Control device for wireless access network

Abstract

A control unit (Non-RT RIC) for a non-real-time system and a control unit (Near-RT RIC) for a near-real-time system are layered, and AI / ML learning functions (11, 12, 16) for generating a learning model on the basis of data collected from an O-RAN base station device (10) and re-learning functions (14, 15, 19, 20) for re-learning the learning model when the occurrence of concept drift is detected on the basis of reasoning performance data are disposed in the control unit for the non-real-time system. Furthermore, an AI / ML inference function (13, 17, 18), which transmits inference performance data to a re-learning function, is disposed in a control unit of the near real-time system, by controlling the wireless access network on the basis of a result of inference using the latest data for the learning model.

Description

Technical Field

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

[0002] The O-RAN Alliance is promoting specification research in the Radio Access Network (RAN) to split the functions of the previously integrated base station into a CU (Centralized Unit) for session processing, a DU (Distributed Unit) for baseband processing, and a RU (Radio Unit) for wireless processing, and to make the interface specifications between each unit public.

[0003] In the Beyond 5G system, it is expected that performance such as throughput, communication latency, and number of connections will be further improved, and a variety of services (such as robot control, connected cars, AR / VR, etc.) will be provided. As a key technology for achieving the above goals, AI (artificial intelligence) / ML (machine learning) has attracted attention.

[0004] In non-patent literature 1 and 2, in order to maximize network performance within limited network resources in wireless access networks, studies have been conducted on the application of AI / ML to various purposes, such as beamforming control, wireless resource allocation, service prediction, and base station function configuration.

[0005] Non-patent document 3 discloses a technology of generating a learning model by learning based on data collected from a radio access network, performing inference using the data collected from the radio access network and the learning model, and controlling the radio access network based on the inference result.

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

[0007] In response to the above technical problems, the inventors of the present invention proposed an AI system that accumulates and monitors data related to AI / ML learning and reasoning from an O-RAN base station device, detects concept drift and performs re-learning, and applied for a patent (Patent Document 1).

[0008] Figure 5 This is a functional block diagram showing a conventional structure of an artificial intelligence 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 reasoning unit 13, and accumulates the collected data in the data accumulation unit 14. The collected data accumulated in the data accumulation 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 model generated in the past by the AI / ML learning unit 12. The AI / ML reasoning unit 13 performs reasoning based on the newly collected data collected by the data collection unit 11 and the learning model, and outputs the reasoning result to the control unit 17 and the reasoning performance measurement unit 18. The control unit 17 controls the O-RAN base station device 10 based on the reasoning result.

[0011] The inference performance measuring 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 inference performance data indicating the determined inference performance in the AI / ML database 15 .

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

[0013] If the AI / ML learning unit 12 is instructed to relearn, it generates a new learning model based on the newly collected data 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 being used by the AI / ML reasoning unit 13 with the new learning model, and outputs the new learning model to the AI / ML reasoning unit 13 if the reasoning performance based on the new learning model is higher than the reasoning performance based on the current learning model.

[0014] The AI / ML reasoning unit 13 then performs reasoning using the new learning model. If the reasoning performance based on the new learning model is lower than the reasoning performance based on the current learning model, the AI / ML model management unit 16 can instruct the AI / ML learning unit 12 to relearn.

[0015] Prior art literature

[0016] Patent Literature

[0017] Patent Document 1: Japanese Patent Application No. 2022-046347

[0018] Non-patent literature

[0019] Non-patent document 1: MEMorocho-Cayamcela, H.Lee and W.Lim, "Machine Learning for 5G / B5G Mobile and Wireless Communications: Potential, Limitations, and Future Directions," in IEEE Access, vol.7, pp.137184-137206, 2019.

[0020] Non-patent literature 2: J.Kaur, MAKhan, M.Iftikhar, M.Imran and Q.Emad Ul Haq, "Machine Learning Techniques for 5G and Beyond," in IEEE Access, vol.9, pp.23472-23488, 2021.

[0021] Non-patent document 3: O-RAN Alliance, "AI / ML workflow description and requirements," O-RAN.WG2.AIML-v01.03, Jul.2021.

