Multi-scale composite robot power supply differentiated intelligent management system and method

Abstract

This invention relates to the field of robot power management technology, and discloses a differentiated intelligent management system and method for power supply of multi-scale composite robots. This method addresses collaborative operation scenarios where macroscopic robots and micro / nano robot groups share an onboard battery and power supply bus. It synchronously collects bus voltage, axis current and temperature, micro / nano terminal voltage and supply current, and pipe insertion depth to construct a cross-domain joint state vector time sequence; calculates the dynamic transmission resistance of flexible micro-cables and forms terminal voltage feedforward compensation; predicts macro and micro power demands based on an attention encoding / decoding model and generates power allocation plans; uses a conditional expectation model to perform spectral residual decomposition and outputs fault warning states; and during the task phase, the state machine triggers predictive switching based on the peak prediction time and supercapacitor charging time, outputting voltage regulation and charge management commands. This invention can achieve differentiated and stable power supply and fault warning during the collaborative operation of multi-scale composite robots.

Description

Technical Field

[0001] This invention relates to the field of robot power management technology, and more specifically, to a multi-scale composite robot power differentiated intelligent management system and method. Background Technology

[0002] In scenarios such as in-service equipment tube bundle inspection, the macroscopic mobile platform needs to move autonomously within the equipment room and drive multi-axis mechanisms, while simultaneously sending multiple micro-nano pipe inspection robots into the pipe via flexible micro-cables to perform scanning tasks. Due to the significant difference in the magnitude between the macroscopic side's driving power requirements and the micro-nano side's precision power supply requirements, and the fact that they often share onboard batteries and power supply buses, a contradiction can easily arise between bus voltage fluctuations and micro-nano terminal voltage stability during collaborative operations.

[0003] Existing technologies typically rely on closed-loop regulation of the secondary power supply bus and passive protection devices to maintain power supply to the micro / nano side. When the macro-side motor starts or accelerates / decelerates rapidly, the timescale of the bus voltage drop is close to the regulator response time, making it difficult for passive methods to compensate in a timely manner. At the same time, the flexible micro-cable continuously deforms with macro-motion and micro / nano advancement, and the dynamic change in its equivalent transmission resistance introduces terminal voltage deviation. Simply relying on fixed parameter compensation is insufficient to cover the entire operating condition.

[0004] Furthermore, load changes on the micro / nano side are coupled into the macro-side current spectrum via a common bus. External harmonics may overlap with the frequency bands characteristic of early faults in the macro-drive system, leading to false alarms in fault diagnosis under coordinated operating conditions. Insufficient or misjudged power supply switching timing may cause power outages on the micro / nano side, task termination, or data loss. Therefore, a management method is needed that can take into account dynamic cable resistance compensation, power prediction and allocation, interference decoupling diagnosis, and predictive switching control. Summary of the Invention

[0005] This invention provides a multi-scale composite robot power supply differentiation intelligent management system and method, which solves the technical problems in related technologies, such as the difficulty in simultaneously meeting the transient power requirements of macroscopic robot drive and the precision power supply requirements of micro-nano robots, the difficulty in real-time compensation of terminal voltage deviation caused by the dynamic transmission resistance of flexible micro-cables, and the technical problems of false alarms in fault diagnosis and insufficient switching control timing caused by external interference under collaborative operation conditions.

[0006] This invention provides a method for differentiated intelligent management of power supplies for multi-scale composite robots, comprising the following steps: S1 collects macroscopic robot bus voltage, axis current, temperature, micro / nano robot end voltage, power supply current and insertion depth, and constructs cross-domain joint state vector time series through real-time frequency conversion spectrum extraction and power statistics fusion; S2, extract joint angles and micro-inlet tube depths from cross-domain joint state vector time sequence, calculate the full cable bending angle distribution through kinematic forward solution and path interpolation, and calculate dynamic transmission resistance by querying pre-calibration mapping table; S3, the cross-domain joint state vector time sequence and joint motion command sequence are fed into the attention encoding and decoding model, and the dynamic transmission resistance is combined to correct the micro-nano equivalent power to generate macro-micro dual-side power demand prediction. The power allocation value of the time step is calculated by recursively using the remaining battery power to obtain the power allocation plan. S4. Extract the micro-nano load state from the cross-domain joint state vector time sequence as a condition variable and the current spectrum of each axis as a decomposition object. Establish a conditional expectation model, fit the external interference spectrum and subtract the measured spectrum to obtain the residual spectrum. Extract the harmonic amplitude and high-frequency energy ratio to form a fault feature vector and output an early warning. S5, the mission phase state machine receives power allocation plans and fault warning states, calculates the preload trigger advance based on the supercapacitor charging time and peak expected time to achieve predictive switching, and outputs voltage regulation and charge management commands.

[0007] Furthermore, the construction of the cross-domain joint state vector time series includes: Using the control cycle of the macroscopic robot servo controller as the global reference clock, the data collected from each channel is aligned and interpolated according to clock counting. Short-time Fourier transform is applied to the three-phase current waveforms of each joint. The current rotation frequency, which is calculated from the real-time angular velocity of each joint, is taken as the fundamental frequency. The amplitude at the fundamental frequency and its preset integer multiples is extracted and spliced ​​to obtain the current spectrum characteristics of each axis. The average power supply, total power of the group, standard deviation, and terminal voltage deviation within the acquisition window of each micro-nano robot are calculated to obtain the micro-nano load status. The macroscopic side bus state, current spectrum and motion state characteristics are sequentially spliced ​​with the micro-nano side load state after standardization to obtain a cross-domain joint state vector time sequence.

[0008] Furthermore, the calculation of the dynamic transmission resistance includes: The flexible micro-cable is divided into a robot body segment and a confined space segment. The body segment is calculated by using the kinematic forward kinematics to deduce the spatial vector of the cable support point and estimate the cumulative bending angle from the joint angles of each joint in the cross-domain joint state vector time sequence. The confined space segment is calculated by interpolating the cable centerline curvature from the micro-inlet pipe depth according to the pre-stored 3D model of the pipe path and converting it into a local bending angle. The two segments are merged to form the full cable bending angle position distribution. Substitute the bending angle of each segment into the offline calibrated bending angle resistance increment coefficient mapping table, calculate the equivalent resistance of each segment from the nominal resistance of each segment and the increment coefficient obtained from the table, and sum them to obtain the dynamic transmission resistance. The expected transmission voltage drop is calculated by multiplying the dynamic transmission resistance by the total power supply current of the micro-nano robot group, and then superimposed on the base voltage setting of the secondary power supply bus to form the terminal voltage feedforward compensation command.

[0009] Furthermore, the generation of the power allocation plan includes: The cross-domain joint state vector time sequence is used as the encoding input, the joint motion command sequence is used as the feedforward input, and the dynamic transmission resistance is used as the micro-nano side power supply efficiency parameter. These are all fed into the attention encoding and decoding model. The encoder of the attention encoding and decoding model applies variable weights to the time aggregation of historical time series, and the decoder generates macro and micro dual-sided power demand predictions through autoregression. The two outputs share weights in the middle layer of the decoder. The transmission loss corresponding to the product of dynamic transmission resistance and predicted supply current is superimposed on the predicted micro / nano body power to obtain the micro / nano equivalent power. With the remaining battery power as the total constraint and the micro / nano side voltage maintaining the nominal operating range as the hard constraint, the power allocation value of each time step is obtained by forward recursion in the prediction time domain according to the micro / nano side priority principle and written into the power allocation plan.

[0010] Furthermore, the formation of the fault feature vector and the output of the fault warning state include: For each servo drive axis of the macro robot, the micro-nano load state extracted from the time sequence of the cross-domain joint state vector is used as the condition variable. Conditional expectation models are established according to the preset discrete interval to which the current frequency of each axis belongs, and the expected distribution of the current spectrum of the corresponding axis is output. Subtract the current spectrum of each axis from the expected distribution of the corresponding current spectrum to obtain the residual spectrum of each axis; extract the harmonic amplitude from the residual spectrum with the current rotation frequency as the fundamental frequency, calculate the proportion of high-frequency energy in the frequency band above the current rotation frequency as the cutoff frequency, and splice it with the corresponding axis temperature to form the fault feature vector of the axis. The components of the fault feature vector are judged based on a three-level threshold system consisting of a first threshold, a second threshold, and a third threshold, and the fault warning status is output.

