Robot health state determination method and apparatus, and storage medium

Abstract

The application discloses a kind of robot health state determination method, device and storage medium.Therein, the method includes: obtaining the motion state data and target vibration data of target robot;Determine the target parameter of target robot based on motion state data;Target vibration data is input into multi-source domain countermove migration model and is processed, obtains the target domain feature encoding of target robot, wherein multi-source domain countermove migration model is initialized based on target parameter, and is obtained using the source domain feature encoding corresponding to a plurality of source domain robots respectively for training, a plurality of source domain robots are respectively in predetermined first state robot, and the source domain feature encoding corresponding to a plurality of source domain robots respectively is obtained based on a plurality of source domain robots respectively corresponding vibration data;Determine the health state of target robot based on target domain feature encoding.The application solves the technical problem that the accuracy of robot health state detection is not ideal in the related art.

Description

Technical Field

[0001] This invention relates to the field of robot system health assessment, and more specifically, to a method, apparatus, and storage medium for determining the health status of a robot. Background Technology

[0002] Collaborative robots, as intricate electromechanical systems with complex motion control, experience performance degradation, health decline, component wear, and increased operational risks during operation, posing significant negative impacts on the safety of manufacturers and users. Related technologies typically require full lifecycle degradation data for model training. However, robot components generally have long rated lifespans, and the equipment operates under normal conditions for extended periods, resulting in limited data on fault states, making full lifecycle data difficult to obtain. Numerous physical factors, including robot motion conditions, component geometric parameters, and environmental noise, lead to variations in data distribution from different sources, complicating the extraction of effective degradation features. Furthermore, existing technologies fail to consider the mechanical impacts of varying robot operating conditions, thus reducing the accuracy of health assessments.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This invention provides a method, apparatus, and storage medium for determining the health status of a robot, thereby at least addressing the technical problem of unsatisfactory accuracy in detecting the health status of robots in related technologies.

[0005] According to one aspect of the present invention, a method for determining the health status of a robot is provided, comprising: acquiring motion state data and target vibration data of a target robot; determining target parameters of the target robot based on the motion state data; inputting the target vibration data into a multi-source domain adversarial transfer model for processing to obtain a target domain feature code of the target robot, wherein the multi-source domain adversarial transfer model is initialized based on the target parameters and trained using source domain feature codes corresponding to multiple source domain robots, the multiple source domain robots being robots in a predetermined first state, and the source domain feature codes corresponding to the multiple source domain robots being obtained based on the vibration data corresponding to the multiple source domain robots; and determining the health status of the target robot based on the target domain feature code.

[0006] Optionally, the target robot includes multiple joints, each of which is equipped with a harmonic reducer. The motion state data is the output frequency of the harmonic reducers for each joint. Determining the target parameters of the target robot based on the motion state data includes: determining the input frequency of the harmonic reducer for the corresponding joint based on the output frequency; obtaining the frequency bands corresponding to each joint based on the input frequencies of the harmonic reducers for each joint and a preset bandwidth; performing a union operation on the frequency bands corresponding to each joint to obtain a target frequency band; and determining the target parameters based on the target frequency band.

[0007] Optionally, determining the target parameters based on the target frequency band includes: inputting the target frequency band into a preset finite impulse response filter for filtering to obtain a finite impulse response; and determining target time-domain filtering coefficients as the target parameters based on the finite impulse response.

[0008] Optionally, the initial model includes a convolutional encoder, a convolutional decoder, and a domain classifier. The method further includes: inputting the vibration data of the plurality of source domain robots into the convolutional encoder to obtain source domain feature codes corresponding to the plurality of source domain robots respectively; inputting the source domain feature codes corresponding to the plurality of source domain robots into the convolutional decoder to obtain reconstructed feature codes corresponding to the plurality of source domain robots respectively; inputting the source domain feature codes corresponding to the plurality of source domain robots into the domain classifier to obtain domain classification codes corresponding to the plurality of source domain robots respectively; determining the total loss of the initial model based on the source domain feature codes, reconstructed feature codes, and domain classification codes corresponding to the plurality of source domain robots respectively; and determining the initial model as the multi-source domain adversarial transfer model if the total loss is less than a predetermined threshold.

[0009] Optionally, before inputting the source domain feature codes corresponding to the plurality of source domain robots into the convolutional decoder to obtain the reconstructed feature codes corresponding to the plurality of source domain robots, the method further includes: setting the number of channels of the first convolutional layer included in the convolutional decoder as the target parameter.

[0010] Optionally, determining the total loss of the initial model based on the source domain feature codes, reconstructed feature codes, and domain classification codes corresponding to the plurality of source domain robots includes: determining a reconstruction loss based on the source domain feature codes and reconstructed feature codes corresponding to the plurality of source domain robots; determining the first-layer convolutional filter coefficient matrix in the initial model and the time-domain coefficient length of a preset finite impulse response filter; determining a filtering loss based on the first-layer convolutional filter coefficient matrix, the predetermined time-series coefficient length, and the predetermined balance coefficient; determining a domain classification loss based on the domain classification codes and vibration data corresponding to the plurality of source domain robots, and the predetermined network coefficients; and determining the total loss based on the reconstruction loss, the filtering loss, and the domain classification loss.

[0011] Optionally, determining the health status of the target robot based on the target domain feature encoding includes: determining a first robot in a predetermined first state and a second robot in a predetermined second state, wherein the health level of the predetermined first state is better than the health level of the predetermined second state; acquiring vibration data of the first robot and vibration data of the second robot; inputting the vibration data of the first robot into the multi-source domain adversarial transfer model for processing to obtain a first feature encoding of the first robot; inputting the vibration data of the second robot into the multi-source domain adversarial transfer model for processing to obtain a second feature encoding of the second robot; performing normalization processing based on the first feature encoding, the second feature encoding, and the target domain feature encoding to obtain a target normalization result of the target domain feature encoding; and determining the health status based on the target normalization result.

