A robot energy scheduling method, system, device and storage medium

Abstract

This disclosure relates to the field of robot energy scheduling technology, and discloses a robot energy scheduling method, system, device, and storage medium. The method includes: real-time acquisition of external scene perception signals and internal body operation status signals of the robot; construction of a high-power event prediction model; feature extraction and fusion of the acquired scene perception signals and body operation status signals to generate a time-series feature vector sequence and input it into the high-power event prediction model, outputting the trigger probability of a high-power event occurring within a specified future time window; generating a corresponding energy pre-scheduling instruction based on the trigger probability; and executing the energy pre-scheduling instruction before the high-power event actually occurs to complete the robot energy scheduling. This disclosure realizes a fundamental shift in energy management from "passive response" to "proactive prediction," significantly improving operational continuity and system reliability.

Description

Technical Field

[0001] This disclosure relates to the field of robot energy scheduling technology, specifically to a robot energy scheduling method, system, device and storage medium, and particularly to a robot energy dynamic scheduling method, system, device and storage medium based on scene and body signal prediction. Background Technology

[0002] Application Scenarios and Background: Highly mobile robots are increasingly used in outdoor or enclosed space operations within the energy sector, including hydropower projects, pumped-storage power station operation and maintenance, wind turbine blade inspection, photovoltaic array cleaning, and energy storage facility maintenance. These robots need to output peak power of 1-3kW instantaneously when performing high-intensity actions such as drilling, grinding, cutting, and de-icing, while maintaining extremely low power consumption for most of the time, such as long-distance inspections and condition monitoring, to ensure endurance. Energy supply typically relies on onboard lithium-ion batteries. The core challenge currently facing energy operation robots is how to dynamically adapt to intermittent, high-intensity power demands with limited battery power.

[0003] Existing technical implementation methods: Currently, the energy management system of this type of robot mainly adopts a feedback control architecture based on the real-time battery status. Its typical implementation is as follows: 1. Power control loop: The motor driver (usually using FOC vector control) adjusts the phase current in real time using a PID algorithm based on the position or torque error fed back by the encoder. The battery management system operates independently, mainly managing output capacity based on the total remaining battery charge (SOC). A common strategy is: when SOC > 50%, maximum current output is allowed; when SOC drops below 30%, the maximum output current is linearly or stepwise limited to 40%-60% of the rated value. 2. Decoupling of task planning and energy management: The upper-level task planning system (such as vision-based path planning or work instruction sequences) and the lower-level power control system are usually independent of each other. Communication between the two is achieved through simple "enable / status" flags or preset power levels, lacking fine-grained information transmission regarding the specific power demands that will occur. 3. Hardware protection mechanism: To ensure the safety of the battery and power devices, the system has hardware-level protection circuits (such as voltage comparators and overcurrent detection ICs). Once the bus voltage is detected to drop below a set threshold (e.g., 10%) due to a sudden increase in load, or the instantaneous current exceeds the hardware limit, the protection circuit will directly shut down the main power MOSFET, causing the robot to stop working immediately.

[0004] Problems with existing technologies: Due to the inherent design and structure of the above implementation methods, the following prominent problems arise: 1. Passivity and lag in energy supply: The system only reacts through the BMS or protection circuit when the motor is actually loaded and the current rises, causing a voltage drop. This "post-event remedy" mode cannot cope with sudden changes in high power demand at the millisecond to second level, inevitably leading to "stuttering" or "abrupt stops" during operation, seriously affecting the efficiency of continuous operation and equipment reliability. 2. Rigid energy scheduling strategy, disconnected from tasks: The "one-size-fits-all" current limiting strategy based on SOC fails to correlate with specific operation scenarios, task criticality, and the robot's real-time motion state. For example, when the battery is low but a critical demolition task is being performed, the system still mechanically limits power output, leading to task failure; conversely, during simple inspections, the system maintains high power output, resulting in energy waste. 3. Lack of adaptive and learning capabilities: The existing architecture cannot learn the relationship between different scenarios (such as cutting steel bars and grinding concrete) and corresponding power demand patterns from historical operation data. Every time a similar scenario is encountered, the system repeats the passive cycle of "loading-protecting-recovering," making it impossible to optimize performance through experience accumulation. This results in low intelligence levels and high maintenance costs. Summary of the Invention

[0005] This disclosure provides a robot energy scheduling method, system, device, and storage medium to address the problems in the prior art.

[0006] According to one aspect of this disclosure, a robot energy scheduling method is provided, comprising: Real-time acquisition of external scene perception signals and internal body operation status signals of the robot; Construct a high-power event prediction model; The collected scene perception signals and body operation status signals are subjected to feature extraction and fusion to generate a time-series feature vector sequence, which is then input into the high-power event prediction model to output the trigger probability of a high-power event occurring within a specified future time window. Based on the trigger probability, a corresponding energy pre-schedule instruction is generated; Before the high-power event actually occurs, the energy pre-scheduling instruction is executed to complete the robot's energy scheduling.

[0007] In one possible implementation, the scene perception signal includes environmental visual information acquired by a visual sensor, environmental sound information acquired by an acoustic sensor, and temperature, humidity, and millimeter-wave detection information acquired by an environmental sensor. The body's operating status signals include joint angles, angular velocities, load torques, and fuselage inertia measurement information obtained by motion status sensors; motor phase currents and DC bus electrical parameters obtained by electrical status sensors; and operating handle and host computer control commands obtained by operating intention sensors.

