A method and system for thermal management of humanoid robot joints based on an agent model
By adopting a thermal management method based on a surrogate model, the problem of the difficulty in incorporating the thermal protection state of humanoid robot joints into motion planning is solved. This method enables early assessment and unified expression of thermal risks, avoids passive handling, and improves the predictive accuracy and reliability of thermal management.
Patent Information
- Application Number
- CN202611619198.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Humanoid robot joints are difficult to predict thermal risks under load. Existing technologies cannot uniformly express the thermal state of different actuators, and it is difficult to provide early warnings when the future load is unknown, resulting in passive handling of thermal protection mechanisms.
A surrogate model-based thermal management method is adopted. By collecting joint operation data, an appropriate I2t surrogate model or winding temperature surrogate model is selected to recursively predict the thermal state. When the future load is unknown, the time when the thermal state reaches the protection threshold and the threshold arrival time are predicted in a rolling manner, providing thermal margin and thermal state constraints.
It enables the assessment of thermal risks before trajectory planning, avoids passive handling after protection triggering, unifies the expression of thermal risks of different actuators, reduces prediction bias caused by differences in initial conditions, and provides lead time for deceleration, task switching, and planning.
Smart Images

Figure CN123103709A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of humanoid robot control technology, specifically relating to a method and system for thermal management of humanoid robot joints based on a proxy model. Background Technology
[0002] In tasks such as continuous handling, loading and unloading, sewing, rapid teleoperation, and human-robot interaction, humanoid robots often experience joints simultaneously subjected to cyclic loads and instantaneous peak loads. To prevent overheating of motor windings, power devices, and internal thermal components of the actuators, joint actuators typically rely on the cumulative effect of the square of the current over time (referred to as Ik). 2 t) Trigger driver protection; when the internal heat load reaches the threshold, the driver may limit current, reduce derating, or in severe cases, trigger the enable.
[0003] Existing thermal protection mechanisms primarily operate within the actuator, suitable for low-level safety protection, but still present information gaps on the robot task planning side: on the one hand, motion libraries or planners may not be able to directly read normalized IT. 2 On the one hand, even if some temperature or analog values can be read, it is often difficult to establish a stable mapping with the actual winding temperature, the moment of de-enabling, and the feasibility of the trajectory. As a result, the planner can only passively process the protection after it is triggered, making it difficult to assess thermal risks in advance during long-term high-load tasks.
[0004] Based on existing publicly available solutions, some joint motor performance optimization methods use temperature as an observation for performance optimization or reinforcement learning, and adjust motion strategies according to temperature, such as the joint motor performance optimization method disclosed in Chinese patent application CN122546678A. These solutions focus on the role of temperature in motion performance optimization, but typically do not address the actual joint torque reconstruction actuator I. 2 The percentage of t is not calculated, nor is the remaining time to reach the protection threshold calculated, making it difficult to directly output the thermal margin to the trajectory planner.
[0005] Other predictive thermal management solutions target vehicle electric drive systems or power battery systems. They collect current, power, speed, and temperatures at multiple locations to predict future thermal states and adjust them through cooling components or thermal management strategies. Examples include the predictive thermal management method for an electric drive system of an electric light truck for urban logistics disclosed in Chinese patent application CN122607091A, and the active thermal management system and control method for a power battery disclosed in Chinese patent application CN112477698A. However, the objects, thermal boundaries, actuators, and control objectives of these solutions differ significantly from those of humanoid robot joints, and they cannot directly solve the problems of inconsistent actuator interfaces and I / O characteristics in small joints. 2 The protection status cannot be directly used for planning, and there are issues such as the unpredictable future load in remote operation scenarios.
[0006] In practical humanoid robot platforms, the differences in actuator interfaces are even more pronounced. For example, some actuators can reliably provide actual joint torques, but the motion library side cannot directly obtain I / O values that can be used for planning. 2 In the t-state, the temperature of another part of the driver may not directly reflect the actual winding temperature, and can only be obtained through analog signals or driver-side signals. Therefore, it is necessary to establish a thermal state proxy mechanism that automatically selects based on available signals, and uniformly expresses the thermal risks of different drivers as the current thermal state, remaining thermal margin, and threshold arrival time. Summary of the Invention
[0007] This invention addresses the problems in existing technologies, such as the difficulty in incorporating joint thermal protection states into motion planning and online control of humanoid robots, inconsistent available signals for different actuators, difficulty in providing early warnings when future loads are unknown, and large prediction deviations across operating conditions due to initial temperature differences. It provides a humanoid robot joint thermal management method and system based on a proxy model.
[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for thermal management of humanoid robot joints based on a proxy model, comprising: Step S1: Collect joint operation data, which includes actual joint torque, timestamp, and sampling period; Step S2: Based on the actuator type and the available signals from the joint operation data acquired in Step S1, select I. 2 t-surrogate model or winding temperature surrogate model; Step S3: Based on the actual joint torque, recursively deduce the thermal state using the selected surrogate model; Step S4: When the future load is unknown, the thermal state obtained by recursion in step S3 is used as the initial state for prediction. Short-term future load assumptions are constructed based on the historical actual joint torque window. The estimated remaining time and estimated threshold arrival time of the thermal state are predicted in a rolling manner according to the preset update cycle. The thermal margin is calculated based on the current thermal state and the protection threshold. Step S5: Based on at least one of the predicted thermal state, thermal margin, estimated remaining time, and estimated threshold arrival time, perform at least one of the following operations: trajectory filtering, cycle time optimization, online speed reduction, and task switching, or perform a simulated thermal management closed loop.
[0009] Furthermore, in step S2, for the actual joint torque that can be obtained but cannot be directly obtained for planning, I... 2 For joints in state t, select I. 2t-surrogate model; for joints where actual joint torque and winding temperature observations can be obtained, a first-order winding temperature surrogate model based on actual joint torque is selected.
