Cloud edge computing power collaborative management method for multi-source heterogeneous embodied intelligent devices

Abstract

The application discloses a cloud edge end computing power collaborative management method for multi-source heterogeneous embodied intelligent devices, and particularly relates to the technical field of computer resource management and collaborative scheduling, and is used for solving the problem that reasoning result transient deviation caused by electrical disturbance cannot be perceived due to the isolation of computing power board state monitoring and embodied device physical action monitoring; by acquiring power component operating parameters and reasoning results of the computing power board, the power component operating parameters are subjected to frequency spectrum conversion and driving current jump frequency distribution is extracted, when there is spectrum overlap, an electrical disturbance event is determined, voltage frequency spectrum characteristics are acquired and the degree of resonance overlap is determined, the operator output characteristic vector is mapped to the characteristic space constructed by the history data of the undisturbed period, and the isolated path length is calculated as the characteristic deviation, the credibility reduction degree of the reasoning result is determined based on the characteristic deviation, and the computing power resource allocation weight of the embodied intelligent device in the cloud edge end collaborative task is adjusted according to the credibility reduction degree.

Description

Technical Field

[0001] This invention relates to the field of computer resource management and collaborative scheduling technology, and more specifically, to a cloud-edge-device computing power collaborative management and control method for multi-source heterogeneous embodied intelligent devices. Background Technology

[0002] In existing technologies, collaborative operation of multi-source heterogeneous embodied intelligent devices in scenarios such as substations and power transmission channels typically employs a three-layer cloud-edge-device architecture for management and control. Edge computing boards are mounted on the embodied devices themselves, such as drones and quadruped robots, responsible for performing on-site inference tasks and transmitting the results back to the edge or cloud for aggregation and analysis. To support collaborative scheduling decisions, the management platform needs to monitor the real-time operating status of each computing board, including online status, chip utilization, and board temperature, and allocate computing resources and orchestrate tasks accordingly. In a typical workflow, the scheduling system assumes that when a computing board is online and all monitored indicators are within their nominal range, its output inference results are considered correct and reliable, and it issues linkage commands to other collaborative nodes accordingly.

[0003] The current management approach has flaws: the monitoring of the computing board's operational status and the monitoring of the physical movements of the embodied device are two isolated pathways. The former focuses on digital domain parameters, while the latter focuses on attitude and motion control, and no correlation or judgment mechanism is established between the two. In actual operation, electrical disturbances generated by the motors and other power components of the embodied device during operation can be coupled to the computing board through a shared power supply circuit, causing transient electrical fluctuations within the processor. At this time, the computing board's reported indicators do not show obvious abnormalities, causing the management platform to continue to assign critical collaborative tasks to the computing board, which may have transient deviations in inference results, as a trusted node based on normal status parameters. Other devices in the link then perform verification or linkage actions accordingly, thus raising unnoticed reliability questions about the basis of the entire collaborative judgment. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices includes: S1: Obtain the operating parameters of the power components of the embodied intelligent device during the execution of the task, as well as the inference results output by the computing board mounted on the embodied intelligent device; S2: Perform spectrum conversion on the operating parameters of the power components, extract the frequency distribution of the drive current jump, and determine that there is an electrical disturbance event when the frequency distribution of the drive current jump overlaps with the operating frequency of the computing unit of the computing board. S3: Obtain the voltage spectrum characteristics of electrical disturbance events and determine the degree of resonance overlap between the energy concentration frequency band and the operating frequency of the computing unit in the voltage spectrum characteristics; S4: When the degree of resonance overlap does not exceed the preset degree threshold, the operator output feature vector corresponding to the operation unit in the inference result is mapped to the feature space constructed by the historical operator output feature vectors in the undisturbed period, and the isolated path length of the operator output feature vector in the feature space is calculated, and the isolated path length is used as the feature deviation. S5: Determine the degree of decrease in the credibility of the inference result based on feature bias; S6: Adjust the weight of computing resources allocated to embodied intelligent devices in cloud-edge-device collaborative tasks based on the degree of credibility decline.

[0006] Furthermore, the operating parameters of the power components of the embodied intelligent device during task execution and the inference results output by the computing board mounted on the embodied intelligent device are obtained, including: The drive current sequence and motor speed sequence are collected in real time from the power driver of the embodied intelligent device, and the drive current sequence and motor speed sequence are used as the operating parameters of the power component; The inference result output stream is collected in real time from the inference framework interface of the computing board, and the output tensor of each inference in the inference result output stream is used as the inference result. The drive current sequence, motor speed sequence, and inference results are aligned according to the timestamp of the acquisition time to form correlation data between the power component operating parameters and the inference results, carrying time tags.

[0007] Furthermore, the association data between the operating parameters of the power components carrying time tags and the inference results is formed, including: taking the output time of each inference in the inference result output stream as the reference time, finding the collection point with the smallest time interval from the reference time in the drive current sequence and the motor speed sequence respectively, binding the values ​​of the collection points of the drive current sequence and the collection points of the motor speed sequence with the inference results corresponding to the reference time into an association data record, traversing each inference in the inference result output stream, and obtaining the association data between the operating parameters of the power components carrying time tags and the inference results.

[0008] Furthermore, the operating parameters of the power components are subjected to spectrum conversion to extract the drive current switching frequency distribution. When the drive current switching frequency distribution overlaps with the operating frequency of the computing unit of the computing board, an electrical disturbance event is determined to exist, including: The drive current spectrum is obtained by performing a fast Fourier transform on the drive current sequence. Frequency components with amplitudes exceeding a preset fluctuation threshold are extracted from the drive current spectrum, and the extracted frequency components are used as the drive current jump frequency distribution. The operating frequency of the computing unit of the computing board is obtained, and each frequency component in the frequency distribution of the drive current jump is compared with the operating frequency of the computing unit. When the frequency deviation of a frequency component from the operating frequency of the computing unit falls within the preset tolerance range, it is determined that there is a spectral overlap between the frequency distribution of the drive current jump and the operating frequency of the computing unit, and an electrical disturbance event is identified.

[0009] Furthermore, the voltage spectrum characteristics of the electrical disturbance event are obtained, and the degree of resonance overlap between the energy concentration frequency band in the voltage spectrum characteristics and the operating frequency of the computing unit is determined, including: During the time period when an electrical disturbance event is determined to exist, the power supply voltage waveform of the computing board is collected, and a fast Fourier transform is performed on the power supply voltage waveform to obtain the voltage spectrum. The voltage spectrum is used as the voltage spectrum feature of the electrical disturbance event. Extract the frequency corresponding to the peak power spectral density from the voltage spectrum as the energy concentration band; The absolute value of the frequency difference between the energy concentration frequency band and the operating frequency of the computing unit is calculated, and the absolute value of the frequency difference is used as the degree of resonance overlap.