[0012] Optionally, acquiring the motion state data and vibration data of the target robot includes: collecting the motion state data and vibration data of the target robot in the target motion state; the target robot has multiple motion states, each corresponding to motion state data and vibration data; determining the health state based on the target normalization result includes: processing each of the multiple motion states as the target motion state to obtain the target normalization result corresponding to each of the multiple motion states; and obtaining the health state based on the number of motions in each of the multiple motion states and the target normalization result corresponding to each of the multiple motion states.

[0013] According to another aspect of the present invention, a robot health status determination device is provided, comprising: an acquisition module for acquiring motion state data and target vibration data of a target robot; a determination module for determining target parameters of the target robot based on the motion state data; a data processing module for inputting the target vibration data into a multi-source domain adversarial transfer model for processing to obtain target domain feature codes of the target robot, wherein the multi-source domain adversarial transfer model is initialized based on the target parameters and trained using source domain feature codes corresponding to multiple source domain robots, the multiple source domain robots being robots in a predetermined first state, and the source domain feature codes corresponding to the multiple source domain robots being obtained based on the vibration data corresponding to the multiple source domain robots; and a health status determination module for determining the health status of the target robot based on the target domain feature codes.

[0014] According to another aspect of the present invention, a non-volatile storage medium is provided, the non-volatile storage medium storing a plurality of instructions adapted for loading by a processor and executing any one of the robot health status determination methods described herein.

[0015] In this embodiment of the invention, motion state data and vibration data of a target robot are acquired; based on the motion state data, target parameters of the target robot are determined; the target vibration data is input into a multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot. The multi-source domain adversarial transfer model is initialized based on the target parameters and trained using source domain feature codes corresponding to multiple source domain robots, each in a predetermined first state. The source domain feature codes corresponding to each of the multiple source domain robots are obtained based on the vibration data corresponding to each of the multiple source domain robots. Based on the target domain feature code, the health state of the target robot is determined. This achieves the goal of combining the robot's motion state data and vibration data to detect its health state, thereby improving the technical accuracy of robot health state identification and solving the technical problem of unsatisfactory robot health state detection accuracy in related technologies. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0017] Figure 1 This is a flowchart of an optional robot health status determination method provided according to an embodiment of the present invention;

[0018] Figure 2 This is a frequency band diagram of an optional robot health status determination method provided according to an embodiment of the present invention;

[0019] Figure 3 This is a schematic diagram of the union of an optional robot health status determination method provided by an embodiment of the present invention;

[0020] Figure 4 This is a schematic diagram of the target time-domain filtering coefficients of an optional robot health state determination method provided by an embodiment of the present invention;

[0021] Figure 5 This is a frequency band flowchart of an optional robot health status determination method provided according to an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram of the algorithm structure of an optional robot health status determination method provided by an embodiment of the present invention;

[0023] Figure 7 This is a schematic diagram of the algorithm structure of an optional robot health status determination method provided by an embodiment of the present invention;

[0024] Figure 8 This is a schematic flowchart of an optional robot health status determination method provided according to an embodiment of the present invention;

[0025] Figure 9 This is a schematic diagram of an optional robot health status determination device provided according to an embodiment of the present invention. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] For ease of description, the following explains some of the nouns or terms used in the embodiments of this application:

[0029] Multi-source Domain Adversarial Transfer Network (MARDN) is a transfer learning method used to address the problem of transferring data from multiple source domains to a target domain. In practical applications, data from different source domains often have different distributions and characteristics, so directly applying source domain data to the target domain may lead to performance degradation. MARDN introduces an adversarial learning framework, enabling the network to adapt to different source domain data distributions and achieve knowledge transfer and fusion between the source and target domains through a shared feature extraction network.

[0030] A harmonic drive is a high-precision speed reduction device commonly used in robot joint transmission systems. It consists of a flexible gear, a rigid gear, and a flexible shaft. The flexible gear, made of a series of elastic materials, has a specific corrugated shape and is called a wave generator. The rigid gear is a rigid disk with a certain number of teeth. During motion, as the drive rotates, the wave generator produces corrugated deformation, generating rolling friction between the rigid and flexible gears. This rolling friction reduces the input speed and features high precision and zero backlash. Harmonic drives offer high precision and high torque transmission capabilities, enabling high-precision position control and repeatability, and can withstand large loads.

[0031] A Finite Impulse Response (FIR) filter is a digital filter whose output takes only a finite number of values ​​for its impulse response and returns to zero within a finite time. The output of an FIR filter depends only on the input signal and the filter coefficients, and not on previous inputs or outputs. A linear-phase filter delays the input signal without distorting its phase, thus obtaining a zero-distortion output signal and better extracting frequency band components.

[0032] Collaborative robots are a new type of industrial robot capable of interacting with humans in shared spaces or working safely near humans. Due to their lightweight and safety features, they have shown promising application prospects in flexible manufacturing, social services, and healthcare. As collaborative robots penetrate various industries, their superior performance significantly improves production or operational efficiency and enhances product quality or service levels. However, as electromechanical systems with intricate mechanical structures and complex motion control, collaborative robots are susceptible to performance degradation, health decline, component wear, and increased operational risks during operation, which can have significant negative impacts on the safety of manufacturers and users. Early detection, real-time analysis, diagnosis, and rapid development of maintenance plans for collaborative robots are of great significance and practical value in improving maintenance levels, reducing failure rates, and ensuring the smooth operation of production lines.

