Ai model monitoring method, terminal device, network side network element, and product

Abstract

The present disclosure provides an AI model monitoring method, a terminal device, a network side network element, and a product. The method comprises: receiving a first parameter and a second parameter configured by the network side network element by means of RRC signaling; and on the basis of the first parameter and the second parameter, controlling the on and off of a first timer and a second timer. According to the present disclosure, the performance of an AI model is evaluated on the basis of timers, so that a terminal device triggers model updating / model switching / rollback to traditional beam management in a timely manner, thereby ensuring communication quality.

Description

An AI model monitoring method, terminal equipment, network side network element and product

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This disclosure is based on and claims the priority of Chinese patent application with application number 202410183591.4 and application date February 19, 2024. The entire content of the Chinese patent application is hereby incorporated into this disclosure by reference. Technical Field

[0003] The present disclosure relates to the technical field of wireless interface physical layer design, and more specifically, to an AI model monitoring method, terminal equipment, network-side network elements, and products. Background Art

[0004] The communication version protocol R18 has launched a project for wireless AI. A typical use case of AI beam management is spatial beam prediction. AI-based spatial beam prediction collects data to monitor the performance of the current AI model and designs AI model monitoring. However, the existing protocol does not yet involve timer-based monitoring of AI models. Summary of the Invention

[0005] In order to solve the above problems, the present disclosure proposes an AI model monitoring method, terminal equipment, network-side network element and product. This solution monitors the performance of the AI ​​model based on a timer.

[0006] The present disclosure provides an AI model monitoring method, the method comprising:

[0007] Receiving a first parameter and a second parameter configured by a network-side network element through RRC signaling; and

[0008] The first timer and the second timer are controlled to be turned on and off according to the first parameter and the second parameter.

[0009] In some embodiments, the first parameter includes at least one of the following: duration of the first timer, start conditions and restart conditions of the first timer, and performance judgment conditions of the first AI model;

[0010] The second parameter includes at least one of the following: the duration of the second timer, the time required to start KPI monitoring within the second timer, the sample size and the number of monitoring moments, and the performance judgment condition of the second AI model.

[0011] In some embodiments, the start condition of the first timer or the second timer includes at least one of the following:

[0012] The reference signal received power or block error rate is less than a preset first threshold for N1 consecutive times within the T1 duration;

[0013] The reference signal received power or block error rate is less than a preset second threshold for a cumulative N2 times within the T2 duration;

[0014] The reference signal received power or block error rate is less than a preset third threshold for N3 consecutive times;

[0015] Wherein, N1, N2 and N3 are positive integers.

[0016] In some embodiments, the performance decision condition of the first AI model is that the first monitoring KPI is less than a preset fourth threshold, or the performance decision condition of the second AI model is that the second monitoring KPI is less than a preset fifth threshold.

[0017] Furthermore, the first monitoring KPI or the second monitoring KPI is at least one of beam prediction accuracy and reference signal received power.

[0018] In some embodiments, controlling the on and off of the first timer and the second timer according to the first parameter and the second parameter includes:

[0019] When it is detected that a start condition of the first timer in the first parameter is met, start the first timer; and

[0020] When it is detected that the performance judgment condition of the first AI model in the first parameter is met, the first timer is turned off and the second timer is turned on according to a pre-configured rule.

[0021] In some embodiments, controlling the on and off of the first timer and the second timer according to the first parameter and the second parameter includes:

[0022] When it is detected that the start condition of the first timer in the first parameter is met, the first timer and the second timer are started at the same time.

[0023] In some embodiments, starting the second timer according to a preconfigured rule includes:

[0024] Starting a second timer and using the second timer to monitor all second AI models;

[0025] When the second timer meets the time or sample size or number of monitoring moments required to start KPI monitoring, start calculating the monitoring KPI of the second AI model; and

[0026] When a second AI model meets the performance judgment condition of the second AI model in the second parameter, or the second timer expires, the second timer is turned off.

[0027] In some embodiments, starting the second timer according to a preconfigured rule includes:

[0028] Serially start multiple second timers, each second timer monitoring a second AI model;

[0029] When the second timer meets the time or sample size or number of monitoring moments required to start KPI monitoring, start calculating the monitoring KPI of the second AI model; and

[0030] When a second AI model meets the performance judgment condition of the second AI model in the second parameter, the second timer corresponding to the second AI model is turned off.

[0031] In some embodiments, the method further comprises:

[0032] When at least one of the first timer and the second timer is closed, the reporting model switches or falls back to traditional beam management.