[0011] Furthermore, the implementation of the predictive switching includes: The charging time of a supercapacitor is obtained by multiplying the shortest charging time, determined by the supercapacitor's rated capacity, target charging voltage, and maximum allowable charging current, by a preset conservative coefficient. The preload trigger advance is obtained by subtracting the supercapacitor charging time from the difference between the peak estimated time and the current time. When the preload trigger advance is greater than zero, the preload charging and discharging circuit is immediately connected. When the preload trigger advance is less than zero, the discharge duration that can be supported is evaluated according to the current state of charge of the supercapacitor. If the discharge duration that can be supported covers the duration corresponding to the peak estimated time, the switching is performed. Otherwise, the macroscopic power limiting command is triggered synchronously. During the task phase, the state machine superimposes the cable resistance feedforward compensation component formed by the dynamic transmission resistance with the bus weight adjustment amount corresponding to the switching state, and outputs voltage regulation and charge management commands.

[0012] Furthermore, the cross-domain joint state vector timing sequence and the joint motion command sequence are output in parallel as two independent data streams and are both refreshed synchronously with the control cycle; The joint motion command sequence is read from the command buffer of the macroscopic robot motion controller. It contains the joint command matrix for each time step in the future planning time domain. The joint command matrix is ​​indexed by each time step, each joint, and the command component. The cross-domain joint state vector time sequence is used for the calculation of dynamic transmission resistance, the historical trajectory input of the attention encoding and decoding model, and the condition variable extraction of the conditional expectation model. The joint motion command sequence is used for the feedforward prediction input of the attention encoding and decoding model.

[0013] Furthermore, the bending angle resistance increment coefficient mapping table is equipped with an online adaptive correction mechanism: The difference between the measured terminal voltage of the micro-nano robot and the expected terminal voltage after feedforward compensation is used as the residual signal. When the absolute mean of the residual signal exceeds the preset residual threshold within several consecutive control cycles, the residual signal is divided by the total power supply current of the micro-nano robot group to obtain the total cable resistance estimation error. The cable resistance estimation error is weighted and distributed to each segment according to the proportion of the bending angle of the segments with the largest bending angle in the distribution of bending angles of the entire cable to the sum of the bending angles. Then, the error is divided by the nominal resistance value of the corresponding segment to obtain the resistance increment coefficient correction amount of each segment. The incremental coefficients of the corresponding bending angle interval in the bending angle resistance incremental coefficient mapping table are updated using a weighted average method according to the preset learning rate.

[0014] Furthermore, in addition to the time-step power allocation values, the power allocation plan also includes a list of planned power limiting trigger times and a macroscopic peak time window: For the time window in which the difference between the power demand of the micro-nano side and the remaining power of the battery in the predicted time domain is less than the preset safety margin, a planned power limiting instruction is generated at the corresponding time. Each planned power limiting instruction carries the corresponding acceleration upper limit and is written to the planned power limiting trigger time list with a timestamp. For the time window in which the macroscopic power demand exceeds the peak switching threshold within the predicted time domain, record the predicted peak time and the corresponding duration and write them into the macroscopic peak time window. The list of planned power limiting trigger times, together with the macroscopic peak time window, is fed into the task phase state machine to trigger predictive switching.

[0015] Furthermore, the multi-scale composite robot power supply differentiation intelligent management system is used to execute the steps in the above-described multi-scale composite robot power supply differentiation intelligent management method, including: The data acquisition module is used to collect the bus voltage of the macro robot, the current of each axis, the temperature, the terminal voltage of the micro / nano robot, the power supply current and the insertion depth of the tube. It constructs a cross-domain joint state vector time series by real-time frequency conversion spectrum extraction and power statistics fusion. The cable resistance compensation module is used to extract joint angles and micro-inlet tube depths from the cross-domain joint state vector time sequence, calculate the bending angle distribution of the entire cable through forward kinematics and path interpolation, and calculate the dynamic transmission resistance by querying the pre-calibration mapping table. The power prediction module is used to feed the cross-domain joint state vector time sequence and joint motion command sequence into the attention encoding and decoding model, combine dynamic transmission resistance to correct micro-nano equivalent power, generate macro-micro dual-side power demand prediction, and recursively calculate the time step power allocation value based on the remaining battery power to obtain the power allocation plan. The fault detection module is used to extract the micro-nano load state as a condition variable and the current spectrum of each axis as a decomposition object from the cross-domain joint state vector time sequence. It establishes a conditional expectation model, fits the external interference spectrum and subtracts the measured spectrum to obtain the residual spectrum, extracts the harmonic amplitude and high-frequency energy ratio to form a fault feature vector and outputs an early warning. The switching control module is used to receive power allocation plans and fault warnings from the state machine during the mission phase. It calculates the preload trigger advance based on the supercapacitor charging time and the peak expected time to achieve predictive switching and outputs voltage regulation and charge management commands.

[0016] The beneficial effects of this invention are as follows: By using a cross-domain joint state vector time sequence, the macroscopic side bus, current spectrum and motion state, as well as the micro-nano side terminal voltage, supply current and insertion depth are uniformly characterized. On this basis, dynamic transmission resistance estimation and terminal voltage feedforward compensation are introduced, and combined with an online adaptive correction mechanism, the transmission voltage drop caused by the change in the flexible micro-cable configuration can be predicted and compensated in advance. This maintains the micro-nano side terminal voltage within the nominal operating range during macroscopic side power fluctuations, reducing the risk of task interruption due to power supply deviation. The attention-based encoding and decoding model enables rolling prediction of power demand on both the macro and micro sides. Combined with the recursive calculation of remaining battery power, a power allocation plan is obtained. In the fault detection stage, the macro-side current spectrum is conditionally decomposed using the micro / nano load state as a conditional variable. Explainable external components are extracted from the measured spectrum, and fault feature vectors are extracted to output graded early warnings. During the mission phase, the state machine integrates the plan and the early warning state. Based on the peak prediction time and the supercapacitor charging time, predictive switching is triggered, and voltage regulation and charge management commands are output to ensure the continuity and controllability of power supply during the switching process. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-scale composite robot power supply differentiation intelligent management method of the present invention; Figure 2 This is a flowchart of the multi-scale composite robot power supply differentiation intelligent management method of the present invention. Figure 1 ; Figure 3 This is a flowchart of the multi-scale composite robot power supply differentiation intelligent management method of the present invention. Figure 2 . Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a method for differentiated intelligent management of power supplies for multi-scale composite robots, such as... Figures 1 to 3 As shown, it includes the following steps: S1 collects robot bus voltage, current and temperature of each axis, end voltage of micro-nano robot, power supply current and insertion depth, and constructs cross-domain joint state vector time sequence through real-time frequency conversion spectrum extraction and power statistics fusion. This step addresses the heterogeneous, multi-source signals generated during collaborative operations between macroscopic and micro / nano robot swarms. It establishes a time-synchronized, multi-channel acquisition system and transforms the acquired raw signals into a joint feature vector that simultaneously reflects the electrical and motion states of both types of robots. This provides a unified data foundation for subsequent steps such as voltage compensation, power prediction, fault detection, and switching decisions. The necessity of constructing a cross-domain joint state vector lies in the following: the power prediction model in S3 needs to simultaneously perceive the motion intention of the macroscopic robot and the current pipeline position of the micro / nano robot to capture the coupling relationship between the power demands of the two types of robots; the conditional residual decomposition in S4 requires the load state on the micro / nano side as a conditional variable to separate external interference from the current spectrum of the macroscopic robot. Neither of these requirements can be met by a single-domain acquisition scheme; therefore, synchronous acquisition and fusion of cross-domain signals must be achieved under a unified time reference.

[0020] On the macroscopic robot side, the data acquisition covers three levels. Bus-level signals continuously record the voltage value of the high-voltage DC bus and its short-window statistics (moving mean and moving standard deviation) at a sampling frequency of at least 1 kHz. This is used to reflect the bus's stable state in real time and capture the timing and amplitude characteristics of power transient events. The lower limit of the 1 kHz sampling frequency is determined by the typical time scale of the bus power transient event (approximately 1 ms) to ensure that transient events are not missed due to undersampling. Joint-level signals continuously record the three-phase current waveforms of each servo drive axis at a sampling rate of at least 10 kHz. This sampling rate lower limit is jointly determined by the upper limit of the target analysis frequency for subsequent short-time Fourier transform (at least twice the upper limit of the target analysis frequency, Nyquist criterion) and the frequency resolution requirement (the frequency resolution corresponding to the analysis window is at least one-third of the fundamental frequency interval). The encoder synchronously reads the angle, angular velocity, and angular acceleration of each joint in units of servo control cycles. The angular velocity is calculated by the difference between the angle sequence and the angular acceleration is calculated by the difference between the angular velocity sequence. Thermal signals are periodically collected at a sampling frequency of no less than 1Hz to measure the temperature of each driver's heat sink and the motor windings, forming a complete state record of the drive system together with the joint-level signals. In addition to the above real-time measurements, the sequence of joint motion commands for the next few control cycles is read from the command buffer of the macroscopic robot motion controller. This sequence includes the target angle, expected angular velocity, and feedforward torque estimates for each joint. This sequence introduces the motion planning layer's predictive information about the system's future behavior into the power management domain in advance.

