Deep learning-based robot arm joint torque offline analysis method and related device

Abstract

The application provides a kind of based on deep learning's mechanical arm joint torque offline analysis method and related equipment, it is related to mechanical arm control technical field.The application is by to the joint torque offline data and joint kinematics offline data that target mechanical arm is actually measured in current running period bidirectional timing inverse dynamics feature fitting learning, to dig the bidirectional timing dependent relationship and parameter physical coupling relationship between joint torque offline data and joint kinematics offline data, obtain corresponding bidirectional dependent fusion dynamics feature, then joint kinematics offline data is used as external constraint condition, calls joint torque prediction network based on bidirectional dependent fusion dynamics feature carries out joint torque offline prediction, to consider the long-term influence of joint kinematics offline data on joint torque prediction result from global angle, to realize high-precision mechanical arm joint torque prediction function in offline analysis scene, improve the information utilization of mechanical arm joint offline data.

Description

Technical Field

[0001] This application relates to the field of robotic arm control technology, and more specifically, to a deep learning-based offline analysis method and related equipment for robotic arm joint torque. Background Technology

[0002] With the continuous development of science and technology, robotics has received widespread attention from various industries due to its immense research and application value. Intelligent robotic arms are a significant application area for robotics in industries such as industrial production, biomedicine, and services. In the practical application of intelligent robotic arms, it is typically necessary to obtain complete historical joint data records of the corresponding intelligent robotic arm performing a specific task for offline data analysis. This data is used to predict the joint torque parameters of the intelligent robotic arm at a future point in time (or several points in between), facilitating the optimization of the robotic arm's control parameters for the next round (e.g., joint motion trajectory, joint angular velocity change curves, etc.).

[0003] However, it is worth noting that the current mainstream offline joint torque analysis solutions for robotic arms are usually built using the "one-way causal reasoning thinking" in the real-time control process of robotic arms (i.e., using historical information to directly predict future information along the time axis). They do not consider the important corrective role of the subsequent information in the complete historical joint data in offline analysis scenarios when understanding the dynamic characteristics of the preceding information. They also often do not consider the strong physical coupling relationship between joint torque and different joint kinematic parameters (e.g., joint angle, joint angular velocity, joint angular acceleration, etc.), which can easily lead to low accuracy in the final offline joint torque prediction results. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and device, computer equipment and readable storage medium for offline analysis of joint torque of robotic arms based on deep learning. It can, on the basis of deep learning technology framework, deeply explore the bidirectional temporal dependency relationship between offline joint torque data and offline joint kinematics data under the same offline observation period (i.e., the current running period). This includes the positive physical causal relationship that conforms to the inverse dynamics equation of the robotic arm (which is used to describe the driving mechanism of joint kinematic parameters on joint torque), and the inverse smoothing correction characteristics of the subsequent steady-state information of the robotic arm when the action is completed on the preceding transient dynamic features during the action execution process (which is used to describe the measurement noise elimination and interpretation ambiguity elimination effect of the subsequent steady-state information on the preceding transient dynamic features) and the physical coupling relationship of parameters. It also considers the long-term impact of offline joint kinematics data on joint torque prediction results from a global perspective, so as to achieve high-precision robotic arm joint torque prediction function by utilizing a global bidirectional view in the offline analysis scenario, and improve the information utilization rate of offline joint data of robotic arms (including offline joint torque data and offline joint kinematics data).

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, this application provides a deep learning-based offline analysis method for joint torque of a robotic arm, the method comprising: Acquire offline joint torque and joint kinematics data of the target robotic arm during the current runtime. Based on the offline joint torque data and the offline joint kinematics data, bidirectional temporal inverse dynamic feature fitting and learning are performed to obtain the corresponding bidirectional dependent fused dynamic features. Using the offline joint kinematics data as an external constraint, a pre-trained joint torque prediction network is invoked to perform offline joint torque prediction based on the bidirectional dependency fused dynamic features, thereby obtaining the joint torque prediction data of the target robotic arm in future running segments, wherein the future running segment is adjacent to the current running segment.

[0006] In an optional implementation, the step of performing bidirectional temporal inverse dynamic feature fitting and learning based on the offline joint torque data and the offline joint kinematics data to obtain the corresponding bidirectional dependent fused dynamic features includes: Time series were constructed from the offline joint torque data and the offline joint kinematics data to obtain the original joint torque sequence and joint motion state sequence with timestamp alignment. The original joint torque sequence and the joint motion state sequence are fused over time to obtain the full joint state sequence corresponding to the inverse dynamics equation of the robotic arm. A bidirectional gated recurrent unit network is invoked to perform bidirectional temporal dependency capture fusion on the full joint state sequence to obtain the bidirectional dependency fused dynamic features.