[0033] Furthermore, the reporting model switching or falling back to traditional beam management includes any one of the following:

[0034] When the first AI model satisfies the performance judgment condition of the first AI model in the first parameter, and each second AI model satisfies the performance judgment condition of the second AI model in the second parameter, reporting to the network side network element to fall back to the default AI model or traditional beam management; or

[0035] When the first AI model satisfies the performance judgment condition of the first AI model in the first parameter, and there is at least one candidate second AI model that does not satisfy the performance judgment condition of the second AI model in the second parameter, reporting model switching information, at least one of the ID of the candidate second AI model and the monitoring KPI to the network side network element, and receiving an AI model switching instruction issued by the network side network element; or

[0036] When the first AI model does not meet the performance judgment condition of the first AI model in the first parameter, and there is at least one candidate second AI model that does not meet the performance judgment condition of the second AI model in the second parameter and whose monitoring KPI is better than the first AI model, the ID of the candidate second AI model with the best monitoring KPI is reported to the network side network element.

[0037] The present disclosure also provides an AI model monitoring method, which is applied to a network-side network element. The method includes:

[0038] The first parameter and the second parameter are configured through RRC signaling, and the first parameter and the second parameter are sent to the terminal, so that the terminal controls the opening and closing of the first timer and the second timer according to the first parameter and the second parameter.

[0039] In some embodiments, the first parameter includes at least one of the following: duration of the first timer, start conditions and restart conditions of the first timer, and performance judgment conditions of the first AI model;

[0040] The second parameter includes at least one of the following: the duration of the second timer, the time required to start KPI monitoring within the second timer, the sample size, the number of monitoring moments, and the performance judgment conditions of the second AI model.

[0041] Furthermore, a start condition of the first timer or the second timer includes at least one of the following:

[0042] The reference signal received power or block error rate is less than a preset first threshold for N1 consecutive times within the T1 duration;

[0043] The reference signal received power or block error rate is less than a preset second threshold for a cumulative N2 times within the T2 duration;

[0044] The reference signal received power or block error rate is less than a preset third threshold for N3 consecutive times;

[0045] Wherein, N1, N2 and N3 are positive integers.

[0046] In some embodiments, the performance decision condition of the first AI model is that the first monitoring KPI is less than a preset fourth threshold, or the performance decision condition of the second AI model is that the second monitoring KPI is less than a preset fifth threshold.

[0047] Furthermore, the first monitoring KPI or the second monitoring KPI is at least one of beam prediction accuracy and reference signal received power.

[0048] In some embodiments, the method further comprises:

[0049] receiving a reporting message indicating that the terminal has fallen back to a default AI model or traditional beam management; and

[0050] receiving the model switching information reported by the terminal, selecting at least one of the ID of the candidate second AI model and the monitoring KPI, and issuing an AI model switching instruction to the terminal;

[0051] Receive the ID of the candidate second AI model with the best monitoring KPI reported by the terminal.

[0052] An embodiment of the present disclosure also provides a terminal device, which is used to execute the AI ​​model monitoring method as described in any of the above embodiments.

[0053] The embodiments of the present disclosure also provide a network-side network element, which is used to execute the AI ​​model monitoring method as described in any of the above embodiments.

[0054] The embodiments of the present disclosure further provide a computer program product, comprising a computer program / instruction, which implements the steps of any of the methods described in the above embodiments when executed by a processor.

[0055] Compared to existing technologies, the present disclosure provides an AI model monitoring method, terminal device, network-side network element, and product. These methods receive first and second parameters configured by the network-side network element via RRC signaling; and control the activation and deactivation of first and second timers based on the first and second parameters. This disclosure implements timer-based evaluation of AI model performance, enabling terminal devices to promptly trigger model updates, model switching, or fallback to traditional beam management, ensuring communication quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] FIG1 is a flow chart of an AI model monitoring method provided by an embodiment of the present disclosure;

[0057] FIG2 is another flow chart of an AI model monitoring method provided by an embodiment of the present disclosure;

[0058] FIG3 is another flowchart of an AI model monitoring method provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0059] The following will clearly and completely describe the technical solutions in the embodiments of the present disclosure in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present disclosure without making any creative efforts shall fall within the scope of protection of the present disclosure.

[0060] 1 , which is a flow chart of an AI model monitoring method provided by an embodiment of the present disclosure, the method includes steps S1 to S2;

[0061] S1, receiving a first parameter and a second parameter configured by a network side network element through RRC signaling;

[0062] S2, controlling the opening and closing of the first timer and the second timer according to the first parameter and the second parameter.

[0063] When implementing this embodiment, the method is executed by a terminal device, and the terminal device uses an AI model monitoring method to monitor the first AI model and the second AI model through a timer.

[0064] Timers are used to determine duration, when a triggering event should start, when a triggering event should stop, and what action to take when the timer expires.

[0065] The AI ​​model is actually a neural network. The final neural network structure and parameters can only be determined after training with a data set. Different data sets correspond to different network structures and parameters. The training methods for the first and second AI models are as follows:

[0066] In the AI-based spatial beam prediction implementation method, training is first performed based on a large dataset, such as a dataset containing millions of samples, each containing the RSRP values ​​of all beams. During training, the RSRP values ​​of a small number of beams are used as input to the AI ​​model (usually a neural network), and the RSRP values ​​of all beams are used as labels for the AI ​​model's output. The AI ​​model's parameters are continuously updated using a gradient descent algorithm until the error between the AI ​​model output and the label is measured using the Normalized Mean Square Error (NMSE). For example, when the NMSE is less than a threshold, it indicates that the model has converged and has good predictive capabilities. In actual use, the RSRP results of the configuration measurement are used as input to the AI ​​model, and the predicted RSRP values ​​for all beams are output. Based on the predicted RSRP values ​​of the top K optimal beams and the beam load, the base station determines the beam to be used for terminal data transmission and provides beam instructions.