[0021] On the micro-nano robot swarm side, the terminal voltage and supply current at the power input of each micro-nano robot are collected. These data are recorded on-chip by the micro power management chip within the robot body at sampling intervals no greater than 5ms, and uploaded to the macro platform main control system in the form of data frames via the signal core of the micro-cable. The 5ms sampling interval is constrained by the signal transmission bandwidth of the micro-cable and the typical timescale of load changes in the micro-nano robots, ensuring that current surges caused by pipe bends are fully captured. Furthermore, the current pipe insertion depth of each micro-nano robot (provided by its built-in odometer) and the current task stage identifier (maintained by its on-chip task state machine) are read as the basis for subsequent task stage perception and cable configuration estimation.

[0022] Regarding time synchronization, the control cycle of the macroscopic robot servo controller is used as the global reference clock. Each micro / nano robot data frame is appended with the current controller clock count during encapsulation. The macroscopic platform main control system aligns and interpolates the data from each channel based on the clock count at the data receiving end, eliminating sampling time deviations caused by inconsistent communication link delays. Time synchronization accuracy directly affects the pairing quality of the conditional variables and spectral feature vectors in S4. If there is a time deviation exceeding one control cycle, the conditional expectation model will be trained on incorrect pairings, resulting in residual external interference components in the residual spectrum. Therefore, time synchronization is a critical quality control step in this process.

[0023] In the feature extraction stage, the current signal of the macroscopic robot is processed first: a short-time Fourier transform with a fixed number of sampling points as the window length is applied to the three-phase current waveform of each joint to obtain the current spectrum of each analysis window; based on the real-time angular velocity of each joint synchronously acquired by S1, the angular velocity (unit: rad / s) is divided by 2π to obtain the current rotational frequency (unit: Hz) of the joint motor, which is used as the fundamental frequency. The amplitude of the fundamental frequency and its 2nd to 5th integer harmonics are extracted from the spectrum to form the current spectrum feature vector of the joint; the real-time rotational frequency is used instead of the fixed fundamental frequency corresponding to the rated speed because the chassis motor of the macroscopic robot frequently accelerates and decelerates during collaborative operation, and the actual speed changes continuously. When the fixed fundamental frequency deviates from the rated value, it will fall at the non-peak position of the spectrum, resulting in the extraction of background noise rather than the true fundamental frequency amplitude; the spectrum feature vectors of all joints are sequentially concatenated to obtain the overall current spectrum feature of the macroscopic robot. The motion state signals are then processed: the Euclidean norm of the angular velocity of each joint is calculated to obtain the comprehensive joint velocity amplitude, which serves as a proxy for the overall motion intensity of the macroscopic robot. The mean of the expected acceleration amplitude and the maximum of the expected feedforward torque amplitude for the next few steps are extracted from the motion command sequence as feedforward power prediction features, directly encoding the near-term expected power demand level of the macroscopic robot. For the micro / nano robot group, the average power supply of each micro / nano robot within the acquisition window (the average of the product of terminal voltage and supply current within the time window) is calculated, and the total power of the group and its standard deviation are statistically analyzed to obtain group power statistical features reflecting the overall load intensity and load distribution dispersion on the micro / nano side. Simultaneously, the deviation between the terminal voltage of each micro / nano robot and the nominal operating voltage is calculated to describe the current power quality status.

[0024] The bus state characteristics, current spectrum characteristics, motion state characteristics, and feedforward prediction characteristics of the macroscopic robot, as well as the power statistics and terminal voltage deviation characteristics of the micro / nano robot group, are uniformly processed by Z-score standardization and then sequentially concatenated to obtain a cross-domain joint operating state vector. This vector integrates the real-time states of the macroscopic and micro / nano robots into the same representation space and is refreshed as a whole with each acquisition window update, forming a joint operating state time sequence indexed by time steps. The complete output of this step contains two independent data streams: one is the aforementioned joint operating state time sequence, used for cable resistance estimation in S2, historical trajectory input of the encoder in S3, and condition variable extraction in S4; the other is the joint motion command sequence read from the motion controller command buffer. The data structure of this sequence is the joint command matrix (each time step × each joint × command component) for each time step in the future planning time domain, which is different from the fixed-dimensional real-time feature structure of the joint state vector. Therefore, it is transmitted separately as an independent data stream for feedforward prediction input of the S3 decoder. Both data streams are refreshed synchronously with the control cycle, together forming the complete information foundation for subsequent steps.

[0025] S2, extract joint angles and micro-inlet tube depths from cross-domain joint state vector time sequence, calculate the full cable bending angle distribution through kinematic forward solution and path interpolation, and calculate dynamic transmission resistance by querying pre-calibration mapping table; The spatial orientation of flexible microcables in multi-scale composite robotic systems is jointly determined by the joint configuration of the macroscopic robot body and the insertion depth of the micro / nano robot within a confined conduit. As the macroscopic robot moves and the micro / nano robot advances, the cable as a whole is in a state of continuous deformation, with the bending angles of its various segments constantly changing, leading to real-time fluctuations in the equivalent transmission resistance. For a 5V-level micro / nano power supply system, at a 100mA supply current, a change in cable resistance of only 1Ω can cause a 100mV voltage deviation at the end, equivalent to 2% of the nominal voltage. In real-world scenarios, as the cable advances from shallow to deep regions, the cumulative increase in equivalent resistance due to the bending angle can reach several ohms, and the corresponding voltage deviation at the end exceeds the normal operating voltage tolerance of the micro / nano robot's on-chip system (typically ±5% of the nominal value). Therefore, this dynamic cable resistance effect must be modeled and actively compensated in real time, rather than relying on the passive tracking of the closed-loop response of a DC-DC regulator.

[0026] This step divides the entire cable into two physical segments and establishes estimation models for each segment because the cable configuration determination mechanisms of the two segments are fundamentally different. The cable routing of the robot body segment (from the cable outlet of the macroscopic robot base to the fixed lead-out point near the end effector) is constrained by the clamps and has a definite geometric relationship with the joint angles, which can be accurately calculated using forward kinematics. The actual cable configuration in the confined space segment (the part extending from the end effector lead-out point to the entrance of the micro / nano robot body) is affected by the combined influence of the pipe curvature, the cable's own stiffness, and the propulsion force. In the shallow section of the pipe (few bends, small propulsion force), the pipe path model can be used for approximation, but in the deep section (many bends, large propulsion force), the cable may be locally coiled, and the estimation error of the path model will increase with the increase of the pipe depth. This is the physical basis for designing the online adaptive correction mechanism in this step.

[0027] In the configuration estimation of the robot body segment, the current joint angle sequence collected by S1 is used as input. The pose transformation matrix of the end effector relative to the base is calculated through the forward kinematics model of the macroscopic robot, and the spatial vector between the positions of each cable support fixture is calculated sequentially along the robot links. Based on the change in the direction of the spatial vector between adjacent support points, the cumulative bending angle contribution of the cable in this segment is estimated, and the bending angle estimate and unfolded length of each segment of the body segment are obtained. In the configuration estimation of the confined space segment, based on the current pipe insertion depth of the micro-nano robot and the three-dimensional model of the pipe path pre-stored in the macroscopic platform main control system, the curvature distribution of the cable centerline is calculated by interpolation from the inlet along the path model. The curvature value is converted into a local bending angle to obtain the bending angle distribution estimate at each position of the confined segment. The two estimation results are combined to form a bending angle-position distribution description of the entire cable.