[0008] In one possible implementation, a high-power event prediction model is constructed, including: During the robot's task execution, a time-series feature vector sequence, the identifier of a high-power event that actually occurs after the time-series feature vector sequence, and the actual energy consumption curve corresponding to the high-power event are synchronously collected and recorded. The collected time-series feature vector sequence, the high-power event identifier, and the actual energy consumption curve are bound together to generate an experience record, and the experience record is stored in the experience memory bank. Based on the experience records stored in the experience memory bank, the long short-term memory network model is trained by using the time-series feature vector sequence as input and the corresponding high-power event identifier or occurrence probability as the learning target, thereby obtaining a high-power event prediction model.

[0009] In one possible implementation, the expression for the high-power event prediction model is: P_trigger = LSTM_Model(F(tn), F(t-n+1), …, F(t)); In the formula, F(t) represents the temporal feature vector at the current timestamp t; n represents the historical time step; Δt represents the length of the predicted future time window; P_trigger represents the trigger probability of a high-power event; and LSTM_Model(·) represents the forward inference process of the long short-term memory network model.

[0010] In one possible implementation, the energy pre-scheduling command includes adjustment commands for the DC bus voltage, the energy reserve of the energy storage unit, and the current limit value of the motor driver.

[0011] In one possible implementation, a corresponding energy pre-scheduling instruction is generated based on the trigger probability, including: When the trigger probability is greater than or equal to the first threshold, a first-level pre-scheduling instruction is generated. The first-level pre-scheduling instruction includes raising the bus voltage to the first target voltage, charging the energy storage unit to the first reserve energy value, and setting the motor driver current limit value to the maximum value. When the trigger probability is less than the first threshold and greater than or equal to the second threshold, a secondary pre-schedule instruction is generated. The secondary pre-schedule instruction includes raising the bus voltage to a second target voltage lower than the first target voltage, charging the energy storage unit to a second reserve energy value lower than the first reserve energy value, and setting the motor driver current limit value to an intermediate limit value between the normal limit value and the maximum value. When the trigger probability is less than the second threshold, the normal energy scheduling mode is maintained.

[0012] In one possible implementation, the energy pre-scheduling instruction is executed before the high-power event actually occurs to complete the robot's energy scheduling, including: By using a bidirectional DC-DC converter, the DC bus voltage is adjusted to the target voltage value before a high-power event occurs, according to the energy dispatch command. By controlling the bidirectional DC-DC converter or an independent charging circuit, the supercapacitor array connected in parallel with the DC bus is pre-charged according to the energy dispatch command to store reserve energy value. The current loop limit value of the motor driver is dynamically configured to the limit value specified by the energy dispatch command.

[0013] In one possible implementation, the robot energy scheduling method further includes: After each task is completed, the robot stores the feature sequence of the scene perception signal and the body operation status signal, the actual high-power event identifier and the corresponding actual energy consumption curve as a piece of empirical data in the empirical memory bank, which is used to update the high-power event prediction model.

[0014] In one possible implementation, the robot energy scheduling method further includes: Periodically synchronize the experience data in the experience memory bank to the cloud server; The cloud server aggregates experience data from multiple robots to train and optimize the global model parameters of the high-power event prediction model. The optimized model parameters are distributed to each robot to update the local high-power event prediction model of each robot.

[0015] According to one aspect of this disclosure, a robot energy scheduling system is provided, comprising: The acquisition unit collects external scene perception signals and internal body operation status signals of the robot in real time; Building blocks are used to construct high-power event prediction models; The prediction unit is used to extract and fuse features from the collected scene perception signals and the body's operating status signals, generate a time-series feature vector sequence, input it into the high-power event prediction model, and output the trigger probability of a high-power event occurring within a specified future time window. The generation unit is used to generate a corresponding energy pre-scheduling instruction based on the triggering probability; The scheduling unit is used to execute the energy pre-scheduling instruction before the high-power event actually occurs, and to complete the robot energy scheduling.

[0016] According to one aspect of this disclosure, an electronic device is provided, comprising: Processor and memory; The memory is used to store computer programs, and the processor calls the computer programs stored in the memory to execute the robot energy scheduling method described above.

[0017] According to one aspect of this disclosure, a computer-readable storage medium is provided, wherein a computer program is stored therein, which, when executed by a processor, enables the processor to perform the robot energy scheduling method described in any of the preceding claims.

[0018] This disclosure offers the following advantages: It achieves a fundamental shift in energy management from "passive response" to "proactive anticipation," significantly improving operational continuity and system reliability. By real-time fusion analysis of scene information and robot motion signals, this disclosure can accurately predict high-power loads seconds before they actually occur and pre-configure energy reserves and circuit parameters. This eliminates, at its root, problems such as voltage collapse and system protective emergency stops caused by instantaneous power supply-demand mismatches in traditional solutions. This allows the robot to perform explosive actions such as demolition, grinding, and de-icing smoothly and without interruption in one go, greatly improving equipment availability and task success rates in critical operational scenarios such as pumped storage and wind power maintenance.