[0010] Furthermore, when the degree of resonance overlap does not exceed a preset threshold, the operator output feature vector corresponding to the operation unit in the inference result is mapped to the feature space constructed by the historical operator output feature vectors within the undisturbed period. The isolated path length of the operator output feature vector in the feature space is calculated, and the isolated path length is used as the feature deviation, including: When the degree of resonance overlap does not exceed the preset threshold, the operator output feature vector corresponding to the operation unit is extracted from the inference result. Multiple sets of feature vectors are randomly selected from the historical operator output feature vectors within the undisturbed period to construct a multidimensional feature space, and multiple binary trees are randomly generated in the multidimensional feature space. The operator output feature vector is mapped to a multi-dimensional feature space. Multiple binary trees are traversed, and the number of edges traversed from the root node to the leaf node in each binary tree is recorded. The average number of edges in each binary tree is taken as the isolated path length. The length of isolated paths is used as a feature bias.

[0011] Furthermore, the mean number of edges in each binary tree is calculated as the isolated path length, including: counting the total number of binary trees, summing the number of edges traversed by the feature vector output by the accumulator in each binary tree from the root node to the leaf node, obtaining the total number of edges, dividing the total number of edges by the total number of trees to obtain the mean number of edges, and using the mean number of edges as the isolated path length.

[0012] Furthermore, the degree of decrease in the credibility of the inference result is determined based on feature bias, including: Obtain the correspondence group of historical feature deviations and historical credibility decline for multiple inferences within a undisturbed period; calculate the similarity measure between the feature deviation and each historical feature deviation in the correspondence group, and select the historical credibility decline corresponding to the top few historical feature deviations with the highest similarity measure as the reference credibility decline set; sort each historical credibility decline in the reference credibility decline set, and take the median value of the sorting result as the credibility decline of the inference result.

[0013] Furthermore, the allocation weight of computing resources for embodied intelligent devices in cloud-edge-device collaborative tasks is adjusted according to the degree of credibility decline, including: Obtain the current computing resource allocation weights of other embodied intelligent devices on the same task link in a cloud-edge-device collaborative task; The degree of credibility decline is compared with the preset untrust threshold. When the degree of credibility decline is lower than the untrust threshold, the current computing power resource allocation weight of the embodied intelligent device is reduced and adjusted, and the amount of weight reduction is distributed according to the proportion of the current computing power resource allocation weight of other embodied intelligent devices. When the level of trustworthiness declines to or exceeds the untrustworthy threshold, the current computing resource allocation weight of the embodied intelligent device will be adjusted to zero, and the embodied intelligent device will be removed from the set of collaborative decision-making nodes for cloud-edge-device collaborative tasks.

[0014] Furthermore, the weight reduction amount due to the adjustment is distributed according to the proportion of the current computing power resource allocation weight of other embodied intelligent devices. This includes: calculating the difference between the current computing power resource allocation weight of the embodied intelligent device before the adjustment and the computing power resource allocation weight after the adjustment to obtain the weight reduction amount; calculating the sum of the current computing power resource allocation weights of other embodied intelligent devices to obtain the total weight of other devices; for each other embodied intelligent device, dividing its current computing power resource allocation weight by the total weight of other devices to obtain the distribution ratio; multiplying the weight reduction amount by the distribution ratio to obtain the weight increase amount of the other embodied intelligent device; and adding the weight increase amount to the current computing power resource allocation weight of the other embodied intelligent device to obtain the distributed computing power resource allocation weight.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This method correlates the operating parameters of the power components of the embodied intelligent device with the inference results of the computing board in a time sequence. It extracts the frequency distribution of the drive current jump through spectrum conversion and determines the spectral overlap with the operating frequency of the computing unit. This identifies electrical disturbance events caused by the coupling of the joint motor current jump through the shared power supply circuit. Furthermore, it obtains the voltage spectrum characteristics of the electrical disturbance event and determines the degree of resonance overlap between the energy concentration frequency band and the operating frequency of the computing unit. This quantifies the coupling strength of the electrical disturbance to the operating clock of the computing unit. When the degree of resonance overlap exceeds a preset threshold, the operator output feature vector is mapped to the feature space constructed from historical data of the disturbance-free period and the isolated path length is calculated. The isolated path length is used as the feature deviation to reflect the actual deviation of the inference result after the disturbance. This establishes a cross-domain correlation judgment mechanism from the physical action of the device to the calculation logic result, enabling the control platform to detect the risk of transient deviation in the inference result before the normal reported indicators of the computing board show any abnormalities.

[0016] 2. After determining the degree of credibility decline of the inference result based on feature deviation, the computing resource allocation weight of the embodied intelligent device in the cloud-edge-device collaborative task is adjusted according to the degree of credibility decline. When the degree of credibility decline is low, the computing resource allocation weight of the problematic device is reduced proportionally and the released weight is distributed to other devices on the same task link to maintain the total computing power supply on the link. When the degree of credibility decline reaches the untrustworthy threshold, the device is directly removed from the collaborative decision node set to prevent untrustworthy inference results from entering the collaborative judgment link. This breaks away from the dependence on superficial indicators such as online status and chip utilization, so that the collaborative scheduling decision is based on the ability to perceive the inference quality. Thus, even when physical disturbances of embodied devices are unavoidable, the reliability of the collaborative judgment basis is still maintained. Attached Figure Description

[0017] Figure 1 This is a flowchart of the cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to the present invention; Figure 2 This is a hardware connection diagram of the cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example: Figure 1 This invention presents a cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices, comprising: S1: Obtain the operating parameters of the power components of the embodied intelligent device during the execution of the task, as well as the inference results output by the computing board mounted on the embodied intelligent device; S2: Perform spectrum conversion on the operating parameters of the power components, extract the frequency distribution of the drive current jump, and determine that there is an electrical disturbance event when the frequency distribution of the drive current jump overlaps with the operating frequency of the computing unit of the computing board. S3: Obtain the voltage spectrum characteristics of electrical disturbance events and determine the degree of resonance overlap between the energy concentration frequency band and the operating frequency of the computing unit in the voltage spectrum characteristics; S4: When the degree of resonance overlap does not exceed the preset degree threshold, the operator output feature vector corresponding to the operation unit in the inference result is mapped to the feature space constructed by the historical operator output feature vectors in the undisturbed period, and the isolated path length of the operator output feature vector in the feature space is calculated, and the isolated path length is used as the feature deviation. S5: Determine the degree of decrease in the credibility of the inference result based on feature bias; S6: Adjust the weight of computing resources allocated to embodied intelligent devices in cloud-edge-device collaborative tasks based on the degree of credibility decline.

