning unit 12 with data for relearning and instructs relearning.

[0013] When the AI / ML learning unit 12 is instructed to relearn, it generates a new learning model based on the collected data newly collected by the data collection unit 11 and outputs it to the AI / ML model management unit 16. The AI / ML model management unit 16 compares the current learning model used by the AI / ML inference unit 13 with the new learning model, and if the inference performance by the new learning model is higher than the inference performance by the current learning model, it outputs the new learning model to the AI / ML inference unit 13.

[0014] Thereafter, the AI / ML inference unit 13 performs inference using the new learning model. If the inference performance by the new learning model is lower than the inference performance by the current learning model, the AI / ML model management unit 16 can instruct the AI / ML learning unit 12 to relearn.

Prior Art Documents

Patent Documents

[0015]

Patent Document 1

Non-Patent Documents

[0016]

Non-Patent Document 1

[0017] The RAN Intelligent Controller (RIC) responsible for the control and optimization of the RAN function has a hierarchical structure of a non-real-time component "Non-RT (Real Time) RIC" and a near-real-time component "Near-RT RIC" with different control cycles, as shown in FIG. 6.

[0018] Here, the Non-RT RIC has a control cycle of 1 second or more and a wide range of control targets, while the Near-RT RIC has different characteristics in that the control cycle is 10 msec to 1 second and the control target is narrow. Therefore, the optimal placement of each functional block related to AI / ML in the Near-RT RIC and Non-RT RIC has been conventionally studied.

[0019] The Non-RT RIC is provided in the premises (data center), and the Near-RT RIC is often placed at the edge site (the rooftop of a building or a rented room in an apartment). Therefore, if the functions related to AI / ML learning are placed in the Non-RT RIC and the functions related to AI / ML inference are placed in the Near-RT RIC, it is possible to create a highly generalizable learning model through learning using a wide range of data within the accommodation range of the premises. In addition, since inference can be performed for each edge site, the processing load of the edge site can be reduced.

[0020] However, if, prioritizing real-time performance, not only the functions related to AI / ML learning but also the functions related to the re-learning of the learning model are placed in the Near-RT RIC, the following technical issues may arise.

[0021] First, the processing load of the Near-RT RIC increases. That is, since the edge site has constraints on power and space, a powerful computer cannot be placed.

[0022] Second, only the information under the Near-RT RIC can be used for concept drift detection. That is, since the information of adjacent areas cannot be used, the detection of concept drift is delayed.

[0023] On the other hand, if all the functions related to re-learning are placed in the Non-RT RIC, the following effects can be expected.

[0024] First, the processing load of the Near-RT RIC is reduced, and re-learning can be performed even if the computing resources of the edge site are limited.

[0025] Second, since the information under the Non-RT RIC can be used for concept drift detection, the adaptability to environmental changes can be enhanced.

[0026] On the other hand, in order to place all functions related to re-learning in the Non-RT RIC and realize re-learning based on concept drift detection, it is necessary to transmit data on the inference performance of the learning model from the Near-RT RIC to the Non-RT RIC, and it is conceivable to use the A1 / O1 interface.

[0027] However, the O1 interface does not stipulate the interface specification for the performance indicators of the learning model. Although there is a specification for the interface of the performance indicators for the base station based on the 3GPP (registered trademark) specification, this cannot be directly used for the inference performance data of the learning model. Also, the A1 interface does not stipulate the interface specification for the performance indicators of the learning model.

[0028] An object of the present invention is to solve the above technical problems and provide a control device for a radio access network that can place all functions related to re-learning of a learning model in the Non-RT RIC and realize re-learning based on concept drift detection.