[0007] In an optional implementation, the bidirectional gated loop unit network includes a forward gated loop unit module and a reverse gated loop unit module. The step of calling the bidirectional gated loop unit network to perform bidirectional temporal dependency capture fusion on the full joint state sequence to obtain the bidirectional dependency fused dynamic features includes: The positive gating loop unit module is invoked to recursively capture the positive physical causal relationship of the full joint state sequence, thereby obtaining positive causal dynamic features that match the full joint state sequence; The reverse gated loop unit module is invoked to recursively capture the reverse smoothing correction characteristics of the full joint state sequence, thereby obtaining the reverse correction dynamic characteristics that match the full joint state sequence. The positive causal dynamic features and the negative correction dynamic features are spliced ​​and fused to obtain the bidirectional dependency fused dynamic features.

[0008] In an optional implementation, the joint torque prediction network includes multiple cascaded TimeXer encoders and torque prediction output modules. The step of using the offline joint kinematics data as an external constraint, calling the pre-trained joint torque prediction network to perform offline joint torque prediction based on the bidirectional dependency fused dynamic features, and obtaining the estimated joint torque data for the target robotic arm during future operation includes: The bidirectional dependency fusion dynamic features are subjected to convolutional slice embedding processing to obtain the corresponding dynamic feature slice embedding vector sequence, and a learnable global label embedding vector is injected into the dynamic feature slice embedding vector sequence to obtain the initial dynamic embedding vector sequence. The joint motion state sequence corresponding to the offline joint kinematics data is inverted and embedded to obtain the corresponding kinematic embedding vector sequence. The kinematic embedding vector sequence is used as the exogenous embedding vector input to each of the plurality of TimeXer encoders, and the initial dynamic embedding vector sequence is used as the endogenous embedding vector input to the first TimeXer encoder among the plurality of TimeXer encoders. Each TimeXer encoder performs cross-variable global spatiotemporal fusion prediction based on the input exogenous embedding vector and endogenous embedding vector, and obtains the target cross-variable fusion prediction feature output by the last TimeXer encoder among the plurality of TimeXer encoders. The cross-variable fusion prediction feature output by the first TimeXer encoder in two adjacent TimeXer encoders is used as the endogenous embedding vector input to the second TimeXer encoder. The torque prediction output module is invoked to perform feature flattening mapping based on the target cross-variable fusion prediction features to obtain the joint torque prediction data.

[0009] In an optional implementation, for each of the plurality of TimeXer encoders, the step of performing cross-variable global spatiotemporal fusion prediction based on the input exogenous embedding vector and endogenous embedding vector to obtain the cross-variable fusion prediction features output by the TimeXer encoder includes: Multi-head self-attention learning is performed on the input endogenous embedding vector using a self-attention mechanism to obtain the corresponding endogenous self-attention learning result, wherein the endogenous self-attention learning result includes a self-attention local slice feature sequence and a global learned feature vector; The cross-attention mechanism is used to perform cross-attention learning on the global learned feature vector and the input exogenous embedding vector to obtain the corresponding cross-attention learning result. The self-attention local slice feature sequence and the cross-attention learning result are spliced ​​and fused, and the joint torque feedforward prediction is performed based on the corresponding feature splicing and fusion result to obtain the corresponding torque feedforward prediction feature; The feature splicing and fusion result and the torque feedforward prediction feature are subjected to feature overlay processing, and the feature overlay result is subjected to layer normalization processing to obtain the corresponding cross-variable fusion prediction feature.

[0010] In an optional implementation, the method further includes: Acquire measured joint torque and joint angular velocity data of the target robotic arm during the future operating period; Calculate the torque residual amplitude data between the estimated joint torque data and the measured joint torque data; For each physical joint involved in the target robotic arm, it is detected whether all target torque residual amplitude values ​​of the physical joint in the torque residual amplitude data are greater than a preset torque residual amplitude threshold, and whether the physical joint continuously maintains a stable joint angular velocity at the measured joint angular velocity data. If all target torque residual amplitude values ​​corresponding to the physical joint are detected to be greater than the preset torque residual amplitude threshold, and the physical joint continues to maintain a stable joint angular velocity, it is determined that there is a wear fault or lubrication failure fault inside the physical joint.