[0010] Further, in step S3, the I 2 The t-surrogate model uses the rated torque to dimensionlessly approximate the actual joint torque, resulting in a normalized load: ; Where k represents the discrete sampling number, The actual joint torque at the k-th sampling point. The rated torque corresponding to this joint is given; and a recursive formula with boundary trimming is used: ; Recursive Normalization I 2 t protection state, where, Normalized I for the k-th sampling point 2 t heat load percentage, Normalized I for the (k-1)th sampling point 2 t heat load percentage, The heat accumulation coefficient is negative. This is the equivalent balance torque ratio when there is no net heat accumulation. This represents the actual time interval between adjacent sampling points. This indicates that the state is limited to the range of 0% to 100%.
[0011] Further, in step S3, the I 2 Parameters of the t proxy model and Through actual measurement of the driver I 2 t is used as a supervisory signal for identification: ; in, Actual test of driver I 2 t, Indicates the actual parameters. To identify the optimal parameters. The value of parameter θ is used to represent the k-th sampled data of p; the identification method adopts bounded nonlinear least squares optimization and performs cross-validation and median fusion on multiple sets of working condition data.
[0012] Furthermore, in step S3, the winding temperature proxy model adopts a first-order thermal inertia element: ; Where t is the continuous physical time, representing the sampling time corresponding to the k-th sampling point. Let be the winding temperature at time t. The actual joint torque at time t. This is the equivalent conversion factor from the square of the torque to the winding temperature rise rate. The equivalent thermal time constant of the winding is... This is the equivalent environmental thermal boundary corresponding to the j-th running process.
[0013] Furthermore, in step S3, the winding temperature proxy model adopts the following discrete recursive relationship: ; ; in, Let the winding temperature be the temperature at the k-th sampling point. For steady-state temperature, The winding temperature is the (k-1)th sampling point.
[0014] Furthermore, in step S3, the initial temperature and equivalent environmental thermal boundary are set for the j-th running process: ; ; in, The winding temperature observation value obtained at the start of the j-th operation process. This is the correction amount for the environmental thermal boundary corresponding to this operation process.
[0015] Furthermore, in step S3, for the temperature proxy model, the parameters include the torque heating coefficient. Thermal time constant And the environmental thermal boundary correction amount corresponding to each operating process. Parameter identification is performed using multiple sets of synchronously acquired actual joint torque and winding temperature data, with the error between the predicted temperature and the measured winding temperature serving as the optimization objective. For example, the following objective function can be used: ; in, The set of parameters to be identified The model predicts temperature. Model parameters can be obtained using a bounded nonlinear least squares method, and physical constraints are set on the parameters based on the torque-temperature rise relationship, thermal time constant, and a reasonable temperature range to reduce the impact of non-physical factors on the prediction results.
[0016] Furthermore, in step S4, for those using I 2 The t-surrogate model's joints are used to construct future normalized loads based on historical actual joint torques. and with the current I 2 t state As the initial state for prediction, calculate I 2 t-forecasted growth rate: ; when and At that time, it is expected to reach I 2 The remaining time for the protection threshold t is: ; The corresponding expected arrival time of the protection threshold is: ; Furthermore, in step S4, for joints employing a winding temperature proxy model, the current winding temperature is used as the reference. As the initial state for temperature prediction, the root mean square value of the actual joint torque within the historical window is used as the future short-term equivalent torque: ; in, This represents the number of sampling points within the historical window. The actual joint torque is used; and the estimated remaining time to reach the temperature protection threshold is calculated using analytical solutions: ; in, This is the temperature protection threshold. This is the equivalent steady-state temperature.
[0017] Furthermore, in step S5, the current thermal state, predicted thermal state, thermal margin, and estimated remaining time are used as motion planning constraints for candidate trajectory selection, speed adjustment, beat optimization, task switching, and closed-loop update of the actual-simulation thermal state.
[0018] Secondly, the present invention also provides a humanoid robot joint thermal management system based on a surrogate model, including a data acquisition module, a actuator type and available signal identification module, a surrogate model module, a temperature boundary correction module, an online rolling prediction module, and an output module; the data acquisition module is used to acquire actual joint torque, timestamps, and sampling periods; the actuator type and available signal identification module is used to determine available signals based on actuator model, object dictionary, analog interface, and data quality; the surrogate model module includes I 2 t-surrogate model and winding temperature surrogate model are used to select I based on the driver type and available signal. 2The module provides a proxy model or a winding temperature proxy model; the temperature boundary correction module is used to correct the initial temperature and the equivalent environmental thermal boundary; the online rolling prediction module is used to use the thermal state obtained by the proxy model module as the initial prediction state, and when the future load is unknown, it constructs a short-term future load hypothesis based on the historical actual joint torque window, and predicts the expected remaining time and expected threshold arrival time of the thermal state according to a preset update cycle; the output module is used to output at least one of the current thermal state, predicted thermal state, thermal margin, expected remaining time and expected threshold arrival time.
[0019] Furthermore, the system also includes a trajectory planning constraint module and a simulation closed-loop module; the trajectory planning constraint module is used to perform at least one of trajectory filtering, cycle time optimization, online speed reduction, and task switching based on thermal margin; the simulation closed-loop module is used to use real machine feedback to correct proxy parameters, initial state, or boundary bias, so that the thermal state of the planning side and the real machine side remains consistent.
[0020] Compared with the prior art, the advantages of the present invention are as follows: (i) This invention transforms thermal state, threshold arrival time and thermal margin into trajectory screening, cycle time optimization and simulation constraints, enabling the planner to assess thermal risks before trajectory execution and avoid passive processing after protection is triggered.
[0021] (ii) Based on the actual joint torque recursive thermal state, this invention converts the joint torque information naturally existing in motion planning and dynamic calculation into a thermal state that can be used for thermal management, and establishes a mapping between the planning side and the thermal protection state without directly using current as model input.