[0028] In the equivalent resistance estimation stage, a bending resistance correction model is applied to each segment. This model is established through offline calibration experiments before system deployment. The specific method of the calibration experiment is as follows: using an adjustable angle bending fixture, the specified model of micro-cable is bent sequentially to seven discrete angles: 0°, 30°, 60°, 90°, 120°, 150°, and 180° (covering the entire bending range in actual use). At each angle, the resistance of the same segment of cable in the straight state is first measured using the four-wire method as a reference value. Then, the resistance in the bent state is measured. The difference between the bent state resistance and the reference value is divided by the reference value to obtain the dimensionless resistance increment coefficient corresponding to that angle (its physical meaning is the proportion of the resistance increment caused by bending to the nominal resistance, and the value range is usually between 0 and 0.3, that is, the resistance increment caused by bending does not exceed 30% of the nominal value). The above measurement is repeated for the seven discrete angles to obtain seven sets of data pairs of bending angle and resistance increment coefficient. A complete bending angle-resistance increment coefficient mapping table is constructed through piecewise linear interpolation. The consistency error of resistance characteristics between batches of cables of the same model is usually within ±5%. This error is within the correction capability of the subsequent online adaptive correction mechanism. Therefore, cables of the same model can share the same calibration table without recalibrating batch by batch. During the online phase, the bending angle of each segment is substituted into the mapping table to obtain the dimensionless resistance increment coefficient for each segment. The equivalent resistance of each segment is calculated according to the following steps: First, the nominal resistance per unit length (unit: Ω / m) of the segment is multiplied by the unfolded length (unit: m) to obtain the nominal resistance value (unit: Ω); Second, the nominal resistance value is multiplied by the dimensionless resistance increment coefficient obtained from the table to obtain the resistance increment caused by bending (unit: Ω); Third, the nominal resistance value is added to the resistance increment to obtain the equivalent resistance (unit: Ω); The equivalent resistances of all segments of the entire cable are summed to obtain the estimated dynamic transmission resistance of the entire cable under the current configuration. (Unit: Ω).

[0029] In the feedforward compensation command generation stage, the estimated value of the dynamic transmission resistance of the entire cable is used. Compared with the total power supply current of the current micro-nano robot group The product of (obtained from the output of S1) is used to calculate the expected transmission voltage drop at the current moment. ,in The unit is Ω. The unit is A. The unit is V; this voltage drop is superimposed on the base voltage setting of the secondary power supply bus to form a terminal voltage feedforward compensation command that includes cable resistance compensation, which is sent to the digital voltage regulation interface of the secondary bus DC-DC converter. This command is updated synchronously with the refresh of the joint state vector in each control cycle, so that the compensation amount tracks the changes in cable configuration in real time. It should be noted that the function of feedforward compensation is to eliminate the main body of voltage deviation caused by cable resistance changes in advance before the closed-loop response of the DC-DC regulator. Its compensation accuracy is limited by kinematic estimation error, and it is expected to control the terminal voltage deviation within ±50mV (corresponding to ±1% of a 5V system). The remaining deviation is further eliminated by an online adaptive correction mechanism.

[0030] In the online adaptive model phase, the difference between the measured terminal voltage of each micro-nano robot output by S1 and the expected terminal voltage after feedforward compensation is used as the residual signal. If the absolute mean of the residual exceeds 20mV within 5 consecutive control cycles (this default value corresponds to the minimum significant deviation that needs to trigger model correction in a 5V system), the mapping table is corrected according to the following steps: First, divide the residual signal (unit: V) by the total power supply current of the current group. (Unit: A) The first step is to obtain the cable resistance estimation error for the entire cable (unit: Ω). The second step is to determine the segments with the largest bending angles in the current cable bending angle-position distribution (i.e., the areas contributing the most to the total cable resistance, referred to as the current main bending areas). The cable resistance estimation error is weighted and distributed to each main bending segment according to the proportion of each segment's bending angle to the sum of all segment bending angles, resulting in the resistance estimation error for each segment (unit: Ω). The third step is to divide the resistance estimation error of each segment by the nominal resistance value of that segment, obtaining the resistance increment coefficient correction amount (dimensionless). The fourth step is to update the resistance increment coefficients of the corresponding bending angle intervals in the mapping table using a weighted average method. The update rule is: the new coefficient equals the old coefficient multiplied by (1 minus the learning rate) plus the correction amount multiplied by the learning rate. The default value of the learning rate is 0.1, which ensures the correction convergence speed while suppressing excessive disturbance of the mapping table by single-step residual noise. The object of the aforementioned online adaptive correction is always the dimensionless resistance increment coefficient in the mapping table, rather than the nominal unit length resistance (nominal resistance is an inherent material parameter of the cable and does not change with bending or aging). This distinction ensures the correct physical meaning of the correction operation. The introduction of the online adaptive correction mechanism upgrades the overall compensation strategy of this step from "pure feedforward" to a two-level structure of "feedforward coarse compensation + adaptive fine correction," effectively covering scenarios where the accuracy of kinematic estimation decreases in deep regions. This step outputs two quantities: the real-time updated secondary bus voltage compensation command, and the current estimated value of dynamic transmission resistance. The latter is passed into S3 as the micro / nano-side power supply efficiency factor to participate in the calculation of the power allocation scheme.

[0031] S3, the cross-domain joint state vector time sequence and joint motion command sequence are fed into the attention encoding and decoding model, and the dynamic transmission resistance is combined to correct the micro-nano equivalent power to generate macro-micro dual-side power demand prediction. The power allocation value of the time step is calculated by recursively using the remaining battery power to obtain the power allocation plan. In collaborative work scenarios, there is an indirect but real coupling relationship between the power requirements of macroscopic robots and micro / nano robots. This is the physical basis for designing the cross-domain joint prediction model in this step. After the macroscopic robot moves to a new working position, the pipeline segment in which the micro / nano robot is located changes accordingly. The bend density and cross-sectional changes on its subsequent propulsion path determine the distribution of the power requirements of the micro / nano side in the future time domain. This means that the current position of the macroscopic robot (encoded in the motion state components of the joint state vector) implicitly contains the pipeline geometry information that the micro / nano robot will encounter, thus determining the future trend of the power requirements of the micro / nano side. This indirect mapping relationship of "macroscopic position → pipeline geometry that the micro / nano robot will encounter → micro / nano power requirements" cannot be directly expressed in kinematic formulas, but it can be learned from historical collaborative work data. Meanwhile, although the motion command sequence of the macroscopic robot contains feedforward torque information for each joint, which can be directly mapped to the ideal power expectation on the macroscopic side, the actual driving power also includes frictional losses between the chassis and the ground (related to ground conditions), transmission efficiency losses of the joint reducers (related to temperature and wear conditions), and compensation current generated by the servo controller due to bus voltage fluctuations. These factors are not visible in the kinematic model, but can be traced in historical operating data. Therefore, relying solely on kinematic calculations cannot obtain sufficiently accurate power predictions; it is necessary to learn the influence of these implicit factors from historical data through data-driven models.

[0032] In the predictive model construction phase, the model receives three types of inputs: First, the joint operating state vector time series output by S1 serves as the historical trajectory encoding input, capturing the actual operating mode of the system over several control cycles. The use of cross-domain joint features enables the model to perceive the historical covariance of macro-micro power demand, which is impossible with single-domain feature inputs. Second, the motion command sequence collected from S1 serves as the feedforward prediction input, containing the expected acceleration and feedforward torque of each joint in the next few steps. These quantities provide prior expectations of macro-side power demand, guiding the model to maintain compliance with kinematic constraints based on data-driven learning. Third, the current dynamic transmission resistance estimate output by S2 serves as the input. As a parameter for power supply efficiency on the micro / nano side, it is used to correct the predicted power demand of the micro / nano robot body to the equivalent power actually drawn from the secondary bus. The correction method is: based on the predicted power demand of the micro / nano robot body, a power efficiency parameter is added... The transmission loss corresponding to the product of the predicted supply current is used to obtain the equivalent power requirement on the secondary bus side.

[0033] The model employs an attention-based sequence encoder-decoder structure. The encoder applies variable weights to the temporal dimension aggregation of historical state sequences, giving greater weight to recent states and historical moments more relevant to the current task stage. This mechanism allows the model to automatically adjust its use of historical information at different task stages, rather than treating all historical moments equally. The decoder uses the context vector output by the encoder and the feature representation of the motion command sequence as joint inputs, and autoregressively generates a sequence of stepwise power demand estimates for the macroscopic robot and the micro / nano robot in the future prediction time domain. The two outputs share weights in the middle layer of the decoder. This design forces the model to learn the covariant structure between the power demands of the two types of robots during training, rather than optimizing the two predictions as independent tasks. This allows the model to capture the cross-domain coupling pattern of "the macroscopic robot moves to a certain position → the micro / nano robot then encounters a specific pipeline segment → the micro / nano power demand changes accordingly."