[0019] A deeply coupled intelligent scheduling strategy based on "task-energy consumption" has been constructed, significantly optimizing overall energy efficiency and endurance. This disclosure abandons the simple linear current limiting method based solely on battery capacity, instead using the importance of specific operational scenarios and the robot's real-time load status as the core decision-making basis for energy scheduling. The system can intelligently determine when to "go all out" and when to "cruise silently," thereby ensuring the effectiveness of high-value tasks while minimizing unnecessary energy consumption. This enables robots equipped with this system to significantly extend the duration or coverage of a single operation with the same battery capacity, meeting the stringent endurance requirements of large-scale, long-distance inspections of energy facilities.

[0020] Through a learning loop of "experience accumulation - cloud collaboration," the system is endowed with the ability to continuously self-optimize, reducing long-term operation and maintenance costs. This disclosure is not merely a one-time static solution, but rather an evolving intelligent agent. The experience gained from each operation is used to optimize the local prediction model, and data from multiple devices is aggregated in the cloud for global model iteration and sharing. This means that the system's prediction accuracy and scheduling rationality will continuously improve with usage time, better adapting to the personalized modes of different power plants, wind turbines, or operators. This "gets smarter with use" characteristic avoids the drawbacks of traditional equipment requiring frequent manual debugging and maintenance, achieving cost amortization and enhanced long-term operational stability.

[0021] This invention employs a synergistic design of "energy buffering" and "computing power elasticity" to achieve flexible protection of core hardware and system-level energy efficiency improvement. On the hardware side, a supercapacitor array is introduced to handle instantaneous power spikes, allowing the battery to provide only average power, effectively reducing battery rate discharge damage and system temperature rise, and extending the lifespan of core energy components. On the software side, by dynamically decomposing and configuring algorithms, precise matching of computing resources and power consumption requirements is achieved. This integrated "power-computing power" elastic management optimizes the system's energy allocation as a whole, improving reliability while achieving green and economical operation goals.

[0022] Details of one or more embodiments of this disclosure are set forth in the following drawings and description. Other features and advantages of this disclosure will become apparent from the accompanying drawings. It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not intended to limit this disclosure. Attached Figure Description

[0023] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0024] Figure 1 This is a flowchart of a robot energy scheduling method according to an exemplary embodiment; Figure 2 This is a system architecture diagram of this exemplary embodiment; Figure 3 This is a flowchart of the process of this exemplary embodiment; Figure 4 This is a flowchart illustrating the prediction and learning process of this exemplary embodiment; Figure 5 This is a block diagram of a robot energy scheduling system according to an exemplary embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device according to an exemplary embodiment of the present invention. Detailed Implementation

[0025] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this disclosure more comprehensive and complete, and to fully convey the concept of the example embodiments to those skilled in the art. The described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided to give a full understanding of embodiments of this disclosure. However, those skilled in the art will recognize that the technical solutions of this disclosure can be practiced with one or more of the specific details omitted, or other methods, components, apparatus, steps, etc., can be employed. In other instances, well-known technical solutions are not shown or described in detail to avoid obscuring various aspects of this disclosure.

[0026] Furthermore, the accompanying drawings are merely illustrative of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware units or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0027] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0028] The terms "first," "second," etc., used in this disclosure, claims, and accompanying drawings are for distinguishing similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein.

[0029] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or device that includes a series of steps or sub-modules is not necessarily limited to those steps or sub-modules that are explicitly listed, but may include other steps or sub-modules that are not explicitly listed or that are inherent to such process, method, product, or device.

[0030] The purpose of this embodiment is to provide a robot energy dynamic scheduling system and method based on scene and ontology signal prediction, aiming to completely solve the core problems existing in the energy management system of energy operation robots, such as response lag, rigid strategy and lack of learning ability.

[0031] Figure 1 This is a flowchart of a robot energy scheduling method according to an exemplary embodiment, such as... Figure 1 As shown, an exemplary embodiment of this disclosure provides a robot energy scheduling method, including: Real-time acquisition of external scene perception signals and internal body operation status signals of the robot; Construct a high-power event prediction model; The collected scene perception signals and body operation status signals are subjected to feature extraction and fusion to generate a time-series feature vector sequence, which is then input into the high-power event prediction model to output the trigger probability of a high-power event occurring within a specified future time window. Based on the trigger probability, a corresponding energy pre-schedule instruction is generated; Before the high-power event actually occurs, the energy pre-scheduling instruction is executed to complete the robot's energy scheduling.

[0032] It is worth noting that high-power events specifically refer to the peak power demand that a robot makes on its energy system in a short period of time when performing a specific task, which is far higher than its basic operating level.

[0033] This embodiment solves the problem of operation interruption caused by passive energy supply response. By integrating environmental scene information and robot motion signals, a high-power event prediction model is constructed, enabling the prediction of peak load demand tens to hundreds of milliseconds in advance. This transforms "post-event remediation" into "pre-event preparation," fundamentally eliminating the phenomena of "operational lag" and "unannounced emergency stop" caused by instantaneous power supply and demand mismatch.

[0034] This embodiment improves the intelligence and task adaptability of the energy scheduling strategy. By deeply coupling specific operational scenario characteristics, the robot's real-time motion status, and energy scheduling decisions, it abandons the "one-size-fits-all" current limiting mode based solely on SOC (State of Charge). This allows the system to dynamically and precisely adjust power output according to the actual importance and urgency of the task, achieving on-demand energy supply that is "powerful when needed and economical when necessary."