[0067] During specific execution, the base station serving as a network-side network element sends RRC signaling to the terminal device, and the RRC signaling is configured with the first parameter of the first AI model timer and the second parameter of the second AI model timer.

[0068] The terminal device receives a first parameter of the first AI model and a second parameter of the second AI model, and controls the on / off timer of the first AI model and the on / off timer of the second AI model based on the first parameter and the second parameter, so as to evaluate the performance of the first AI model and the second AI model respectively through the timers.

[0069] This application uses a timer to evaluate the performance of multiple first and second AI models in a terminal device, allowing the terminal device to promptly trigger model updates, model switching, or fallback to traditional beam management based on the detected performance. This solution is applicable not only to monitoring different models, but also to monitoring different functions. The second AI model timer configuration may be the same as the first AI model timer configuration, but it is applied to monitoring different models.

[0070] In another embodiment provided in the present disclosure, the first parameter configured by the base station through RRC signaling includes at least one of the following: the duration of the first timer, the start condition and restart condition of the first timer, and the performance judgment condition of the first AI model.

[0071] The second parameter configured by the base station through RRC signaling includes at least one of the following: the duration of the timer of the second AI model, the time, sample size, and number of monitoring moments required to start KPI monitoring within the timer of the second AI model, and the performance judgment condition of the second AI model.

[0072] It should be noted that the configurations of the first AI model timer and the second AI model timer can be different. If different second AI models correspond to a timer respectively, the above configurations of each timer can be different.

[0073] The time / sample size / number of monitoring moments required to start KPI monitoring within the timer of the second AI model indicates that periodic monitoring is performed while the timer is turned on, and there is one monitoring moment in each period.

[0074] By configuring the model's timer duration, timer start conditions, restart conditions, and model performance judgment conditions, you can accurately control the start and stop of the timer of the first AI model and the timer of the second AI model.

[0075] In another embodiment provided by the present disclosure, the start conditions of the first timer or the second timer are specifically one, two or three of the following three items:

[0076] The reference signal received power or block error rate is less than a preset first threshold for a cumulative N1 times within the T1 duration;

[0077] The reference signal received power or block error rate is less than a preset second threshold for a cumulative N2 times within the T2 duration;

[0078] The reference signal received power or block error rate is less than a preset third threshold for N3 consecutive times;

[0079] Wherein, N1, N2 and N3 are positive integers.

[0080] In a specific embodiment of the present disclosure, when the activation condition includes only one item, the activation condition includes the following situations:

[0081] When the reference signal received power is less than X1 or the block error rate is less than Y1 during N1 consecutive detections within the preset T1 duration, the first timer or the second timer starts.

[0082] When the reference signal received power is less than X1 or the block error rate is less than Y1 during N2 detections accumulated within the preset T2 duration, the first timer or the second timer starts.

[0083] When the reference signal received power is less than X1 or the block error rate is less than Y1 during N3 consecutive detections, the first timer or the second timer is started.

[0084] In another specific embodiment of the present disclosure, when the activation conditions include only two items, the activation conditions include the following situations:

[0085] When the reference signal receiving power is less than X1 or the block error rate is less than Y1 during N1 consecutive detections within the preset T1 duration, and the reference signal receiving power is less than X1 or the block error rate is less than Y1 for N2 cumulative times within the preset T2 duration, the first timer or the second timer starts.

[0086] When the reference signal receiving power is less than X1, or the block error rate is less than Y1 during N2 detections within the preset T2 duration, and the reference signal receiving power is less than X1, or the block error rate is less than Y1 during N3 consecutive detections, the first timer or the second timer starts.

[0087] When the reference signal receiving power is less than X1, or the block error rate is less than Y1, during N1 consecutive detections within the preset T1 duration, and the reference signal receiving power is less than X1, or the block error rate is less than Y1 for N3 consecutive times, the first timer or the second timer starts.

[0088] In another specific embodiment of the present disclosure, when the activation condition includes three items, the activation condition is as follows:

[0089] When the reference signal received power is less than X1, or the block error rate is less than Y1, during N1 consecutive detections within the preset T1 duration, and the reference signal received power is less than X1, or the block error rate is less than Y1, during N2 cumulative detections within the preset T2 duration, and the reference signal received power is less than X1, or the block error rate is less than Y1, during N3 consecutive detections, the first timer or the second timer starts.

[0090] Whether the timer is started is determined by detecting the reference signal received power or the block error rate.

[0091] In another embodiment provided by the present disclosure, the first monitoring KPI being less than a preset fourth threshold is used as a performance judgment condition for the first AI model, or the second monitoring KPI being less than a preset fifth threshold is used as a performance judgment condition for the second AI model.

