A hybrid-expert-based pdcch power consumption scheduling method

Abstract

The application discloses a PDCCH power consumption scheduling method based on a hybrid expert. A wireless terminal constructs a decision state according to historical scheduling observation, recent grant density, last monitoring action, component power consumption state and reliability control quantity; a plurality of expert strategy networks generate a conditional distribution of a PDCCH monitoring action and a component power consumption state, a gating network allocates expert weights and forms a hybrid strategy for each scheduling time slot; a feasibility projection is performed on the original action generated by the hybrid strategy, so that a radio frequency receiving chain and a physical layer receiving module required for performing PDCCH monitoring are kept activated. A reliability-aware reward is constructed according to component operation energy consumption, state switching energy consumption, model inference calculation energy consumption and equivalent cost of missed detection grant, and a critic network, a clipped policy gradient and a reliability control quantity are used to update the model. The method can adapt to a heterogeneous grant arrival mode, and control PDCCH monitoring and modem component power consumption state under the condition of limiting long-term effective grant miss detection rate.

Description

Technical Field

[0001] This invention belongs to the fields of wireless communication and Internet of Things (IoT) technology, and specifically relates to a PDCCH power scheduling method based on hybrid experts. Background Technology

[0002] With the development of 5G New Radio (NR) systems and future mobile communication systems, IoT terminals, mobile terminals, and other battery-powered devices need to maintain continuous connectivity under wider bandwidth, more flexible frame structures, and more complex protocol processes. Modem components in terminals, such as the RF receiver chain, physical layer receiver module, digital signal processor, and memory, require frequent activation, making terminal-side energy consumption a significant factor affecting device battery life, service availability, and deployment and maintenance costs.

[0003] The base station transmits Downlink Control Information (DCI) via the Physical Downlink Control Channel (PDCCH) to indicate downlink receive resources, uplink transmit resources, modulation and coding schemes, Hybrid Automatic Repeat Request (HARQ) process, and other scheduling parameters. The base station-side Medium Access Control (MAP) scheduler dynamically generates scheduling results based on traffic queues, radio channels, quality of service requirements, retransmission status, and multi-user contention. Before the start of the current scheduling slot, the radio terminal typically cannot determine whether the base station has allocated a valid scheduling grant; therefore, it needs to perform PDCCH monitoring and blind detection according to the configured control resource set and search space.

[0004] PDCCH monitoring and blind detection require the terminal to activate the RF receiver chain and baseband processing module, and perform reference signal processing, channel estimation, candidate control channel decoding, cyclic redundancy check, and descrambling of temporary radio network identifiers. For intermittent services, low-activity IoT services, and bursty services, there may be no effective scheduling authorization for the terminal in a large number of scheduling slots, but the terminal will still consume the corresponding RF and baseband processing energy.

[0005] Existing methods such as discontinuous reception, connected-state discontinuous reception, wake-up signals, partial bandwidth switching, and search space adaptation can reduce some PDCCH monitoring opportunities, but they usually rely on pre-configured periods, additional network-side signaling, or partial signal reception processing, making it difficult to fully utilize the temporal correlation in historical grant sequences before the start of each scheduling slot. Existing predictive dynamic power management methods can predict subsequent grant states based on historical scheduling information, but they typically employ a single prediction model, fixed scenario selection, or a structure that separates the prediction module from component power control. This makes it difficult to continuously adapt to periodic, bursty, sparse, and mixed service modes, and it cannot directly guarantee that the actions output by the model meet the hardware activation conditions of the components required for PDCCH monitoring.

[0006] Furthermore, some existing methods primarily rely on classification accuracy or fixed penalty for missed detections as evaluation criteria. When the proportion of effective scheduling grants in all scheduling slots is low, the model may achieve a high overall classification accuracy even if it classifies most slots as unauthorized, but this will result in a high rate of missed effective grants. If the terminal excessively pursues energy saving, it may frequently skip scheduling slots with actual grants; if a conservative monitoring strategy is consistently adopted, it will be difficult to obtain actual energy-saving benefits. Therefore, how to directly generate PDCCH monitoring actions and component power consumption status that meet hardware constraints for heterogeneous grant arrival patterns, and how to limit the long-term missed grant rate through closed-loop reliability control, are technical problems that need to be solved in terminal power management. Summary of the Invention

[0007] To address the above problems, this invention proposes a PDCCH power scheduling method based on hybrid experts.