[0011] Secondly, this application provides a deep learning-based offline analysis device for robotic arm joint torque, the device comprising: The joint data acquisition module is used to acquire offline joint torque data and offline joint kinematics data that are actually measured by the target robotic arm during the current running period. The bidirectional dependency learning module is used to perform bidirectional temporal inverse dynamic feature fitting and learning based on the offline joint torque data and the offline joint kinematics data to obtain the corresponding bidirectional dependency fused dynamic features. The torque offline prediction module is used to use the joint kinematics offline data as an external constraint, call the pre-trained joint torque prediction network to perform joint torque offline prediction based on the bidirectional dependency fused dynamic features, and obtain the joint torque prediction data of the target robotic arm in the future operation period.

[0012] In an optional embodiment, the device further includes a residual amplitude calculation module, a joint state detection module, and a joint fault output module; The joint data acquisition module is also used to acquire measured joint torque data and measured joint angular velocity data of the target robotic arm during the future operating period; The residual amplitude calculation module is used to calculate the torque residual amplitude data between the estimated joint torque data and the measured joint torque data; The joint state detection module is used to detect, for each physical joint involved in the target robotic arm, whether all target torque residual amplitude values ​​of the physical joint in the torque residual amplitude data are greater than a preset torque residual amplitude threshold, and whether the physical joint continuously maintains a stable joint angular velocity state at the measured joint angular velocity data. The joint fault output module is used to determine that there is a wear fault or lubrication failure fault inside the physical joint when it is detected that all target torque residual amplitude values ​​corresponding to the physical joint are greater than the preset torque residual amplitude threshold and the physical joint continues to maintain a stable joint angular velocity.

[0013] Thirdly, this application provides a computer device, including a processor and a memory, wherein the memory stores a computer program executable by the processor, and the processor can execute the computer program to implement the deep learning-based offline analysis method for joint torque of a robotic arm as described in any of the foregoing embodiments for a target robotic arm.

[0014] Fourthly, this application provides a readable storage medium storing a computer program thereon, which, when executed by a computer device, implements the deep learning-based offline analysis method for robotic arm joint torque as described in any of the foregoing embodiments.

[0015] In this case, the beneficial effects of the embodiments of this application may include the following: After obtaining the offline joint torque and kinematics data of the target robotic arm during the current operating period, this application performs bidirectional temporal inverse dynamics feature fitting and learning based on the offline joint torque and kinematics data to deeply explore the bidirectional temporal dependency and parameter physical coupling relationship between the offline joint torque and kinematics data under the same offline observation period, and obtains the corresponding bidirectional dependency fused dynamic features. Then, the offline kinematics data is used as an external constraint condition, and the joint torque prediction network is called to perform offline joint torque prediction based on the bidirectional dependency fused dynamic features, so as to consider the long-term impact of the offline kinematics data on the joint torque prediction results from a global perspective, and obtain the joint torque prediction data of the target robotic arm in the future operating period. Thus, in the offline analysis scenario, a high-precision robotic arm joint torque prediction function can be achieved by using deep learning technology and a global bidirectional vision, thereby improving the information utilization rate of the robotic arm joint offline data.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of the composition of a computer device provided in the embodiments of this application; Figure 2 One of the flowcharts for the offline analysis method of robot arm joint torque based on deep learning provided in the embodiments of this application; Figure 3 for Figure 2 A flowchart illustrating the sub-steps included in step S220; Figure 4 for Figure 2 A flowchart illustrating the sub-steps included in step S230; Figure 5 The second flowchart illustrates the offline analysis method for robot joint torque based on deep learning provided in this application embodiment. Figure 6 One of the schematic diagrams of the composition of the deep learning-based robotic arm joint torque offline analysis device provided in the embodiments of this application; Figure 7This is a second schematic diagram of the composition of the deep learning-based robotic arm joint torque offline analysis device provided in the embodiments of this application.

[0019] Icons: 10-Computer equipment; 11-Memory; 12-Processor; 13-Communication unit; 100-Offline torque analysis device for robotic arm joints; 110-Joint data acquisition module; 120-Bidirectional dependency learning module; 130-Offline torque prediction module; 140-Residual amplitude calculation module; 150-Joint status detection module; 160-Joint fault output module. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0021] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0022] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0023] In the description of this application, it should be understood that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are used only for the convenience of describing this application and simplifying the description, and are not intended to indicate or imply that the equipment or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0024] In the description of this application, it should also be noted that, unless otherwise expressly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0025] Furthermore, it is understood in the description of this application that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, 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 said element. Those skilled in the art will understand the specific meaning of the above terms in this application based on the specific circumstances.