Means for Solving the Problems

[0029] To achieve the above object, the present invention provides a control device for a radio access network in which a non-real-time control unit and a quasi-real-time control unit are hierarchically arranged, a learning unit that generates a learning model based on data collected from the radio access network, an inference unit that controls the radio access network based on the result of inference by applying the collected data to the learning model, and a re-learning unit that detects whether a concept drift has occurred based on the collected data and causes the learning unit to re-learn the learning model when the occurrence of the concept drift is detected. The inference unit is arranged in the quasi-real-time control unit, and the learning unit and the re-learning unit are arranged in the non-real-time control unit.

Advantages of the Invention

[0030] According to the present invention, only the functions related to AI / ML inference are arranged in the quasi-real-time control unit, and the functions related to AI / ML learning and relearning are arranged in the non-real-time control unit. Therefore, the following effects are achieved.

[0031] (1) Since the functions related to AI / ML learning can be learned using a wide range of data within the accommodation range of the premises, a highly generalized learning model can be created.

[0032] (2) Since AI / ML inference can be performed for each edge site, the processing load on the edge site can be reduced.

[0033] (3) Since all the functions related to relearning are arranged in the non-real-time control unit, the processing load on the quasi-real-time control unit is reduced, and relearning can be implemented even if the computing resources at the edge site are limited.

[0034] (4) Since the information under the non-real-time control unit can be used for concept drift detection, the adaptability to environmental changes can be enhanced.

Brief Description of the Drawings

[0035]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Embodiments for Carrying Out the Invention

[0036] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a functional block diagram showing the configuration of the main part of an O-RAN control device according to an embodiment of the present invention, where illustrations of configurations unnecessary for the description of the present invention are omitted. Also, the same reference numerals as above represent the same or equivalent parts. This embodiment is characterized in that functions related to AI / ML inference are arranged in the Near-RT RIC, and functions related to AI / ML learning and retraining of the learning model are arranged in the Non-RT RIC.

[0037] The O-RAN control device is composed of an O-CU / O-DU, a Near-RT RIC, and a Non-RT RIC, and each function can communicate with each other via various interfaces including the O1 interface, A1 interface, and E2 interface defined by the O-RAN Alliance. The O-RAN base station device 10 is arranged in the O-CU / O-DU.

[0038] In the Near-RT RIC, as the main functions related to AI / ML inference, an AI / ML inference unit 13, a control unit 17, and an inference performance measurement unit 18 are arranged. In the Non-RT RIC, as the main functions related to AI / ML learning, a data collection unit 11, an AI / ML learning unit 12, and an AI / ML model management unit 16 are arranged, and further, as the main functions related to retraining, a data storage unit 14, an AI / ML database 15, a concept drift detection unit 19, and a retraining control unit 20 are arranged.

[0039] In the Non-RT RIC, the data collection unit 11 collects the latest data from the O-RAN base station device 10 of the O-CU / O-DU via the O1 interface and provides it to the data storage unit 14 and the AI / ML learning unit 12. The latest data is managed in the AI / ML database 15. The learning model created by the AI / ML learning unit 12 based on the collected data is registered in the AI / ML management unit 16 and provided to the AI / ML inference unit 13 of the Near-RT RIC via the A1 interface.

[0040] In the Near-RT RIC, the AI / ML inference unit 13 provides the inference results obtained by applying the latest data to the learning model acquired via the A1 interface to the control unit 17 and the inference performance measurement unit 18. The inference performance measurement unit 18 transmits the inference performance data to the AI / ML database 15 and the AI / ML management unit 16 of the Non-RT RIC via the A1 / O1 interface.

[0041] In the Non-RT RIC, when the concept drift detection unit 19 detects a concept drift, the re-learning control unit 20 instructs the AI / ML learning unit 12 to re-learn the learning model. The re-learned learning model is updated and registered in the AI / ML management unit 16 and provided to the AI / ML inference unit 13 of the Near-RT RIC via the A1 interface.

[0042] Thus, in this embodiment, since the functions related to AI / ML learning, AI / ML inference, and re-learning of the learning model are distributed between the Near-RT RIC and the Non-RT RIC, the following two messages (a) and (b) are added to send and receive information related to re-learning between the Near-RT RIC and the Non-RT RIC.