[0035] This embodiment establishes an energy management mechanism with continuous evolution capabilities. By introducing experience learning and a cloud-based collaborative architecture, the system can continuously accumulate and optimize the "signal-energy consumption" mapping model under different operating scenarios. As the amount of operating data increases, its prediction accuracy, scheduling rationality, and cross-scenario generalization ability will continue to improve, significantly enhancing the robot's long-term operational stability and overall energy efficiency.

[0036] The implementation idea of ​​this embodiment is as follows: The core idea of ​​this embodiment is to construct a dynamic energy management closed loop of "perception-prediction-pre-scheduling-execution-learning". The system abandons the traditional passive response energy management model and instead guides the pre-configuration of energy resources through active prediction. Specifically, as follows... Figure 3 As shown, the system collects real-time information about the robot's external operating environment (visual, acoustic, environmental) and internal operating status signals (joint movement, motor load, body dynamics) from multiple sensors. An embedded intelligent predictive model analyzes this fused time-series data to determine if a high-power demand event will occur within the next few seconds. Once a high-probability, high-load operation is predicted, the system immediately triggers proactive energy scheduling. It adjusts the bus voltage in advance via a bidirectional DC-DC converter and stores and prepares the required peak energy for release using a supercapacitor array. Simultaneously, it dynamically configures the current-limiting parameters of the power driver. When the actual operation is triggered, the energy is ready, ensuring the instantaneous and stable power output. Furthermore, the system continuously optimizes the predictive model through an experience-learning mechanism, forming a positive cycle that becomes more accurate with use.

[0037] Specific implementation process: The specific implementation of this embodiment can be decomposed into three levels: system structure, data flow and workflow.

[0038] Specifically, the scene perception signal includes environmental visual information acquired by a visual sensor, environmental sound information acquired by an acoustic sensor, and temperature, humidity, and millimeter-wave detection information acquired by an environmental sensor. The body's operating status signals include joint angles, angular velocities, load torques, and fuselage inertia measurement information obtained by motion status sensors; motor phase currents and DC bus electrical parameters obtained by electrical status sensors; and operating handle and host computer control commands obtained by operating intention sensors.

[0039] like Figure 2 As shown, the system described in this embodiment mainly consists of four functional layers: a multi-source perception layer, an edge prediction and decision-making layer, a dynamic energy execution layer, and an adaptive computing resource management layer.

[0040] Multi-source sensing layer: This layer is responsible for collecting all raw signals related to energy demand prediction. Visual sensor module: Employing a fusion of binocular cameras and LiDAR, this module acquires real-time point cloud data of the working environment for identifying target objects (such as steel bars, concrete cracks, and ice), measuring distances, and assessing surface material hardness. Acoustic sensor module: Composed of a directional microphone array, this module picks up characteristic sound events in the working environment, such as "metal impact sounds," "drill bit roars," and "ice cracking sounds." The sound pressure level threshold is settable (e.g., >82 dB). This module assists in determining the type and intensity of the work based on sound patterns. Environmental sensor module: Includes temperature and humidity sensors, millimeter-wave radar, etc. The temperature and humidity sensors monitor environmental conditions (e.g., blade internal humidity >85%RH, temperature <-2℃ as one of the criteria for icing), while the millimeter-wave radar can be used for non-contact detection of vital signs or to compensate for measurement errors under extreme temperatures.

[0041] Body signal sensing unit: Motion state sensing module: The core consists of encoders and high-precision torque sensors installed at each joint of the robot to measure joint angles, angular velocities, and load torques in real time. Simultaneously, a six-axis inertial measurement unit (IMU) is installed in the robot's torso to measure the body's acceleration and angular velocity, thereby sensing the robot's overall motion trend and vibration intensity.

[0042] Electrical status sensing module: This includes the phase current sampling circuit for the motor driver and the voltage and current sampling circuit for the DC bus. These signals directly reflect the real-time power demand of the motor and the energy throughput status of the system.

[0043] Operation Intent Sensing Module: Reads raw control signals (such as throttle opening and lever angle) from the operating handle or host computer commands, as a direct precursor signal to determine that the operator's intention is about to be converted into a high-power action.

[0044] Specifically, a high-power event prediction model is constructed, including: During the robot's task execution, a time-series feature vector sequence, the identifier of a high-power event that actually occurs after the time-series feature vector sequence, and the actual energy consumption curve corresponding to the high-power event are synchronously collected and recorded. The collected time-series feature vector sequence, the high-power event identifier, and the actual energy consumption curve are bound together to generate an experience record, and the experience record is stored in the experience memory bank. Based on the experience records stored in the experience memory bank, the long short-term memory network model is trained by using the time-series feature vector sequence as input and the corresponding high-power event identifier or occurrence probability as the learning target, thereby obtaining a high-power event prediction model.

[0045] Specifically, the expression for the high-power event prediction model is: P_trigger = LSTM_Model(F(tn), F(t-n+1), …, F(t)); In the formula, t represents the current timestamp; F(t) represents the temporal feature vector at the current timestamp t; n represents the historical time step; Δt represents the length of the predicted future time window; P_trigger represents the trigger probability of a high-power event; and LSTM_Model(·) represents the forward inference process of the long short-term memory network model.