[0008] The technical solution adopted in this invention is as follows: A mobile communication system including a base station and a wireless terminal is constructed. The base station sends scheduling authorization to the wireless terminal via PDCCH within continuous scheduling time slots. The wireless terminal forms a decision state based on historical scheduling observations, PDCCH monitoring actions in the previous scheduling time slot, component power consumption status in the previous scheduling time slot, and current reliability control parameters. A hybrid expert deep reinforcement learning model composed of multiple expert policy networks, gating networks, and commentator networks is established to directly output PDCCH monitoring actions and modem component power consumption status. Feasibility projection is performed on the model output to keep the RF receiver chain and physical layer receiver module required for PDCCH monitoring active.

[0009] The basic working principle of this invention is as follows: multiple expert policy networks learn the power consumption scheduling rules in different authorization arrival modes and different energy-saving and reliability state regions respectively; the gating network allocates expert weights according to the current decision state in each scheduling time slot and forms a hybrid policy; the wireless terminal executes the PDCCH monitoring action and component power consumption control action after feasibility projection, and forms an instantaneous comprehensive cost based on the component operating energy consumption, state switching energy consumption, model inference calculation energy consumption, and the equivalent cost of missed valid authorizations; the reliability control quantity is updated according to the deviation between the actual authorization missed detection situation and the preset reliability threshold, so that the hybrid expert policy can limit the effective scheduling authorization missed detection rate while reducing the long-term comprehensive cost.

[0010] The PDCCH power scheduling method proposed in this invention aims to minimize the long-term average integrated cost of the wireless terminal, under the constraints of long-term effective license miss rate, PDCCH monitoring action coupling with component state, and binary action constraints. It jointly determines the PDCCH monitoring action and modem component power consumption state for each scheduling slot. The established mathematical problem model is expressed as follows:

[0011]

[0012] in, Indicates the scheduling slot index. Indicates the length of the statistical time domain. This represents the power scheduling strategy determined by a hybrid expert deep reinforcement learning model. Indicates the first The actual authorized status of each scheduling slot Indicates the first PDCCH monitoring actions for each scheduling time slot, Indicates the first The modem component in the first The power consumption status of each scheduling slot. This represents the set of components that must be activated to complete PDCCH reception and blind detection. (In the objective function) This indicates the operating power consumption of the modem component in either the active or low-power state. This represents the switching energy consumption caused by changes in component power consumption state between adjacent scheduling time slots. This represents the computational energy consumption generated by the hybrid expert model performing forward inference. This represents the equivalent cost coefficient for retransmissions, additional activation time, throughput loss, or latency loss caused by missing valid scheduling authorizations. This indicates the effective authorization miss rate. The first constraint limits the long-term effective authorization miss rate to no more than a preset reliability threshold. ,in The first constraint is the denominator protection constant; the second constraint indicates that when the wireless terminal decides to perform PDCCH monitoring, it belongs to... All necessary components must be active; the third set of constraints indicates that both PDCCH monitoring actions and component power consumption states are binary variables. Therefore, the objective function is used to reduce the long-term overall terminal cost, the reliability constraint is used to prevent the model from obtaining energy-saving results without communication availability by excessively skipping PDCCH monitoring, and the hardware coupling constraint is used to ensure that the output actions can be actually executed by the terminal modem.

[0013] In the above mathematical problem, the current authorization state is unknown when the wireless terminal makes a monitoring decision. The authorization arrival process is affected by the service type, wireless channel, retransmission status and base station scheduling strategy. Furthermore, the component states in adjacent scheduling slots are coupled with each other through switching energy consumption. At the same time, the monitoring action and component state are discrete variables, and the long-term reliability constraints span multiple scheduling slots. Therefore, this problem belongs to the constrained sequential decision problem that is difficult to solve directly using conventional convex optimization methods on a slot-by-slot basis.