[0033] Research on robot health assessment has been conducted at both the component and system levels. The component level typically focuses on monitoring the health status of individual components; the system level assesses the overall operational status of the robot system by considering system architecture, system functions, and process-related parameters. The signal analysis and modeling methods employed vary depending on the type of signals collected during the health assessment process. Vibration signal-based methods are widely used in robot health monitoring. However, vibration signal data is non-process data and cannot directly display degradation trends; therefore, it needs to be converted into health indices to quantitatively assess the robot's degradation. Related technologies generally use common signal analysis methods such as wavelet transform to process the signals and construct relevant statistics to extract degradation features. Meanwhile, neural network models, due to their powerful nonlinear fitting and feature self-extraction capabilities, are also widely used in health assessment.

[0034] On the one hand, related technologies rely on full lifecycle degradation data for model training. However, robot components generally have long rated lifespans, and the equipment operates in normal conditions for extended periods, resulting in limited data on fault states, making full lifecycle data difficult to obtain. On the other hand, considering the similarity of degradation processes for similar models, historical data from multiple robots in different health states can be collected to compensate for the insufficient full lifecycle data. However, numerous physical factors, including operating conditions, component geometry parameters, and environmental noise, will lead to differences in data distribution among robots from different sources, increasing the difficulty of extracting effective degradation features and thus reducing the accuracy of health assessments. Furthermore, data-driven assessment methods often neglect the reflection of the robot's motion state on its mechanical state, failing to leverage prior mechanical domain knowledge to improve the accuracy of health state identification.

[0035] To address the aforementioned problems, this invention provides a method embodiment for determining the health status of a robot. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0036] Figure 1 This is a flowchart of a robot health status determination method according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0037] Step S102: Obtain motion state data and target vibration data of the target robot;

[0038] It is understandable that the motion state data of the target robot can reflect its mechanical state. Unlike other fixed equipment, the target robot can move, and the motion state data obtained will be different according to different motion states. Combining the target vibration data can help determine a more accurate health status.

[0039] Optionally, the aforementioned target vibration data is a high-frequency vibration signal. This can be acquired by placing a vibration sensor external to the target robot's base, collecting high-frequency vibration signals along the X, Y, and Z axes during robot operation. In robotics, robot motion is divided into three directions: the X-axis, Y-axis, and Z-axis. These axes describe the robot's direction and position in three-dimensional space. When it is necessary to collect high-frequency vibration signals from the target robot, collecting data along the X, Y, and Z axes allows for the acquisition of vibration information in different directions, facilitating vibration analysis, fault detection, and other applications.

[0040] Step S104: Determine the target parameters of the target robot based on the motion state data;

[0041] It is understandable that the target parameters can be determined by processing the collected motion state data of the target robot.

[0042] In one optional embodiment, the target robot includes multiple joints, each of which is equipped with a harmonic reducer. The motion state data is the output frequency of the harmonic reducers for each joint. Based on the motion state data, the target parameters of the target robot are determined, including: determining the input frequency of the harmonic reducer for each joint based on the output frequency; obtaining the frequency bands corresponding to each joint based on the input frequencies of the harmonic reducers for each joint and a preset bandwidth; performing a union operation on the frequency bands corresponding to each joint to obtain the target frequency band; and determining the target parameters based on the target frequency band.

[0043] It is understandable that the target robot itself can move, and it has multiple movable joints. Each of these joints is equipped with a harmonic reducer, which improves the accuracy and repeatability of the motion. The collected motion state data of the target robot can be used as the output frequency of the harmonic reducers for each joint. Based on the output frequency, the input frequency of the corresponding harmonic reducer can be determined. The physical meaning of these two frequencies describes the speed relationship between the input and output shafts of the harmonic reducer during transmission; that is, the reducer converts the high-speed, low-torque speed of the input shaft into the low-speed, high-torque speed of the output shaft. The output frequency determines the movement speed of the robot joints, while the input frequency determines the working state of the reducer and the operating requirements of the motor. Based on the input frequencies of the harmonic reducers for each joint and the preset bandwidth, the frequency bands corresponding to each joint of the target robot can be obtained. By performing a union operation on the frequency bands corresponding to each joint, the target frequency band is obtained. By collecting the input frequencies of the harmonic reducers of each joint of the robot and converting them into frequency bands, the key frequency band information, i.e., the target frequency band, can be obtained. This process is called spectrum analysis. Based on the target frequency band, target parameters are determined. By analyzing the target frequency band, anomalies in the machine can be identified, such as abnormal vibrations and resonance phenomena. The mechanical state can be effectively reflected through the robot's motion state, which is beneficial for model recognition and processing.

[0044] It should be noted that by installing harmonic reducers at multiple joints of the robot, torque output can be improved, accuracy and repeatability can be enhanced, size and weight can be reduced, and system reliability can be increased. These effects enable the robot to perform various tasks better, improving work efficiency and performance.

[0045] Optionally, the aforementioned output frequency is obtained by acquiring the speed signals of the output terminals of each joint harmonic reducer through sensor signals inside the low-frequency control cabinet.

[0046] Optionally, the input frequency is the rotational speed of the input shaft (usually the motor shaft) of the harmonic reducer, that is, the rotational speed supplied by the motor to the harmonic reducer. The output frequency is the rotational speed of the output shaft (usually the robot joint shaft) of the harmonic reducer, that is, the rotational speed output by the harmonic reducer to the robot joint. The output frequency is calculated from the input frequency using the reduction ratio, which determines the proportional relationship between the input and output frequencies. The input frequency can be obtained in the following way:

[0047] f in =i*f out

[0048] Where i is the reduction ratio of the harmonic reducer, f in The input frequency of the harmonic reducer is expressed in Hertz (Hz). out The frequency of the output of the harmonic reducer of each joint of the robot is collected through the control cabinet, and the unit is Hertz.