[0046] Specifically, the energy pre-scheduling command includes adjustment commands for DC bus voltage, energy storage unit reserve energy, and motor driver current limit value.

[0047] Specifically, based on the trigger probability, a corresponding energy pre-scheduling instruction is generated, including: When the trigger probability is greater than or equal to the first threshold, a first-level pre-scheduling instruction is generated. The first-level pre-scheduling instruction includes raising the bus voltage to the first target voltage, charging the energy storage unit to the first reserve energy value, and setting the motor driver current limit value to the maximum value. When the trigger probability is less than the first threshold and greater than or equal to the second threshold, a secondary pre-schedule instruction is generated. The secondary pre-schedule instruction includes raising the bus voltage to a second target voltage lower than the first target voltage, charging the energy storage unit to a second reserve energy value lower than the first reserve energy value, and setting the motor driver current limit value to an intermediate limit value between the normal limit value and the maximum value. When the trigger probability is less than the second threshold, the normal energy scheduling mode is maintained.

[0048] This embodiment discloses an edge prediction and decision layer, which is deployed in the edge computing unit of the robot body (such as a platform based on the RK3588 chip) and is responsible for signal processing, feature extraction, model inference, and decision generation. The signal fusion and feature extraction module preprocesses (filters, normalizes) and extracts features from the raw data from the perception layer. For example, it extracts the rate of change and trend from joint torque signals, the size and contour features of the target object from the visual point cloud, and the energy of specific frequency bands from the sound signal. Finally, these heterogeneous features are fused into a unified multi-dimensional temporal feature vector.

[0049] F(t) = [f_visual(t), f_audio(t), f_torque(t), f_current(t), f_IMU(t),...]; Where t represents the current timestamp, and F(t) represents the temporal feature vector at the current timestamp t; f_ The following represent the feature values ​​extracted from each dimension: f_visual(t) represents the feature value extracted from the visual sensor data at the current timestamp t; f_audio(t) represents the feature value extracted from the acoustic sensor data at the current timestamp t; f_torque(t) represents the feature value extracted from the joint torque sensor data at the current timestamp t; f_current(t) represents the feature value extracted from the electrical state sensor data at the current timestamp t; and f_IMU(t) represents the feature value extracted from the inertial measurement unit (IMU) data at the current timestamp t.

[0050] Experience Memory Bank: A lightweight embedded database (such as SQLite) is used to store historical work experience. Each experience record is a triple: Experience Record = {Feature Vector Sequence F_history, Corresponding High-Power Event Label E_label, Actual Energy Consumption Curve P_curve}. For example, E_label could be "dismantling", "grinding", "de-icing", etc. These records constitute the sample library for the system to perform similarity matching and model training. High-Power Event Prediction Module: This is the "intelligent brain" of the system. Its core is a lightweight temporal prediction model, preferably a Long Short-Term Memory network. This model takes the temporal feature vector F(tn : t) generated in the previous stage as input and outputs the probability of a high-power event occurring within a future time window Δt (e.g., 5 seconds), i.e., the trigger probability P_trigger of the high-power event.

[0051] P_trigger = LSTM_Model ( F(tn), F(t-n+1), ..., F(t) ) Where n is the historical time step. The model's training data comes from an experience memory database, and the learning objective is to establish a mapping relationship from F_history to E_label.

[0052] Dynamic scheduling decision module: Based on the value of P_trigger, it generates specific execution instructions according to a preset strategy. If P_trigger ≥ θ_high (first threshold, e.g., 0.8), it generates first-level pre-scheduling instructions, including target bus voltage V_target, supercapacitor pre-charge E_cap, and motor driver current limit I_limit_max. If θ_mid (second threshold, e.g., 0.6) ≤ P_trigger < θ_high (first threshold), it generates second-level pre-scheduling instructions, performing pre-adjustments on some parameters. If P_trigger < θ_mid, it maintains the normal scheduling mode.

[0053] Specifically, before the high-power event actually occurs, the energy pre-scheduling instruction is executed to complete the robot's energy scheduling, including: By using a bidirectional DC-DC converter, the DC bus voltage is adjusted to the target voltage value before a high-power event occurs, according to the energy dispatch command. By controlling the bidirectional DC-DC converter or an independent charging circuit, the supercapacitor array connected in parallel with the DC bus is pre-charged according to the energy dispatch command to store reserve energy value. The current loop limit value of the motor driver is dynamically configured to the limit value specified by the energy dispatch command.

[0054] This embodiment discloses a dynamic energy execution layer. This layer is the physical execution mechanism for decision-making, responsible for converting energy commands into actual voltage and current outputs. The energy buffer and fast response unit includes: a bidirectional DC-DC converter connected to the main battery pack and the DC bus. Receiving the V_target command from the decision layer, it can raise the bus voltage from the nominal value (e.g., 48V) to the target value (e.g., 54V) within 200ms, or quickly pull it back after the operation ends. Its "bidirectional" characteristic allows excess energy to be fed back to the battery or supercapacitor during braking or idle periods. A supercapacitor array is connected in parallel with the DC bus. As the system's "power depot," its charge and discharge rate is much higher than that of the battery. During the pre-scheduling phase, it rapidly charges and stores energy according to the E_cap command; during high-power bursts, it undertakes the main instantaneous power output and smooths the battery discharge current.