[0014] To solve the above problems, this invention constructs the power scheduling process as a Markov decision process. Multiple expert policy networks are used to learn the action condition distribution under different authorization modes. A gating network is used to form a hybrid policy based on the current state, and a commentator network is used to evaluate long-term rewards. Feasibility projection is performed on the original actions generated by the hybrid policy. A reward function is constructed by negating the comprehensive cost and adding a reliability constraint penalty. The expert policy network and the gating network are updated using a pruning policy gradient. Gating weight entropy regularization is used to avoid expert collapse, and the reliability control quantity is updated using a projection subgradient form. Finally, a PDCCH power scheduling policy that meets reliability requirements and can be executed online is obtained.

[0015] The beneficial effects of this invention are as follows: This invention performs state-related soft adaptation on periodic, bursty, sparse, and hybrid licensing modes through multiple expert policy networks and gating networks, avoiding the long-term maintenance of fixed scheduling rules by a single policy when service modes change; it generates PDCCH monitoring actions and component power consumption status as unified actions, and ensures the hardware activation conditions of necessary receiving components through feasibility projection; it directly introduces the effective licensing miss detection rate into the training and online control process through reliability control variables, rather than just using it as an offline evaluation indicator; and it can adaptively control the PDCCH monitoring and receiving chain power consumption status based on historical scheduling information and component status without requiring the wireless terminal to know the current scheduling results in advance. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the reliability-constrained PDCCH power scheduling system model of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings and simulation embodiments.

[0019] like Figure 1 As shown, in a preferred embodiment, the mobile communication system includes a base station and at least one battery-powered wireless terminal. The base station includes a Media Access Control (MAP) scheduler and a PDCCH transmission module; the wireless terminal includes a historical observation buffer, a PDCCH monitoring and decision module, an RF receiver chain, a physical layer receiver module, a digital signal processor, a memory, a hybrid expert deep reinforcement learning module, an action feasibility projection module, a reliability control module, and a component power consumption control module. In each scheduling slot, the base station-side MAP scheduler determines whether to allocate uplink or downlink resources to the wireless terminal based on the service queue, radio channel state, quality of service requirements, hybrid automatic repeat request (HARRequest) status, and multi-user contention, and transmits the corresponding scheduling parameters via downlink control information (DCI) carried on the PDCCH. Figure 1This diagram illustrates the information interaction relationships between the base station scheduler, the time-slot scheduling process, the historical observation window, the wireless channel, the wireless terminal's PDCCH monitoring decision, the modem component power consumption status, and the communication results. The current true license state generated by the base station is unknown before the wireless terminal performs PDCCH monitoring. The wireless terminal obtains scheduling information such as past license results, modulation and coding schemes, transport block sizes, adjacent license intervals, and license densities from the historical observation window, and forms the decision state for the current scheduling time slot based on this scheduling information.

[0020] The wireless terminal chooses to perform PDCCH monitoring or skip PDCCH monitoring based on its current decision state. When the wireless terminal chooses to perform PDCCH monitoring, necessary components in the RF receiver chain, physical layer receiver module, digital signal processor, and memory are switched to an active state to complete reference signal processing, channel estimation, candidate control channel decoding, cyclic redundancy check, and radio network temporary identifier descrambling. When the wireless terminal chooses to skip PDCCH monitoring, the corresponding components maintain or switch to a low-power state. If a valid grant actually exists in the current scheduling slot but the wireless terminal skips PDCCH monitoring, a valid grant missed detection event occurs, which may cause transmission delay, retransmission, or throughput loss. Figure 1 As shown below, the overall cost generated by each scheduling slot includes at least the power consumption of the modem components, the power consumption of component power state switching between adjacent scheduling slots, and the equivalent cost of missed valid authorizations. In one embodiment, the computational power consumption generated by the hybrid expert deep reinforcement learning model performing forward inference is also included in the overall cost.

[0021] Divide the system time into continuous scheduling time slots, let the first... The actual authorized status of each scheduling slot is:

[0022]

[0023] in, Generated by the base station-side scheduler, it cannot be directly observed before the wireless terminal performs PDCCH monitoring in the current scheduling time slot. At that time, the PDCCH contains valid downlink control information that can indicate the wireless terminal's subsequent reception or transmission resources; At that time, there is no valid downlink control information for the wireless terminal.