[0020] Figure 2 A hardware connection diagram of the cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to the present invention is provided. The embodied intelligent devices are connected to the cloud management platform via a communication link. The cloud management platform breaks down the cloud-edge-device collaborative task into multiple sub-tasks and distributes them to each embodied intelligent device, and manages the addition and removal of embodied intelligent devices participating in collaborative decision-making within the collaborative decision-making node set. The embodied intelligent device is internally divided into five functional domains: main control and clock domain, power domain, computing domain, sensing domain, and power supply path.

[0021] The master control and clock domain includes the master controller and the system clock. The master controller is the central processing unit built into the embodied intelligent device. The system clock is provided by the master controller's real-time clock. The master controller broadcasts the current system clock count to the power domain and computing domain via a clock synchronization signal, ensuring that the timestamps of the drive current sequence, the motor speed sequence, and the inference result output stream belong to the same clock domain. The master controller and computing domain are connected via a serial bus, which is used to transmit inference results and control commands. The master controller and power domain are connected via a controller area network (CAN) bus, which is used to transmit power domain feedback data and master controller drive commands.

[0022] The power domain includes a power drive and a joint motor. The power drive is a joint motor driver, connected to the joint motor via a three-phase power line and sensor signal lines. The power drive outputs three-phase drive current to the joint motor, driving its rotation. Simultaneously, it acquires the bus current through a current sensor built into the joint motor and the rotor speed through a position sensor built into the joint motor. The instantaneous sampled sequence of the bus current is used as the drive current sequence, and the instantaneous sampled sequence of the rotor speed is used as the motor speed sequence. The drive current sequence and motor speed sequence are acquired from the power drive's feedback interface at a fixed sampling frequency, and the system clock timestamps each acquired data point.

[0023] The perception domain includes a visible light camera, which is installed at the front end of the embodied intelligent device and connected to the computing board in the computing domain via a wired interface. The wired interface transmits the raw pixel data of the inspection image.

[0024] The computing domain includes a computing board, a computation unit, a neural network inference framework, a power input terminal, and a voltage sampling circuit. The computing board is an embedded board integrating a neural network inference acceleration unit. The neural network inference framework is deployed on the computing board, performing target detection inference on the inspection images transmitted by the visible light camera. It outputs the inference result stream through an application programming interface (API). The inference result output stream contains the output tensors of each inference step arranged in chronological order, and the system clock timestamps the output time of each inference step. The computation unit is the hardware unit within the neural network inference acceleration unit that performs convolution or matrix multiplication operations. The computation unit operates at a fixed frequency, which is read from the computing board's technical specifications or obtained by reading the register value of the clock management chip on the computing board. The voltage sampling circuit is connected in parallel to the power input terminal of the computing board. The voltage sampling circuit is a combination of a resistor divider network and an analog-to-digital converter (ADC). The ADC samples the power input voltage at a fixed sampling frequency, which is set to an integer multiple of the computation unit's operating frequency to ensure that voltage fluctuations near the computation unit's operating frequency are captured. The sampled signal is the power supply voltage waveform.

[0025] The power supply path includes a shared power supply circuit, a power bus, and a filter capacitor. One end of the power bus is connected to the battery inside the embodied intelligent device, and the other end is connected to the power supply terminal of the joint motor driver and the power input terminal of the computing board, respectively, realizing unified power supply from the battery to the joint motor driver and the computing board. The filter capacitor is connected in parallel between the positive and negative terminals of the power bus to filter out low-frequency ripple on the power bus. When the joint motor is running under startup, commutation, or sudden load conditions, a steep current jump occurs in the drive current of the joint motor. The high-frequency component in the current jump can reach the vicinity of the operating frequency of the computing unit. The high-frequency component is coupled to the power bus through the parasitic capacitance of the switching transistor inside the joint motor driver. Due to the large equivalent series resistance of the filter capacitor to the high-frequency component, the high-frequency component cannot be completely bypassed by the filter capacitor. The high-frequency component is conducted along the power bus to the power input terminal of the computing board, superimposing an electrical disturbance on the power supply voltage of the computing unit. The electrical disturbance resonates and couples with the operating clock of the computing unit, causing transient deviations in the logic circuit inside the computing unit, thereby affecting the correctness of the inference results.

[0026] In a specific implementation of S1, the embodied intelligent device is a robotic dog equipped with a computing board, which performs substation equipment inspection tasks. The robotic dog's power driver is a joint motor driver, which is a DC brushless motor driver. When driving the robotic dog, the joint motor driver outputs a drive current sequence and a motor speed sequence in real time through its built-in current sensor and position sensor. The drive current sequence and motor speed sequence are collected from the feedback interface of the joint motor driver at a fixed sampling frequency. The fixed sampling frequency is determined based on the highest current change frequency of the joint motor driver and is set to an integer multiple of the pulse width modulation carrier frequency of the joint motor driver to avoid sampling aliasing. The collected drive current sequence and motor speed sequence are used as operating parameters of the power components. The drive current sequence is a sequence of instantaneous sampled values ​​of the bus current of the joint motor driver, and the motor speed sequence is a sequence of instantaneous sampled values ​​of the rotor speed of the joint motor.

[0027] The computing board is an embedded board integrating a neural network inference acceleration unit, which is either a graphics processing unit (GPU) or a neural network processor. An inference framework runs on the computing board. This inference framework is a software platform used to load and execute trained neural network models for inference computation. The inference framework performs target detection inference on inspection images captured by a visible light camera mounted on the robotic dog. The visible light camera connects to the computing board via a wired interface, transmitting inspection images to the computing board frame by frame. The inference framework provides an application programming interface (API) from which the inference result output stream is read in real time. The inference result output stream contains output tensors for each inference iteration arranged chronologically. Each output tensor is a multidimensional array, and its dimensions include the bounding box coordinates, target category index, and confidence score. The output tensor of each inference iteration in the inference result output stream is used as the inference result at the corresponding time point.

[0028] The acquisition times of the drive current sequence and motor speed sequence, and the output times of the inference result output stream, are timestamped by the robot dog's system clock. The system clock is provided by the real-time clock of the robot dog's main controller, which is the robot dog's built-in central processing unit. The main controller broadcasts a clock synchronization signal to the computing board and the joint motor driver. The clock synchronization signal is a periodically sent data message containing the current system clock count value. After receiving the clock synchronization signal, the computing board and the joint motor driver align their respective local clock counters with the system clock count value in the data message. The computing board is connected to the robot dog's main controller via a serial bus, and the joint motor driver is connected to the robot dog's main controller via a controller area network bus. Using the output time of each inference in the inference result output stream as the base time, the sampling point with the smallest time interval from the base time is found in the drive current sequence. The value of the sampling point with the smallest time interval from the base time is used as the drive current value corresponding to the base time. Similarly, the sampling point with the smallest time interval from the base time is found in the motor speed sequence. The value of the sampling point with the smallest time interval from the base time is used as the motor speed value corresponding to the base time. The drive current value, motor speed value, and inference result corresponding to the base time are bound together as a single associated data record. This binding operation is repeated for each inference in the inference result output stream to obtain the associated data of the power component operating parameters with time tags and the inference results.