[0055] The configurable power drive unit, i.e., the intelligent motor driver, no longer has a fixed current loop limit value I_limit. Instead, it is dynamically set by receiving the I_limit_max parameter (the maximum current limit value of the motor driver) from the decision layer. The limit is unlocked during the pre-scheduling phase to prepare for peak current output.

[0056] Specifically, the robot energy scheduling method further includes: like Figure 4 As shown, after each task is completed, the feature sequence of the scene perception signal and the body running status signal, the actual high-power event identifier and the corresponding actual energy consumption curve during the task are stored as a piece of empirical data in the empirical memory bank, which is used to update the high-power event prediction model.

[0057] Specifically, the robot energy scheduling method further includes: Periodically synchronize the experience data in the experience memory bank to the cloud server; The cloud server aggregates experience data from multiple robots to train and optimize the global model parameters of the high-power event prediction model. The optimized model parameters are distributed to each robot to update the local high-power event prediction model of each robot.

[0058] This embodiment discloses an adaptive computing resource management layer: to collaboratively achieve overall energy efficiency optimization, this system also flexibly manages the robot's task computing resources. The main algorithm (such as a large CNN vision model) responsible for high-power computing tasks such as environmental understanding and human-computer interaction is designed with a dynamically pruned structure. For example, it is divided into a core feature extractor and an optional higher-order inference module. According to the current energy mode of the system (such as burst mode or cruise mode), the computational precision of the algorithm is dynamically adjusted (such as using 8-bit quantization) and the depth of the activated model (such as skipping certain network layers or using an Early-exit strategy), thereby significantly reducing the computing power consumption during standby while ensuring basic functionality.

[0059] Example: Taking the grinding operation of the inclined shaft flow channel of a pumped storage power station as an example: Step 1: Experience Accumulation (Learning Phase) During the first three days of debugging, whenever the operator prepared to grind the rough concrete surface, the system recorded the following scene signals: visual recognition of pits with a diameter greater than 40mm and acoustic recognition of a "chisel-striking" characteristic sound. Body signals: the joint torque sensor detected a continuous pressure trend in the manipulator arm, and the IMU detected a specific vibration pattern in the body due to contact. Energy results: the electric spindle then started, and the load power jumped to 2.2kW within 2 seconds, lasting for approximately 4 seconds. The system bound this time-correlated signal sequence to the energy consumption curve, encoded it as the "inclined well grinding" scene template T_grind, and stored it in the local experience memory.

[0060] Step 2: Real-time perception and feature fusion (online operation) On the fourth day of formal operation, the robot cruised along the inclined shaft (low-power mode), with the perception layer continuing to operate. Vision again detected a similar indentation, and the joint torque sensors showed that the torque at each joint was increasing slowly and steadily, indicating that the robotic arm was approaching the work surface and preparing to load. The acoustic module detected a specific resonant frequency in the ambient background sound.

[0061] Step 3: Forecasting and Decision Making The edge prediction layer calculates the similarity (e.g., cosine similarity) between the temporal feature vector F_current from the current period (approximately 30 seconds) and the template T_grind in the experience base, and simultaneously inputs it into the LSTM (Long Short-Term Memory) model. After comprehensive evaluation, the model outputs P_trigger = 0.88.

[0062] Since P_trigger > 0.8, the dynamic scheduling decision module immediately generates a first-level pre-scheduling instruction: {V_target: 54V, E_cap: 130J, I_limit_max: 100%}.

[0063] Step 4: Energy Pre-scheduling The command is issued to the energy execution layer, and the bidirectional DC-DC converter raises the bus voltage from 42V to 54V within 150ms. Simultaneously, the excess power charges the 100F / 48V supercapacitor array, allowing it to store an additional 130J of energy. The current limit of the electric spindle driver is set to its maximum value.

[0064] Step 5: High Peak Power Execution The operator confirms the location and starts the grinding program. The electric spindle instantly demands 2.2kW of power, and the large current required at this time is quickly provided by the supercapacitor, which has already stored sufficient charge. Because the bus voltage has been pre-raised, even if the voltage drops due to capacitor discharge (e.g., from 54V to 47V), it is still far above the system hardware protection threshold (e.g., 42V). Therefore, the grinding process is smooth, uninterrupted, and completed in one go.

[0065] Step Six: State Restoration and Learning Loop After the grinding process ended, the system's signals showed that torque and current had returned to low values. The system detected that the P_trigger had dropped to a low level and issued a command to adjust the DC-DC converter back to the 42V cruising voltage. The computing resource management layer switched the vision processing network to a low-power "Early-exit" mode. New "signal-energy consumption" data from this operation was recorded, encrypted, and uploaded to the cloud server during off-peak hours at night. This data was used to aggregate data from hundreds of robots, train a more accurate global prediction model, and update and distribute it to each edge device via OTA (Over-the-Air) updates, completing a learning loop.

[0066] Essential core structure: Body signal sensing units (especially joint torque / encoder and motor current sampling): These are fundamental to achieving "predictability." Without real-time and accurate sensing of the robot's own motion state and load, prediction loses its most important input source.

[0067] High-power event prediction algorithm module (implemented as an embedded, runnable LSTM or other lightweight timing model): This is the core of the intelligent decision-making of this invention, undertaking the key function of inferring future needs from historical and current data. Without it, the system reverts to a passive reaction mode.