[0026] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0027] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the device composition of the computer device 10 provided in this application embodiment. In this application embodiment, the computer device 10 can achieve high-precision joint torque prediction for any intelligent robotic arm in an offline analysis scenario. Based on the predicted joint torque results, it can further accurately analyze whether wear or lubrication failures occur within the various physical joints involved in the intelligent robotic arm, or further optimize the intended motion trajectory of the intelligent robotic arm offline. The computer device 10 can be integrated with the intelligent robotic arm, or it can be an independent electronic device, including but not limited to servers, personal computers, and laptops.

[0028] In this embodiment of the application, the computer device 10 may include a memory 11, a processor 12, and a communication unit 13. The memory 11, the processor 12, and the communication unit 13 are electrically connected to each other directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other via one or more communication buses or signal lines.

[0029] In this embodiment, the memory 11 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc. The memory 11 is used to store computer programs, and the processor 12 can execute the computer programs accordingly after receiving execution instructions.

[0030] In this embodiment, the processor 12 can be an integrated circuit chip with signal processing capabilities. The processor 12 can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in this embodiment.

[0031] In this embodiment, the communication unit 13 is used to establish a communication connection between the computer device 10 and other electronic devices via a network, and to send and receive data via the network, wherein the network includes wired communication networks and wireless communication networks. For example, the computer device 10 can obtain offline joint data of the target robotic arm that needs to perform offline joint torque analysis through the communication unit 13 (which includes measured joint torque information of the target robotic arm at any time point (composed of the joint torque values ​​actually measured by each of the physical joints of the target robotic arm at that time point) and measured joint kinematic parameters (composed of the joint angles, joint angular velocities, and joint angular accelerations actually measured by each of the physical joints of the target robotic arm at that time point)).

[0032] In this embodiment, the computer device 10 may pre-store a specific computer program related to the offline analysis function of the robotic arm joint torque in the memory 11. By driving the processor 12 to execute the specific computer program, the device introduces a deep learning technology framework in the offline analysis scenario to deeply explore the bidirectional temporal dependency and parameter physical coupling relationship between the offline joint torque data and the offline joint kinematics data under the same offline observation period. The device also considers the long-term impact of the offline joint kinematics data on the joint torque prediction results from a global perspective, so as to realize a high-precision robotic arm joint torque prediction function by utilizing a global bidirectional view and improve the information utilization rate of the robotic arm joint offline data.

[0033] Understandable Figure 1 The block diagram shown is only a schematic diagram of one configuration of the computer device 10. The computer device 10 may also include components such as... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown. Figure 1 The components shown can be implemented using hardware, software, or a combination thereof.

[0034] In this application, to ensure that the computer device 10 can achieve high-precision prediction of robotic arm joint torque for any type of intelligent robotic arm in an offline analysis scenario, this application embodiment provides a deep learning-based offline analysis method for robotic arm joint torque to achieve the aforementioned objective. The offline analysis method for robotic arm joint torque provided in this application will be described in detail below.

[0035] Please refer to Figure 2 , Figure 2 This is one of the flowcharts illustrating the offline torque analysis method for robotic arms based on deep learning provided in this application embodiment. In this application embodiment, Figure 2The offline analysis method for the joint torque of the robotic arm shown may include steps S210 to S230.

[0036] Step S210: Obtain offline joint torque data and offline joint kinematics data of the target robotic arm during the current running period.

[0037] In this embodiment, the target robotic arm is an intelligent robotic arm that needs to perform offline joint torque analysis. The target robotic arm can measure offline joint torque data and offline joint kinematics data involving multiple equally spaced sampling timestamps within any operating segment according to a fixed data sampling rate (e.g., 100Hz) and a fixed time window length (e.g., 100). The offline joint torque data includes measured joint torque information for each sampling timestamp within the current operating segment (wherein, the measured joint torque information for each sampling timestamp consists of the joint torque values ​​actually measured by each of the physical joints of the target robotic arm at that sampling timestamp). The offline joint kinematics data includes measured joint kinematic parameters for each sampling timestamp within the current operating segment (wherein, the measured joint kinematic parameters for each sampling timestamp consists of the joint angle, joint angular velocity, and joint angular acceleration actually measured by each of the physical joints of the target robotic arm at that sampling timestamp).

[0038] Step S220: Based on the offline joint torque data and the offline joint kinematics data, perform bidirectional temporal inverse dynamic feature fitting and learning to obtain the corresponding bidirectional dependent fused dynamic features.