[0022] Non-patent literature 4: Hendrycks, D. and Gimpel, K. 2017. A BASELINE FOR DETECTINGMISCLASSIFIED AND OUT-OF-DISTRIBUTION EXAMPLES IN NEURAL NETWORKS.ICLR.(2017).

[0023] Non-patent literature 5: Hendrycks, D. and Dietterich, T. 2019. DEEP ANOMALY DETECTION WITH OUTLIER EXPOSURE. ICLR. (2019).

[0024] Non-patent literature 6: Liang, S., Li, Y. and Srikant, R. 2018. ENHANCING THERELIABILITY OF OUT-OF-DISTRIBUTION IMAGE DETECTION IN NEURAL NETWORKS.ICLR.(2018).

[0025] Non-patent literature 7: Lee, K., Lee, K., Lee, H. and Shin, J. 2018. A simple unified framework for detecting out-of-distribution samples and adversarial lattacks. NeurIPS. (2018).

[0026] Non-patent literature 8: Liu, W., Wang, X., Owens, J. D. and Li, Y. 2020. Energy-based Out-of-distribution Detection. NeurIPS. (2020).

[0027] Non-patent literature 9: Huang, R., Geng, A. and Li, Y. 2021. On the Importance of Gradients for Detecting Distributional Shifts in the Wild. NeurIPS (2021). Summary of the invention

[0028] Problem that the invention aims to solve

[0029] The RAN Intelligent Controller (RIC) is responsible for controlling and optimizing RAN functions. Figure 6 As shown, the hierarchical structure includes a non-real-time system component "Non-RT (Real Time) RIC (Non-real-time Radio Access Network Intelligent Controller)" with different control cycles and a near-real-time system component "Near-RT RIC (Near Real-time Radio Access Network Intelligent Controller)".

[0030] Here, since Non-RT RIC has a control period of more than 1 second and a wide control object, while Near-RT RIC has a control period of 10 milliseconds to 1 second and a narrow control object, the optimal configuration of each functional block related to AI / ML to Near-RT RIC and Non-RT RIC has been studied.

[0031] Non-RT RIC is installed in the station building (data center), and Near-RT RIC is often configured at the edge site (the roof of the building, the equipment room of the apartment). Therefore, if the functions related to AI / ML learning are configured in Non-RT RIC, and the functions related to AI / ML reasoning are configured in Near-RT RIC, in addition to being able to create a highly versatile learning model by learning with a wide range of data as the storage range of the station building, it is possible to perform reasoning for each edge site, thereby reducing the processing load of the edge site.

[0032] However, if real-time performance is prioritized, not only the functions involved in AI / ML learning but also the functions involved in re-learning of the learning model are configured in the Near-RT RIC, which may cause the following technical problems.

[0033] First, the processing load of Near-RT RIC increases. That is, edge sites cannot be equipped with sufficient computers due to power and space limitations.

[0034] Second, only the information under the Near-RT RIC can be used for concept drift detection. That is, since the information of the neighboring regions cannot be used, the detection of concept drift is delayed.

[0035] On the other hand, if all functions related to relearning are arranged in Non-RT RIC, the following effects can be expected.

[0036] First, relearning can be implemented even if the processing load of the Near-RT RIC is reduced and the computer resources of the edge site are limited.

[0037] Second, since the information under Non-RT RIC can be used for concept drift detection, the ability to follow environmental changes can be improved.

[0038] On the other hand, in order to configure all functions involved in relearning in Non-RT RIC and realize relearning based on concept drift detection, it is necessary to send data on the reasoning performance of the learning model from Near-RT RIC to Non-RT RIC, and the use of A1 / O1 interface is considered.

[0039] However, for the O1 interface, there is no interface specification for the performance indicators of the learning model. There is a provision for the interface of the performance indicators of the base station based on the 3GPP (registered trademark) specification, but it cannot be used as is for the inference performance data of the learning model. In addition, for the A1 interface, there is no interface specification for the performance indicators of the learning model.

[0040] Means used to solve problems

[0041] The object of the present invention is to solve the above technical problems and to provide a control device for a wireless access network that configures all functions related to relearning of a learning model in a Non-RT RIC and can implement relearning based on concept drift detection.