[0024] Order No. The PDCCH monitoring actions for each scheduling slot are as follows:

[0025]

[0026] when At that time, the wireless terminal activates the radio frequency and baseband components required to complete PDCCH reception and blind detection; when At this time, the wireless terminal does not perform blind PDCCH detection for the current scheduling time slot. If at this time If this happens, the wireless terminal cannot obtain the corresponding scheduling authorization in a timely manner, resulting in a valid authorization missed detection event. The missed detection indication quantity is defined as follows:

[0027]

[0028] In the formula, if and only if and From time to time It exists in all three scenarios: unauthorized and monitored, unauthorized and dormant, and authorized and monitored. .therefore, It can directly link PDCCH power consumption control actions with communication reliability loss.

[0029] set up This refers to the set of modem components in a wireless terminal that participate in PDCCH monitoring and power management. This set of components may include an RF receiver chain, a physical layer receiver module, a digital signal processor, and a memory. Indicates the first The component in the first Power consumption status of each scheduling slot: This indicates that the m-th component is in an active state, and a value of 0 indicates that the m-th component is in a low-power state. Let... Let PDCCH represent the set of components essential for performing PDCCH reception and blind detection. Then, the PDCCH monitoring action and component power consumption state satisfy the following:

[0030]

[0031] The physical meaning of this constraint is: as long as All belonging to All components must meet ;when At this time, the constraint does not force all components to be shut down. The terminal may keep some components active based on the cost of state switching, subsequent business forecasts, or hardware timing requirements.

[0032] No. The energy consumption of a component operating in a scheduling time slot is expressed as follows:

[0033]

[0034] in, Indicates the duration of a scheduling slot. Indicates the first Power of each component when it is in an active state Indicates the first The power of a component when it is in a low-power state is typically: .when At that time, the component consumes time in the current scheduling slot. ;when Time, consumption .

[0035] Since switching a component from an active state to a low-power state or back to an active state requires processes such as power gating, clock recovery, and state saving, the energy consumption of state switching between adjacent scheduling slots is expressed as follows:

[0036]

[0037] in, Indicates the first The energy required for a component to undergo one power state transition. When When the absolute value is zero, no switching power consumption is generated for that component; when the two are not equal, the absolute value is one, indicating a state transition has occurred. This item can prevent the strategy from frequently switching components in adjacent scheduling slots in order to obtain short-term low power consumption benefits.

[0038] Hybrid expert models incur computational energy consumption when performing gating computation, expert forward inference, and action sampling in each scheduling slot. Let Indicates the first The number of floating-point operations per scheduling slot, Let the equivalent energy consumption of a unit floating-point operation be represented, then the energy consumption for inference calculation can be expressed as:

[0039]

[0040] In implementations that do not require a precise processor power consumption model, It can also be obtained by multiplying the average inference power of the terminal processor by the inference duration. Including the inference computation energy consumption in the objective function can prevent the computational energy consumption caused by the increase in model complexity from offsetting the energy-saving benefits of PDCCH monitoring.

[0041] Taking into account the equivalent losses caused by component operation, component switching, model inference, and missed valid authorization, the first The total cost of a scheduling slot is expressed as:

[0042]

[0043] in, It doesn't necessarily have to be equivalent to actual electrical energy. It can represent the weighted equivalent cost of retransmission energy consumption, subsequent additional activation time, throughput loss, and latency loss caused by missed authorization. When the system only focuses on electrical energy, it can be set as the equivalent additional energy consumption corresponding to the missed event; when the system also focuses on service quality, this coefficient can be determined according to the normalized weights of each performance indicator.

[0044] Within a finite statistical window, the effective authorization false negative rate is defined as:

[0045]

[0046] Where the numerator represents the number of valid authorizations missed within the statistical window, and the denominator represents the total number of valid authorizations actually granted within the statistical window. Used to avoid a denominator of zero when no valid authorization appears within the statistical window. Unlike overall classification accuracy, It measures only the proportion of truly valid authorizations that are missed, making it suitable for describing reliability losses in authorization-sparse business scenarios.