[0068] Energy buffering and fast-response hardware units (including at least a supercapacitor array and a DC-DC converter with fast response capability): These are the physical guarantees for translating predictions into action and solving the "time lag" problem. Without them, prediction results cannot be converted into usable energy reserves in a timely manner.

[0069] Enhancement structures for achieving better results: Cloud-based collaborative learning architecture (including cloud-based model training platform and secure OTA update mechanism): This structure is not necessary for single-machine operation, but it can significantly improve the generalization ability and evolution speed of prediction models by aggregating data from multiple devices and multiple scenarios. It is a key enhancement to achieve "the more you use it, the smarter it becomes".

[0070] Multimodal scene perception fusion (such as millimeter-wave radar and high-precision temperature and humidity sensors): Adding these modules to the core visual and acoustic perception can further improve the accuracy of scene recognition and prediction in complex and harsh environments (such as dense fog and low temperature).

[0071] A fine-grained, elastic computing management framework: Large algorithms are broken down into core and optional components, and dynamic switching between multiple precision / power levels is implemented. This is an optimization layer designed to work in conjunction with energy scheduling to maximize system-level energy efficiency, and can be used as an optional optimization on resource-constrained platforms.

[0072] Advanced energy recovery and reuse circuits (such as more precise regenerative braking control and micro-devices that utilize waste heat to generate electricity): These designs further explore energy-saving potential on the basis of basic bidirectional DC-DC converters, representing enhanced designs that raise the upper limit of energy efficiency.

[0073] The key point of this embodiment is: The core of this implementation is a feedforward control mechanism based on "scene and body signal fusion prediction". The essential breakthrough of this embodiment lies in changing the triggering logic of energy scheduling. Its core is to use an algorithm model to deeply fuse and analyze the external environment perception with the robot's internal motion state, operation intention and other body signals, to actively predict energy demand, rather than passively responding to changes in voltage and current.

[0074] The system employs a millisecond-level rapid response link for "prediction-pre-scheduling". At the software level, it forms an efficient instruction flow from the prediction module to the scheduling decision module. At the hardware level, it relies on a fast energy buffer channel composed of bidirectional DC-DC converters and supercapacitors. This enables the system to complete bus voltage adjustment and energy preparation in a very short time after prediction confirmation, filling the critical time window from "foresight" to "execution".

[0075] The continuous learning loop of "experience memory - cloud evolution" transforms each task process into a reusable experience template, and leverages cloud computing to achieve cross-device and cross-scenario knowledge aggregation and model optimization. This enables the system to break free from the limitations of fixed strategies and possess the ability to evolve to adapt to complex and ever-changing on-site conditions, which is the foundation for achieving long-term intelligence.

[0076] Flexible and coordinated management of the "power link" and the "computing link". It innovatively incorporates the resource scheduling of high-power computing tasks into the overall energy management framework, so that the computing power configuration and power output level change in tandem, achieving optimal energy efficiency of the entire system from energy supply to information processing, rather than energy saving in a single link.

[0077] Figure 5 This is a block diagram of a robot energy scheduling system according to an exemplary embodiment of the present invention, such as Figure 5 As shown, an exemplary embodiment of this disclosure provides a robot energy scheduling system, including: The acquisition unit collects external scene perception signals and internal body operation status signals of the robot in real time; Building blocks are used to construct high-power event prediction models; The prediction unit is used to extract and fuse features from the collected scene perception signals and the body's operating status signals, generate a time-series feature vector sequence, input it into the high-power event prediction model, and output the trigger probability of a high-power event occurring within a specified future time window. The generation unit is used to generate a corresponding energy pre-scheduling instruction based on the triggering probability; The scheduling unit is used to execute the energy pre-scheduling instruction before the high-power event actually occurs, and to complete the robot energy scheduling.

[0078] Figure 6 This is a schematic diagram of the structure of an electronic device according to an exemplary embodiment of this invention. Figure 6 As shown, corresponding to the robot energy scheduling method provided above, this disclosure also provides an electronic device. Since the embodiment of this device is similar to the method embodiment described above, the description is relatively simple; relevant details can be found in the description of the method embodiment section above. The device described below is merely illustrative. This device may include: a processor 1, a memory 2, a communication bus (i.e., the aforementioned device bus), and a lookup engine. The processor 1 and memory 2 communicate with each other via the communication bus and communicate with external systems via a communication interface. The processor 1 can call logical instructions in the memory 2 to execute the robot energy scheduling method.

[0079] Furthermore, the logical instructions in the aforementioned memory 2 can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as memory chips, USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0080] On the other hand, this disclosure also provides a processor-readable storage medium storing a computer program 3, which, when executed by a processor 1, is implemented to perform the robot energy scheduling method provided in the above embodiments.

[0081] The processor-readable storage medium can be any available medium or data storage device that the processor 1 can access, including but not limited to magnetic memory (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical memory (e.g., CD, DVD, BD, HVD), and semiconductor memory (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0082] The above are merely preferred embodiments of this disclosure. The scope of protection of this disclosure is not limited to the above embodiments. All technical solutions falling within the scope of this disclosure are protected. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this disclosure should be considered within the scope of protection of this disclosure.