[0034] The offline training of the model employs a transfer learning strategy, conducted in two phases. The first phase involves pre-training using simulation data generated from a task simulation environment. This simulation data must cover three key operating conditions: a high-power conflict scenario where a macroscopic robot moves at high speed while a micro / nano robot operates in deep environments; a pure micro / nano-side power fluctuation scenario where the macroscopic robot is stationary while the micro / nano robot encounters a pipe bend; and a baseline scenario where both are in a low-power state. Simulation data covering these three conditions is sufficient to establish the model's basic parameter distribution. The second phase involves fine-tuning using a small amount of real collaborative operation data to adapt the model to hardware characteristics in real systems that are difficult to accurately model in simulations, such as ground friction and reducer efficiency. For new factories or new pipe layouts, the distribution of power demand on the micro / nano side will change due to different pipe path geometry. Therefore, the second phase of fine-tuning needs to be performed again using a small amount of real data from the new scenario, while the parameters pre-trained in the first phase can be reused. The training loss function includes both macroscopic and micro / nano-side prediction mean square error terms. An additional weight penalty is applied to the prediction error at the micro / nano-side power peak moment (i.e., the moment the micro / nano robot encounters a pipe bend) to ensure higher prediction accuracy at the moment when the micro / nano power supply most needs protection. The online inference phase is executed in a fixed-step rolling manner. The input window shifts forward with each inference as the joint state vector is updated, and the output is the joint power demand prediction sequence in the future prediction time domain.

[0035] Based on the joint prediction results, a power allocation plan is calculated. The forward recursion state variable is the remaining available battery power at the current time step, with the initial value being the maximum discharge power corresponding to the remaining charge of the current vehicle battery. At each time step in the prediction time domain, the remaining available battery power at that time step is used as the total constraint, and the voltage at the micro / nano robot end is kept within the nominal operating range as a hard constraint (refer to the expected end voltage after cable resistance compensation in S2). Following the principle of prioritizing the hard constraints on the micro / nano side, the predicted power demand of the micro / nano side is guaranteed first, and the remaining available power is allocated to the macro side to obtain the upper limit of the macro side power at that time step. The sum of the macro side allocated power and the micro / nano side predicted power at that time step is subtracted from the remaining available battery power to obtain the remaining available battery power at the next time step. This process is repeated until the end of the prediction time domain to form a sequence of upper limit values ​​for the macro side and micro / nano side power allocation at each time step. For time windows in the prediction time domain where the difference between the predicted power value on the micro / nano side and the current power supply margin exceeds the safety margin, a planned power limiting instruction is generated for the corresponding time. This instruction is timestamped and stored in the power allocation plan, and executed by the S5 state machine at the corresponding time. Its function is to moderately limit the upper limit of the macro robot's acceleration to obtain a stable power supply margin on the micro / nano side. This type of planned power limiting instruction is planned in advance during the power allocation plan generation stage. It differs from the emergency power limiting instruction triggered by S5 when the prediction lead is insufficient in both triggering timing and execution path. The former is scheduled and executed by the S5 state machine according to the plan timestamp, while the latter is directly sent to the motion controller by the S5 state machine when it determines in real time that the preload conditions are not met. At the same time, the time window in the prediction time domain where the macro power demand exceeds the peak switching threshold is marked, and its expected occurrence time and duration are recorded as the time basis for predictive energy preload and switching triggering in the S5 step. The peak switching threshold here is the same threshold used in S5 to determine whether to trigger predictive switching, and it is uniformly set during the system parameter configuration stage. The final power allocation plan output in this step includes: the sequence of upper limit values ​​for power allocation on the macro and micro / nano sides at each time step, the list of planned power limiting trigger times and the upper limit value of acceleration for each limiting event, the expected occurrence time and duration of each peak event, and all information is transmitted to S5 to participate in redundancy switching decisions.

[0036] S4. Extract the micro-nano load state from the cross-domain joint state vector time sequence as a condition variable and the current spectrum of each axis as a decomposition object. Establish a conditional expectation model, fit the external interference spectrum and subtract the measured spectrum to obtain the residual spectrum. Extract the harmonic amplitude and high-frequency energy ratio to form a fault feature vector and output an early warning. In multi-scale composite robot collaborative operation scenarios, a specific interference mechanism exists that is not present in other application scenarios: when a micro-nano robot encounters a pipe bend or narrowing of its cross-section, the driving current surges within a timescale of 50ms to 200ms, causing a momentary power surge in the secondary power supply bus, which in turn leads to a brief voltage drop in the primary bus. The servo controllers of each axis of the macro-robot, aiming to maintain joint torque, compensate for the bus voltage fluctuation by increasing the phase current. This compensation action superimposes harmonic components corresponding to the bus fluctuation rhythm onto the current spectrum of the macro-robot. From a frequency characteristic perspective, the bearing fault characteristic frequencies (outer ring fault frequency BPFO and inner ring fault frequency BPFI) of the macro-robot chassis drive motor (speed range 500rpm to 1500rpm) are typically distributed in the range of 25Hz to 250Hz. The sudden current surge event when the micro-nano robot encounters a pipe bend is transmitted to the macro-robot current spectrum through the path of bus voltage drop → servo compensation current, generating a response within the servo controller bandwidth (typically 100Hz to 500Hz), where the components overlapping with the bearing fault characteristic frequencies are real. More importantly, during collaborative operations, the frequency with which micro- and nano-robots encounter bends depends on the density of bends in the pipeline and their propulsion speed. These values ​​may dynamically change with the progress of the task, resulting in the non-fixed position of external interference components in the frequency spectrum, which cannot be eliminated by a pre-set notch filter. The existence of this interference mechanism leads to an increased false alarm rate during periods of heavy load for micro- and nano-robots when directly diagnosing faults in the macroscopic robot current spectrum in collaborative operation scenarios. Existing fault diagnosis methods have not designed decoupling mechanisms specifically for this particular interference path.

[0037] In the establishment phase of the conditional expectation reference model, data collected during the historical operation of the system during which the drive system was confirmed to be in a healthy state were used to establish a conditional expectation estimation model for each servo drive axis of the macroscopic robot. The training samples for each axis are composed of the load state vector of the micro-nano robot group at the same moment (extracted from the S1 output, including the average power supply, power change rate and total power fluctuation index of each micro-nano robot) and the current spectrum feature vector of that axis. The time alignment quality of the two directly depends on the time synchronization accuracy of the S1 step. The reason for establishing separate models for different axes is that the motor parameters (rated speed, number of pole pairs) of each axis are different, the corresponding fundamental frequency and fault characteristic frequency are different, and the response gain and bandwidth of each axis servo controller to bus fluctuations are also different, resulting in differences in the magnitude of external interference experienced by each axis, making it impossible to cover all axes with a single model. During training data preparation, the motion state of each axis (the real-time rotation frequency of the axis, calculated from the real-time angular velocity output by S1) was divided into several discrete intervals. A conditional expectation estimation model was established for each interval. This allowed the model to naturally eliminate the influence of the axis's own rotational speed changes on the spectrum while fitting the micro / nano load-spectrum relationship, avoiding misjudging spectral variations caused by rotational speed changes as external interference. The conditional expectation estimation model for each axis used the micro / nano load state vector as input, the expected value of the current spectrum of the corresponding axis as the output target, and was trained using mean square error as the basic loss function. During training, greater training weight was given to the spectral components corresponding to known external interference frequency intervals (frequency bands corresponding to the bus fluctuation rhythm, i.e., frequency bands overlapping with the micro / nano load change rhythm within the servo controller bandwidth). This guided the model to focus on fitting the conditional expectation of external interference components, while ignoring the spectral components caused by changes in the internal state of the drive system (including fault characteristics). This ensured that the model only captured "externally interpretable" spectral variations, while retaining internal fault characteristics in the residuals.

[0038] During the online detection phase, for each servo drive axis, the conditional expectation estimation model to be invoked and the corresponding spectral expectation distribution are jointly determined using the current micro / nano load state vector and the discrete interval to which the axis's current real-time frequency belongs. The measured current spectrum feature vector of the axis output by S1 is subtracted from the expectation distribution to obtain the conditional residual spectrum vector of the axis, which is the deviation component in the current spectrum of the axis that exceeds the normal conditional expectation at the current moment. The above operation is performed on all axes to obtain a multi-axis conditional residual spectrum set. The external harmonic contribution caused by micro / nano load fluctuations has been eliminated from the conditional residual spectrum of each axis. The remaining spectral deviation mainly comes from the internal state changes of each axis drive system, including the emergence of early fault symptoms. It is worth noting that the input quality of the conditional expectation model depends on the cable resistance compensation accuracy of S2: if the compensation of S2 is inaccurate, the measured terminal voltage of the micro-nano robot will deviate from the true load state, causing the micro-nano load state vector extracted from S1 to be distorted, which in turn affects the quality of the conditional variables of the conditional expectation model of each axis, leaving external interference components in the residual spectrum; therefore, there is an indirect technical dependency between S2 and S4, and the compensation accuracy of S2 is a prerequisite for the decoupling effect of S4.