[0047] Based on the above energy consumption and reliability models, a long-term comprehensive cost minimization problem is established. The objective function and constraints are expressed as follows:

[0048]

[0049]

[0050]

[0051] The objective function is used to find a power scheduling strategy with low long-term average integrated cost; reliability constraints limit the effective license missed detection ratio; hardware coupling constraints ensure that monitoring actions can be actually implemented by the receiver chain; binary constraints reflect the discrete power consumption state of the terminal components. Because... Since the decision-making process is unknowable, the transfer rules of the authorization process are unknown, the action affects the switching energy consumption across time slots, and the reliability constraints span the entire statistical window, deep reinforcement learning is used to learn the causal mapping from state to action.

[0052] The above sequential decision-making process can be described as a Markov decision-making process. .in, Representing the state space, Represents the action space. This represents an unknown state transition pattern. Represents the reward function, This represents the discount factor. Since the internal state of the base station scheduler cannot be directly observed, this state is constructed from historical terminal observations to approximately reflect the time correlation of the authorized arrival process.

[0053] make Indicates the first The scheduling observation vector that can be obtained after the end of a scheduling time slot may include at least the grant detection result, and may further include the modulation and coding scheme, transport block size, channel quality indicator, hybrid automatic repeat request feedback, and service buffer status. The decision state before the start of each scheduling slot is represented as follows:

[0054]

[0055] in, Indicates the length of the historical observation window; Indicates recent The authorization density within a scheduling slot is used to reflect the recent scheduling activity level. This indicates the time interval since the most recent valid authorization, used to reflect periodic or long periods of inactivity; This indicates the monitoring action taken in the previous scheduling time slot; Represents the component state vector of the previous scheduling slot; This represents the current reliability control value. Because... and Since the switching cost when performing a new action can be determined, it is included in the state.

[0056] The recent authorization density and the time interval since the most recent authorization can be expressed as follows:

[0057]

[0058]

[0059] If no valid authorization is available in the history window, you can Set to the preset upper limit value. For authorized tags that are temporarily missing due to skipped monitoring during the online phase, predicted values, unknown tags, or subsequently obtained delayed tags can be used to fill in the gaps.

[0060] The action vector is represented as:

[0061]

[0062] The action vector simultaneously determines whether to perform PDCCH monitoring and whether each component remains active. Compared to schemes that only output the authorization category and then map the component state using fixed rules, this joint action allows the reinforcement learning policy to directly consider the component's runtime energy consumption and switching energy consumption during the optimization process.

[0063] Establish containing A hybrid expert model of an expert policy network. An expert strategy network based on state Output action condition distribution:

[0064]

[0065] in, Indicates the first Each expert has a policy network parameter. Since the action components are binary variables, each expert can output the PDCCH monitoring probability and component activation probability, and generate the original action according to the Bernoulli distribution. Different experts can adapt to periodic services, bursty services, sparse services, reliability-first states, and energy-saving-first states during training, but the service category corresponding to each expert is not fixed in advance.

[0066] The gating network outputs the weights of each expert based on the current state:

[0067]

[0068]

[0069] in, Indicates that the gating network is for the first The expert's rating, This represents the gating network parameters. The gating weight is not a fixed service label, but changes continuously in each scheduling slot based on the current historical authorization, component status, and reliability pressure.

[0070] Multiple expert strategies are combined into a hybrid strategy through gating weights:

[0071]

[0072] in, This represents all parameters on the actor side. During online execution, actions can be sampled directly from the hybrid strategy, or expert indices can be sampled first based on gating weights, and then actions can be sampled from the corresponding expert strategy.

[0073] By hybrid strategy Generate original actions The original action is represented as ,in, This represents the original PDCCH monitoring action output by the hybrid strategy. The output of the hybrid strategy is the first... The raw power state of each modem component. Let... And perform a feasibility projection on the original power state of each component based on the following relationship:

[0074]

[0075] in, This represents the set of components that must be activated to complete PDCCH reception and blind detection. When the original action requires PDCCH monitoring, i.e. However, the original power consumption state of a certain necessary receiving component is At that time, through Correct the power state of this component to For those that do not belong to The components retain the original power consumption state of the hybrid strategy output. After feasibility projection, the executable actions are obtained: .

[0076] To introduce long-term reliability constraints into reinforcement learning, the first... Constraint deviation for each scheduling slot:

[0077]

[0078] When there is valid authorization and the terminal skips monitoring When valid authorization exists and the terminal performs monitoring, When no valid authorization exists, .therefore, It can reflect the deviation between the actual missed detection rate and the target missed detection rate after being accumulated over multiple scheduling time slots.