[0043] (a) Request for inference performance data (b) Transmission of the requested inference performance data

[0044] Here, in the (a) request for inference performance data, the following three pieces of information (1)-(3) are specified. (1) Inference performance data In O-RAN WG2, the metrics exemplified in FIG. 2 are listed as metrics related to the inference performance of AI / ML, and at least one metric is specified. In this metric, metrics for binary classification problems, metrics for multiclass classification problems, and metrics for regression classification problems are defined respectively. In the present embodiment, the average value and median value for the specified period can be obtained.

[0045] (2) Object The learning model for which inference performance data is to be obtained is specified.

[0046] (3) Data acquisition interval The acquisition interval of the inference performance data is specified.

[0047] By the way, all of the metrics exemplified in FIG. 2 are metrics that use correct answer data (true value), but there may be cases where correct answer data cannot be obtained from the system. Therefore, in the present embodiment, as described in detail below, it is possible to select including inference performance metrics when correct answer data cannot be obtained.

[0048] Generally, the performance of the model deteriorates for out-of-distribution (OOD) data. Therefore, an OOD score based on a learned model is added as an inference performance metric when correct answer data cannot be obtained.

[0049] For multiclass classification problems, Maximum over softmax probabilities (MSP) disclosed in Non-Patent Document 4, Outlier Exposure disclosed in Non-Patent Document 5, or ODIN disclosed in Non-Patent Document 6 can be specified as the maximum value when the output of the learned model is normalized by Softmax in the class direction.

[0050] Also, like Mahalanobis disclosed in Non-Patent Document 7, a Gaussian distribution of feature amounts can be calculated for each class using the second-to-last feature amount and the label of the learned model, and a value obtained by attaching a minus sign to the Mahalanobis distance of the class with the minimum Mahalanobis distance can also be specified.

[0051] Furthermore, like Energy disclosed in Non-Patent Document 8, a value obtained by attaching a minus sign to the free energy when the energy function is the output of the learned model with a minus sign attached can also be specified.

[0052] Furthermore, like GradNorm disclosed in Non-Patent Document 9, the norm of the gradient when backpropagating the KLD between the Softmax output of the learned model and the uniform distribution can also be specified.

[0053] On the other hand, for a binary classification problem, an index for a multi-class classification problem can be specified.

[0054] Regarding the transmission of the inference performance data in (b), in the present embodiment, as shown by an example in FIG. 3, inference performance data (p1 to pm) regarding the m indices specified by the request for information to the (1) AI / ML database are transmitted in a table format and at specified time intervals for each learning model (No.).

[0055] FIG. 4 is a sequence flow showing the operation of the present embodiment, and here, the communication among the O-CU / O-DU, the Near-RT RIC, and the Non-RT RIC will be described with attention focused thereon. In the present embodiment, the communication between the O-CU / O-DU and the Near-RT RIC is performed via the E2 interface, and the communication between the Near-RT RIC and the Non-RT RIC is performed via the O1 interface or the A1 interface.

[0056] The O-CU / O-DU repeatedly transmits the latest data of the O-RAN base station device 10 to the Near-RT RIC and the Non-RT RIC at a predetermined period. In this embodiment, at time t1, the O-CU / O-DU transmits the latest data to the Near-RT RIC via the E2 interface and to the Non-RT RIC via the O1 interface, respectively. In the Non-RT RIC, the latest data is acquired by the data collection unit 11.

[0057] In the Near-RT RIC, the AI / ML inference unit 13 applies the latest data to the current learning model to execute inference, and notifies the inference result to the control unit 17 and the inference performance measurement unit 18. The control unit 17 instructs the O-RAN base station device 10 of the O-CU / O-DU based on the inference result via the E2 interface at time t2.