[0039] In this embodiment, the computer device 10 can pre-store a bidirectional gated recurrent unit network and configure it to implicitly fit the state of the inverse dynamics equation of the robotic arm. This allows the bidirectional gated recurrent unit network to learn the bidirectional coupling relationship between the kinematic state of the robotic arm (i.e., the joint angles, angular velocities, and angular accelerations of each joint of the target robotic arm) and the dynamic response of the robotic arm (i.e., the joint torques of each joint of the target robotic arm). Therefore, after obtaining the timestamp-aligned offline joint torque data and offline joint kinematics data within the current runtime, the computer device 10 can call the bidirectional gated recurrent unit network to capture the bidirectional temporal dependency and parametric physical coupling relationship between the offline joint torque data and the offline joint kinematics data for comprehensive dynamic feature characterization, thereby obtaining a high-dimensional dynamic feature flow fused with bidirectional temporal dependencies (i.e., the bidirectional dependency fused dynamic features).

[0040] Alternatively, please refer to Figure 3 , Figure 3 yes Figure 2A flowchart illustrating the sub-steps included in step S220. In this embodiment, step S220 may include sub-steps S221 to S223 to deeply explore the bidirectional temporal dependency and parameter physical coupling relationship between the timestamp-aligned offline joint torque data and offline joint kinematics data, ensuring that the learned bidirectional dependency fused dynamic features can truly reflect the overall dynamic evolution of the target robotic arm during the current operating period.

[0041] Sub-step S221 involves constructing time series data from offline joint torque data and offline joint kinematics data to obtain timestamp-aligned original joint torque sequences and joint motion state sequences.

[0042] In this embodiment, the sampling timestamps involved in the original joint torque sequence and the joint motion state sequence are aligned with each other. The original joint torque sequence includes the measured joint torque information of each of the multiple sampling timestamps that are equally distributed within the current running segment, and the joint motion state sequence includes the measured joint kinematic parameter information of each of the multiple sampling timestamps that are equally distributed within the current running segment.

[0043] Sub-step S222 involves time-series fusion of the original joint torque sequence and joint motion state sequence to obtain the full joint state sequence corresponding to the inverse dynamics equation of the robotic arm.

[0044] In this embodiment, the sampling timestamps involved in the full joint state sequence, the original joint torque sequence, and the joint motion state sequence are aligned with each other. Each sequence element in the full joint state sequence is formed by splicing the measured joint torque information and measured joint kinematic parameter information of a single sampling timestamp, so that each sequence element of the full joint state sequence substantially covers all physical quantities involved in the inverse dynamics equation of the robotic arm.

[0045] Sub-step S223 calls a bidirectional gated recurrent unit network to perform bidirectional temporal dependency capture and fusion on the full joint state sequence to obtain bidirectional dependency fused dynamic features.

[0046] In this embodiment, the bidirectional gated loop unit network includes a forward gated loop unit module and a reverse gated loop unit module. The forward gated loop unit module is used to recursively capture the physical causal relationship that conforms to the inverse dynamics equation of the robotic arm (which describes the driving mechanism of joint kinematic parameters on joint torque, i.e., joint torque is essentially a strongly coupled function result of joint angle, joint angular velocity, and joint angular acceleration) along the forward time axis. The reverse gated loop unit module is used to recursively capture the dynamic smoothing correction characteristics (which describes the measurement noise cancellation and interpretation ambiguity cancellation effect of the subsequent steady-state information of the robotic arm when the action is completed on the preceding transient dynamic characteristics during the action execution process, where interpretation ambiguity is often caused by nonlinear friction) along the reverse time axis, so as to ensure that the complete feature vector sequence (i.e., the bidirectional dependent fused dynamic characteristics) finally output by the bidirectional gated loop unit network can truly reflect the overall dynamic evolution law of the target robotic arm in the current running segment.

[0047] At this point, it can be understood that the specific execution steps of the above sub-step S223 may include: The positive gating loop unit module is invoked to recursively capture the positive physical causal relationship of the full joint state sequence, thereby obtaining positive causal dynamic features that match the full joint state sequence; The reverse gated loop unit module is invoked to recursively capture the reverse smoothing correction characteristics of the full joint state sequence, thereby obtaining the reverse correction dynamic characteristics that match the full joint state sequence. The positive causal dynamic features and the negative correction dynamic features are spliced ​​and fused to obtain the bidirectional dependency fused dynamic features.