[0092] Specifically, see Figure 2, which is another flow chart of the AI ​​model monitoring method provided by an embodiment of the present disclosure.

[0093] The terminal device has four models: Model ID 1 is the first AI model, and Model IDs 2, 3, and 4 are the second AI models. The terminal device infers and monitors the performance of the first AI model. The duration of the timer for the first AI model is equal to that of the timer for the second AI model, both 10 seconds.

[0094] The start condition / restart condition of the first AI model timer is that the top1 beam prediction accuracy is less than the threshold value X1=85% for N1=4 consecutive times.

[0095] The first AI model performance judgment condition is that the top1 beam prediction accuracy rate is less than the threshold value X2=80% for N2=4 consecutive times.

[0096] The sample size required to start KPI monitoring within the second AI model timer = 50.

[0097] The second AI model performance judgment condition is that the top1 beam prediction accuracy is greater than the threshold value X3=90% for N3=4 consecutive times.

[0098] In another embodiment provided by the present disclosure, the first monitoring KPI and the second monitoring KPI are beam prediction accuracy and reference signal received power, respectively.

[0099] The beam prediction accuracy and the reference signal received power are used as the first monitoring KPI and the second monitoring KPI, respectively, to monitor the first AI model and the second AI model.

[0100] In another embodiment provided by the present disclosure, considering the processing / storage capabilities of the terminal device, a terminal device with low processing capabilities needs to serially start multiple model timers. When the timers are serially started, step S2 specifically includes:

[0101] The terminal device monitors the performance of the first AI model, and when a condition for triggering the start of the first AI model timer is detected, controls the first AI model timer to start.

[0102] Because terminal devices with low processing power need to serially start multiple model timers, after starting the first AI model timer, when starting the timers of other models, it is necessary to monitor the first AI model that is started;

[0103] When the first AI model performance judgment condition is met, indicating that the first AI model has poor performance, the first AI model timer is turned off and a second AI model timer is serially started according to pre-configured rules.

[0104] While the second AI model timer is on, multiple second AI models are monitored simultaneously through the timer of the second AI model. When the performance of a second AI model meets the judgment conditions, it can be considered that the model performance is better and can be switched to the second AI model.

[0105] By turning on the model timer in a serial switching manner and then turning on the timer of the second AI model after the first AI model timer is turned off, the needs of low-processing power terminal devices can be met. By selecting the model timer with the best performance, the performance of the timer can be maximized, thereby improving the monitoring capability.

[0106] In another embodiment provided by the present disclosure, considering the processing / storage capabilities of the terminal device, a high-processing-capacity terminal device needs to start multiple model timers in parallel. When starting the timers in parallel, step S2 specifically includes:

[0107] The terminal device monitors the performance of the first AI model. When a condition for triggering the start of the first AI model timer is detected, the first AI model timer is started and the second AI model timer is started at the same time.

[0108] The timers of the models are turned on by parallel switching. According to the needs of high-processing terminal devices, the timers of all models are turned on in parallel at the same time to ensure the optimal performance of the timers and improve the monitoring capabilities.

[0109] In another embodiment provided by the present disclosure, when the terminal detects that the performance judgment condition of the first AI model is still not met when the first AI model timer expires and is closed, there are two processing methods.

[0110] The first processing method is that when the performance judgment condition of the first AI model is still not met when the first AI model timer expires and closes, it means that the performance of the first AI model is good. The terminal can not report and still use the first AI model, and the second AI model timer is closed; or

[0111] The second processing method is to start a second timer when the performance judgment condition of the first AI model is still not met when the first AI model timer expires and closes.

[0112] By monitoring the first AI model, when system performance fluctuates, the model can be switched autonomously, thus realizing intelligent AI model monitoring.

[0113] In another embodiment provided by the present disclosure, the configuration rules for serially starting the second AI model timer specifically include:

[0114] After the first AI model timer is closed, a second AI model timer is started.

[0115] The timer started by the second AI model monitors multiple second AI models simultaneously during the start-up period, and starts calculating the monitoring KPI of the second AI model after the time / sample size / monitoring time required for KPI monitoring is met.

[0116] When a second AI model that calculates a KPI meets a performance judgment condition, or a timer expires, the timer of the enabled second AI model is turned off.

[0117] Among them, when a second AI model for calculating a KPI meets the performance judgment condition of the second AI model, it is considered that the model has good performance and can be used as a candidate model for model switching.

[0118] Referring to Figure 2, when the top1 beam prediction accuracy of the first AI model is 75% for N2=4 consecutive times, which is less than the threshold value X2=80%, it indicates that the performance of the first AI model is poor, the timer of the first AI model is turned off, and the second timer 1 of the first and second AI models is serially turned on.