[0022] (iii) This invention selects different proxy models by driver type and signal availability, and provides a unified thermal state output interface to express the thermal risks of different drivers as the current thermal state, thermal margin and threshold arrival time.
[0023] (iv) The present invention independently sets the initial temperature and equivalent environmental thermal boundary correction for each operation process, thereby reducing the prediction deviation caused by differences in initial conditions and changes in heat dissipation conditions.
[0024] (v) This invention utilizes historical actual joint torque windows to construct short-term future load assumptions and predicts I online. 2 The estimated time and remaining time for the winding temperature to reach the corresponding thermal protection threshold can still provide a risk time estimate when the future load is unknown, providing lead time for deceleration, task switching and replanning. Attached Figure Description
[0025] Figure 1This is a diagram of the overall structure of a joint thermal management prediction system. In the diagram: 1-Joint operation data acquisition module; 2-Actuator type and available signal identification module; 3-I 2 1- Proxy model module; 2- Winding temperature proxy model module; 3- Temperature boundary correction module; 4- Online rolling prediction module; 5- Thermal protection risk assessment module; 6- Trajectory planning constraint module; 7- Early warning display and control interface module; Figure 2 For torque type I 2 Flowchart of the t-proxy model calculation; Figure 3 This is a structural diagram of a first-order temperature proxy model based on actual joint torques; Figure 4 A structural diagram for predicting the online rolling thermal state under unknown future loads; Figure 5 The flowchart of the closed-loop simulation trajectory constraint driven by thermal state prediction is shown. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the invention and not to limit it. Those skilled in the art will recognize that the invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the invention by illustrating examples of it.
[0027] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0028] Example 1: Method Example.
[0029] For ease of explanation, in this embodiment, t represents continuous physical time, k represents discrete sampling number, and t represents the actual time corresponding to the k-th sampling point. This represents the actual time interval between adjacent sampling points. When the sampling period remains constant, it can be denoted as... , and have For continuous-time models, the following is adopted: States or inputs are represented in the form of [variable name]; for discrete recursive calculations in the controller, [variable name] is used. The state or input corresponding to the k-th sampling point can be represented in the form of , etc.
[0030] This embodiment provides a method for thermal management of humanoid robot joints based on a proxy model, including the following steps: Step S1: Collect joint operation data.
[0031] Joint operation data includes actual joint torque, timestamps, and sampling period. Actual joint torque is obtained from feedback from the actuator side, timestamps are provided by the controller or actuator, and the sampling period is determined based on the control cycle, such as 1ms, 5ms, or 10ms. In joints where winding temperature observations are available, the joint operation data also includes the winding temperature observations; in joints where the internal thermal state of the actuator is available, the joint operation data also includes the internal I / O of the actuator. 2 t represents a percentage or a simulated temperature value. Data acquisition is performed periodically according to a preset sampling period. A set of data is recorded at each sampling time, along with a data quality flag to indicate whether the data at that sampling time is valid, abnormal, or missing.
[0032] The principle behind this step is that the actual joint torque is a physical quantity that naturally exists in motion planning and dynamics calculations, and it does not depend on the internal current of the actuator or the estimated torque. Therefore, it can be used as a universal input between the planning side and the thermal protection state. The timestamp is used for time alignment of subsequent multi-channel data. The sampling period determines the discrete time step of the thermal state recursion. The data quality flag can avoid abnormal data from causing errors in the surrogate model recursion and improve the reliability of prediction.
[0033] Step S2: Based on the actuator type and the available signals from the joint operation data acquired in Step S1, select I. 2 t proxy model or winding temperature proxy model.
[0034] Specifically, the available signals are determined based on the driver model, object dictionary, analog interface, and data quality. If the driver can provide actual joint torque but cannot directly provide I / O signals usable for planning... 2 In state t, then select I. 2 A surrogate model is selected based on the actual joint torque and winding temperature observations if the actuator can provide these values. For different joints within the same robot, different surrogate models can be selected based on their actuator type and available signals. The selection results are recorded as model identifiers for use in subsequent steps.
[0035] The principle behind this step is that different drivers have different available signals. By selecting a surrogate model based on the available signals, the thermal risks of different drivers can be uniformly expressed as the current thermal state, thermal margin, and threshold arrival time, thereby solving the problems of inconsistent driver interfaces and I / O.2 The protection state cannot be directly used for planning. At the same time, two proxy models are used to adapt to two types of typical driver signal conditions, avoiding the problem that a single model cannot cover multiple drivers.
[0036] Step S3: Based on the actual joint torque, recursively deduce the thermal state using the selected surrogate model.
[0037] When selecting I 2 When using the t-surrogate model, the actual joint torque is dimensionless based on the rated torque to obtain the normalized load: ; in, The actual joint torque at the k-th sampling point. This is the rated torque corresponding to the joint. Then, the normalized I is calculated using a recursive formula with boundary trimming. 2 t protection status: ; in, Normalized I for the k-th sampling point 2 t heat load percentage, The heat accumulation coefficient is negative. This is the equivalent balance torque ratio when there is no net heat accumulation. The sampling interval is... This indicates that the state is limited to the range of 0% to 100%. and Through actual measurement of the driver I 2 t is used as a supervisory signal for identification: ; in, Actual test of driver I 2 The identification method employs bounded nonlinear least squares optimization and performs cross-validation and median fusion on multiple sets of working condition data. Initial state recursion is then performed. Take 0 or according to the current I of the driver 2 t state setting.
[0038] The principle behind this step is as follows: normalized load allows for comparison of thermal loads on different joints and with different rated torques on the same scale; the recursive formula with boundary clipping enables a unified description of overload heat accumulation and low-load recovery with relatively low computational cost; parameter identification allows the model to adapt to the rated torque and heat dissipation conditions of different joints; and the measured I of the actuator... 2 t serves as a monitoring signal, ensuring consistency between the proxy model output and the internal protection state of the driver.