[0048] The forward causal dynamics feature is essentially a forward hidden state vector sequence calculated recursively by the forward gated loop unit module based on the full joint state sequence, and the forward hidden state vector sequence is aligned with the timestamp of the full joint state sequence. The reverse correction dynamics feature is essentially a reverse hidden state vector sequence calculated recursively by the reverse gated loop unit module based on the full joint state sequence, and the reverse hidden state vector sequence is aligned with the timestamp of the full joint state sequence. The bidirectional dependency fusion dynamics feature is essentially the result of fusion of the forward hidden state vector sequence and the reverse hidden state vector sequence. The dynamics feature vector of the bidirectional dependency fusion dynamics feature at each sampling timestamp is formed by concatenating the forward hidden state vector and the reverse hidden state vector of the corresponding timestamp, so that the dynamics feature vector of each sampling timestamp can simultaneously contain the historical causal information before the corresponding sampling timestamp and the future evolution trend after the corresponding sampling timestamp, ensuring that the entire bidirectional dependency fusion dynamics feature can truly reflect the overall dynamic evolution law of the target robotic arm in the current running period.

[0049] Therefore, by executing the above sub-steps S221 to S223, this application can deeply explore the bidirectional temporal dependency and parameter physical coupling relationship between the timestamp-aligned offline joint torque data and the offline joint kinematics data, ensuring that the learned bidirectional dependency fused dynamic features can truly reflect the overall dynamic evolution law of the target robotic arm in the current running segment.

[0050] Step S230: Using the offline joint kinematics data as an external constraint, the pre-trained joint torque prediction network is called to perform offline joint torque prediction based on bidirectional dependency fusion dynamic features, thereby obtaining the joint torque prediction data of the target robotic arm in the future operation period.

[0051] In this embodiment, the future runtime segment is adjacent to the current runtime segment. The joint torque prediction data includes joint torque estimation information corresponding to multiple future timestamps that are equally spaced within the future runtime segment (which consists of the joint torque estimation values ​​of all physical joints of the target robotic arm at each future timestamp). The number of future timestamps involved in the future runtime segment may be the same as or different from the number of sampling timestamps involved in the current runtime segment. The joint torque prediction network is used to predict the joint torque distribution of the target robotic arm under ideal conditions within the future runtime segment. The joint torque prediction network may include multiple cascaded TimeXer (Time Series Transformer with exogenous variables) encoders and torque prediction output modules (consisting of a flatten head and a linear layer) to treat the offline joint kinematics data as exogenous variables and the bidirectional dependent fused dynamic features as endogenous variables. This allows for a global consideration of the long-term impact of the offline joint kinematics data on the joint torque time series prediction results, ensuring that the final joint torque time series prediction results (i.e., the joint torque prediction data) have good accuracy and reliability.

[0052] Alternatively, please refer to Figure 4 , Figure 4 yes Figure 2 A flowchart illustrating the sub-steps included in step S230 is provided. In this embodiment, step S230 may include sub-steps S231 to S234 to introduce offline joint kinematics data into the joint torque time series prediction process to constrain the dynamic behavior of the robotic arm at the physical motion level, thereby ensuring that the final joint torque time series prediction result has good accuracy and reliability.

[0053] Sub-step S231 involves performing convolutional slice embedding on the bidirectional dependency fusion dynamic features to obtain the corresponding dynamic feature slice embedding vector sequence, and injecting a learnable global label embedding vector into the dynamic feature slice embedding vector sequence to obtain the initial dynamic embedding vector sequence.

[0054] In this embodiment, a one-dimensional convolutional layer can be used to map continuous and high-frequency bidirectional dependency fusion dynamic features into low-frequency high-dimensional semantic slices. This ensures that the final generated dynamic feature slice embedding vector sequence can effectively preserve local dynamic features while compressing the overall sequence length to 1 / N times the original sequence length. Here, N represents the preset semantic slice length, and the kernel size and stride of the one-dimensional convolutional layer are both set to the preset semantic slice length. The global label embedding vector can be concatenated to the end of the dynamic feature slice embedding vector sequence to form the initial dynamic embedding vector sequence. This facilitates the aggregation of local information from all dynamic feature slice embedding vectors using the global label embedding vector to generate a global summary representation of the initial dynamic embedding vector sequence.

[0055] Sub-step S232 involves performing inverted embedding processing on the joint motion state sequence corresponding to the offline joint kinematics data to obtain the corresponding kinematic embedding vector sequence.