[0042] In order to achieve the above-mentioned purpose, the present invention is a control device for a wireless access network formed by hierarchical control units of a non-real-time system and a near-real-time system, wherein the control device for the wireless access network comprises: a learning unit, which generates a learning model based on data collected from the wireless access network; an inference unit, which controls the wireless access network based on the result of inference using the collected data for the learning model; and a re-learning unit, which detects whether a concept drift has occurred based on the collected data, and if the occurrence of a concept drift is detected, causes the learning unit to re-learn the learning model, the inference unit is configured in the control unit of the near-real-time system, and the learning unit and the re-learning unit are configured in the control unit of the non-real-time system.

[0043] Effects of the Invention

[0044] According to the present invention, only the functions involved in AI / ML reasoning are configured in the control unit of the near real-time system, and the functions involved in AI / ML learning and the functions involved in relearning are configured in the control unit of the non-real-time system, thereby achieving the following effects.

[0045] (1) Functions involved in AI / ML learning can be learned using a wide range of data within the capacity of the station building, so a highly versatile learning model can be created.

[0046] (2) Since AI / ML inference can be performed at each edge site, the processing load of the edge site can be reduced.

[0047] (3) Since all functions involved in relearning are configured in the control unit of the non-real-time system, the processing load of the control unit of the near-real-time system is reduced, and relearning can be implemented even if the computer resources of the edge site are limited.

[0048] (4) Information from subordinates of the control unit of the non-real-time system can be used for concept drift detection, thereby improving the ability to follow environmental changes.

[0049] Other features and advantages of the present invention will become apparent from the following description with reference to the accompanying drawings. In addition, in the accompanying drawings, the same reference numerals are used for the same or similar structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are incorporated in and constitute a part of the specification, illustrate embodiments of the present invention, and together with the description, serve to explain the principles of the present invention.

[0051] Figure 1 This is a functional block diagram showing the configuration of main parts of an O-RAN control device according to an embodiment of the present invention.

[0052] Figure 2 is a diagram showing how to specify metrics when requesting inference performance data.

[0053] Figure 3 This is a diagram showing an example of transmitting the inference performance data in a table format.

[0054] Figure 4 It is a time sequence flow showing the operation of the present invention.

[0055] Figure 5 This is a functional block diagram showing a conventional structure of an artificial intelligence system that detects concept drift and performs relearning.

[0056] Figure 6 It is a functional block diagram of a radio access network intelligent controller (RIC). DETAILED DESCRIPTION

[0057] Hereinafter, the embodiments are described in detail with reference to the accompanying drawings. In addition, the following embodiments do not limit the invention involved in the technical solution. In addition, the combination of features described in the embodiments may not all be necessary for the invention. Two or more of the multiple features described in the embodiments may also be combined arbitrarily. In addition, the same or similar configurations are marked with the same reference numerals, and repeated descriptions are omitted.

[0058] Figure 1 This is a functional block diagram showing the configuration of the main parts of the O-RAN control device involved in one embodiment of the present invention, and the illustration of the configuration that is not necessary for the description of the present invention is omitted. In addition, the same reference numerals as those described above represent the same or equivalent parts. The feature of this embodiment is that the functions involved in AI / ML reasoning are configured in Near-RTRIC, and the functions involved in AI / ML learning and re-learning of the learning model are configured in the point of Non-RT RIC.

[0059] The O-RAN control device is composed of O-CU / O-DU, Near-RT RIC and Non-RT RIC, and each function can communicate with each other via various interfaces including O1 interface, A1 interface and E2 interface specified by the O-RAN alliance. The O-RAN base station device 10 is configured in the O-CU / O-DU.

[0060] The Near-RT RIC is equipped with an AI / ML reasoning unit 13, a control unit 17, and an inference performance measurement unit 18 as main functions related to AI / ML reasoning. The Non-RT RIC is equipped with a data collection unit 11, an AI / ML learning unit 12, and an AI / ML model management unit 16 as main functions related to AI / ML learning, and a data accumulation unit 14, an AI / ML database 15, a concept drift detection unit 19, and a relearning control unit 20 as main functions related to relearning.

[0061] In the Non-RT RIC, the data receiving 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 accumulation unit 14 and the AI / ML learning unit 12. The latest data is managed by 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 reasoning unit 13 of the Near-RT RIC via the A1 interface.