[0058] The Non-RT RIC requests inference performance data from the Near-RT RIC at a predetermined period. In this embodiment, at time t3, when the AI / ML model management unit 16 of the Non-RT RIC requests the inference performance data from the Near-RT RIC via the O1 interface, in the Near-RT RIC, at time t4, the inference performance measurement unit 18 responds to the request and transmits the measurement result of the inference performance data to the Non-RT RIC in the table format via the O1 interface.

[0059] In the Non-RT RIC, the concept drift detection unit 19 measures the concept drift based on the inference performance data and the data stored in the AI / ML database 15. When the concept drift is detected at time t5, at time t6, the re-learning control unit 20 instructs the AI / ML learning unit 12 to perform re-learning. The AI / ML learning unit 12 performs re-learning to generate a learning model and updates and registers it in the AI / ML model management unit 16.

[0060] At time t7, the re-learned learning model is transmitted from the AI / ML model management unit 16 of the Non-RT RIC to the AI / ML inference unit 13 of the Near-RT RIC via the A1 interface. Therefore, hereafter, every time the latest data is collected, control based on the re-learned learning model is performed.

[0061] According to this embodiment, since the function related to AI / ML inference is arranged in the Near-RT RIC, while the functions related to AI / ML learning and re-learning are arranged in the Non-RT RIC, the processing load of the Near-RT RIC can be reduced. Therefore, in an environment where there are constraints on the computer resources at the edge site and concept drift does not occur frequently, it is possible to detect concept drift immediately and enhance the adaptability to environmental changes.

[0062] As a result, according to the embodiment, it becomes possible to contribute to Goal 9, "Build resilient infrastructure, promote inclusive and sustainable industrialization," and Goal 11, "Make cities inclusive, safe, resilient and sustainable," of the Sustainable Development Goals (SDGs) led by the United Nations.

Explanation of Reference Numerals

[0063] 10…O-RAN base station device, 11…Data collection unit, 12…AI / ML learning unit, 13…AI / ML inference unit, 14…Data storage unit, 15…AI / ML database, 16…AI / ML model management unit, 17…Control unit, 18…Inference performance measurement unit, 19…Concept drift detection unit, 20…Re-learning control unit

Claims

1. In a control device for a wireless access network in which a non-real-time control unit and a quasi-real-time control unit are hierarchically arranged, a learning unit that generates a learning model based on data collected from the wireless access network, an inference unit that controls the wireless access network based on the result of applying and inferring the collected data to the learning model, a relearning unit that detects whether concept drift has occurred based on the collected data, and when detecting the occurrence of concept drift, causes the learning unit to relearn the learning model, wherein the inference unit is arranged in the quasi-real-time control unit, and the learning unit and the relearning unit are arranged in the non-real-time control unit. A control device for a wireless access network characterized by this.

2. The inference unit includes inference performance measurement means for measuring inference performance based on the collected data and the result of inference, the relearning unit includes concept drift detection means for detecting the occurrence of concept drift based on the inference performance, and relearning control means for causing the learning unit to relearn the learning model when the occurrence of concept drift is detected, the non-real-time control unit requests inference performance data to the quasi-real-time control unit via the O1 interface, the quasi-real-time control unit responds to the request and transmits the inference performance data to the non-real-time control unit via the O1 interface, The non-real-time control unit transmits the relearned learning model to the quasi-real-time control unit via either the O1 interface or the A1 interface. The control device for a wireless access network according to claim 1.

3. The request for the inference performance data includes specifications of an index of inference performance, a target learning model, and an acquisition interval of data. The control device for a wireless access network according to claim 2.

4. The quasi-real-time control unit transmits the requested inference performance data in a table format describing data of each of the specified indexes for each of the specified learning models. The control device for a wireless access network according to claim 3.

5. The control device for a radio access network according to claim 3 or 4, characterized in that the index of the inference performance includes an out-of-distribution score assuming a learned model.

Images