[0039] When selecting a winding temperature proxy model, a first-order thermal inertia element is used: ; in, For winding temperature, This represents the actual joint torque. This is the equivalent conversion factor from the square of the torque to the winding temperature rise rate. The equivalent thermal time constant of the winding is... Let be the equivalent environmental thermal boundary corresponding to the j-th running process. The discrete recursive relation is: ; ; in, Let the winding temperature be the temperature at the k-th sampling point. For steady-state temperature, This represents the winding temperature at the previous sampling point. Model parameters. , And the environmental thermal boundary correction amount corresponding to each operating process. Identification was performed using multiple sets of synchronously acquired actual joint torque and winding temperature data, with the optimization objective being: ; in, During online runtime, the model state is initialized with the currently obtained winding temperature observations: ; The principle behind this step is that: a first-order thermal inertial element can describe the dynamic process of winding temperature changing with load with fewer parameters; analytical discrete recursion can be used to directly complete the state recursion using closed-form solutions, reducing the impact of sampling period changes on numerical integration accuracy; using the square of the actual joint torque as the equivalent heat input, the winding temperature can be predicted without directly using current; setting the initial temperature and equivalent environmental thermal boundary independently for each operating process can reduce prediction deviations caused by differences in initial conditions and changes in heat dissipation conditions.
[0040] Step S4: When the future load is unknown, the thermal state obtained by recursion in step S3 is used as the initial state for prediction. Short-term future load assumptions are constructed based on the historical actual joint torque window. The estimated remaining time and estimated threshold arrival time of the thermal state are predicted in a rolling manner according to the preset update cycle. The thermal margin is calculated based on the current thermal state and the protection threshold.
[0041] The system stores historical actual joint torques within a certain time range prior to the current moment, with a historical window length of, for example, 1 second. The prediction process is executed in a rolling fashion according to a preset update cycle, for example, an update cycle of 0.05 seconds and a prediction step size of 0.01 seconds. The prediction window can be set to 5 seconds, 10 seconds, 20 seconds, or 30 seconds. Future loads can be constructed using historical window averages, root mean square values, high quantile values, or repetitions of historical load segments.
[0042] The principle behind this step is that: recent load levels can reflect the current task load trend and provide an interpretable risk time estimate when the future load is unknown; rolling execution can promptly correct the prediction results when the load changes; and using the historical window average or root mean square value as the assumption for future load is simple to calculate and has a certain representativeness, making it suitable for scenarios such as teleoperation and human-computer interaction where the future torque trajectory cannot be fully obtained in advance.
[0043] For using I 2 The t-surrogate model's joints are used to construct future normalized loads based on historical actual joint torques. and with the current I 2 t state As the initial state for prediction, calculate I. 2 t-forecasted growth rate: ; when and At that time, it is expected to reach I 2 The remaining time for the protection threshold t is: ; The corresponding expected arrival time of the protection threshold is: ; when When it is within the current prediction window, output I 2 Heat risk warning; when the predicted growth rate is not positive, or the expected threshold arrival time exceeds the current prediction window, continue rolling monitoring, and in the next prediction cycle, adjust the forecast based on the updated actual torque and I. 2 The state t is recalculated.
[0044] For joints employing a winding temperature proxy model, the current winding temperature... As the initial state for temperature prediction, and based on historical actual joint moments, the future short-term thermal load is constructed. The root mean square value of the actual joint moments within the historical window is used as the future short-term equivalent torque. ; in, This represents the number of sampling points within the historical window. This represents the actual joint torque. Within the prediction window, let the future equivalent torque be: ; The corresponding equivalent steady-state temperature is: ; Future temperatures are predicted using a first-order thermal model: ; Let the winding temperature protection threshold be... When the predicted temperature reaches or exceeds the threshold, the expected threshold arrival time is obtained, and the remaining time from the current time to reaching the temperature protection threshold is calculated. This assumes a constant equivalent torque for future short-term loads and satisfies... At that time, the estimated remaining time to reach the temperature threshold is calculated using analytical solutions: ; The corresponding expected temperature threshold arrival time is: ; when When the current temperature has reached or exceeded the protection threshold, it is determined that the temperature has been set to [a certain value]. ;when When this occurs, it indicates that the winding temperature will not reach the protection threshold under the current future load assumptions, and rolling monitoring will continue.
[0045] The effect of this step is that, through the above analytical solution, the estimated remaining time and the estimated threshold arrival time can be quickly given when the future load is unknown, providing advance notice for online early warning; at the same time, the current thermal state, the predicted thermal state, the thermal margin, the estimated remaining time and the estimated threshold arrival time are output, providing a unified thermal state constraint for subsequent planning and control.
[0046] Step S5: Based on at least one of the predicted thermal state, thermal margin, estimated remaining time, and estimated threshold arrival time, perform at least one of the following operations: trajectory filtering, cycle time optimization, online speed reduction, and task switching, or perform a simulated thermal management closed loop.
[0047] The current thermal state, predicted thermal state, thermal margin, and estimated remaining time are used as constraints for motion planning. Specifically: Trajectory selection: The future torque sequence is calculated using inverse dynamics for candidate trajectories. This sequence is then input into the selected surrogate model to predict the thermal state and threshold arrival time of the entire trajectory. If the predicted threshold arrival time of a candidate trajectory is less than the task execution time, the trajectory is filtered out or its priority is reduced. The principle behind this is that assessing thermal risk before trajectory execution avoids selecting trajectories that could trigger protection.
[0048] Cyclic cadence optimization: The trajectory execution speed is adjusted based on the thermal margin. When the thermal margin is sufficient, the original cadence is maintained; when the thermal margin is insufficient, the cadence is reduced or pauses are added. The principle is that by adjusting the cadence, the time distribution of joint load can be changed, thereby reducing the rate of heat accumulation.