[0119] The second timer 1 is turned on to monitor the second AI model with ID = 2. After the second timer 1 is turned on, the KPI calculation starts after 50 samples are met. During the period when the second timer 1 is turned on, when the top1 beam prediction accuracy of N3 = 4 consecutive times is 92% and is greater than the threshold X3 = 90%, the second timer 2 with ID = 3 is considered to be a candidate model for model switching, the first timer 1 is turned off, and the second timer 2 is turned on.

[0120] The second timer 2 is turned on to monitor the second AI model with ID=3. After the second timer 2 is turned on, the KPI calculation starts after 50 samples are met. During the period when the second timer 2 is turned on, the top1 beam prediction accuracy of 88% is less than the threshold X3=90%. It is considered that the performance of the second timer 2 with ID=3 is poor, and the second timer 2 is turned off after expiration, and the third timer 3 is turned on.

[0121] The second timer 3 is turned on to monitor the second AI model with ID=4. After the second timer 3 is turned on, the KPI calculation starts after 50 samples are met. During the period when the second timer 3 is turned on, when the top1 beam prediction accuracy of N3=4 consecutive times is 95% and is greater than the threshold X3=90%, the second timer 3 with ID=4 is considered to be a candidate model for model switching, and the second timer 3 is turned off.

[0122] In addition, the terminal device reports the ID of the candidate second AI model and its monitoring KPI to the base station as the network side network element, and the base station instructs to switch to the model with higher prediction accuracy.

[0123] The present disclosure can monitor the first AI model and multiple second AI models based on a timer, so as to promptly detect changes in model performance, promptly trigger model update / model switching / fallback to traditional beam management, and ensure communication quality.

[0124] In another embodiment provided by the present disclosure, as an alternative method to the above embodiment, the configuration rules for serially starting the second AI model timer specifically include:

[0125] The terminal device monitors the performance of the first AI model. When the condition for triggering the start of the first AI model timer is detected, the first timer of the first AI model is started, and the second timer of the second AI model is also started. Each second timer monitors a second AI model.

[0126] During the period when the first timer is turned on, multiple second AI models are monitored simultaneously. After the time / sample size / number of monitoring moments required for KPI monitoring are met, the monitoring KPI of the second AI model is calculated.

[0127] The second timer closing condition is that the performance of a second AI model meets the judgment condition.

[0128] In addition, when the first AI model performance judgment condition is met, it indicates that the performance of the first AI model is poor, and the first timer of the first AI model is turned off.

[0129] When the performance of a second AI model meets the judgment conditions, the model is considered to have better performance and can be used as a candidate model for model switching.

[0130] Specifically, refer to Figure 3, which is another flow chart of the AI ​​model monitoring method provided by an embodiment of the present disclosure.

[0131] The terminal device has four models: model ID 1 is the first AI model, and model IDs 2, 3, and 4 are the second AI models. The terminal device monitors the performance of the first AI model and starts the first and second timers when it detects that the top 1 beam prediction accuracy of 82% is less than the threshold X1 = 85% for N1 = 4 consecutive times.

[0132] When the top1 beam prediction accuracy of the first AI model for N2=4 consecutive times is 75% and is less than the threshold value X2=80%, it indicates that the performance of the first AI model is poor, and the first timer of the first AI model is closed.

[0133] After the second AI model timer is turned on and 50 samples have been met, the KPI calculation for models ID = 2, 3, and 4 begins. Model ID = 2 has a top 1 beam prediction accuracy of 92% for N3 = 4 consecutive times, which is greater than the threshold X3 = 90%. Model ID = 2 is considered a candidate for model switching. Model ID = 3 has a top 1 beam prediction accuracy of less than the threshold X3 = 90% during the timer period. Model ID = 3 is considered to have poor performance. Model ID = 4 has a top 1 beam prediction accuracy of 95% for N3 = 4 consecutive times, which is greater than the threshold X3 = 90%. Model ID = 4 is considered a candidate for model switching.

[0134] The terminal reports the candidate second AI model ID and its monitoring KPI, and the base station instructs to switch to the model with higher prediction accuracy.

[0135] In another embodiment provided by the present disclosure, after the first timer and / or the second timer is closed, the monitoring KPI is monitored and the model is switched or returned to the traditional beam management according to the monitoring KPI control.

[0136] The switchable AI model is determined based on the size of the monitoring KPI. When the monitoring KPIs of all AI models do not meet the conditions, it falls back to traditional beam management.

[0137] By monitoring the AI ​​model, when system performance fluctuates, the model can be switched or rolled back autonomously, thus realizing intelligent AI model monitoring.

[0138] In another embodiment provided by the present disclosure, switching or falling back to traditional beam management according to the monitoring KPI reporting model includes:

[0139] When the first AI model meets the performance judgment condition of the first AI model in the first parameter, and each second AI model meets the performance judgment condition of the second AI model in the second parameter, that is, when there is no candidate model, the terminal reports and falls back to traditional beam management.