[0049] Optionally, based on the aforementioned input frequency, the characteristic frequencies of the flexible gear meshing and the rigid gear meshing can be obtained, and can be determined in the following way:

[0050] f G1 =2*f in

[0051]

[0052] Among them, f G1 The characteristic frequency of the flexure meshing is expressed in Hertz (Hz). G2 The characteristic frequency of the gear meshing is given in Hertz.

[0053] For one joint among multiple joints, the frequency band corresponding to that joint is obtained based on the characteristic frequency of the flexible gear meshing, the characteristic frequency of the rigid gear meshing, and a preset bandwidth, using the following method:

[0054]

[0055] Among them, F i For frequency band, f c The center frequency is the fundamental frequency and its harmonics, which are the meshing frequencies of the rigid or flexible gear. Δf is a predetermined bandwidth, preferably Δf = 20 Hz. Using this method, frequency bands corresponding to multiple joints can be obtained, and the frequency bands F at different center frequencies can be calculated. i If there are overlapping frequency bands, they are merged in the form of a union to finally obtain the target frequency band.

[0056] like Figure 2 The frequency curve is illustrated in the figure. Figure 2The vertical line above the arrow in the image represents the center frequency, and the square selection area below the arrow represents the selected frequency band. For example... Figure 3 The diagram illustrates the union processing of frequency bands. Figure 3 The image above the middle arrow shows overlapping frequency bands. After union fusion, the result is as follows: Figure 3 The target frequency band in the image below the middle arrow.

[0057] In one optional embodiment, determining the target parameters based on the target frequency band includes: inputting the target frequency band into a preset finite impulse response filter for filtering to obtain a finite impulse response; and determining the target time-domain filter coefficients as target parameters based on the finite impulse response.

[0058] It is understandable that, in order to obtain the target parameters, the target frequency band is input into a preset finite impulse response filter to obtain the finite impulse response, and the target time-domain filter coefficients are determined for use in subsequent models.

[0059] Optionally, based on the target frequency band, the target parameter, denoted as h, is determined in the following manner. r (n), that is, the non-zero value of the finite impulse response:

[0060]

[0061] Where h(.) represents the ideal impulse response of the filter with infinite length, ω l For the target frequency band F i The initial frequency, ω h For the target frequency band F i The termination frequency, [ω l ,ω h [F] represents the frequency passband, i.e., the entire target frequency band. i N is the length of the filter's time-domain coefficients, w r (n) is a window function, preferably a rectangular window, where n represents the time index.

[0062] The above w r (n) can be represented in the following form:

[0063] The finite impulse response h obtained by the above method r (n), about Symmetry ensures the filter's single-peak and linear properties. This guarantees the filter's linearity in the frequency domain without introducing additional computational effects. It ensures near-zero distortion extraction of components within the critical frequency band, thus better extracting fault-related features. The linear-phase filter delays the input signal without distorting its phase, resulting in a zero-distortion output signal and enabling better extraction of component elements within the frequency band. For example... Figure 4As shown, for multiple joints, their target frequency bands can be determined separately, and the corresponding target time-domain filter coefficients can be obtained. Figure 4 The processing results of the filter are represented by #1 to #15, where the specific values ​​are for illustrative purposes only.

[0064] The above processing flow is as follows Figure 5 As shown, based on the output rotation speeds of the harmonic reducers set at each joint of the target robot, the output rotation speeds are processed to obtain the rigid wheel meshing characteristic frequencies and flexible wheel meshing characteristic frequencies of the harmonic reducers, generating a key frequency band, i.e., the target frequency band. Based on the target frequency band obtained above, a finite impulse response filter (FIR filter) is used for filtering to obtain the target time-domain filter coefficients. Ensuring the linearity of the filter is crucial for subsequent model feature extraction, enabling more accurate extraction of components within the target frequency band, better identification of relevant robot fault features by the model, and thus improving the accuracy of health status.

[0065] Step S106: Input the target vibration data into the multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot. The multi-source domain adversarial transfer model is initialized based on the target parameters and trained using the source domain feature codes corresponding to multiple source domain robots. The multiple source domain robots are robots in a predetermined first state. The source domain feature codes corresponding to the multiple source domain robots are obtained based on the vibration data corresponding to the multiple source domain robots.

[0066] It is understandable that multi-source domain adversarial transfer models can handle data from different data sources, meaning that the feature codes corresponding to multiple source domain robots represent multiple different source domains, and these data sources may have different distributions and characteristics. By inputting feature data from different source domains into the multi-source domain adversarial transfer model for training, cross-source domain knowledge transfer and fusion can be achieved. The multi-source domain adversarial transfer model learns the shared feature representations of the source and target domains, extracting features useful for both domains, thereby achieving cross-domain transfer learning. Furthermore, initializing the multi-source domain adversarial transfer model with target parameters, and initializing the convolutional layer channels of the multi-source domain adversarial transfer model, can lead to faster convergence speed, reduced overfitting risk, enhanced model robustness, and improved transfer learning performance. This is beneficial for extracting the target domain feature codes of the target robot based on the input target vibration data.

[0067] Optionally, the robot in the first predetermined state is the normal state.

[0068] In an optional embodiment, before inputting the source domain feature codes corresponding to the multiple source domain robots into the convolutional decoder to obtain the reconstructed feature codes corresponding to the multiple source domain robots, the method further includes setting the number of channels of the first convolutional layer included in the convolutional decoder as a target parameter.

[0069] Understandably, it's necessary to initialize the model parameters by setting the number of channels in the first convolutional layer of the convolutional decoder in the multi-source domain adversarial transfer model as the target parameter. The target parameter reflects the target robot's operational state and mechanical characteristics, providing prior knowledge to help the network better learn and adapt to the features of the target domain.