[0049] Online speed reduction: In remote operation or online control, when the predicted remaining time is lower than the preset warning time, a speed reduction command is output to the motion control module to reduce the joint speed or torque. The principle is that early speed reduction can reduce the thermal load and avoid reaching the protection threshold.
[0050] Task switching: When the predicted thermal risk level reaches the warning or protection level, switch to a low-load task or pause the current task. The principle is that task switching can change the load level, allowing the joint to enter a low-load recovery state.
[0051] The simulation-based thermal management closed loop continuously receives torque, current, and temperature feedback from the controller during actual machine operation, initializing the next recursion based on the current measured state. The simulation side utilizes the actual machine feedback to correct surrogate parameters, initial states, or boundary biases, ensuring consistency between the thermal states of the planning and actual machine sides. The principle is that actual machine feedback can correct model deviations on the simulation side, improving subsequent prediction accuracy.
[0052] Example 2: System Example like Figure 1 As shown, this embodiment provides a humanoid robot joint thermal management system based on a proxy model, including a joint operation data acquisition module 1, a actuator type and available signal identification module 2, and an I... 2 The module comprises: 3. Proxy model module; 4. Winding temperature proxy model module; 5. Temperature boundary correction module; 6. Online rolling prediction module; 7. Thermal protection risk assessment module; 8. Trajectory planning constraint module; and 9. Early warning display and control interface module. These modules can be deployed in computing units such as robot controllers.
[0053] The functions and collaborative working process of each module are as follows: The joint operation data acquisition module 1 is responsible for acquiring actual joint torque, winding temperature observations, optional internal thermal state of the actuator or other auxiliary observations, timestamps, and sampling period. This module periodically acquires data according to a preset sampling period and transmits the data, along with a data quality flag, to the actuator type and available signal identification module 2.
[0054] The driver type and available signal identification module 2 determines the available signals based on the driver model, object dictionary, analog interface, and data quality, and selects the appropriate I / O signal accordingly. 2 Choose either the proxy model module 3 or the winding temperature proxy model module 4. The selection result is recorded in the form of a model identifier for subsequent modules to use.
[0055] When selecting I 2 When using the proxy model module 3, this module performs dimensionless transformation on the actual joint torque using the rated torque to obtain the normalized load, and calculates the normalized I using a recursive formula with boundary trimming. 2 t Protection state. Model parameters. and Through actual measurement of the driver I 2 t is used as a supervisory signal for identification. Specific formulas and methods are given in formulas (1) to (3) of Example 1.
[0056] When the winding temperature proxy model module 4 is selected, this module uses a first-order thermal inertia element to describe the dynamic process of winding temperature change with actual joint torque. For specific formulas and methods, see formulas (4) to (8) in Example 1.
[0057] Temperature boundary correction module 5 corrects the equivalent environmental thermal boundary based on the initial temperature during operation. For the first... Each running process sets the initial temperature and equivalent ambient thermal boundary: ; ; in, For the first The winding temperature observations were obtained at the start of each operation process. This is the correction amount for the environmental thermal boundary corresponding to this operation process. The principle of this module is that the initial winding temperature and heat dissipation conditions differ in different operation processes, and independent correction can reduce the prediction bias caused by the differences in initial conditions. Without experimental data, we can demonstrate the prediction bias caused by inconsistent initial conditions solely from the temperature proxy model. Assuming k=1, then: ; in This is the initial winding temperature. Note this. The initial conditions are different each time the system starts running, meaning the predicted values will also be different. Therefore, the temperature boundary correction module can correct these initial conditions to reflect the actual initial winding temperature during operation, allowing the surrogate model to make effective predictions.
[0058] Online rolling prediction module 6 with I 2 The thermal state obtained by the t proxy model module 3 or the winding temperature proxy model module 4 is used as the initial state for prediction. Based on the historical actual joint torque window, short-term future load assumptions are constructed, and the expected remaining time and expected threshold arrival time of the thermal state are predicted in a rolling manner according to the preset update cycle.
[0059] For I 2 The t-proxy model calculates the predicted growth rate and estimates the achievement of I. 2The remaining time for the protection threshold; for the winding temperature proxy model, the root mean square value of the actual joint torque within the historical window is used as the future short-time equivalent torque, and the expected remaining time to reach the temperature threshold is calculated using the analytical solution. For specific formulas and methods, see formulas (9) to (17) in Example 1.
[0060] The thermal protection risk assessment module 7 outputs the risk level based on the current thermal state, the predicted threshold arrival time, and the protection status flag. When the predicted remaining time is less than the preset warning time or the temperature margin decreases, a thermal risk warning is output.
[0061] The trajectory planning constraint module 8 performs at least one of the following based on thermal margin: trajectory selection, cycle time optimization, online speed reduction, and task switching, or outputs a manual intervention warning. The principle of this module is that by using thermal state, thermal margin, and threshold arrival time as motion planning constraints, thermal risks can be assessed before trajectory execution, avoiding passive handling after protection is triggered.
[0062] The early warning display and control interface module 9 outputs unified results to the host computer, motion library or human-machine interface, including the current thermal state, predicted thermal state, thermal margin, estimated remaining time, estimated threshold arrival time and thermal risk level.
[0063] like Figure 2 As shown, when the driver cannot provide a directly usable I / O to the motion library... 2 t percentage, but when the actual joint torque can be obtained, I 2 Proxy model module 3 uses the rated torque to make the actual torque dimensionless, resulting in a normalized load: ; set up For normalization I 2 t heat load percentage, The heat accumulation coefficient, This represents the equivalent equilibrium torque multiplier when there is no net heat accumulation. In continuous form, it is expressed as: ; In this invention Take a negative value. When Greater than At that time, the right end is positive. As time goes on; when Less than When the right side is negative, It decreases over time. Therefore, the same equation of state can be applied to both overload heat accumulation and low load recovery.