[0056] In this embodiment, the inverted embedding operation can be used to describe the continuous change characteristics of the parameter values ​​of the corresponding joint motion state sequence when facing different kinematic parameter types (e.g., joint angle, joint angular velocity, or joint angular acceleration). This includes the numerical change trend, periodic change characteristics, and change frequency characteristics of the corresponding kinematic parameter types, so that the final kinematic embedding vector sequence has good temporal integrity and multivariate heterogeneity compatibility.

[0057] Sub-step S233: The kinematic embedding vector sequence is used as the exogenous embedding vector input to each of the multiple TimeXer encoders, and the initial dynamic embedding vector sequence is used as the endogenous embedding vector input to the first TimeXer encoder among the multiple TimeXer encoders. Each TimeXer encoder performs cross-variable global spatiotemporal fusion prediction based on the input exogenous embedding vector and endogenous embedding vector, and obtains the target cross-variable fusion prediction feature output by the last TimeXer encoder among the multiple TimeXer encoders.

[0058] In this embodiment, the cross-variable fusion prediction feature output by the preceding TimeXer encoder among any two adjacent TimeXer encoders is used as the endogenous embedding vector input to the following TimeXer encoder. This allows the global information learned by the preceding TimeXer encoder from its own input kinematic embedding vector sequence to be inherited and used by the following TimeXer encoder, ensuring that the last TimeXer encoder can learn the long-term impact of offline joint kinematics data on the joint torque time series prediction results from a global perspective.

[0059] Specifically, for each of the multiple TimeXer encoders, the TimeXer encoder can, based on its own network model architecture, utilize the organic combination of self-attention mechanism, cross-attention mechanism, and feedforward neural network to achieve cross-variable global spatiotemporal fusion prediction of endogenous and exogenous embedding vectors at the level of robotic arm dynamics. This ensures that its output cross-variable fusion prediction features can reflect the influence of offline joint kinematics data on joint torque time series prediction from a global perspective. The step of each TimeXer encoder performing cross-variable global spatiotemporal fusion prediction based on the input exogenous and endogenous embedding vectors to obtain the output cross-variable fusion prediction features can be divided into sub-steps A to D as shown below, according to the network model architecture of the TimeXer encoder.

[0060] Sub-step A: Use the self-attention mechanism to perform multi-head self-attention learning on the input endogenous embedding vector to obtain the corresponding endogenous self-attention learning result, wherein the endogenous self-attention learning result includes the self-attention local slice feature sequence and the global learned feature vector.

[0061] In this embodiment, the internal temporal dependencies of the input endogenous embedding vector can be learned using a self-attention mechanism. Combined with residual connection operations and layer normalization operations, this ensures that the globally learned feature vector in the corresponding endogenous self-attention learning result can effectively aggregate all dynamic features of the relevant self-attention local slice feature sequence. Specifically, for any TimeXer encoder, its input endogenous embedding vector can be divided into a known local slice embedding vector sequence… (e.g., the sequence of dynamic feature slice embedding vectors in the initial dynamic embedding vector sequence) and the known global label vector. (For example, the global label embedding vector in the initial dynamic embedding vector sequence), and the specific learning process of the endogenous self-attention learning result of this TimeXer encoder can be represented as: ; in, This is used to represent the endogenous self-attention learning result of the TimeXer encoder. This is used to represent the self-attention local slice feature sequence in the corresponding endogenous self-attention learning result. Used to represent the global learned feature vector in the corresponding endogenous self-attention learning result. This is used to represent the endogenous embedding vector of the TimeXer encoder input. Used to represent the layer normalization function. Used to represent the self-attention calculation function.

[0062] Sub-step B: Use the cross-attention mechanism to perform cross-attention learning on the global learning feature vector and the input exogenous embedding vector to obtain the corresponding cross-attention learning result.

[0063] In this embodiment, the globally learned feature vector can be used as a query term, and matching key and value terms can be constructed based on the input exogenous embedding vector. This ensures the accuracy of the corresponding cross-attention learning results through the organic combination of cross-attention computation operations, residual connection operations, and layer normalization operations. "It is possible to globally learn the physical motion mapping relationship between the joint torque change trend and the joint kinematic state at the time dimension. The specific learning process of the cross-attention learning result can be represented as follows:" ; in, Used to represent the cross-attention calculation function. This is used to represent the exogenous embedding vector of the TimeXer encoder input.

[0064] Sub-step C: Perform feature splicing and fusion on the self-attention local slice feature sequence and the cross-attention learning result, and perform joint torque feedforward prediction based on the corresponding feature splicing and fusion result to obtain the corresponding torque feedforward prediction feature.