[0140] When the first AI model satisfies the performance judgment condition of the first AI model in the first parameter, and there is at least one candidate second AI model that does not satisfy the performance judgment condition of the second AI model in the second parameter, that is, when there is a candidate model that can be used as a model switch. The terminal reports at least one of the candidate second AI model ID and the monitoring KPI to the base station as a network-side network element, so that the base station selects one of the models according to the monitoring KPI and instructs switching to the model; or,

[0141] When the first AI model does not meet the performance judgment condition of the first AI model in the first parameter, and there is at least one candidate second AI model that does not meet the performance judgment condition of the second AI model in the second parameter and whose monitoring KPI is better than the first AI model, the terminal device selects the candidate model with the best KPI based on the monitoring KPI, and reports the candidate model with the best KPI to the base station, instructing the base station to switch to the candidate model with the best KPI.

[0142] The terminal device reports to the network side network element, which determines the switching model. The terminal device determines the switching model by itself as a bottom line measure to ensure the stability of the model switching.

[0143] The present disclosure also provides an AI model monitoring method, which is applied to a network-side network element. The method includes:

[0144] The first parameter and the second parameter are configured through RRC signaling, and the first parameter and the second parameter are sent to the terminal, so that the terminal controls the opening and closing of the first timer and the second timer according to the first parameter and the second parameter.

[0145] The method is performed by a network side network element, which configures the first parameter and the second parameter through RRC signaling and sends them to the terminal so that the terminal controls the opening and closing of the first timer and the second timer according to the first parameter and the second parameter.

[0146] Timers are used to determine the duration of a timer, when a triggering event should start, when the event should stop, and what action to take when the timer expires.

[0147] During specific execution, the base station serving as a network-side network element sends RRC signaling to the terminal device, and the RRC signaling is configured with the first parameter of the first AI model timer and the second parameter of the second AI model timer.

[0148] The terminal device receives a first parameter of the first AI model and a second parameter of the second AI model, and controls the start and stop of the timer of the first AI model and the start and stop of the timer of the second AI model based on the first parameter and the second parameter, so as to evaluate the performance of the first AI model and the second AI model respectively through the timers.

[0149] This disclosure uses a timer to evaluate the performance of multiple first and second AI models in a terminal device, allowing the terminal device to promptly trigger model updates, model switching, or fallback to traditional beam management based on the detected performance. This solution is applicable not only to monitoring different models, but also to monitoring different functions. The second AI model timer configuration may be the same as the first AI model timer configuration, but is applied to monitoring different models.

[0150] In another embodiment provided in the present disclosure, the first parameter configured by the base station through RRC signaling includes at least one of the following: the duration of the first timer, the start condition and restart condition of the first timer, and the performance judgment condition of the first AI model.

[0151] The second parameter configured by the base station through RRC signaling includes at least one of the following: the duration of the timer of the second AI model, the time, sample size, and number of monitoring moments required to start KPI monitoring within the timer of the second AI model, and the performance judgment condition of the second AI model.

[0152] It should be noted that the configurations of the first AI model timer and the second AI model timer can be different. If different second AI models correspond to a timer respectively, the above configurations of each timer can be different.

[0153] The time / sample size / number of monitoring moments required to start KPI monitoring within the timer of the second AI model indicates that periodic monitoring is performed while the timer is turned on, and there is one monitoring moment in each period.

[0154] By configuring the model's timer duration, timer start conditions, restart conditions, and model performance judgment conditions, you can accurately control the start and stop of the timer of the first AI model and the timer of the second AI model.

[0155] In another embodiment provided by the present disclosure, the start conditions of the first timer or the second timer are specifically one, two or three of the following three items:

[0156] The reference signal received power or block error rate is less than a preset first threshold for a cumulative N1 times within the T1 duration;

[0157] The reference signal received power or block error rate is less than a preset second threshold for a cumulative N2 times within the T2 duration;

[0158] The reference signal received power or block error rate is less than a preset third threshold for N3 consecutive times;

[0159] Wherein, N1, N2 and N3 are positive integers.

[0160] In a specific embodiment of the present disclosure, when the activation condition includes only one item, the activation condition includes the following situations:

[0161] When the reference signal received power is less than X1 or the block error rate is less than Y1 during N1 consecutive detections within the preset T1 duration, the first timer or the second timer starts.

[0162] When the reference signal received power is less than X1 or the block error rate is less than Y1 during N2 detections accumulated within the preset T2 duration, the first timer or the second timer starts.

[0163] When the reference signal received power is less than X1 or the block error rate is less than Y1 during N3 consecutive detections, the first timer or the second timer is started.

[0164] In another specific embodiment of the present disclosure, when the activation conditions include only two items, the activation conditions include the following situations:

[0165] When the reference signal receiving power is less than X1 or the block error rate is less than Y1 during N1 consecutive detections within the preset T1 duration, and the reference signal receiving power is less than X1 or the block error rate is less than Y1 for N2 cumulative times within the preset T2 duration, the first timer or the second timer starts.

[0166] When the reference signal receiving power is less than X1, or the block error rate is less than Y1 during N2 detections within the preset T2 duration, and the reference signal receiving power is less than X1, or the block error rate is less than Y1 during N3 consecutive detections, the first timer or the second timer starts.