[0064] To simplify the calculation, a recursive formula with boundary clipping is used: ; Model parameters and To optimize variables, I was measured through the driver. 2 t is used as a supervisory signal for identification: ; The optimizer can employ nonlinear least squares optimization methods such as L-BFGS-B, and... and Set a physically reasonable range of values. For multiple sets of training data under different speeds, loads, and trajectories, fit the data separately and then perform multi-fold cross-validation; finally, use median fusion for the parameters to reduce the impact of a single working condition on the model.
[0065] In one specific embodiment, a certain joint has a rated torque of 13.5 N·m, a peak torque of 57 N·m, and a peak duration of 3 s. Based on nominal parameters, Approximately -1.98, Take 1; after multi-condition identification, the following is obtained: , In another specific embodiment, the rated torque of the other joint is 3 N·m, which is identified. , The nominal parameter is approximately , The above values are used to illustrate the implementation of the model and do not limit the scope of protection of this invention.
[0066] like Figure 2 As shown, in scenarios where the reference trajectory is known, or where the future torque sequence can be obtained through inverse dynamics, the system inputs the future torques into I in chronological order. 2 The t-surrogate model recursively obtains the thermal state changes of the entire trajectory and finds the sampling point that first reaches the protection threshold: ; To improve the accuracy of the protection threshold arrival time estimation, linear interpolation can be performed between adjacent sampling points where the threshold is first exceeded: ; When adjacent sampling points have the same thermal state, making linear interpolation impossible, or when the sampling timestamp is abnormal, or when the recursive initial state has reached the protection threshold, the system stops normal interpolation calculation and determines the threshold arrival time using the discrete sampling time that first reached the threshold, the most recent valid timestamp, or a preset conservative time estimation rule, while simultaneously outputting the corresponding data quality flag. Therefore, even under abnormal sampling conditions or boundary conditions, the system can still provide the planner with thermal protection trigger judgments for candidate trajectories, the estimated threshold arrival time, and the remaining thermal margin.
[0067] like Figure 3 As shown, when the driver cannot provide normalized I...2 For joints in state t, but capable of obtaining actual joint torque and winding temperature observations on the driver side, this invention constructs a first-order winding temperature surrogate model based on the actual joint torque. This model uses winding temperature... As a thermal state, the square of the actual joint torque is used as the equivalent heat input, and a first-order thermal inertial element is used to describe the heat dissipation process of the winding to the equivalent environmental thermal boundary.
[0068] Considering that the initial winding temperature and surrounding heat dissipation conditions may differ during different operating processes, this invention sets the initial temperature and equivalent environmental thermal boundary for different experiments or operating processes. For the first... The process includes: ; ; The first-order continuous temperature surrogate model based on actual joint torque is expressed as follows: ; in, The actual joint torque obtained from the driver side; This is the equivalent conversion factor from the square of the torque to the winding temperature rise rate, and its unit can be expressed as °C·(s). -1 ·N -2 ·m -2 ); The equivalent thermal time constant of the winding; For the first The equivalent environmental thermal boundary corresponding to each operating process.
[0069] Within one sampling period, when the actual joint torque remains approximately constant, the steady-state temperature corresponding to this thermal model is: ; By further utilizing the analytical solution of the first-order thermal system, the discrete-time temperature recurrence relation can be obtained: ; in, This refers to the winding temperature status at the previous sampling point. The time interval between adjacent sampling points is denoted as . Compared with numerical integration using the Euler method, the above analytical discrete form can directly utilize the closed-form solution of the first-order thermal system to complete the state recursion and reduce the impact of sampling period variation on the accuracy of numerical integration.
[0070] Model parameters include torque heating coefficient Thermal time constant And the environmental thermal boundary correction amount corresponding to each operating process. Parameter identification is performed using multiple sets of synchronously acquired actual joint torque and winding temperature data, with the error between the predicted temperature and the measured winding temperature used as the optimization objective. For example, the following objective function can be used: ; in, The set of parameters to be identified Predict temperatures for the model. Model parameters can be obtained using a bounded nonlinear least squares method, with physical constraints set for the parameters based on the torque-temperature rise relationship, thermal time constant, and a reasonable temperature range.
[0071] During online operation, the system first initializes the model state using the currently obtained winding temperature observations: ; Subsequently, the future winding temperature is recursively inferred based on the actual joint torque, sampling interval, and identified model parameters fed back by the driver in real time. When there is a known future torque trajectory, the winding temperature during the entire trajectory execution period can be predicted point by point according to the future torque sequence; when the future load is unknown, a short-term future load can be constructed by combining the historical torque window, and rolling temperature prediction can be performed.
[0072] Based on the predicted winding temperature and the preset temperature protection threshold The relationship between these factors allows us to define a temperature margin: ; When the predicted temperature reaches or exceeds the temperature protection threshold, the system outputs a corresponding thermal risk flag; when the predicted temperature has not yet reached the threshold, the temperature margin can be further used for trajectory screening, speed adjustment, cycle optimization, and thermal risk warning.
[0073] To verify the first-order winding temperature proxy model based on actual joint torque, actual joint torque and winding temperature data were synchronously collected at the joint on the driver side. Parameter identification was performed using partial operating data, and model verification was performed using independent operating data. In two independent verification conditions in one embodiment, the root mean square error of the first-order winding temperature proxy model based on actual joint torque was 0.576°C and 0.245°C, respectively; the mean absolute error was 0.463°C and 0.192°C, respectively; and the maximum absolute error was 1.872°C and 0.911°C, respectively. The above implementation results show that, under the condition of synchronous signal acquisition on the driver side, the first-order temperature proxy model based on actual joint torque can accurately describe the dynamic process of joint winding temperature change with load, and can predict winding temperature based on actual joint torque without directly using current as model input.