[0070] In one optional embodiment, the initial model includes a convolutional encoder, a convolutional decoder, and a domain classifier. The method further includes: inputting vibration data from multiple source domain robots into the convolutional encoder to obtain source domain feature codes corresponding to the multiple source domain robots; inputting the source domain feature codes corresponding to the multiple source domain robots into the convolutional decoder to obtain reconstructed feature codes corresponding to the multiple source domain robots; inputting the source domain feature codes corresponding to the multiple source domain robots into the domain classifier to obtain domain classification codes corresponding to the multiple source domain robots; determining the total loss of the initial model based on the source domain feature codes, reconstructed feature codes, and domain classification codes corresponding to the multiple source domain robots; and determining the initial model as a multi-source domain adversarial transfer model if the total loss is less than a predetermined threshold.

[0071] It can be understood that the adversarial learning framework in the multi-source domain adversarial transfer model can be viewed as a competition between a convolutional encoder and a domain classifier, making the output result approximate the true result. Vibration data from multiple source domain robots are input into the convolutional encoder to obtain the source domain feature codes corresponding to each robot. These source domain feature codes are then input into the convolutional decoder to obtain the reconstructed feature codes. Finally, these codes are input into the domain classifier to obtain the domain classification codes. Through this process, the source domain feature codes, reconstructed feature codes, and domain classification codes can be determined, which are used to obtain the total loss of the initial model. During the initial model training process, gradient backpropagation and parameter updates are performed on the initial network based on the loss value. Training is considered complete when the total loss of the initial model is less than a predetermined threshold, resulting in the multi-source domain adversarial transfer model.

[0072] Optionally, the vibration data of the multiple source robots are obtained using the same acquisition method as the vibration data of the target robot.

[0073] Optionally, such as Figure 6As shown, since multiple source domain robots are all in the predetermined first state, they can be considered as normal robots. The convolutional encoder is denoted as E. i The convolutional decoder is denoted as D. i Let G be the domain classifier, and let x be the vibration data corresponding to the i-th source domain robot among multiple source domain robots. i The following explanation focuses on the i-th source domain robot, where i can take values ​​from 1 to K, and K is the total number of source domain robots. The vibration data x corresponding to the i-th source domain robot is... i Input convolution encoder E i The source domain feature code corresponding to the i-th source domain robot is denoted as h. i h i =E i (x i Then, the source domain feature encoding h corresponding to the i-th source domain robot is performed. i Input convolutional decoder D i The reconstructed feature code corresponding to the i-th source domain robot is denoted as . Encode the source domain feature h corresponding to the i-th source domain robot. i Inputting a domain classifier G, we obtain the domain classification code corresponding to the i-th source domain robot, denoted as y. i y i =G(h i ).

[0074] In one optional embodiment, the total loss of the initial model is determined based on the source domain feature codes, reconstructed feature codes, and domain classification codes corresponding to the multiple source domain robots, including: determining the reconstruction loss based on the source domain feature codes and reconstructed feature codes corresponding to the multiple source domain robots; determining the first-layer convolutional filter coefficient matrix in the initial model and the time-domain coefficient length of the preset finite impulse response filter; determining the filtering loss based on the first-layer convolutional filter coefficient matrix, the predetermined time-series coefficient length, and the predetermined balance coefficient; determining the domain classification loss based on the domain classification codes and vibration data corresponding to the multiple source domain robots, and the predetermined network coefficients; and determining the total loss based on the reconstruction loss, filtering loss, and domain classification loss.

[0075] It is understandable that the total loss mentioned above includes reconstruction loss, filtering loss, and domain classification loss. The reconstruction loss is obtained based on the source domain feature encodings and reconstructed feature encodings corresponding to multiple source domain robots. The filtering loss is based on the first-layer convolutional filter coefficient matrix, the temporal coefficient length, and the balance coefficients. The domain classification loss is obtained from the domain classification encodings and vibration data corresponding to multiple source domain robots, as well as the network coefficients. Using these loss functions can help the model achieve better performance and generalization ability, enabling it to learn more useful representations and parameters to better complete the health status recognition task.

[0076] Optionally, the reconstruction loss is obtained by the following method and denoted as:

[0077]

[0078] Where, m i Let K be the total number of samples in source domain i (i.e., the sample from the i-th source domain), and K be the total number of source domains (i.e., the total number of robots in multiple source domains). For the k-th vibration data corresponding to the i-th source domain robot, Encode the k-th reconstructed feature corresponding to the i-th source domain robot.

[0079] The filtering loss is obtained as follows:

[0080]

[0081] Where C is the number of channels in the convolutional layer (number of convolutional kernels), N is the length of the time-domain coefficients of the filter, and λ is the balance coefficient. It is the first layer convolutional filter coefficient matrix, where f can take values ​​from 1 to C, and l can take values ​​from 1 to N.

[0082] The domain classification loss is obtained as follows:

[0083]

[0084] Where θ is the network coefficient, x i Let y be the vibration data corresponding to the i-th source domain robot. i The domain classifier G outputs the domain classification code corresponding to the i-th source domain robot, y itrue These are the true domain labels, from which the domain classification encoding actually originates. During model training, the initial model is backpropagated and its parameters are updated based on the total loss described above. During backpropagation, the gradient of the domain classification loss of the domain classifier is automatically inverted before being backpropagated to the convolutional encoder.

[0085] Step S108: Determine the health status of the target robot based on the target domain feature encoding.

[0086] It is understandable that by outputting the target feature encoding of the target robot through the multi-source domain adversarial transfer model, further data processing can be performed to ultimately determine the health status of the target robot.