[0167] When the reference signal receiving power is less than X1, or the block error rate is less than Y1, during N1 consecutive detections within the preset T1 duration, and the reference signal receiving power is less than X1, or the block error rate is less than Y1 for N3 consecutive times, the first timer or the second timer starts.

[0168] In another specific embodiment of the present disclosure, when the activation condition includes three items, the activation condition is as follows:

[0169] When the reference signal received power is less than X1, or the block error rate is less than Y1, during N1 consecutive detections within the preset T1 duration, and the reference signal received power is less than X1, or the block error rate is less than Y1, during N2 cumulative detections within the preset T2 duration, and the reference signal received power is less than X1, or the block error rate is less than Y1, during N3 consecutive detections, the first timer or the second timer starts.

[0170] Whether the timer is started is determined by detecting the reference signal received power or the block error rate.

[0171] In another embodiment provided by the present disclosure, the first monitoring KPI being less than a preset fourth threshold is used as a performance judgment condition for the first AI model, or the second monitoring KPI being greater than a preset fifth threshold is used as a performance judgment condition for the second AI model.

[0172] In another embodiment provided in the present disclosure, the first monitoring KPI or the second monitoring KPI is at least one of beam prediction accuracy and reference signal received power.

[0173] In a specific example, the present disclosure uses the beam prediction accuracy and the reference signal received power as the first monitoring KPI and the second monitoring KPI, respectively, to monitor the first AI model and the second AI model.

[0174] In another embodiment provided by the present disclosure, the AI ​​model monitoring method performed by the network-side network element further includes:

[0175] Switch to or fall back to traditional beam management based on the monitoring KPI reporting model, including:

[0176] When the terminal determines that there is no candidate model, it reports back to the traditional beam management to the network side network element.

[0177] When the terminal determines that there is a candidate model that can be used as a model switch, the terminal reports at least one of the candidate second AI model ID and the monitoring KPI to the base station as the network side network element. The base station selects one of the models according to the monitoring KPI and instructs switching to the model.

[0178] The terminal device selects the candidate model with the best KPI based on the monitored KPI, and reports the candidate model to the base station, instructing the base station to switch to the model.

[0179] The terminal device reports to the network side network element, which determines the switching model and uses the switching model as a backup measure to ensure the stability of the model switching.

[0180] Another embodiment of the present disclosure provides a terminal device, which is used to execute the AI ​​model monitoring method as described in any of the above embodiments.

[0181] The terminal device provided in this embodiment is capable of executing all steps and functions of the AI ​​model monitoring method executed by the terminal device in any of the above embodiments, and the specific functions of the terminal device are not described in detail here.

[0182] Another embodiment of the present disclosure provides a network-side network element, which is used to execute the AI ​​model monitoring method as described in any of the above embodiments.

[0183] The network side network element provided in this embodiment can execute all the steps and functions of the AI ​​model monitoring method executed by the network side network element in any of the above embodiments. The specific functions of the network side network element are not described here.

[0184] The embodiments of the present disclosure further provide a computer program product, comprising a computer program / instruction, which implements the steps of any of the methods described in the above embodiments when executed by a processor.

[0185] It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present disclosure, and these improvements and modifications are also considered to be within the scope of protection of the present disclosure.

Claims

1. An AI model monitoring method, applied to a terminal, comprising: Receiving a first parameter and a second parameter configured by a network-side network element through RRC signaling; as well as According to the first parameter and the second parameter, the first timer and the second timer are controlled to be turned on and off.

2. The AI ​​model monitoring method according to claim 1, wherein: The first parameter includes at least one of the following: the duration of the first timer, the start condition and / or restart condition of the first timer, and the performance judgment condition of the first AI model; The second parameter includes at least one of the following: the duration of the second timer, the time or sample size or number of monitoring moments required to start KPI monitoring within the second timer, and the performance judgment condition of the second AI model.

3. The AI ​​model monitoring method according to claim 2, wherein: The start condition of the first timer or the second timer includes at least one of the following: The reference signal received power or block error rate is less than a preset first threshold for N1 consecutive times within the T1 duration; The reference signal received power or block error rate is less than a preset second threshold for a cumulative N2 times within the T2 duration; The reference signal received power or block error rate is less than a preset third threshold for N3 consecutive times; Wherein, N1, N2 and N3 are positive integers.

4. The AI ​​model monitoring method according to claim 2, wherein: The performance judgment condition of the first AI model is that the first monitoring KPI is less than a preset fourth threshold, or the performance judgment condition of the second AI model is that the second monitoring KPI is less than a preset fifth threshold.

5. The AI ​​model monitoring method according to claim 4, wherein: The first monitoring KPI or the second monitoring KPI is at least one of beam prediction accuracy and reference signal received power.

6. The AI ​​model monitoring method according to claim 1, wherein: The controlling the opening and closing of the first timer and the second timer according to the first parameter and the second parameter includes: When it is detected that a start condition of the first timer in the first parameter is met, start the first timer; and When it is detected that the performance judgment condition of the first AI model in the first parameter is met, the first timer is turned off and the second timer is turned on according to a pre-configured rule.