[0074] like Figure 4 As shown, for scenarios where the future torque trajectory cannot be fully obtained in advance, such as teleoperation, human-computer interaction, and continuous task execution, this invention saves the historical actual joint torque within a certain time range before the current moment and uses the historical torque window to construct short-term future load assumptions. The prediction process is executed in a rolling manner according to a preset cycle. Each time, the latest actual joint torque and current thermal state are read, the historical window is updated, and the future thermal state and the estimated remaining time to reach the corresponding protection threshold are predicted based on the surrogate model used by the current joint.
[0075] In one implementation, the historical window length is 1 second, the prediction update period is 0.05 seconds, the prediction step size is 0.01 seconds, and the prediction window can be set to 5 seconds, 10 seconds, 20 seconds, or 30 seconds. The future load can be constructed using historical window averages, root mean square values, high quantile values, or repetitions of historical load segments.
[0076] For using I 2 The t-proxy model's joints are used to construct future normalized loads based on historical actual joint torques. and with the current I 2 t state As the initial state for prediction. In one implementation, the equivalent normalized load obtained from a historical window is used as the short-term future load, and I is calculated. 2 t-forecasted growth rate: ; when , At that time, it is expected to reach I 2 The remaining time for the protection threshold t is: ; The corresponding expected arrival time of the protection threshold is: ; when When the value is within the current prediction window, the system outputs I. 2 Heat risk warning; when the predicted growth rate is not positive, or the expected threshold arrival time exceeds the current prediction window, the system continues to monitor and, in the next prediction cycle, adjusts the forecast based on the updated actual torque and I. 2 The state t is recalculated. The system can simultaneously output the current I. 2 t state, prediction I 2 t state, I 2 t represents the heat margin, the expected threshold arrival time, and the expected remaining time.
[0077] For joints employing a first-order winding temperature proxy model based on actual joint torque, the system uses the current winding temperature... As the initial state for temperature prediction, and based on historical actual joint torques, the future short-term thermal load is constructed. Considering that the equivalent heating of the windings is related to the square of the actual joint torque, in one embodiment, the root mean square value of the actual joint torque within the historical window is used as the future short-term equivalent torque: ; in, This represents the number of sampling points within the historical window. This represents the actual joint torque. Within the prediction window, let the future equivalent torque be: ; The corresponding equivalent steady-state temperature is: ; Future temperatures are predicted using a first-order thermal model: ; Let the winding temperature protection threshold be... When the predicted temperature reaches or exceeds the threshold, the system obtains the expected threshold arrival time and calculates the remaining time from the current moment until the temperature protection threshold is reached. When the future short-term load adopts the constant equivalent torque assumption and satisfies... In this case, the estimated remaining time to reach the temperature threshold can be directly calculated based on the analytical solution of the first-order temperature model: ; The corresponding expected temperature threshold arrival time is: ; when When the system determines that the current temperature has reached or exceeded the protection threshold, it will... .when This indicates that under the current future load assumptions, the winding temperature will not reach the protection threshold, and the system continues rolling monitoring. Simultaneously, the system calculates the temperature margin based on the current winding temperature: ; and temperature margin With the estimated remaining time Together, they serve as indicators for temperature risk assessment. When the temperature margin decreases or the estimated remaining time is less than the preset warning time, the system outputs a thermal risk warning and provides control constraints to the trajectory planning module or motion control module, such as reducing speed, reducing load, adjusting the beat, replanning the trajectory, or switching tasks.
[0078] In one embodiment, during the first rolling alert It is 36.9s, and the current I 2 t is 11.65%, the predicted normalized torque is 1.31, and I 2The growth rate of t is 3.3% / s, and the enable time is predicted. The lead time is 63.75 seconds, with a warning lead time of approximately 26.76 seconds. In another calculation example based on measured parameters, the torque-temperature rise coefficient is identified based on synchronously collected data of actual joint torque and winding temperature. and thermal time constant Suppose the current winding temperature at a certain rolling prediction moment is 72.4°C. Based on the most recent historical torque window, the future short-term equivalent torque is 9.0 N·m. The current equivalent environmental thermal boundary is taken as 30.0°C. Therefore, the corresponding equivalent steady-state temperature is calculated to be approximately 125.51°C. When the temperature protection threshold is set to 100°C, the estimated remaining time to reach the temperature threshold is approximately 48.43 s. If the current moment is 41.20 s, the estimated time to reach the temperature protection threshold is approximately 89.63 s, allowing for a temperature risk warning to be issued approximately 48.43 s in advance.
[0079] like Figure 5 As shown, after the planner generates candidate task trajectories, inverse dynamics can be used to calculate the future torque sequence corresponding to each joint. For the use of I... 2 The t-surrogate model's joints, the system inputs the future torque sequence into I. 2 t-proxy model, predicting I during candidate trajectory execution process 2 The system calculates the arrival time of the t-state and protection threshold. For joints using a winding temperature surrogate model, the system constructs an equivalent thermal input based on the square of the torque based on the future torque sequence, and uses a first-order winding temperature surrogate model to predict the winding temperature and the arrival time of the temperature threshold. Subsequently, based on the predicted thermal state, thermal margin, and estimated remaining time, the system performs at least one of the following operations on candidate trajectories: retention, filtering, reducing the cycle time, adjusting local speed, switching tasks, or outputting a manual intervention warning. During actual operation, the system continuously receives torque, current, and temperature feedback from the controller and initializes the next iteration with the current measured state. On the simulation side, the system uses actual feedback to correct surrogate parameters, initial states, or boundary biases, ensuring that the thermal states on the planning side and the actual side are consistent, forming a closed-loop thermal management system from actual operation to simulation.