[0039] In the fault feature extraction stage, the conditional residual spectrum of each axis is used as the operation object, and two types of features are extracted for each axis. The first type is the harmonic residual amplitude sequence: the current rotation frequency obtained by converting the real-time angular velocity of the axis output by S1 is taken as the fundamental frequency, and the residual amplitude at the fundamental frequency and its 2nd to 5th integer harmonics is calculated one by one in the residual spectrum to form the harmonic residual amplitude sequence of the axis. This sequence is responsive to electromagnetic fault precursors such as rotor eccentricity and magnetic pole wear. The second category is the high-frequency energy ratio index: using 5 times the current rotational frequency as the cutoff frequency, the ratio of the energy in the residual spectrum above this cutoff frequency (i.e., the frequency band above the 5th harmonic frequency; the bearing fault characteristic frequencies BPFO and BPFI are usually distributed in the range of 3 to 10 times the rotational frequency, mainly falling within this high-frequency band) to the total energy of the entire frequency band is obtained. The reason for using 5 times the current rotational frequency as the cutoff frequency instead of a fixed Hz value is that the bearing fault characteristic frequency changes with the rotational speed. A fixed cutoff frequency would cause the fault characteristic frequency to fall below the cutoff frequency and be missed when the rotational speed deviates from the design point. This index is sensitive to high-frequency impacts caused by damage to the bearing raceway surface. The harmonic residual amplitude sequence and high-frequency energy ratio of each axis are combined with the corresponding axis driver heat sink temperature characteristics obtained from the S1 output to form the fault feature vector of that axis. The fault feature vectors of all axes are sequentially concatenated to obtain a multi-dimensional multi-axis fault feature vector. The statistical mean of each component of the vector is taken within a sliding time window to form a smoothed fault feature representation, which is used to suppress misjudgments caused by single accidental interference.

[0040] During the multi-level early warning output phase, the distribution of fault feature vectors statistically analyzed during system healthy operation is used as a reference benchmark, and a three-level threshold system is used for decision-making. When any component in the smoothed fault feature vector exceeds the first threshold for the first time, a level one early warning is triggered. The system increases the acquisition frequency and records the feature timing for analysis, without changing the current task scheduling. When any component continuously exceeds the first threshold for several consecutive control cycles, or exceeds the second threshold once, a level two early warning is triggered. The early warning status is simultaneously notified to the task phase state machine of S5, which then applies a power conservative margin limit to the axis driver in the subsequent power allocation plan. At the same time, it is recommended that the task scheduling layer arrange priority maintenance after completing the current operation. When any component exceeds the third threshold or multiple axes simultaneously exhibit level two or higher early warnings, a level three early warning is triggered, and the redundancy switching procedure of S5 is immediately initiated. This step outputs the fault feature vector and the level three early warning status at the current moment. These two quantities are jointly transmitted to S5 as one of the key criteria for triggering the switching plan.

[0041] S5, the mission phase state machine receives power allocation plans and fault warning status, calculates the preload trigger advance based on the supercapacitor charging time and peak expected time to achieve predictive switching, and outputs voltage regulation and charge management commands. During ultrasonic scanning at the deepest part of the pipeline, any power interruption exceeding the permissible range will cause the loss of on-chip scan data buffer and motor driver reset, resulting in the loss of the micro-nano robot's current position record and inability to resume the detection task in situ, leading to high redeployment costs. The passive protection scheme suffers from a fundamental timing defect in this scenario: when the macro-robot chassis drive motor starts at high speed, the bus voltage drop timescale is approximately tens of milliseconds. The delay in the entire passive response chain, from detecting the bus voltage drop to the supercapacitor completing the switching connection, is typically on the same order of magnitude. The power interruption of the micro-nano robot occurs before the passive scheme completes its response. The core design idea of ​​this step is to advance the switching timing from "responding after detecting the bus voltage drop" to "preloading when a peak event is predicted." Through the power peak event forecast provided by S3, the supercapacitor's charging preparation is completed before the peak arrives, and the discharge circuit is connected in advance, ensuring that the power supply to the micro-nano side is completely independent of the bus during power surges, thus eliminating the timing defect of the passive scheme from a fundamental mechanism perspective.

[0042] The method for determining the preload time window of a supercapacitor is as follows. Preload required time. The rated capacity of the supercapacitor (Unit: F) Target charging voltage (Unit: V) and maximum allowable charging current (Unit: A) The estimation method is: the amount of charge transferred to charge the supercapacitor from its current charging voltage to the target voltage (equal to...) The shortest charging time is obtained by dividing the product of the voltage difference and the maximum charging current. Considering the characteristic that the current gradually decreases at the end of the charging process, the actual preload time is taken as 1.5 times the shortest charging time as a conservative estimate. The default value of this coefficient is 1.5. Preload trigger advance. The method for determining this is as follows: take the difference between the predicted peak event time of S3 and the current time, and subtract the conservative preload time. If the difference is greater than zero, preloading is triggered immediately at the current moment; if the difference is less than zero, it indicates that the prediction lead is insufficient to complete the full preloading. In this case, the system assesses the discharge duration that can be supported based on the actual state of charge of the supercapacitor. If the supported duration covers the predicted peak event duration, the switching is still performed; otherwise, a macroscopic power limiting command is triggered simultaneously to compress the peak amplitude and duration. To address the timing deviation caused by the S3 prediction error, during the macroscopic platform travel phase and the collaborative operation phase, the supercapacitor always maintains a minimum charge level of no less than 30% of its rated capacity (this default value corresponds to being able to support approximately 100ms of emergency discharge without preloading), as a fallback guarantee against prediction errors.

[0043] The task phase state machine divides the collaborative operation process into five typical phases and defines the transition conditions between adjacent phases. The transition exit condition for the macroscopic platform travel phase is: the positional error of the macroscopic robot's end effector reaching the target pipe opening coordinates is lower than the positioning accuracy requirement, and the macroscopic platform's movement speed drops to near zero. The transition exit condition for the micro / nano robot deployment phase is: all micro / nano robots to be deployed in the current batch have completed pipe entry initiation, and the pipe entry depth exceeds the initial safe pipe entry length threshold. The transition condition between the collaborative operation phase and the deep-area operation phase is: the pipe entry depth of any micro / nano robot exceeds 60% of the pre-stored pipe length (this default ratio corresponds to the critical depth at which the dynamic cable resistance begins to affect power supply quality; that is, above this depth, the end voltage deviation caused by the equivalent resistance increase due to bending exceeds 1% of the nominal value). The transition exit condition for the deep-area operation phase is: each micro / nano robot has completed the scanning task of the current batch of pipe bundles and initiated the withdrawal procedure. The transition exit condition for the micro / nano robot withdrawal phase is: each micro / nano robot has completely withdrawn from the pipe bundle, and the pipe entry depth returns to zero.

[0044] The charge management strategy for the local supercapacitor in each stage is as follows. During the macroscopic platform travel phase and the micro / nano robot deployment phase, the supercapacitor maintains a low charge level, supplemented only by trickle charging from the secondary bus to reduce self-discharge during standby, while maintaining a minimum charge level of no less than 30% of its rated capacity. During the collaborative operation phase (when the micro / nano robot's pipe entry depth is between the initial safety threshold and 60% of the pipe length), the supercapacitor also maintains a minimum charge level of no less than 30% of its rated capacity. During this phase, the macroscopic robot may have already begun moving towards the next batch of pipe openings. If a macroscopic-side power peak event is marked in the power allocation plan of S3, the state machine triggers predictive switching according to the preload time window calculation method, with the switching logic being the same as in the deep-area operation phase. During this phase, the impact of dynamic cable resistance on power quality has not yet reached a significant level, and the feedforward compensation of S2 is sufficient to maintain the terminal voltage within the allowable range. The main function of the supercapacitor switching is to cope with bus power surges, rather than cable resistance compensation. When the state machine determines that the collaborative operation phase is about to transition to the deep-area operation phase (by monitoring the growth rate of the micro-nano robot's entry depth and estimating the remaining time before triggering the 60% depth threshold), it sends a supercapacitor fast charging command to the secondary bus DC-DC converter in advance at the time determined by the preload time window calculation method mentioned above. This charges the supercapacitors in each micro-nano robot to full charge, ensuring that the preload charging is completed before the state machine completes the transition to the deep-area operation phase.