[0079] Negating the overall cost and adding a Lagrange reliability penalty yields the immediate reward:

[0080]

[0081] Reinforcement learning algorithms minimize the overall cost of the original problem by maximizing the long-term cumulative reward. However, when there are too many missed authorizations, the cost becomes significantly higher. By reducing the reward for the corresponding action, the expert strategy can gradually increase the probability of PDCCH monitoring; when the reliability constraint has a large margin, the strategy can make more use of low-power opportunities.

[0082] Reliability control quantities are updated using projection:

[0083]

[0084] in, This indicates the update step size for reliability control variables. This projection guarantees Non-negative. In practice, it can also be updated based on the window average constraint deviation after a training batch or a statistical window ends. This is to reduce random fluctuations in a single time slot.

[0085] Critics Network Used to estimate from state The cumulative long-term discount rewards that can be obtained by starting the current strategy, of which This represents the network parameters for the commentator. The time-series difference error is expressed as:

[0086]

[0087] Temporal difference error measures the deviation of the sum of the current immediate reward and the next state value from the current state value estimate. Dominance estimation can be calculated based on the temporal difference errors of multiple consecutive time slots. It is used to judge the quality of the current action relative to the average action in the current state.

[0088] To limit the magnitude of a single policy update, the probability ratio between the old and new policies is defined as follows:

[0089]

[0090] The goal is to update the expert policy network and the gating network using a pruning strategy:

[0091]

[0092] in, This represents the pruning factor. The pruning operation prevents the new policy from changing too much compared to the old policy, thereby reducing the risk of reliability loss due to sudden and drastic changes in monitoring probabilities during training.

[0093] To avoid the gating network selecting only a single expert in the long run, the gating weight entropy is defined as follows:

[0094]

[0095] The training objectives for actors are expressed as follows:

[0096]

[0097] in, This represents the gating weight entropy regularization coefficient. In the early stages of training, this regularization term encourages multiple experts to obtain effective samples, avoiding the situation where all states are handled by the same expert; in the later stages of training, each expert can still form a specialized strategy based on different states.

[0098] The critic network updates by minimizing the value regression error:

[0099]

[0100] in, This represents the experience-discounted reward calculated from the training trajectory. The expert policy network, gating network, critic network, and reliability control are updated alternately until the training objective and authorized false negative rate tend to stabilize.

[0101] The training and update process of the PDCCH power scheduling method based on hybrid expert deep reinforcement learning described in this invention specifically includes the following steps:

[0102] First, multiple expert policy networks, gating networks, critic networks, reliability control variables, and experience replay buffers are initialized. Within each training round, the current decision state is constructed sequentially according to the scheduling slots, the gating weights corresponding to each expert are calculated, a hybrid expert aggregation policy is generated, and original actions are sampled from the aggregation policy. Subsequently, feasibility projection is performed on the original actions to obtain executable actions that satisfy the terminal hardware constraints, and the actual authorization state, immediate reward, and next decision state are obtained based on the action execution results. After completing the interaction of one or more scheduling slots, training samples are obtained from the experience replay buffer, and experience discount rewards and advantage estimates are calculated. The expert policy networks and gating networks are updated based on the pruning policy objective, the critic network is updated based on the value regression error, and the reliability control variable is updated based on the authorization miss constraint deviation. The above process is repeated until the preset training termination condition is met, and the trained expert policy network, gating network, and critic network are output.

[0103] Step S1: Obtain historical scheduling trajectories with valid authorization tags from base station scheduling logs, wireless terminal modem logs, or simulated scheduling environment, and set the historical observation window length. Number of experts Maximum Permissible Authorized Missed Detection Rate Discount Factor Strategy clipping factor The learning rates of each network are determined, and the expert policy network, gating network, critic network, and reliability control variables are initialized.

[0104] Step S2, in the Before each scheduling slot begins, a decision state is constructed based on historical scheduling observations, the most recent authorization density, the time interval since the most recent authorization, the previous monitoring action, the previous component power consumption status, and the current reliability control parameters. .