Claims

1. A robot energy scheduling method, characterized in that, include: Real-time acquisition of external scene perception signals and internal body operation status signals of the robot; Construct a high-power event prediction model; The collected scene perception signals and body operation status signals are subjected to feature extraction and fusion to generate a time-series feature vector sequence, which is then input into the high-power event prediction model to output the trigger probability of a high-power event occurring within a specified future time window. Based on the trigger probability, a corresponding energy pre-schedule instruction is generated; Before the high-power event actually occurs, the energy pre-scheduling instruction is executed to complete the robot's energy scheduling.

2. The robot energy scheduling method according to claim 1, characterized in that: The scene perception signals include environmental visual information acquired by a visual sensor, environmental sound information acquired by an acoustic sensor, and temperature, humidity, and millimeter-wave detection information acquired by an environmental sensor. The body's operating status signals include joint angles, angular velocities, load torques, and fuselage inertia measurement information obtained by motion status sensors; motor phase currents and DC bus electrical parameters obtained by electrical status sensors; and operating handle and host computer control commands obtained by operating intention sensors.

3. The robot energy scheduling method according to claim 1, characterized in that, Constructing a high-power event prediction model includes: During the robot's task execution, a time-series feature vector sequence, the identifier of a high-power event that actually occurs after the time-series feature vector sequence, and the actual energy consumption curve corresponding to the high-power event are synchronously collected and recorded. The collected time-series feature vector sequence, the high-power event identifier, and the actual energy consumption curve are bound together to generate an experience record, and the experience record is stored in the experience memory bank. Based on the experience records stored in the experience memory bank, the long short-term memory network model is trained by using the time-series feature vector sequence as input and the corresponding high-power event identifier or occurrence probability as the learning target, thereby obtaining a high-power event prediction model.

4. The robot energy scheduling method according to claim 1, characterized in that, The expression for the high-power event prediction model is: P_trigger = LSTM_Model(F(tn), F(t-n+1), …, F(t)); In the formula, F(t) represents the temporal feature vector at the current timestamp t; n represents the historical time step; Δt represents the length of the predicted future time window; P_trigger represents the trigger probability of a high-power event; and LSTM_Model(·) represents the forward inference process of the long short-term memory network model.

5. The robot energy scheduling method according to claim 1, characterized in that, The energy pre-scheduling instructions include adjustment instructions for DC bus voltage, energy storage unit reserve energy, and motor driver current limit values.

6. The robot energy scheduling method according to claim 1, characterized in that, Based on the trigger probability, a corresponding energy pre-schedule instruction is generated, including: When the trigger probability is greater than or equal to the first threshold, a first-level pre-scheduling instruction is generated. The first-level pre-scheduling instruction includes raising the bus voltage to the first target voltage, charging the energy storage unit to the first reserve energy value, and setting the motor driver current limit value to the maximum value. When the trigger probability is less than the first threshold and greater than or equal to the second threshold, a secondary pre-schedule instruction is generated. The secondary pre-schedule instruction includes raising the bus voltage to a second target voltage lower than the first target voltage, charging the energy storage unit to a second reserve energy value lower than the first reserve energy value, and setting the motor driver current limit value to an intermediate limit value between the normal limit value and the maximum value. When the trigger probability is less than the second threshold, the normal energy scheduling mode is maintained.

7. The robot energy scheduling method according to claim 1, characterized in that, Before the high-power event actually occurs, the energy pre-scheduling instruction is executed to complete the robot's energy scheduling, including: By using a bidirectional DC-DC converter, the DC bus voltage is adjusted to the target voltage value before a high-power event occurs, according to the energy dispatch command. By controlling the bidirectional DC-DC converter or an independent charging circuit, the supercapacitor array connected in parallel with the DC bus is pre-charged according to the energy dispatch command to store reserve energy value. The current loop limit value of the motor driver is dynamically configured to the limit value specified by the energy dispatch command.

8. The robot energy scheduling method according to claim 1, characterized in that, Also includes: After each task is completed, the robot stores the feature sequence of the scene perception signal and the body operation status signal, the actual high-power event identifier and the corresponding actual energy consumption curve as a piece of empirical data in the empirical memory bank, which is used to update the high-power event prediction model.

9. The robot energy scheduling method according to claim 8, characterized in that, Also includes: Periodically synchronize the experience data in the experience memory bank to the cloud server; The cloud server aggregates experience data from multiple robots to train and optimize the global model parameters of the high-power event prediction model. The optimized model parameters are distributed to each robot to update the local high-power event prediction model of each robot.

10. A robot energy scheduling system, characterized in that, include: The acquisition unit collects external scene perception signals and internal body operation status signals of the robot in real time; Building blocks are used to construct high-power event prediction models; The prediction unit is used to extract and fuse features from the collected scene perception signals and the body's operating status signals, generate a time-series feature vector sequence, input it into the high-power event prediction model, and output the trigger probability of a high-power event occurring within a specified future time window. The generation unit is used to generate a corresponding energy pre-scheduling instruction based on the triggering probability; The scheduling unit is used to execute the energy pre-scheduling instruction before the high-power event actually occurs, and to complete the robot energy scheduling.

11. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer programs, and the processor calls the computer programs stored in the memory to execute the robot energy scheduling method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, enables the processor to perform the robot energy scheduling method according to any one of claims 1 to 9.

Images