[0062] In the Near-RT RIC, the AI / ML reasoning unit 13 provides the control unit 17 and the reasoning performance measurement unit 18 with the reasoning results obtained by using the latest data for the learning model obtained via the A1 interface. The reasoning performance measurement unit 18 sends the reasoning 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.

[0063] In the Non-RT RIC, when the concept drift detection unit 19 detects concept drift, the relearning control unit 20 instructs the AI / ML learning unit 12 to relearn the learning model. The relearned learning model is updated and registered in the AI / ML management unit 16 and provided to the AI / ML reasoning unit 13 of the Near-RT RIC via the A1 interface.

[0064] In this way, because in this embodiment, the functions involved in AI / ML learning, AI / ML reasoning and re-learning of the learning model are distributed in the Near-RT RIC and the Non-RT RIC, in order to send and receive information related to re-learning between the Near-RT RIC and the Non-RTRIC, the following two messages (a) and (b) are added.

[0065] (a) Request for inference performance data

[0066] (b) Sending of requested inference performance data

[0067] Here, in the request for (a) inference performance data, the following three pieces of information (1) to (3) are specified.

[0068] (1) Inference performance data

[0069] In O-RAN WG2, Figure 2 The indicators exemplified in are listed as indicators related to the reasoning performance of AI / ML, and at least one indicator is specified. In the indicators, indicators suitable for binary classification problems (Binary classification problems, binary classification problems), indicators suitable for multi-class classification problems (Multiclass classification problems, multi-class classification problems) and indicators suitable for regression problems (Regression classification problems, regression classification problems) are respectively specified. In this embodiment, the average value and median value of the specified period can be obtained.

[0070] (2) Object

[0071] Specifies the learning model for the object that gets inference performance data.

[0072] (3) Data acquisition interval

[0073] Specifies the interval for obtaining inference performance data.

[0074] However, the Figure 2 The indices shown in the examples are all indices using correct answer data (true value), but there are cases where correct answer data cannot be obtained from the system. Therefore, in the present embodiment, as described in detail below, inference performance indices for cases where correct answer data cannot be obtained are also included and can be selected.

[0075] Generally, the performance of a model decreases for out-of-distribution (OOD) data. Therefore, an out-of-distribution score based on a learned model is supplemented as an inference performance indicator when correct answer data is not obtained.

[0076] If it is suitable for a multi-class classification problem, as the maximum value when the output of the learned model is normalized in the hierarchical direction through a normalized exponential function (Softmax), the maximum normalized exponential function probability (Maximum over softmax probabilities, MSP) disclosed in non-patent document 4, the outlier exposure (OutlierExposure) disclosed in non-patent document 5, or the ODIN disclosed in non-patent document 6 can be specified.

[0077] Alternatively, as in Mahalanobis disclosed in non-patent document 7, a Gaussian distribution of feature quantities calculated for each level using the penultimate feature quantity and label of the learned model may be specified, and a value obtained by adding a negative sign to the Mahalanobis distance of the level where the Mahalanobis distance is the smallest.

[0078] Furthermore, as with the energy (Energy) disclosed in Non-Patent Document 8, it is also possible to specify a value obtained by adding a negative sign to the free energy when the energy function is a value obtained by adding a negative sign to the output of the learned model.

[0079] Furthermore, it is also possible to specify the norm of the angle when backpropagating (backprop) the KLD distributed in the same way as the normalized exponential function output of the learned model, as in the gradient normalization (GradNorm) disclosed in non-patent document 9.

[0080] On the other hand, if it is suitable for binary classification problems, it is possible to specify metrics suitable for multi-class classification problems.

[0081] Regarding the transmission of the inference performance data (b), in this embodiment Figure 3 As an example, inference performance data (p1 to pm) related to m indicators specified in the request for information on the AI / ML database (1) is sent in a table format at a specified time interval for each learning model (No.).

[0082] Figure 4 This is a timing flow of the actions of this embodiment, and here, the communication between O-CU / O-DU, Near-RT RIC and Non-RT RIC is explained. In this embodiment, the communication between O-CU / O-DU and Near-RT RIC is performed via the E2 interface, and the communication between Near-RT RIC and Non-RT RIC is performed via the O1 interface or the A1 interface.