[0029] The drive current values ​​in the associated data records reflect the instantaneous load state of the mechanical dog's joint motors at the corresponding inference time, the motor speed values ​​reflect the instantaneous movement speed of the joint motors at the corresponding inference time, and the inference result reflects the target detection output of the computing board to the inspection image at the corresponding inference time. When the joint motors are in start-up, commutation, or sudden load conditions, current jumps occur in the drive current sequence. The frequency components of the current jumps are conducted to the power input terminal of the computing board through the mechanical dog's shared power supply circuit, forming electrical disturbances. The shared power supply circuit is a power bus that supplies power to both the joint motor driver and the computing board from the mechanical dog's internal battery. A filter capacitor is connected in parallel on the power bus, but the high-frequency components in the current jumps can still be conducted to the computing board through the power bus. The associated data records provide a data foundation for subsequent analysis of the timing correspondence between electrical disturbances and inference result deviations.

[0030] In a specific implementation of S2, a Fast Fourier Transform (FFT) is performed on the driving current sequence to obtain the driving current spectrum. The FFT is an algorithm that converts a time-domain signal into a frequency-domain signal. After performing the FFT on the driving current sequence, a driving current spectrum is obtained with frequency as the abscissa and amplitude as the ordinate. Each frequency point in the driving current spectrum corresponds to an amplitude, with the amplitude unit being amperes and the frequency unit being hertz.

[0031] Frequency components with amplitudes exceeding a preset fluctuation threshold are extracted from the drive current spectrum. These extracted frequency components are used as the drive current jump frequency distribution. The preset fluctuation threshold is calculated as: T = μ + 3 × σ; where T represents the preset fluctuation threshold, μ represents the mean amplitude of each frequency component in the undisturbed drive current spectrum, and σ represents the standard deviation of each frequency component amplitude in the undisturbed drive current spectrum. The undisturbed drive current spectrum is obtained by acquiring the undisturbed drive current sequence during the period when the mechanical dog is in a stable walking state and no electrical disturbance event is determined to exist. The undisturbed drive current spectrum is obtained by performing a Fast Fourier Transform on the undisturbed drive current sequence. Frequency components with amplitudes exceeding the preset fluctuation threshold are extracted from the drive current spectrum. These extracted frequency components constitute the drive current jump frequency distribution, which is the set of frequency components corresponding to current jumps caused by the joint motor under starting, commutation, or sudden load conditions in the drive current sequence.

[0032] Obtain the operating frequency of the computing unit on the computing board. The operating frequency of the computing unit is the main clock frequency of the neural network inference acceleration unit inside the computing board. The operating frequency can be obtained by reading the nominal value of the computing unit's operating frequency from the computing board's technical specifications document, or by reading the register value of the clock management chip on the computing board. The operating frequency of the computing unit is a fixed value, for example, 800 MHz.

[0033] Each frequency component in the drive current switching frequency distribution is compared one by one with the operating frequency of the computing unit. The preset tolerance range is set to 1% of the operating frequency of the computing unit. The comparison process is as follows: for each frequency component in the drive current switching frequency distribution, the absolute value of the frequency difference between the frequency value of the frequency component and the operating frequency of the computing unit is calculated. When the absolute value of the frequency difference is less than the preset tolerance range, it is determined that the frequency deviation between the frequency component and the operating frequency of the computing unit falls within the preset tolerance range. When at least one frequency component's frequency deviation from the operating frequency of the computing unit falls within the preset tolerance range, it is determined that there is spectral overlap between the drive current switching frequency distribution and the operating frequency of the computing unit of the computing board, thus confirming the existence of an electrical disturbance event.

[0034] Spectral overlap indicates that the drive current switching frequency distribution contains a frequency component close to the operating frequency of the computing unit. This frequency component, after being conducted through the shared power supply circuit of the robot to the power input terminal of the computing board, resonates and couples with the operating clock of the computing unit, creating an electrical disturbance to the internal logic circuit of the computing unit, causing transient deviations during inference calculations. When the frequency deviations of each frequency component in the drive current switching frequency distribution from the operating frequency of the computing unit exceed the preset tolerance range, it is determined that there is no spectral overlap between the drive current switching frequency distribution and the operating frequency of the computing unit of the computing board, thus confirming that there is no electrical disturbance event.

[0035] In a specific implementation of S3, during the time period when an electrical disturbance event is determined to exist, the power supply voltage waveform of the computing board is collected. The power supply voltage waveform is a signal of the voltage at the power supply input terminal of the computing board changing over time. The method for collecting the power supply voltage waveform is as follows: a voltage sampling circuit is connected in parallel at the power supply input terminal of the computing board. The voltage sampling circuit is a combination of a resistor voltage divider network and an analog-to-digital converter. The analog-to-digital converter samples the voltage at the power supply input terminal at a fixed sampling frequency. The fixed sampling frequency is set according to the operating frequency of the computing unit and is set to an integer multiple of the operating frequency of the computing unit. For example, when the operating frequency of the computing unit is 800 MHz, the fixed sampling frequency is set to 8 GHz to ensure that voltage fluctuations near the operating frequency of the computing unit are captured. A fast Fourier transform is performed on the collected power supply voltage waveform to obtain the voltage spectrum, which is used as the voltage spectrum feature of the electrical disturbance event. The voltage spectrum is represented in the frequency domain with frequency as the abscissa and voltage amplitude as the ordinate. Each frequency point in the voltage spectrum corresponds to a voltage amplitude, with the voltage amplitude unit being volts and the frequency unit being hertz.

[0036] The frequency corresponding to the peak power spectral density in the voltage spectrum is extracted as the energy concentration band. The power spectral density is the square of the voltage amplitude of each frequency component in the voltage spectrum divided by the frequency resolution. The frequency resolution is the frequency interval between adjacent frequency points after the Fast Fourier Transform, which is determined by the fixed sampling frequency and the number of points in the Fast Fourier Transform.

[0037] The power spectral density is calculated as: P(f) = (|V(f)| × |V(f)|) / Δf; where P(f) represents the power spectral density at frequency f, V(f) represents the voltage amplitude at frequency f, and Δf represents the frequency resolution. The peak value of the power spectral density is the maximum value point on the power spectral density curve, and the frequency corresponding to the peak value is the frequency position where energy is most concentrated in the voltage spectrum. The maximum value point is found in the calculated power spectral density curve, and the frequency corresponding to the maximum value point is taken as the energy concentration band.