[0045] During the deep-area operation phase, the state machine continuously checks two types of triggering conditions. The first category comes from S3: If the expected occurrence time of the macroscopic power peak event marked in the current power allocation plan is within the preload time window after the current time, and the peak amplitude exceeds the peak switching threshold, then the predictive switching process is triggered according to the above preload time window calculation method. Before the expected event time, the discharge control switch of the local supercapacitor of each micro-nano robot is turned on in advance, so that the supercapacitor actively injects current into the micro-nano robot drive circuit, forming a parallel power supply state of the local supercapacitor and the secondary bus to the micro-nano robot; at the same time, the secondary bus DC-DC converter switches to the standby mode with low output weight to reduce the interference of its output oscillation caused by bus fluctuations to the micro-nano side; after the macroscopic power peak event actually ends (based on the bus voltage returning to the normal range), the output weight of the secondary bus is gradually increased, and the power supply of the micro-nano robot is smoothly switched back to the secondary bus as the main body. At the same time, the supercapacitor is started to supplement the charging in the background with the maximum allowable charging current. The target charge level of the supplementary charging is the full charge state. The charging process is carried out in parallel with the remaining output capacity of the secondary bus while supplying power to the micro-nano robot normally, without the need to switch the charging mode. During the recharging process, S3 continuously performs rolling inference. If the supercapacitor has not yet recovered to a fully charged state when the next peak event is predicted, the state machine re-executes the preload time window calculation based on the current actual charge state. If the discharge duration that the current charge state can support covers the expected duration of the next peak event, then the predictive switching is triggered according to the normal process. If it is insufficient, the planned power limiting command corresponding to the S3 plan is triggered synchronously (if it has been marked in the plan) or an emergency power limiting command is generated immediately to compress the peak amplitude and duration, so that the supercapacitor in the current charge state can cover the compressed peak event, thereby maintaining continuous protection of the power supply to the micro-nano side in the scenario of continuous peak events. The second category comes from S4: If S4 triggers a level 2 warning, the state machine does not wait for the peak event forecast of S3, but directly switches all micro-nano robots still in the deep zone to local supercapacitor power supply, applies a conservative upper limit constraint on the corresponding axis of the macro robot, and issues a task adjustment suggestion to the task scheduling layer to prioritize the withdrawal of micro-nano robots after ensuring that the current operation is completed; if S4 triggers a level 3 warning, the current motion planning of the macro robot is immediately stopped, the drive power is reduced in a safe deceleration mode, and the controlled withdrawal program of all micro-nano robots is triggered at the same time. The state machine monitors the withdrawal progress throughout the process until each micro-nano robot safely exits the control.

[0046] This step ultimately outputs two types of control commands: a comprehensive voltage regulation command for the secondary power supply bus DC-DC converter, which superimposes the cable resistance feedforward compensation calculated by S2 and the bus weight adjustment determined in this step based on the switching state, ensuring a smooth transition of the micro / nano robot's terminal voltage before and after the switching process. "Zero interruption" means that the terminal voltage remains within the nominal operating range (±5% of the nominal value) throughout the switching process, rather than remaining completely unchanged. The second type is a charge management command sequence for the supercapacitor controllers of each micro / nano robot body, containing complete control information for timing actions such as preload charging trigger, discharge circuit connection, weight switching, and background supplementary charging initiation. Both types of commands are uniformly scheduled and output by the state machine, forming a complete redundant power supply control chain covering the entire process from predictive preload to bumpless switching and recovery supplementation. This chain works in conjunction with the feedforward compensation output of S2 to jointly ensure differentiated and highly reliable power supply for the macroscopic and micro / nano robots throughout the collaborative operation of the multi-scale composite robots.

[0047] In one embodiment of the present invention, the inspection task of air cooler tube bundles in a refining and chemical plant is taken as the implementation object. Specific parameters are: tube inner diameter 22mm, tube length 4.5m, totaling 320 tubes. Four micro-nano ultrasonic inspection robots are deployed for each inspection task. The nominal operating voltage of the micro-nano robots is 5V, with an allowable operating voltage range of 4.75V to 5.25V (±5% of the nominal value). During task execution, after the macroscopic mobile platform completes the deployment of the micro-nano robots for the first batch of four tubes, the micro-nano robots advance into the tubes at a stable speed for inspection. Simultaneously, the macroscopic platform initiates a lateral displacement action towards the opening of the second batch of tubes, with a displacement distance of approximately 0.8m. The chassis drive motor accelerates and generates a noticeable bus power pulse. The following two tables show typical operating data collected during this time period, with each time step corresponding to one control cycle; data examples are shown in Tables 1 and 2. Table 1. Examples of power supply circuit data collected from micro-nano robot swarms

[0048] Table 2. Examples of Operating Data for the Macroscopic Robot Chassis Drive System

[0049] At time t1, step S3 predicts, based on the motion command sequence, that the macroscopic robot chassis motor will enter the acceleration phase during the time period from t2 to t3. It is expected that the peak power of the bus will exceed the first threshold, triggering the predictive switching process: During the control cycle between t1 and t2, the discharge circuit of the local supercapacitor of each micro-nano robot is connected in advance, and the supercapacitor begins to actively inject current into the micro-nano robot drive circuit; At the same time, the feedforward compensation command of step S2 adjusts the secondary bus voltage setting value synchronously upward based on the dynamic transmission resistance estimated by the current cable configuration, to offset the increase in equivalent voltage drop caused by the increase in power supply current. During time steps t2 to t3, the macroscopic robot chassis motor was in the acceleration phase, with the bus current increasing from 12.7A to 17.6A and the bus voltage decreasing from 48.3V to 47.4V, a drop of 0.9V. Since the supercapacitor had already completed preloading and was connected in advance at time t1, the power supply for the micro-nano robot during this period was dominated by the local supercapacitor. The voltage at the No. 1 robot only fluctuated slightly from 5.02V to 4.96V, always remaining within the allowable operating range (4.75V to 5.25V), verifying the effective protection of the micro-nano side power supply by predictive preload switching.

[0050] When processing the macroscopic robot current spectrum at times t2 and t3 in step S4, the micro / nano load state vector increases at time t2 relative to t1 (the group power standard deviation increases from 38.2mW to 44.5mW). The conditional expectation estimation model updates the expected spectrum distribution based on this change, removing the external harmonic components introduced by the micro / nano load fluctuations through the bus coupling path from the measured spectrum. This ensures that the residual spectrum only retains the spectral deviation caused by the internal state changes of the macroscopic robot drive system. The high-frequency energy proportion at the corresponding time increases slightly from 0.031 to 0.035, a change within the normal fluctuation range of a healthy state, without triggering any level of warning. This aligns with the objective reality that the macroscopic robot drive system is actually in a healthy state, verifying the effectiveness of the conditional residual decomposition mechanism in eliminating external interference and suppressing false alarms. From time t4 to t5, the macroscopic robot chassis motor decelerates to a stable state, and the bus voltage returns to the normal range. In step S5, the output weight of the secondary bus is gradually increased, and the power supply of the micro-nano robot is smoothly switched back to being dominated by the secondary bus. At the same time, the supercapacitor is started to charge in the background. During the entire switching process, the voltage at the micro-nano robot end does not fluctuate beyond the allowable range, achieving zero-interruption redundant switching.

[0051] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multi-scale composite robot power supply differentiated intelligent management method, characterized in that, Includes the following steps: S1 collects macroscopic robot bus voltage, axis current, temperature, micro / nano robot end voltage, power supply current and insertion depth, and constructs cross-domain joint state vector time series through real-time frequency conversion spectrum extraction and power statistics fusion; S2, extract joint angles and micro-inlet tube depths from cross-domain joint state vector time sequence, calculate the full cable bending angle distribution through kinematic forward solution and path interpolation, and calculate dynamic transmission resistance by querying pre-calibration mapping table; S3, the cross-domain joint state vector time sequence and joint motion command sequence are fed into the attention encoding and decoding model, and the dynamic transmission resistance is combined to correct the micro-nano equivalent power to generate macro-micro dual-side power demand prediction. The power allocation value of the time step is calculated by recursively using the remaining battery power to obtain the power allocation plan. S4. Extract the micro-nano load state from the cross-domain joint state vector time sequence as a condition variable and the current spectrum of each axis as a decomposition object. Establish a conditional expectation model, fit the external interference spectrum and subtract the measured spectrum to obtain the residual spectrum. Extract the harmonic amplitude and high-frequency energy ratio to form a fault feature vector and output an early warning. S5, the mission phase state machine receives power allocation plans and fault warning states, calculates the preload trigger advance based on the supercapacitor charging time and peak expected time to achieve predictive switching, and outputs voltage regulation and charge management commands.