[0087] In one optional embodiment, determining the health status of a target robot based on target domain feature encoding includes: determining a first robot in a predetermined first state and a second robot in a predetermined second state, wherein the health level of the predetermined first state is better than the health level of the predetermined second state; acquiring vibration data of the first robot and vibration data of the second robot; inputting the vibration data of the first robot into a multi-source domain adversarial transfer model for processing to obtain a first feature encoding of the first robot; inputting the vibration data of the second robot into a multi-source domain adversarial transfer model for processing to obtain a second feature encoding of the second robot; performing normalization processing based on the first feature encoding, the second feature encoding, and the target domain feature encoding to obtain a target normalization result of the target domain feature encoding; and determining the health status based on the target normalization result.

[0088] It is understandable that the maximum mean difference is used to calculate the feature code distribution distance between the target robot and the first and second robots. After normalization, the health status of the target robot is obtained. The first robot in the predetermined first state is considered a normal robot in a healthy state, and the second robot in the predetermined second state is considered a robot that has reached the engineering limit of acceptable degradation. The vibration data of the first and second robots remain unchanged after being determined during the model building phase. During the model's evaluation of the target robot, normalization is performed using the first feature code, the second feature code, and the target domain feature code to obtain the target normalization result, which represents the degree of state deviation of the target robot, thus determining the health status of the target robot.

[0089] Optionally, before performing normalization, it is also necessary to determine the feature distribution distance of the target domain feature encoding, which can be obtained in the following way:

[0090]

[0091] Where p represents the feature encoding of multiple source domain robots, i.e., h i =[h 1 ,h 2 ,…,h K ], q represents the feature code h of the target robot t E x～p [f(h iE represents the distribution of feature encodings of multiple source domain robots. x～q [f(h t [)] represents the distribution that the feature encoding of the target robot follows. MMD[·] represents the maximum mean difference operator. Let f be a mapping function space, and f be a mapping function whose norm in the reproducing Hilbert space should be less than or equal to 1, i.e. Alternatively, the first feature code and the second feature code can be processed using the methods described above.

[0092] Optionally, normalization processing using the first feature encoding, the second feature encoding, and the target domain feature encoding can be represented as follows:

[0093]

[0094] Wherein d(h t ) represents the feature distribution distance of the target robot, s t d represents the target normalization result corresponding to the target robot. min The first feature encoding represents robot data that is considered to be in a healthy and normal state in engineering, d max The second feature encoding represents robot data that is considered to have reached the limit of acceptable degradation in engineering.

[0095] In one optional embodiment, acquiring the motion state data and vibration data of the target robot includes: collecting motion state data and vibration data of the target robot in the target motion state; the target robot has multiple motion states, each corresponding to motion state data and vibration data; determining the health state based on the target normalization result includes: processing each of the multiple motion states as a target motion state to obtain the target normalization result corresponding to each of the multiple motion states; and obtaining the health state based on the number of motions in the multiple motion states and the target normalization result corresponding to each of the multiple motion states.

[0096] It is understandable that the target robot's motion state differs under different working conditions. Data is collected according to different motion states, including motion state data and vibration data. By processing each of these multiple motion states as a separate target motion state, the target normalization results corresponding to each of these states can be obtained. Based on the number of motion states and the target normalization results corresponding to each state, the health status can be determined.

[0097] Optionally, based on the number of movements in various motion states and the target normalization results corresponding to each motion state, the health status can be obtained in the following way:

[0098]

[0099] Where A represents the number of motion states included in the target robot's data acquisition and execution program. HI represents the target normalization result obtained by the target robot (labeled as t) in the a-th motion state. t This represents the final health state of the target robot t.

[0100] It should be noted that after training the multi-source domain adversarial transfer model, the normal-state robot data and the robot data with a near-engineering degradation threshold used for model construction are input into the trained model to obtain d. min With d max During the health status determination process, data is collected from the target robot and input into a pre-trained multi-source domain adversarial transfer model to calculate the distance d(h). t After normalization and averaging, the health status is obtained.

[0101] Through step S102, the motion state data and vibration data of the target robot are acquired; in step S104, the target parameters of the target robot are determined based on the motion state data; in step S106, the target vibration data is input into a multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot. The multi-source domain adversarial transfer model is initialized based on the target parameters and trained using the source domain feature codes corresponding to multiple source domain robots, each in a predetermined first state. The source domain feature codes corresponding to the multiple source domain robots are obtained based on the vibration data corresponding to each of the multiple source domain robots; in step S108, the health status of the target robot is determined based on the target domain feature code. This method combines the robot's motion state data and vibration data to detect its health status, achieving the technical effect of improving the accuracy of robot health status identification, and thus solving the technical problem of unsatisfactory robot health status detection accuracy in related technologies.

[0102] Based on the above embodiments and optional embodiments, the present invention proposes an optional implementation method, such as... Figure 7 As shown, vibration data can be collected from multiple source domain robots. Based on the vibration data corresponding to the multiple source domain robots, the data can be input into an untrained initial model for iteration until the total loss of the initial model is less than a predetermined threshold, thus obtaining a trained multi-source domain adversarial transfer model.

[0103] Multi-source domain adversarial transfer model (MDL) can be applied to determine the health status of a target robot. By collecting the output frequencies of harmonic reducers at multiple joints of the target robot and performing a series of data processing steps, the target frequency band can be determined. Filtering based on the target frequency band yields the target time-domain filter coefficients, which are used to initialize the MDL. The initialized MDL can then process the input vibration data of the target robot and output the target domain feature code. Given multiple motion states of the target robot, the target normalization results corresponding to these states can be determined, thus yielding the final health status of the target robot.