[0038] The absolute value of the frequency difference between the energy concentration band and the operating frequency of the computing unit is calculated, and this absolute value is used as the degree of resonance overlap. The absolute value of the frequency difference is calculated as: D = |fpeak - funit|; where D represents the degree of resonance overlap, fpeak represents the energy concentration band, and funit represents the operating frequency of the computing unit. The unit of resonance overlap is Hertz. The smaller the degree of resonance overlap, the closer the energy concentration band and the operating frequency of the computing unit are, the stronger the resonant coupling between voltage fluctuations from electrical disturbances and the operating clock of the computing unit, and the more severe the interference to the computing unit from electrical disturbances. When the degree of resonance overlap is zero, the energy concentration band and the operating frequency of the computing unit completely overlap. At this time, the power supply voltage fluctuation of the computing board is directly superimposed on the operating clock of the computing unit, the timing margin of the internal logic circuit of the computing unit is compressed to the maximum extent, and the probability of transient deviations in inference calculation is the highest. The degree of resonance overlap is used as the basis for determining whether further evaluation of the deviation of the inference result is needed.

[0039] In a specific implementation of S4, when the degree of resonance overlap does not exceed a preset threshold, the operator output feature vector corresponding to the computing unit is extracted from the inference result. The preset threshold is set as follows: during the period when the robot dog is in a stable walking state and no electrical disturbance event is determined to exist, the degree of resonance overlap is calculated multiple times, and the maximum value of the degree of resonance overlap during the disturbance-free period is obtained. The preset threshold is then set to this maximum value. The fact that the degree of resonance overlap does not exceed the preset threshold indicates that the resonant coupling strength between the voltage fluctuation of the electrical disturbance event and the operating clock of the computing unit has exceeded the maximum fluctuation level under normal operating conditions, requiring further evaluation of the reliability of the inference result.

[0040] The operator output feature vector is the output data vector of the operator layer corresponding to the computation unit in the inference result. The computation unit refers to the hardware unit in the neural network inference acceleration unit inside the computing board that performs convolution or matrix multiplication operations. The operator layer corresponding to the computation unit is the convolutional layer or fully connected layer allocated to the computation unit in the neural network model. The method for extracting the operator output feature vector corresponding to the computation unit from the inference result is as follows: when the inference framework performs inference calculation, after the computation unit has completed the calculation of the corresponding operator layer, the output feature tensor of the corresponding operator layer is read from the internal cache of the inference framework, and the output feature tensor is expanded into a one-dimensional vector to obtain the operator output feature vector.

[0041] Multiple sets of feature vectors are randomly selected from the historical operator output feature vectors within the undisturbed period to construct a multi-dimensional feature space. The historical operator output feature vectors within the undisturbed period are obtained by performing multiple inferences and extracting them in the same way as extracting operator output feature vectors during periods when the robot dog is in a stable walking state and no electrical disturbance events are determined. The total number of historical operator output feature vectors within the undisturbed period is determined as follows: for historical correlation data generated when the robot dog performs inspection tasks in a stable walking state, the inference results corresponding to periods when no electrical disturbance events are determined are selected. Operator output feature vectors are extracted from the inference results until the number of extracted operator output feature vectors reaches a preset sample size limit. The preset sample size limit is set according to the dimension of the operator output feature vectors; for example, if the dimension of the operator output feature vectors is 256, the preset sample size limit is set to 2560. When randomly selecting multiple sets of feature vectors from the historical operator output feature vectors within the undisturbed period, the ratio of the number of randomly selected feature vectors each time to the total number of historical operator output feature vectors within the undisturbed period is set to 80%. The multidimensional feature space is a vector space composed of multiple selected feature vectors. The dimension of the multidimensional feature space is consistent with the dimension of the feature vectors output by the operator. The method for randomly generating multiple binary trees in the multidimensional feature space is as follows: A dimension is randomly selected as the partition dimension in the multidimensional feature space. Within the range of values ​​for the partition dimension, a value is randomly selected as the partition value. Feature vectors in the current node's feature vector set whose value on the partition dimension is less than the partition value are assigned to the left child node. Feature vectors whose value on the partition dimension is greater than or equal to the partition value are assigned to the right child node. This process of randomly selecting the partition dimension and partition value and partitioning child nodes is repeated until the number of feature vectors in the child node is 1 or the preset binary tree depth is reached. Each time partitioning stops, a binary tree is generated. This binary tree generation process is repeated until the number of binary trees reaches the preset number. The preset number of binary trees is set based on the total number of historical operator output feature vectors during the undisturbed period. The preset number of binary trees is proportional to the square root of the total number of historical operator output feature vectors during the undisturbed period.

[0042] The operator output feature vector is mapped to a multi-dimensional feature space. Multiple binary trees are traversed, and the number of edges traversed from the root node to a leaf node in each binary tree is recorded. Mapping the operator output feature vector to a multi-dimensional feature space means treating the operator output feature vector as a point in the multi-dimensional feature space; the projection value of this point in each dimension is the component value of the operator output feature vector in each dimension. Starting from the root node, based on the comparison result between the component value of the operator output feature vector in the partition dimension corresponding to the current node and the partition value, the node enters either the left or right child node. Each entry into a child node is counted as one edge, until a leaf node is reached. The number of edges traversed from the root node to the leaf node is counted. This traversal from the root node to the leaf node is repeated for each of the multiple binary trees, and the number of edges traversed from the root node to the leaf node in each binary tree is recorded.

[0043] The isolated path length is calculated by taking the average number of edges in each binary tree. This process includes: counting the total number of binary trees; summing the number of edges traversed by the operator's output feature vector from the root node to a leaf node in each binary tree; dividing the sum of edges by the total number of binary trees to obtain the average number of edges; and using this average number of edges as the isolated path length. The formula for calculating the isolated path length is: L = (h1 + h2 + ... + hn) / N; where L represents the isolated path length, h1 represents the number of edges traversed by the operator's output feature vector from the root node to a leaf node in the first binary tree, h2 represents the number of edges traversed by the operator's output feature vector from the root node to a leaf node in the second binary tree, hn represents the number of edges traversed by the operator's output feature vector from the root node to a leaf node in the Nth binary tree, and N represents the total number of binary trees.

[0044] The length of isolated paths is used as a feature bias. Feature bias reflects the degree of anomaly of the operator's output feature vector within the multidimensional feature space constructed from historical operator output feature vectors during undisturbed periods. The smaller the isolated path length, the easier it is for the operator's output feature vector to be isolated in the multidimensional feature space, resulting in a greater difference in distribution from historical operator output feature vectors during undisturbed periods, and a higher degree of bias in the inference result due to electrical disturbances. Feature bias is used as the basis for determining the degree of decrease in the reliability of the inference result in subsequent steps.