[0105] Step S3: Input the decision state into multiple expert policy networks and a gating network respectively. Each expert policy network outputs the action condition distribution, and the gating network outputs the expert weights. A hybrid strategy is then formed according to the gating weights. .

[0106] Step S4, from the hybrid strategy Mid-sampling original action ;make ,for The necessary components, according to Correct component power consumption status; for The components make This yields executable actions that satisfy the PDCCH monitoring hardware activation constraints. .

[0107] Step S5: Control the wireless terminal to execute the executable action; calculate the operating energy consumption and switching energy consumption based on the actual power consumption status of the components, calculate the inference calculation energy consumption based on the forward inference complexity of the model, and determine the effective authorization missed detection indication and comprehensive cost when the authorization tag is available.

[0108] Step S6: Calculate the reliability-perceived reward based on the comprehensive cost and constraint bias, and use the critic network to calculate the state value, temporal difference error, and advantage estimate.

[0109] Step S7: Update each expert policy network and gating network using the pruning strategy objective and gating weight entropy regularization, and update the critic network using the value regression error.

[0110] Step S8: Update the reliability control quantity according to the authorized missed detection constraint deviation in a single time slot or statistical window, so that when the missed detection frequency is higher than the preset threshold, the tendency for subsequent monitoring is increased, and when the missed detection frequency is lower than the preset threshold, more energy-saving space is released.

[0111] Step S9: Determine whether the training reward, policy parameter change, and experience-granted false negative rate meet the preset termination conditions; if not, return to step S2 to continue iteration; if so, save the trained expert policy network, gating network, and critic network, and execute state construction, hybrid policy reasoning, feasibility projection, and component power consumption control in each scheduled time slot during the online operation phase.

[0112] For scheduling slots where PDCCH monitoring is performed, the wireless terminal can directly obtain the grant tag based on the PDCCH decoding result. For scheduling slots that skip PDCCH monitoring, the delayed grant tag can be obtained through base station delay feedback, periodic full monitoring windows, subsequent retransmission indications, or upper-layer service feedback. Once the delayed tag arrives, it is used to correct historical observation buffers, calculate the granted missed detection rate, and adjust reliability control parameters. If the delayed tag is unavailable within the current control period, the online decision-making process can retain the unknown tag and not use the slot as an immediate monitoring tag; a retrospective update will be performed once the tag becomes available.

[0113] In another embodiment, the model may jointly control only the RF receiver chain and physical layer receiver module, or it may further control the digital signal processor, memory, clock domain, and other hardware modules related to PDCCH processing. The activation power, low-power power, and state-switching energy consumption of different components can be obtained through chip power consumption models, hardware measurements, or offline calibration.

[0114] In another implementation, the gating network can operate using full expert soft fusion, sparse expert selection, or sampling based on expert weights; the expert policy network can use the same or different structures; and the reliability control variables can be updated on a time-slot-by-time, batch-by-batch, or statistical window-by-window basis. These alternative methods do not change the fundamental technical concept of selecting or fusing multiple experts based on the current state and controlling PDCCH power consumption scheduling through closed-loop reliability constraints.

[0115] In one simulation embodiment, periodic granting, burst granting, sparse granting, and hybrid granting trajectories can be constructed and compared with continuous monitoring, fixed-period monitoring, random monitoring, a single supervised learning predictor, a single reinforcement learning strategy, and a scene clustering predictor. Evaluation metrics may include total terminal energy consumption, net energy saving ratio relative to continuous monitoring, PDCCH monitoring ratio, effective granting miss rate, number of component state transitions, and inference computation cost per scheduling slot. The simulation implementation is used to illustrate the feasibility of the technical solution of this invention and does not constitute a limitation on the scope of protection.