2. The multi-scale composite robotic power supply differentiated intelligent management method of claim 1, wherein, The construction of the cross-domain joint state vector time series includes: Using the control cycle of the macroscopic robot servo controller as the global reference clock, the data collected from each channel is aligned and interpolated according to clock counting. Short-time Fourier transform is applied to the three-phase current waveforms of each joint. The current rotation frequency, which is calculated from the real-time angular velocity of each joint, is taken as the fundamental frequency. The amplitude at the fundamental frequency and its preset integer multiples is extracted and spliced ​​to obtain the current spectrum characteristics of each axis. The average power supply, total power of the group, standard deviation, and terminal voltage deviation within the acquisition window of each micro-nano robot are calculated to obtain the micro-nano load status. The macroscopic side bus state, current spectrum and motion state characteristics are sequentially spliced ​​with the micro-nano side load state after standardization to obtain a cross-domain joint state vector time sequence.

3. The method of claim 1, wherein, The calculation of the dynamic transmission resistance includes: The flexible micro-cable is divided into a robot body segment and a confined space segment. The body segment is calculated by using the kinematic forward kinematics to deduce the spatial vector of the cable support point and estimate the cumulative bending angle from the joint angles of each joint in the cross-domain joint state vector time sequence. The confined space segment is calculated by interpolating the cable centerline curvature from the micro-inlet pipe depth according to the pre-stored 3D model of the pipe path and converting it into a local bending angle. The two segments are merged to form the full cable bending angle position distribution. Substitute the bending angle of each segment into the offline calibrated bending angle resistance increment coefficient mapping table, calculate the equivalent resistance of each segment from the nominal resistance of each segment and the increment coefficient obtained from the table, and sum them to obtain the dynamic transmission resistance. The expected transmission voltage drop is calculated by multiplying the dynamic transmission resistance by the total power supply current of the micro-nano robot group, and then superimposed on the base voltage setting of the secondary power supply bus to form the terminal voltage feedforward compensation command.

4. The method of claim 1, wherein, The generation of the power allocation plan includes: The cross-domain joint state vector time sequence is used as the encoding input, the joint motion command sequence is used as the feedforward input, and the dynamic transmission resistance is used as the micro-nano side power supply efficiency parameter. These are all fed into the attention encoding and decoding model. The encoder of the attention encoding and decoding model applies variable weights to the time aggregation of historical time series, and the decoder generates macro and micro dual-sided power demand predictions through autoregression. The two outputs share weights in the middle layer of the decoder. The transmission loss corresponding to the product of dynamic transmission resistance and predicted supply current is superimposed on the predicted micro / nano body power to obtain the micro / nano equivalent power. With the remaining battery power as the total constraint and the micro / nano side voltage maintaining the nominal operating range as the hard constraint, the power allocation value of each time step is obtained by forward recursion in the prediction time domain according to the micro / nano side priority principle and written into the power allocation plan.

5. The multi-scale composite robot power supply differentiation intelligent management method according to claim 1, characterized in that, The formation of the fault feature vector and the output of the fault warning status include: For each servo drive axis of the macro robot, the micro-nano load state extracted from the time sequence of the cross-domain joint state vector is used as the condition variable. Conditional expectation models are established according to the preset discrete interval to which the current frequency of each axis belongs, and the expected distribution of the current spectrum of the corresponding axis is output. Subtract the current spectrum of each axis from the expected distribution of the corresponding current spectrum to obtain the residual spectrum of each axis; extract the harmonic amplitude from the residual spectrum with the current rotation frequency as the fundamental frequency, calculate the proportion of high-frequency energy in the frequency band above the current rotation frequency as the cutoff frequency, and splice it with the corresponding axis temperature to form the fault feature vector of the axis. The components of the fault feature vector are judged based on a three-level threshold system consisting of a first threshold, a second threshold, and a third threshold, and the fault warning status is output.

6. The method of claim 1, wherein, The implementation of the predictive switching includes: The charging time of a supercapacitor is obtained by multiplying the shortest charging time, determined by the supercapacitor's rated capacity, target charging voltage, and maximum allowable charging current, by a preset conservative coefficient. The preload trigger advance is obtained by subtracting the supercapacitor charging time from the difference between the peak estimated time and the current time. When the preload trigger advance is greater than zero, the preload charging and discharging circuit is immediately connected. When the preload trigger advance is less than zero, the discharge duration that can be supported is evaluated according to the current state of charge of the supercapacitor. If the discharge duration that can be supported covers the duration corresponding to the peak estimated time, the switching is performed. Otherwise, the macroscopic power limiting command is triggered synchronously. During the task phase, the state machine superimposes the cable resistance feedforward compensation component formed by the dynamic transmission resistance with the bus weight adjustment amount corresponding to the switching state, and outputs voltage regulation and charge management commands.

7. The multi-scale composite robot power supply differentiation intelligent management method according to claim 2, characterized in that, The cross-domain joint state vector timing sequence and the joint motion command sequence are output as two independent data streams in parallel and are both refreshed synchronously with the control cycle. The joint motion command sequence is read from the command buffer of the macroscopic robot motion controller. It contains the joint command matrix for each time step in the future planning time domain. The joint command matrix is ​​indexed by each time step, each joint, and the command component. The cross-domain joint state vector time sequence is used for the calculation of dynamic transmission resistance, the historical trajectory input of the attention encoding and decoding model, and the condition variable extraction of the conditional expectation model. The joint motion command sequence is used for the feedforward prediction input of the attention encoding and decoding model.

8. The multi-scale hybrid robot power supply differentiated intelligent management method of claim 3, wherein, The bending angle resistance increment coefficient mapping table is equipped with an online adaptive correction mechanism: The difference between the measured terminal voltage of the micro-nano robot and the expected terminal voltage after feedforward compensation is used as the residual signal. When the absolute mean of the residual signal exceeds the preset residual threshold within several consecutive control cycles, the residual signal is divided by the total power supply current of the micro-nano robot group to obtain the total cable resistance estimation error. The cable resistance estimation error is weighted and distributed to each segment according to the proportion of the bending angle of the segments with the largest bending angle in the distribution of bending angles of the entire cable to the sum of the bending angles. Then, the error is divided by the nominal resistance value of the corresponding segment to obtain the resistance increment coefficient correction amount of each segment. The incremental coefficients of the corresponding bending angle interval in the bending angle resistance incremental coefficient mapping table are updated using a weighted average method according to the preset learning rate.

9. The method of claim 4, wherein, In addition to the time-step power allocation values, the power allocation plan also includes a list of planned power limiting trigger times and a macroscopic peak time window. For the time window in which the difference between the power demand of the micro-nano side and the remaining power of the battery in the predicted time domain is less than the preset safety margin, a planned power limiting instruction is generated at the corresponding time. Each planned power limiting instruction carries the corresponding acceleration upper limit and is written to the planned power limiting trigger time list with a timestamp. For the time window in which the macroscopic power demand exceeds the peak switching threshold within the predicted time domain, record the predicted peak time and the corresponding duration and write them into the macroscopic peak time window. The list of planned power limiting trigger times, together with the macroscopic peak time window, is fed into the task phase state machine to trigger predictive switching.

10. A multi-scale composite robot power supply differentiated intelligent management system for performing the steps of the multi-scale composite robot power supply differentiated intelligent management method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to collect the bus voltage of the macro robot, the current of each axis, the temperature, the terminal voltage of the micro / nano robot, the power supply current and the insertion depth of the tube. It constructs a cross-domain joint state vector time series by real-time frequency conversion spectrum extraction and power statistics fusion. The cable resistance compensation module is used to extract joint angles and micro-inlet tube depths from the cross-domain joint state vector time sequence, calculate the bending angle distribution of the entire cable through forward kinematics and path interpolation, and calculate the dynamic transmission resistance by querying the pre-calibration mapping table. The power prediction module is used to feed the cross-domain joint state vector time sequence and joint motion command sequence into the attention encoding and decoding model, combine dynamic transmission resistance to correct micro-nano equivalent power, generate macro-micro dual-side power demand prediction, and recursively calculate the time step power allocation value based on the remaining battery power to obtain the power allocation plan. The fault detection module is used to extract the micro-nano load state as a condition variable and the current spectrum of each axis as a decomposition object from the cross-domain joint state vector time sequence. It establishes a conditional expectation model, fits the external interference spectrum and subtracts the measured spectrum to obtain the residual spectrum, extracts the harmonic amplitude and high-frequency energy ratio to form a fault feature vector and outputs an early warning. The switching control module is used to receive power allocation plans and fault warnings from the state machine during the mission phase. It calculates the preload trigger advance based on the supercapacitor charging time and the peak expected time to achieve predictive switching and outputs voltage regulation and charge management commands.

Images