[0045] In a specific implementation of S5, a correspondence set of historical feature deviations and historical reliability degradation levels corresponding to multiple inferences during a disturbance-free period is obtained. This correspondence set represents a one-to-one correspondence between multiple historical feature deviations and multiple historical reliability degradation levels obtained by simulating different degrees of electrical disturbances during a period when the robot dog is in a stable walking state and no electrical disturbance event is determined to exist. The correspondence set is constructed as follows: During the disturbance-free period, a simulated voltage ripple with a known amplitude is injected into the power input terminal of the computing board. The frequency of the simulated voltage ripple is set to be the same as the operating frequency of the computing unit. The amplitude of the simulated voltage ripple increases incrementally from zero, with each increment set to 1% of the computing board's power supply voltage. Multiple inferences are performed at each simulated voltage ripple amplitude level. For each inference, the historical feature deviation is calculated in the same way as in step S4. The percentage decrease in the average detection accuracy of multiple inferences at each simulated voltage ripple amplitude level relative to the baseline detection accuracy without simulated voltage ripple is taken as the historical reliability degradation level. The historical reliability degradation level is a dimensionless value ranging from 0 to 1. 0 indicates that the inference result is not affected and the detection accuracy does not decrease, while 1 indicates that the inference result is completely unreliable and the detection accuracy decreases to zero. A correspondence record is created by mapping the historical characteristic deviation calculated for each level of simulated voltage ripple amplitude to the corresponding historical reliability degradation level. All correspondence records for all simulated voltage ripple amplitude levels are then grouped into a correspondence set.

[0046] Calculate the similarity measure between the feature deviation and each historical feature deviation in the corresponding relationship group. The similarity measure measures the closeness between the feature deviation and each historical feature deviation in the corresponding relationship group. The similarity measure is expressed as the reciprocal of the absolute value of the difference between the feature deviation and the historical feature deviation. The formula for calculating the similarity measure is: S = 1 / (|B - Bhist| + ε); where S represents the similarity measure, B represents the feature deviation, Bhist represents a historical feature deviation in the corresponding relationship group, and ε represents a small positive value to prevent the denominator from being zero. ε is taken as one-tenth of the smallest non-zero value of the absolute value of the difference between the feature deviation and the historical feature deviation. Calculate the similarity measure for each historical feature deviation in the corresponding relationship group to obtain a set of similarity measure values. Select the historical reliability decline degree corresponding to the top few historical feature deviations with the highest similarity measures as the reference reliability decline degree set. The number of the top few is set as follows: the number of the top few is set to 10% of the total number of historical feature deviations in the corresponding relationship group. When the total number of historical feature deviations in the corresponding relationship group is less than 10, the number of the top few is set to 1.

[0047] The historical confidence decline levels in the reference confidence decline level set are sorted, and the median value of the sorted results is taken as the confidence decline level of the inference result. The sorting is based on the numerical order of the historical confidence decline levels from smallest to largest. The median value is the Kth value of the sorted results. K is calculated as follows: when the number M of historical confidence decline levels in the reference confidence decline level set is odd, K is (M+1) / 2; when the number M is even, K is M / 2. Taking the median value of the sorted results avoids the interference of individual outliers in the reference confidence decline level set on the confidence decline level, making the determined confidence decline level more stably reflect the true extent to which the inference result is affected by electrical disturbances.

[0048] In a specific implementation of S6, the current computing resource allocation weights of other embodied intelligent devices on the same task link in the cloud-edge-device collaborative task are obtained. The cloud-edge-device collaborative task is an inspection task requiring the cooperation of multiple embodied intelligent devices. The task is broken down into multiple sub-tasks by the cloud management platform and distributed to each embodied intelligent device. Each embodied intelligent device executes its assigned sub-task and summarizes the results to the cloud management platform for collaborative decision-making. The same task link refers to a set of embodied intelligent devices in the cloud-edge-device collaborative task where all sub-tasks have data or temporal dependencies. The computing resource allocation weight is a quantitative indicator of allocation priority determined by the cloud management platform based on the credibility of the inference results of each embodied intelligent device when allocating sub-tasks. The initial value of the computing resource allocation weight is set to 1, and its range is between 0 and 1. Other embodied intelligent devices refer to embodied intelligent devices on the same task link other than those corresponding to the degree of credibility decline. The current computing resource allocation weights of other embodied intelligent devices are read from the scheduling records of the cloud management platform.

[0049] The degree of credibility degradation is compared with a preset untrust threshold. The preset untrust threshold is set based on the minimum tolerance limit for the credibility of the inference results in the cloud-edge-device collaborative task. The preset untrust threshold is set by obtaining the historical credibility degradation levels corresponding to multiple inferences within a period when the robotic dog is in a stable walking state and no electrical disturbance event is determined to exist, calculating the maximum historical credibility degradation level, and setting the preset untrust threshold as the sum of the maximum historical credibility degradation level and the preset tolerance margin. The preset tolerance margin is determined based on the security level of the cloud-edge-device collaborative task; the higher the security level, the smaller the preset tolerance margin. For example, the preset tolerance margin is set to 0.05 for high-risk equipment inspection and 0.10 for routine equipment inspection. A credibility degradation level below the preset untrust threshold indicates that although the inference results of the embodied intelligent device are affected by electrical disturbances and deviate, the degree of deviation is still within the tolerable range of the cloud-edge-device collaborative task, and the inference results of the embodied intelligent device can still partially participate in collaborative decision-making. If the degree of credibility decline reaches or exceeds the preset untrust threshold, it means that the deviation of the reasoning results of the embodied intelligent device has exceeded the tolerance range of the cloud-edge-device collaborative task, and the reasoning results of the embodied intelligent device cannot be used for collaborative decision-making.

[0050] When the degree of trust reduction is lower than the preset untrust threshold, the current computing resource allocation weight of the embodied intelligent device is adjusted by reducing it. The adjusted computing resource allocation weight is calculated as: Wnew = Wcurrent × (1 - C); where Wnew represents the adjusted computing resource allocation weight, Wcurrent represents the current computing resource allocation weight of the embodied intelligent device before the adjustment, and C represents the degree of trust reduction. The reduction in weight allocation is distributed according to the proportion of the current computing resource allocation weight of other embodied intelligent devices.

[0051] The weight reduction is calculated as: ΔW = Wcurrent - Wnew; where ΔW represents the weight reduction. The total weight of other devices is calculated as: Wsum = W1 + W2 + ... + Wm; where Wsum represents the total weight of other devices, W1 represents the current computing resource allocation weight of the first other embodied intelligent device, W2 represents the current computing resource allocation weight of the second other embodied intelligent device, Wm represents the current computing resource allocation weight of the mth other embodied intelligent device, and m represents the total number of other embodied intelligent devices. For each other embodied intelligent device, the allocation ratio is calculated as: Ri = Wi / Wsum; where Ri represents the allocation ratio of the i-th other embodied intelligent device, and Wi represents the current computing resource allocation weight of the i-th other embodied intelligent device. The weight increase is calculated as: ΔWi = ΔW × Ri; where ΔWi represents the weight increase of the i-th other embodied intelligent device. The weighted allocation of computing resources after distribution is calculated as: Wi_new = Wi + ΔWi; where Wi_new represents the weighted allocation of computing resources for the i-th other embodied intelligent device after distribution. Weight distribution reduces the computing load borne by embodied intelligent devices with decreased credibility in collaborative tasks, and the released computing resources are taken over by other embodied intelligent devices with normal credibility, ensuring that the total weighted allocation of computing resources in the same task link is not lost.