[0083] The O-CU / O-DU repeatedly sends 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 sends the latest data to the Near-RT RIC via the E2 interface and to the Non-RT RIC via the O1 interface. In the Non-RT RIC, the latest data is obtained by the data collection unit 11.

[0084] In the Near-RT RIC, the AI / ML inference unit 13 applies the latest data to the current learning model to perform 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 to perform control based on the inference result via the E2 interface at time t2.

[0085] The Non-RT RIC requests the reasoning performance data from the Near-RT RIC at a predetermined period. In this embodiment, at time t3, if the AI / ML model management unit 16 of the Non-RT RIC requests the reasoning performance data from the Near-RT RIC via the O1 interface, then in the Near-RT RIC, at time t4, the reasoning performance measurement unit 18 responds to the request and sends the measurement result of the reasoning performance data to the Non-RT RIC via the O1 interface in the form of the table.

[0086] In Non-RT RIC, the concept drift detection unit 19 measures the concept drift based on the inference performance data and the data accumulated in the AI / ML database 15. If the concept drift is detected at time t5, the relearning control unit 20 instructs the AI / ML learning unit 12 to relearn at time t6. The AI / ML learning unit 12 performs relearning to generate a learning model, and updates and registers the learning model in the AI / ML model management unit 16.

[0087] At time t7, the relearned learning model is sent from the AI / ML model management unit 16 of the Non-RT RIC to the AI / ML reasoning unit 13 of the Near-RT RIC via the A1 interface. Therefore, control based on the relearned learning model is performed every time the latest data is collected.

[0088] According to this embodiment, since the functions involved in AI / ML reasoning are configured in the Near-RT RIC, on the other hand, the functions involved in AI / ML learning and re-learning are configured in the Non-RT RIC, the processing load of the Near-RT RIC can be reduced. Therefore, if the computer resources of the edge site are limited and concept drift does not occur frequently, the concept drift can be detected in real time, and the tracking performance with respect to environmental changes can be improved.

[0089] As a result, depending on the implementation method, it will be possible to contribute to Goal 9 of the Sustainable Development Goals (SDGs) led by the United Nations, "Build resilient infrastructure and promote integrated sustainable industrialization" and Goal 11, "Make cities inclusive, safe, resilient, and sustainable."

[0090] The present invention is not limited to the above-described embodiments, and various changes and modifications can be made without departing from the spirit and scope of the present invention. Therefore, in order to disclose the scope of the present invention, the following claims are attached.

[0091] This application claims priority based on Japanese patent application No. 2022-153230 filed on September 27, 2022, and all the contents described are incorporated herein by reference.

Claims

1. A control device for a wireless access network, wherein a control unit of a non-real-time system and a control unit of a near-real-time system are hierarchically connected, wherein: The control device of the wireless access network comprises: a learning unit that generates a learning model based on data collected from the wireless access network; an inference unit configured to control a wireless access network based on an inference result obtained by applying the collected data to the learning model; as well as a re-learning unit that detects whether a concept drift has occurred based on the collected data, and if the occurrence of a concept drift is detected, causes the learning unit to re-learn the learning model, The inference unit is configured in a control unit of the near real-time system, and the learning unit and the re-learning unit are configured in a control unit of the non-real-time system.

2. The control device for a wireless access network according to claim 1, wherein: The inference unit includes an inference performance measuring unit that measures inference performance based on the collected data and the inference result. The relearning unit includes: a concept drift detection mechanism that detects the occurrence of concept drift based on the inference performance; and a relearning control mechanism that causes the learning unit to relearn the learning model when the occurrence of the concept drift is detected. The control unit of the non-real-time system requests the inference performance data from the control unit of the near-real-time system via the O1 interface. The control unit of the near real-time system responds to the request by sending the inference performance data to the control unit of the non-real-time system via the O1 interface. The control unit of the non-real-time system transmits the relearned learning model to the control unit of the near-real-time system via one of the O1 interface and the A1 interface.

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

4. The control device for a wireless access network according to claim 3, wherein: The control unit of the near real-time system transmits the requested inference performance data in a table format in which data of each of the designated indicators is described for each of the designated learning models.

5. The control device for a wireless access network according to claim 3 or 4, wherein: The inference performance metric includes an out-of-distribution score premised on the learned model.

Images