[0104] To demonstrate the effectiveness of the above processing, a health status verification was performed using a pre-selected robot model. Table 1 shows the robot data collection results, including the total runtime of 11 robots of the same model and rough health stage labels obtained based on expert knowledge. Specific machine identifiers and values ​​in Table 1 are for illustrative purposes only.

[0105] Table 1

[0106]

[0107]

[0108] Health status is displayed using indicators, with values ​​ranging from [0,1], where 1 represents a fully healthy state and 0 represents a set fault threshold. Figure 8 This is an image display of Table 1, such as... Figure 8 As shown, the obtained health indicators are basically consistent with the expert labels, and exhibit good trend and monotonicity with the total running time. They can clearly and accurately characterize the current health status of the robot, demonstrating the effectiveness of the health assessment method proposed in this invention.

[0109] The above-described optional implementation methods achieve at least the following effects: They overcome the dependence of robot health assessment on full lifecycle degradation data. By measuring the operational monitoring data of multiple robots of the same model in different health states (i.e., the aforementioned source domain robots), an unsupervised health assessment model framework is constructed, effectively solving the data limitations in practical applications of the algorithm. By constructing a multi-source domain adversarial transfer network based on a convolutional autoencoder, the distribution differences between data from multiple robots are eliminated, fully extracting the common degradation features of robots of the same model to achieve robot health assessment. By initializing the network model parameters with target frequency band filter coefficients, mechanically relevant domain knowledge is embedded and fused to represent the motion state. Furthermore, by designing key frequency band filter losses to ensure the effectiveness of the filters, knowledge-assisted improvement of the model's assessment effect is achieved.

[0110] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0111] This embodiment also provides a robot health status determination device, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the terms "module" and "device" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0112] According to embodiments of the present invention, an apparatus embodiment for implementing a robot health status determination method is also provided. Figure 9 This is a schematic diagram of a robot health status determination device according to an embodiment of the present invention, such as... Figure 9 As shown, the above-mentioned robot health status determination device includes an acquisition module 902, a determination module 904, a data processing module 906, and a health status determination module 908. The device will be described below.

[0113] The acquisition module 902 is used to acquire motion state data and vibration data of the target robot;

[0114] The determination module 904, connected to the acquisition module 902, is used to determine the target parameters of the target robot based on motion state data.

[0115] The data processing module 906, connected to the determination module 904, is used to input the target vibration data into the multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot. The multi-source domain adversarial transfer model is initialized based on the target parameters and trained using the source domain feature codes corresponding to multiple source domain robots. The multiple source domain robots are robots in a predetermined first state, and the source domain feature codes corresponding to the multiple source domain robots are obtained based on the vibration data corresponding to the multiple source domain robots.

[0116] The health status determination module 908, connected to the data processing module 906, is used to determine the health status of the target robot based on the target domain feature encoding.

[0117] In a robot health status determination device provided by this invention, an acquisition module 902 is used to acquire motion state data and target vibration data of the target robot; a determination module 904, connected to the acquisition module 902, is used to determine the target parameters of the target robot based on the motion state data; a data processing module 906, connected to the determination module 904, is used to input the target vibration data into a multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot. The multi-source domain adversarial transfer model is initialized based on the target parameters and trained using the source domain feature codes corresponding to multiple source domain robots, each in a predetermined first state. The source domain feature codes corresponding to the multiple source domain robots are obtained based on the vibration data corresponding to each of the multiple source domain robots; a health status determination module 908, connected to the data processing module 906, is used to determine the health status of the target robot based on the target domain feature code. This achieves the goal of combining the robot's motion state data and vibration data to detect its health status, thereby improving the technical effect of improving the accuracy of robot health status identification and solving the technical problem of unsatisfactory robot health status detection accuracy in related technologies.

[0118] It should be noted that the above modules can be implemented by software or hardware. For example, for the latter, it can be implemented in the following ways: the above modules can be located in the same processor; or the above modules can be located in different processors in any combination.

[0119] It should be noted that the acquisition module 902, determination module 904, data processing module 906, and health status determination module 908 mentioned above correspond to steps S102 to S108 in the embodiments. The instances and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in the above embodiments. It should be noted that the above modules, as part of the device, can run in a computer terminal.

[0120] It should be noted that the optional or preferred implementation methods of this embodiment can be found in the relevant descriptions in the embodiments, and will not be repeated here.

[0121] The aforementioned robot health status determination device may further include a processor and a memory. The acquisition module 902, determination module 904, data processing module 906, health status determination module 908, etc., are all stored as program units in the memory, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0122] The processor contains a core that retrieves the corresponding program unit from memory. One or more cores may be configured. Memory may include non-persistent memory in computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory includes at least one memory chip.

[0123] This invention provides a non-volatile storage medium storing a program that, when executed by a processor, implements a method for determining the health status of a robot.

[0124] This invention provides an electronic device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it performs the following steps: acquiring motion state data and target vibration data of a target robot; determining target parameters of the target robot based on the motion state data; inputting the target vibration data into a multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot. The multi-source domain adversarial transfer model is initialized based on the target parameters and trained using the source domain feature codes corresponding to multiple source domain robots, each in a predetermined first state. The source domain feature codes corresponding to the multiple source domain robots are obtained based on the vibration data corresponding to each of the multiple source domain robots; and determining the health state of the target robot based on the target domain feature code. The device described herein can be a server, PC, etc.

[0125] This invention also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program with the following steps: acquiring motion state data and target vibration data of a target robot; determining target parameters of the target robot based on the motion state data; inputting the target vibration data into a multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot, wherein the multi-source domain adversarial transfer model is initialized based on the target parameters and trained using the source domain feature codes corresponding to multiple source domain robots, the multiple source domain robots being robots in a predetermined first state, and the source domain feature codes corresponding to the multiple source domain robots being obtained based on the vibration data corresponding to the multiple source domain robots; and determining the health state of the target robot based on the target domain feature code.