[0052] When the credibility decline reaches or exceeds a preset untrust threshold, the current computing resource allocation weight of the embodied intelligent device is adjusted to zero, and the embodied intelligent device is removed from the collaborative decision-making node set of the cloud-edge-device collaborative task. The collaborative decision-making node set consists of all embodied intelligent devices participating in collaborative decision-making within the same task chain. The reasoning results of each embodied intelligent device in the collaborative decision-making node set are adopted by the cloud management platform as the basis for collaborative decision-making. Removing an embodied intelligent device from the collaborative decision-making node set means that the cloud management platform will no longer adopt the reasoning results of that embodied intelligent device in subsequent collaborative decisions until the credibility decline of that embodied intelligent device recovers to below the preset untrust threshold, after which it will be re-added to the collaborative decision-making node set.

[0053] To illustrate this more clearly, let's look at the following scenario: A multi-machine collaborative inspection task was carried out in a substation, which was completed by three robotic dogs equipped with computing boards. The first robotic dog was responsible for visible light image acquisition and on-site inference of the transformer area, the second robotic dog was responsible for infrared thermal image acquisition and on-site inference of the transformer area, and the third robotic dog served as a mobile relay node, standing by in the area.

[0054] Three robotic dogs are connected to the cloud management platform via a 5G communication link. The cloud management platform breaks down the inspection task into a sub-task of identifying transformer bushing appearance defects and a sub-task of detecting abnormal transformer bushing temperature. These sub-tasks are then distributed to the first and second robotic dogs, respectively. The reasoning results of the first and second robotic dogs are designated as the basis for mutual verification, and the third robotic dog serves as an auxiliary node for collaborative decision-making.

[0055] During the mission, the first robotic dog needed to cross a section of gravel when approaching the transformer bushing. Its articulated motors repeatedly switched drive modes on the gravel to maintain stability. The drive current sequence showed current jump components with frequencies concentrated around 800 MHz. The computing unit within the first robotic dog's computing board operates at 800 MHz. The frequency distribution of the drive current jumps overlapped with the operating frequency of the computing unit, confirming the presence of an electrical disturbance event.

[0056] Further data was collected on the power supply voltage waveform of the computing board and the degree of resonance overlap was calculated. The degree of resonance overlap did not exceed the preset threshold, and the confidence level of the first mechanical dog's reasoning result was calculated to be 0.35. The confidence level of 0.35 was compared with the preset untrust threshold of 0.30. The confidence level of 0.35 reached the preset untrust threshold. The current computing resource allocation weight of the first mechanical dog was adjusted to zero, and the first mechanical dog was removed from the collaborative decision-making node set.

[0057] In subsequent collaborative decision-making, the cloud-based management platform no longer adopts the reasoning results of the first robotic dog as the basis for judging the appearance defects of the transformer bushings. Instead, it increases the weight of the second robotic dog's computing resources allocation, allowing the second robotic dog's infrared thermal imaging reasoning results to independently complete the anomaly judgment of the transformer bushing status. The platform then schedules the third robotic dog to move to the transformer area to take over the visible light image acquisition and on-site reasoning tasks from the first robotic dog. After crossing the gravel ground, the first robotic dog resumes a stable walking state, and the current jump component in the driving current sequence disappears. The credibility of the recalculated reasoning results decreases to 0.05, below the preset untrustworthy threshold of 0.30. The cloud-based management platform then re-adds the first robotic dog to the collaborative decision-making node set, restoring the verification relationship between the first and second robotic dogs, and the collaborative task continues to execute normally.

[0058] All calculations involved in the embodiments are performed using dimensionless numerical values, and the preset parameters and thresholds in the calculations can be set by those skilled in the art according to actual conditions.

[0059] This technical solution can be flexibly deployed, for example, as embedded software running on device hardware, or installed on personal computers or other smart terminals with user interfaces, thus adapting to various hardware environments and usage requirements.

[0060] The above solutions can be implemented in software, hardware, firmware, or a combination thereof. When implemented in software, they are presented, in whole or in part, as a computer program product, containing one or more computer instructions or programs. When these instructions or programs are loaded and executed on a computer, results are produced corresponding to the processes or functions of the embodiments of this application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one medium to another via wired or wireless means, such as from a website, server, or data center via wired means such as fiber optic cables, twisted-pair cables, or coaxial cables, or wireless means such as infrared or microwaves to another site. A computer-readable storage medium refers to any usable medium that a computer can access or a data storage device such as a server or data center that contains one or more usable media, including magnetic media such as floppy disks, hard disks, and magnetic tapes, optical media such as DVDs, and semiconductor media such as solid-state drives.

[0061] The specific working process of the system, device and module can be found in the method embodiment, and will not be repeated here.

[0062] The disclosed systems, devices, and methods can be implemented in other ways. The device embodiments are for illustrative purposes only, and the module division is only a logical division. In practice, different divisions can be implemented, such as merging or integrating multiple modules or components, or omitting some features. Coupling, direct coupling, or communication connections between the components can be achieved through interfaces, while indirect coupling or communication connections can take electrical, mechanical, or other forms.

[0063] The modules described as separate components may or may not be physically separated. The components shown as modules can be physical hardware or software, and can be deployed centrally or distributed across multiple network nodes. Some or all of the modules can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0064] The functional modules in each embodiment can be integrated into one processing module, or they can exist independently, or two or more modules can be integrated into one.

[0065] If the functionality is implemented as a software module and used as an independent product, it can be stored in a computer-readable storage medium. Under this understanding, the substantial contribution of the technical solution of this application can be embodied in the form of a software product. This computer software product is stored in a storage medium and contains instructions to cause a computer device, such as a personal computer, server, or network device, to execute all or part of the steps of the methods in the embodiments of this application. The storage medium includes any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory, random access memory, magnetic disk, or optical disk.

[0066] The above are merely specific embodiments of this application, and the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should fall within the scope of protection of this application.