Claims

1. A PDCCH power scheduling method based on hybrid experts, used in a mobile communication system comprising a base station and at least one wireless terminal; characterized in that, The base station's Media Access Control scheduler allocates downlink or uplink resources to the wireless terminal within consecutive scheduling slots and sends downlink control information via PDCCH. The wireless terminal has a modem component. Before the start of the current scheduling slot, the wireless terminal cannot directly know whether there is a valid scheduling authorization for itself in the current scheduling slot. PDCCH monitoring and modem component power consumption control are established as a long-term comprehensive cost minimization problem under reliability constraints. ， in, This represents the power scheduling strategy determined by a hybrid expert deep reinforcement learning model; Indicates the number of time slots counted; Indicates the scheduling slot index; Indicates the power consumption of the modem component; Indicates the power consumption during modem component state switching; This represents the energy consumption of inference computation generated by the hybrid expert model in completing gating computation, expert forward inference, and action generation. This represents the equivalent cost coefficient for missed valid scheduling authorizations; Indicates the first The actual scheduling authorization status of each scheduling slot This indicates that there is a valid scheduling authorization for the wireless terminal. This indicates that there is no valid scheduling authorization; This indicates that the PDCCH monitoring action has been initiated. This indicates that PDCCH monitoring is being performed. This indicates that PDCCH monitoring is skipped. Represents the denominator protection constant; Indicates the maximum allowed authorized false negative rate; Indicates the first Power status of each modem component Indicates an active state. Indicates a low-power state; This represents the set of components that must be activated to complete PDCCH reception and blind detection. Represents a set of modem components; where and The specific expression is: ， ， in, Indicates the duration of the scheduling slot. and They represent the first The active power and low power of each modem component; Indicates the first The energy required for a modem component to undergo a single power state transition; By solving the problem of minimizing the long-term comprehensive cost under reliability constraints, the PDCCH monitoring actions and modem component power consumption status of each scheduling slot are obtained.

2. The PDCCH power scheduling method based on hybrid experts according to claim 1, characterized in that, The specific method for solving the long-run comprehensive cost minimization problem under reliability constraints is as follows: S1. Obtain historical scheduling trajectories with valid scheduling authorization tags from base station scheduling logs, wireless terminal modem logs, or simulated scheduling environments, and set the historical observation window length. Number of expert strategy networks Maximum Permissible Authorized Missed Detection Rate Discount Factor Strategy clipping factor And the learning rates of each network, and initialized Expert strategy network, gating network, critic network, and reliability control quantity ; S2, in the Construct a decision state before each scheduling slot begins: ， in, Indicates the first The scheduling observation vector of a historical scheduling time slot includes at least one of the following: historical grant status, modulation and coding scheme, transport block size, channel quality indicator and hybrid automatic repeat request feedback. Indicates the most recently authorized density; Indicates the time interval since the most recent valid scheduling authorization; Represents the component power state vector of the previous scheduling slot; Indicates the first Reliability control quantity for each scheduling time slot; S3, No. An expert strategy network outputs an action condition distribution based on the decision state. ,in, Indicates the first Parameters of an expert policy network This indicates a power consumption scheduling action; the gating network outputs expert weights based on the decision state. ， in, Indicates the first Gating scores from a single expert, Indicate the gating network parameters; form a hybrid strategy: ， The hybrid strategy generates the original power scheduling action. ; S4, Order and in accordance with: ， A feasibility projection is performed on the original power scheduling action to obtain an executable power scheduling action that meets the hardware operating conditions; S5. Execute the executable power scheduling action to obtain the operating energy consumption, state transition energy consumption, inference computing energy consumption, and actual or delayed scheduling authorization tag of the current scheduling time slot, and calculate the constraint deviation. and reliability perception rewards ,in, ; S6. Utilizing Critics Networks Estimate state value and calculate time series difference error and advantage estimation Based on the probability ratio of the old and new mixed strategies: ， A pruning strategy gradient objective is used to update the expert policy networks and gating networks, and the critic network is updated based on the state value regression error. At the same time, a gating weight entropy regularization term is added to restrict the gating network from using only the same expert in the long run. S7. According to: ， Update reliability control variables, among which, This indicates the update step size for reliability control variables. ; Step S8: Determine whether the average cumulative reward change, model parameter change, authorization miss rate within the continuous evaluation window, or training rounds meet the preset termination conditions; if not, return to step S2 to continue iteration; if so, save the trained expert policy network, gating network, and critic network; during online operation, execute the hybrid policy and feasibility projection according to the current decision state, control the power consumption status of PDCCH monitoring and modem components, and update the authorization miss rate and reliability control quantity after the actual or delayed scheduling authorization tag becomes available.

Images