[0080] Therefore, those skilled in the art should understand that although the present invention has been described with reference to preferred embodiments, the present invention is not limited to the specific embodiments described above. Various modifications and variations can be made to the present invention without departing from the spirit and scope of the present invention.
Claims
1. A method for thermal management of humanoid robot joints based on a proxy model, characterized in that, include: Step S1: Collect joint operation data, which includes actual joint torque, timestamp, and sampling period; Step S2: Based on the actuator type and the available signals from the joint operation data acquired in Step S1, select I. 2 t-surrogate model or winding temperature surrogate model; Step S3: Based on the actual joint torque, recursively deduce the thermal state using the selected surrogate model; Step S4: When the future load is unknown, the thermal state obtained by recursion in step S3 is used as the initial state for prediction. Short-term future load assumptions are constructed based on the historical actual joint torque window. The estimated remaining time and estimated threshold arrival time of the thermal state are predicted in a rolling manner according to the preset update cycle. The thermal margin is calculated based on the current thermal state and the protection threshold. Step S5: Based on at least one of the predicted thermal state, thermal margin, estimated remaining time, and estimated threshold arrival time, perform at least one of the following operations: trajectory filtering, cycle time optimization, online speed reduction, and task switching, or perform a simulated thermal management closed loop.
2. The method according to claim 1, characterized in that, In step S3, the I 2 The t-surrogate model uses the rated torque to dimensionlessly approximate the actual joint torque, resulting in a normalized load: ; Where k represents the discrete sampling number, The actual joint torque at the k-th sampling point. The rated torque corresponding to this joint is given; and a recursive formula with boundary trimming is used: ; Recursive Normalization I 2 t protection state, where, Normalized I for the k-th sampling point 2 t heat load percentage, The heat accumulation coefficient is negative. This is the equivalent balance torque ratio when there is no net heat accumulation. The sampling interval is... This indicates that the state is limited to the range of 0% to 100%.
3. The method according to claim 2, characterized in that, In step S3, the I 2 Parameters of the t proxy model and Through actual measurement of the driver I 2 t is used as a supervisory signal for identification: ; in, Actual test of driver I 2 t, Indicates the actual parameters. To identify the optimal parameters, the identification method employs bounded nonlinear least squares optimization, and performs cross-validation and median fusion on multiple sets of working condition data. This indicates that the k-th sampled data of p uses the value of parameter θ.
4. The method according to claim 3, characterized in that, In step S3, the winding temperature proxy model adopts a first-order thermal inertia element: ; Where t is the continuous physical time, representing the sampling time corresponding to the k-th sampling point. For winding temperature, This represents the actual joint torque. This is the equivalent conversion factor from the square of the torque to the winding temperature rise rate. The equivalent thermal time constant of the winding is... This is the equivalent environmental thermal boundary corresponding to the j-th running process.
5. The method according to claim 4, characterized in that, In step S3, the winding temperature proxy model adopts the following discrete recursive relationship: ; ; in, Let the winding temperature be the temperature at the k-th sampling point. For steady-state temperature, This is the winding temperature at the previous sampling point.
6. The method according to claim 4, characterized in that, In step S3, the initial temperature and equivalent environmental thermal boundary are set for the j-th running process: ; ; in, The winding temperature observation value obtained at the start of the j-th operation process. This is the correction amount for the environmental thermal boundary corresponding to this operation process.
7. The method according to claim 2, characterized in that, In step S4, for those using I 2 The t-surrogate model's joints are used to construct future normalized loads based on historical actual joint torques. and with the current I 2 t state As the initial state for prediction, calculate I 2 t-forecasted growth rate: ; when and At that time, it is expected to reach I 2 The remaining time for the protection threshold t is: ; The corresponding expected arrival time of the protection threshold is: 。 8. The method according to claim 5, characterized in that, In step S4, for joints using a winding temperature proxy model, the current winding temperature is used as the reference. As the initial state for temperature prediction, the root mean square value of the actual joint torque within the historical window is used as the future short-term equivalent torque: ; in, This represents the number of sampling points within the historical window. The actual joint torque is used; and the estimated remaining time to reach the temperature protection threshold is calculated using analytical solutions: ; in, This is the temperature protection threshold. This is the equivalent steady-state temperature.
9. The method according to claim 1, characterized in that, In step S5, the current thermal state, predicted thermal state, thermal margin, and estimated remaining time are used as motion planning constraints for candidate trajectory selection, speed adjustment, beat optimization, task switching, and closed-loop update of the actual machine-simulation thermal state.
10. A humanoid robot joint thermal management system based on an agent model, characterized in that, It includes a data acquisition module, a driver type and available signal identification module, a surrogate model module, a temperature boundary correction module, an online rolling prediction module, and an output module; the data acquisition module is used to acquire the actual joint torque, timestamp, and sampling period; the driver type and available signal identification module is used to determine the available signals based on the driver model, object dictionary, analog interface, and data quality; The proxy model module includes I 2 t-surrogate model and winding temperature surrogate model are used to select I based on the driver type and available signal. 2 The module provides a proxy model or a winding temperature proxy model; the temperature boundary correction module is used to correct the initial temperature and the equivalent environmental thermal boundary; the online rolling prediction module is used to use the thermal state obtained by the proxy model module as the initial prediction state, and when the future load is unknown, it constructs a short-term future load hypothesis based on the historical actual joint torque window, and predicts the expected remaining time and expected threshold arrival time of the thermal state according to a preset update cycle; the output module is used to output at least one of the current thermal state, predicted thermal state, thermal margin, expected remaining time and expected threshold arrival time.
Citation Information
Patent Citations
Power battery active thermal management system and control method thereof
CN112477698A
Performance optimization method and optimization device of joint motor and electronic equipment
CN122546678A
Predictive thermal management method for electric driving system of urban logistics electric light truck
CN122607091A