Claims

1. A cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices, characterized in that, include: S1: Obtain the operating parameters of the power components of the embodied intelligent device during the execution of the task, as well as the inference results output by the computing board mounted on the embodied intelligent device; S2: Perform spectrum conversion on the operating parameters of the power components, extract the drive current jump frequency distribution, and determine the existence of an electrical disturbance event when the drive current jump frequency distribution overlaps with the operating frequency of the computing unit of the computing board, including: The drive current spectrum is obtained by performing a fast Fourier transform on the drive current sequence. Frequency components with amplitudes exceeding a preset fluctuation threshold are extracted from the drive current spectrum, and the extracted frequency components are used as the drive current jump frequency distribution. The operating frequency of the computing unit of the computing board is obtained, and each frequency component in the frequency distribution of the drive current jump is compared with the operating frequency of the computing unit. When the frequency deviation between the frequency value of a frequency component and the operating frequency of the computing unit falls within the preset tolerance range, it is determined that there is a spectral overlap between the frequency distribution of the drive current jump and the operating frequency of the computing unit, and an electrical disturbance event is identified. S3: Obtain the voltage spectrum characteristics of electrical disturbance events and determine the degree of resonance overlap between the energy concentration frequency band and the operating frequency of the computing unit in the voltage spectrum characteristics; S4: When the degree of resonance overlap does not exceed the preset degree threshold, the operator output feature vector corresponding to the operation unit in the inference result is mapped to the feature space constructed by the historical operator output feature vectors in the undisturbed period, and the isolated path length of the operator output feature vector in the feature space is calculated, and the isolated path length is used as the feature deviation. S5: Determine the degree of decrease in the credibility of the inference result based on feature bias; S6: Adjust the weight of computing resources allocated to embodied intelligent devices in cloud-edge-device collaborative tasks based on the degree of credibility decline.

2. The cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to claim 1, characterized in that, Obtain the operating parameters of the power components of the embodied intelligent device during task execution, as well as the inference results output by the computing board mounted on the embodied intelligent device, including: The drive current sequence and motor speed sequence are collected in real time from the power driver of the embodied intelligent device, and the drive current sequence and motor speed sequence are used as the operating parameters of the power component; The inference result output stream is collected in real time from the inference framework interface of the computing board, and the output tensor of each inference in the inference result output stream is used as the inference result. The drive current sequence, motor speed sequence, and inference results are aligned according to the timestamp of the acquisition time to form correlation data between the power component operating parameters and the inference results, carrying time tags.

3. The cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to claim 2, characterized in that, The process involves generating time-stamped data linking the operating parameters of the power components with the inference results. This includes: using the output time of each inference in the inference result output stream as the reference time, finding the collection point with the smallest time interval from the reference time in the drive current sequence and the motor speed sequence, binding the values ​​of the collection points in the drive current sequence and the motor speed sequence with the inference results corresponding to the reference time into a single data record; and iterating through each inference in the inference result output stream to obtain the time-stamped data linking the operating parameters of the power components with the inference results.

4. The cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to claim 1, characterized in that, Obtain the voltage spectrum characteristics of electrical disturbance events, and determine the degree of resonance overlap between the energy concentration frequency band in the voltage spectrum characteristics and the operating frequency of the computing unit, including: During the time period when an electrical disturbance event is determined to exist, the power supply voltage waveform of the computing board is collected, and a fast Fourier transform is performed on the power supply voltage waveform to obtain the voltage spectrum. The voltage spectrum is used as the voltage spectrum feature of the electrical disturbance event. Extract the frequency corresponding to the peak power spectral density from the voltage spectrum as the energy concentration band; The absolute value of the frequency difference between the energy concentration frequency band and the operating frequency of the computing unit is calculated, and the absolute value of the frequency difference is used as the degree of resonance overlap.

5. The cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to claim 1, characterized in that, When the degree of resonance overlap does not exceed a preset threshold, the operator output feature vector corresponding to the operation unit in the inference result is mapped to the feature space constructed by the historical operator output feature vectors in the undisturbed period. The isolated path length of the operator output feature vector in the feature space is calculated, and the isolated path length is used as the feature deviation, including: When the degree of resonance overlap does not exceed the preset threshold, the operator output feature vector corresponding to the operation unit is extracted from the inference result. Multiple sets of feature vectors are randomly selected from the historical operator output feature vectors within the undisturbed period to construct a multidimensional feature space, and multiple binary trees are randomly generated in the multidimensional feature space. The operator output feature vector is mapped to a multi-dimensional feature space. Multiple binary trees are traversed, and the number of edges traversed from the root node to the leaf node in each binary tree is recorded. The average number of edges in each binary tree is taken as the isolated path length. The length of isolated paths is used as a feature bias.

6. The cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to claim 5, characterized in that, The average number of edges in each binary tree is used as the isolated path length. This includes: counting the total number of binary trees, summing the number of edges traversed from the root node to the leaf node in each binary tree by the feature vector output by the accumulator operator, obtaining the total number of edges, dividing the total number of edges by the total number of trees to obtain the average number of edges, and using the average number of edges as the isolated path length.

7. The cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to claim 1, characterized in that, The degree of credibility reduction in inference results is determined based on feature bias, including: Obtain the correspondence group of historical feature deviations and historical credibility decline for multiple inferences within a undisturbed period; calculate the similarity measure between the feature deviation and each historical feature deviation in the correspondence group, and select the historical credibility decline corresponding to the top few historical feature deviations with the highest similarity measure as the reference credibility decline set; sort each historical credibility decline in the reference credibility decline set, and take the median value of the sorting result as the credibility decline of the inference result.

8. The cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to claim 1, characterized in that, Adjusting the weight of computing resources allocated to embodied intelligent devices in cloud-edge-device collaborative tasks based on the degree of credibility decline includes: Obtain the current computing resource allocation weights of other embodied intelligent devices on the same task link in a cloud-edge-device collaborative task; The degree of credibility decline is compared with the preset untrust threshold. When the degree of credibility decline is lower than the untrust threshold, the current computing power resource allocation weight of the embodied intelligent device is reduced and adjusted, and the amount of weight reduction is distributed according to the proportion of the current computing power resource allocation weight of other embodied intelligent devices. When the level of trustworthiness declines to or exceeds the untrustworthy threshold, the current computing resource allocation weight of the embodied intelligent device will be adjusted to zero, and the embodied intelligent device will be removed from the set of collaborative decision-making nodes for cloud-edge-device collaborative tasks.

9. The cloud-edge-device computing power collaborative management method for multi-source heterogeneous embodied intelligent devices according to claim 8, characterized in that, The reduction in weight allocation due to the reduction adjustment is distributed according to the proportion of the current computing power resource allocation weight of other embodied intelligent devices. This includes: calculating the difference between the current computing power resource allocation weight of the embodied intelligent device before the reduction adjustment and the computing power resource allocation weight after the reduction adjustment to obtain the weight reduction amount; calculating the sum of the current computing power resource allocation weights of other embodied intelligent devices to obtain the total weight of other devices; for each other embodied intelligent device, dividing its current computing power resource allocation weight by the total weight of other devices to obtain the distribution ratio; multiplying the weight reduction amount by the distribution ratio to obtain the weight increase amount of the other embodied intelligent device; and adding the weight increase amount to the current computing power resource allocation weight of the other embodied intelligent device to obtain the distributed computing power resource allocation weight.

Images