7. The AI ​​model monitoring method according to claim 1, wherein: The controlling the opening and closing of the first timer and the second timer according to the first parameter and the second parameter includes: When it is detected that the start condition of the first timer in the first parameter is met, the first timer and the second timer are started at the same time.

8. The AI ​​model monitoring method according to any one of claims 1 to 7, further comprising: When the first timer is closed and the performance judgment condition of the first AI model in the first parameter is still not met, the second timer is started.

9. The AI ​​model monitoring method according to claim 6, wherein: The starting of the second timer according to the pre-configured rule includes: Starting a second timer and using the second timer to monitor all second AI models; When the second timer meets the time or sample size or number of monitoring moments required to start KPI monitoring, start calculating the monitoring KPI of the second AI model; and When a second AI model meets the performance judgment condition of the second AI model in the second parameter, or the second timer expires, the second timer is turned off.

10. The AI ​​model monitoring method according to claim 6, wherein: The starting of the second timer according to the pre-configured rule includes: Serially start multiple second timers, each second timer monitoring a second AI model; When the second timer meets the time or sample size or number of monitoring moments required to start KPI monitoring, start calculating the monitoring KPI of the second AI model; and When a second AI model meets the performance judgment condition of the second AI model in the second parameter, the second timer corresponding to the second AI model is turned off.

11. The AI ​​model monitoring method according to any one of claims 1 to 7, further comprising: When at least one of the first timer and the second timer is closed, the reporting model is activated or deactivated or switched or falls back to traditional beam management.

12. The AI ​​model monitoring method according to claim 11, wherein: The reporting model activation or deactivation or switching or fallback to traditional beam management includes any of the following: When the first AI model satisfies the performance judgment condition of the first AI model in the first parameter, and each second AI model satisfies the performance judgment condition of the second AI model in the second parameter, reporting to the network-side network element to fall back to the default AI model or traditional beam management; When the first AI model satisfies the performance judgment condition of the first AI model in the first parameter, and there is at least one candidate second AI model that does not satisfy the performance judgment condition of the second AI model in the second parameter, reporting model switching information, at least one of the ID of the candidate second AI model and the monitoring KPI to the network side network element, and receiving an AI model switching instruction issued by the network side network element; or When the first AI model does not meet the performance judgment condition of the first AI model in the first parameter, and there is at least one candidate second AI model that does not meet the performance judgment condition of the second AI model in the second parameter and has a monitoring KPI better than the first AI model, report model switching information, the ID of the candidate second AI model and at least one of the monitoring KPIs to the network side network element, and receive the AI ​​model switching instruction issued by the network side network element.

13. An AI model monitoring method, applied to the network side, comprising: The first parameter and the second parameter are configured through RRC signaling, and the first parameter and the second parameter are sent to the terminal, so that the terminal controls the opening and closing of the first timer and the second timer according to the first parameter and the second parameter.

14. The AI ​​model monitoring method according to claim 13, wherein: The first parameter includes at least one of the following: the duration of the first timer, the start condition and / or restart condition of the first timer, and the performance judgment condition of the first AI model; The second parameter includes at least one of the following: the duration of the second timer, the time or sample size or number of monitoring moments required to start KPI monitoring within the second timer, and the performance judgment condition of the second AI model.

15. The AI ​​model monitoring method according to claim 14, wherein: The start condition of the first timer or the second timer includes at least one of the following: The reference signal received power or block error rate is less than a preset first threshold for N1 consecutive times within the T1 duration; The reference signal received power or block error rate is less than a preset second threshold for a cumulative N2 times within the T2 duration; The reference signal received power or block error rate is less than a preset third threshold for N3 consecutive times; Wherein, N1, N2 and N3 are positive integers.

16. The AI ​​model monitoring method according to claim 14, wherein: The performance judgment condition of the first AI model is that the first monitoring KPI is less than a preset fourth threshold, or the performance judgment condition of the second AI model is that the second monitoring KPI is less than a preset fifth threshold.

17. The AI ​​model monitoring method according to claim 16, wherein: The first monitoring KPI or the second monitoring KPI is at least one of beam prediction accuracy and reference signal received power.

18. The AI ​​model monitoring method according to claim 13, further comprising: receiving a reporting message indicating that the terminal has fallen back to a default AI model or traditional beam management; as well as Receive the model switching information reported by the terminal, select at least one of the ID of the candidate second AI model and the monitoring KPI, and issue an AI model switching instruction to the terminal.

19. A terminal device, configured to execute the AI ​​model monitoring method according to any one of claims 1 to 12.

20. A network-side network element, configured to execute the AI ​​model monitoring method according to any one of claims 13 to 18.

21. A computer program product comprising a computer program / instruction, wherein the computer program / instruction implements the steps of the method according to any one of claims 1 to 18 when executed by a processor.

Images