[0126] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0127] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0128] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0129] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0130] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0131] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0132] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0133] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0134] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0135] The above are merely embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for determining the health status of a robot, characterized in that, include: Acquire motion state data and vibration data of the target robot; Based on the motion state data, the target parameters of the target robot are determined; The target vibration data is input into a multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot. The multi-source domain adversarial transfer model is initialized based on the target parameters and trained using the source domain feature codes corresponding to multiple source domain robots. The multiple source domain robots are robots in a predetermined first state, and the source domain feature codes corresponding to the multiple source domain robots are obtained based on the vibration data corresponding to the multiple source domain robots. The health status of the target robot is determined based on the target domain feature encoding.

2. The method according to claim 1, characterized in that, The target robot includes multiple joints, each of which is equipped with a harmonic reducer. The motion state data is the output frequency of the harmonic reducers in each joint. Determining the target parameters of the target robot based on the motion state data includes: Based on the output frequency, determine the input frequency of the harmonic reducer set at the corresponding joint; Based on the input frequency of the harmonic reducers set at the multiple joints respectively, and the preset bandwidth, the frequency bands corresponding to the multiple joints are obtained respectively. The target frequency band is obtained by performing a union operation on the frequency bands corresponding to the multiple joints respectively. The target parameters are determined based on the target frequency band.

3. The method according to claim 2, characterized in that, Determining the target parameters based on the target frequency band includes: The target frequency band is input into a preset finite impulse response filter for filtering to obtain a finite impulse response; Based on the finite impulse response, the target time-domain filter coefficients are determined as the target parameters.

4. The method according to any one of claims 1 to 3, characterized in that, The initial model includes a convolutional encoder, a convolutional decoder, and a domain classifier, and the method further includes: The vibration data of the multiple source domain robots are respectively input into the convolutional encoder to obtain the source domain feature codes corresponding to the multiple source domain robots. The source domain feature codes corresponding to the multiple source domain robots are respectively input into the convolutional decoder to obtain the reconstructed feature codes corresponding to the multiple source domain robots. The source domain feature codes corresponding to the multiple source domain robots are respectively input into the domain classifier to obtain the domain classification codes corresponding to the multiple source domain robots. Based on the source domain feature codes corresponding to the multiple source domain robots, the feature codes are reconstructed, and the domain classification codes are determined to determine the total loss of the initial model. If the total loss is less than a predetermined threshold, the initial model is determined to be the multi-source domain adversarial migration model.

5. The method according to claim 4, characterized in that, Before inputting the source domain feature codes corresponding to the plurality of source domain robots into the convolutional decoder to obtain the reconstructed feature codes corresponding to the plurality of source domain robots, the method further includes: Set the number of channels of the first convolutional layer included in the convolutional decoder to the target parameter.

6. The method according to claim 4, characterized in that, The step of determining the total loss of the initial model based on the source domain feature encoding, reconstructed feature encoding, and domain classification encoding corresponding to the multiple source domain robots includes: Based on the source domain feature encoding and reconstruction feature encoding corresponding to the multiple source domain robots, the reconstruction loss is determined; Determine the first layer convolutional filter coefficient matrix in the initial model, and the time-domain coefficient length of the preset finite impulse response filter; The filtering loss is determined based on the first layer convolutional filter coefficient matrix, the predetermined time coefficient length, and the predetermined balance coefficient. Based on the domain classification codes and vibration data corresponding to the multiple source domain robots, and the predetermined network coefficients, the domain classification loss is determined; The total loss is determined based on the reconstruction loss, the filtering loss, and the domain classification loss.

7. The method according to claim 1, characterized in that, Determining the health status of the target robot based on the target domain feature encoding includes: A first robot is identified as being in the predetermined first state, and a second robot is identified as being in the predetermined second state, wherein the health status of the predetermined first state is better than the health status of the predetermined second state; Obtain vibration data of the first robot and vibration data of the second robot; The vibration data of the first robot is input into the multi-source domain adversarial transfer model for processing to obtain the first feature code of the first robot; The vibration data of the second robot is input into the multi-source domain adversarial transfer model for processing to obtain the second feature code of the second robot; Based on the first feature encoding, the second feature encoding, and the target domain feature encoding, normalization processing is performed to obtain the target normalization result of the target domain feature encoding; The health status is determined based on the target normalization result.

8. The method according to claim 7, characterized in that, The acquisition of motion state data and vibration data of the target robot includes: collecting motion state data and vibration data of the target robot in the target motion state; The target robot has multiple motion states, each corresponding to motion state data and target vibration data. Determining the health state based on the target normalization result includes: processing each of the multiple motion states as the target motion state to obtain the target normalization result corresponding to each of the multiple motion states; and obtaining the health state based on the number of motions in each of the multiple motion states and the target normalization result corresponding to each of the multiple motion states.

9. A device for determining the health status of a robot, characterized in that, include: The acquisition module is used to acquire motion state data and vibration data of the target robot; The determination module is used to determine the target parameters of the target robot based on the motion state data; The data processing module is used to input the target vibration data into a multi-source domain adversarial transfer model for processing to obtain the target domain feature code of the target robot. The multi-source domain adversarial transfer model is initialized based on the target parameters and trained using the source domain feature codes corresponding to multiple source domain robots. The multiple source domain robots are robots in a predetermined first state, and the source domain feature codes corresponding to the multiple source domain robots are obtained based on the vibration data corresponding to the multiple source domain robots. A health status determination module is used to determine the health status of the target robot based on the target domain feature encoding.

10. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores multiple instructions, which are adapted to be loaded by a processor and executed by the robot health status determination method according to any one of claims 1 to 8.

Images