[0065] In this embodiment, the torque feedforward prediction feature of a single TimeXer encoder The result of feature splicing and fusion The torque feedforward prediction feature is obtained through joint torque prediction processing using a feedforward neural network. The specific feedforward prediction process of the torque feedforward prediction feature can be expressed as follows: .in, Used to represent the sequence of predicted local slice embedding vectors calculated by the feedforward neural network. Used to represent the estimated global label vector calculated by the feedforward neural network. Used to represent feedforward prediction functions.

[0066] Sub-step D: Perform feature overlay processing on the feature splicing and fusion result and the torque feedforward prediction feature, and perform layer normalization processing on the feature overlay result to obtain the corresponding cross-variable fusion prediction feature.

[0067] In this embodiment, the cross-variable fusion prediction features output by a single TimeXer encoder The sequence of local slice embedding vectors to be output and the global tag vector to be output These are respectively used as the known local slice embedding vector sequence and the known global label vector in the corresponding endogenous embedding vector of the next TimeXer encoder. The specific fusion process of the cross-variable fusion prediction feature can be expressed as follows: .

[0068] In sub-step S234, the torque prediction output module is called to perform feature flattening mapping based on the target cross-variable fusion prediction features to obtain joint torque prediction data.

[0069] In this embodiment, the torque prediction output module can perform feature tensor flattening on the received target cross-variable fusion prediction features through FlattenHead, and use a linear layer to map the feature tensor flattening result to a pre-set prediction time dimension (which can be constrained by the number of future timestamps mentioned above) and a robotic arm joint torque feature dimension (which can be constrained by the number of ideal joint torque dimensions associated with the number of all physical joints of the target robotic arm), thereby obtaining the joint torque time series prediction result (i.e., joint torque estimation data) corresponding to the future running segment.

[0070] Therefore, by executing the above sub-steps S231 to S234, this application introduces offline joint kinematics data to constrain the dynamic behavior of the robotic arm at the physical motion level during the joint torque time series prediction process, thereby ensuring that the final joint torque time series prediction result has good accuracy and reliability.

[0071] Meanwhile, by executing the above steps S210 to S230, this application, based on the introduction of a deep learning technology framework in the offline analysis scenario, can deeply explore the bidirectional temporal dependency and parameter physical coupling relationship between offline joint torque data and offline joint kinematics data under the same offline observation period, and consider the long-term impact of offline joint kinematics data on joint torque prediction results from a global perspective, so as to realize a high-precision robotic arm joint torque prediction function by utilizing a global bidirectional view, and improve the information utilization rate of offline data of robotic arm joints.

[0072] Alternatively, please refer to Figure 5 , Figure 5 This is the second flowchart illustrating the offline torque analysis method for robotic arms based on deep learning provided in this application embodiment. In this application embodiment, [the method is related to...]. Figure 2 Compared to the offline analysis method for robot arm joint torque shown, Figure 5 The offline analysis method for joint torque of the robotic arm shown may also include steps S240 to S270, so as to accurately diagnose whether the target robotic arm is involved in joint wear failure or joint lubrication failure in the offline analysis scenario by combining the prediction results of the joint torque of the robotic arm.

[0073] Step S240: Obtain measured joint torque and measured joint angular velocity data of the target robotic arm during future operation periods.

[0074] In this embodiment, the measured joint torque data includes measured joint torque information for each future timestamp of the future running segment (wherein, the measured joint torque information for each future timestamp consists of the measured joint torque values ​​actually measured for each of the physical joints of the target robotic arm at that future timestamp); the measured joint angular velocity data includes measured joint angular velocity information for each future timestamp of the future running segment (wherein, the measured joint angular velocity information for each future timestamp consists of the measured joint angular velocity values ​​actually measured for each of the physical joints of the target robotic arm at that future timestamp).

[0075] Step S250: Calculate the torque residual amplitude data between the estimated joint torque data and the measured joint torque data.

[0076] In this embodiment, the torque residual amplitude data includes the absolute value information of the joint torque residual for each future timestamp of the future running period (wherein, the absolute value information of the joint torque residual for each future timestamp is composed of the absolute value of the joint torque residual for each physical joint of the target robotic arm at that future timestamp (which is obtained by subtracting the estimated value of the joint torque from the measured value of the joint torque of the same physical joint at that future timestamp and then performing an absolute value calculation).

[0077] Step S260: For each physical joint involved in the target robotic arm, detect whether all target torque residual amplitude values ​​in the torque residual amplitude data of the physical joint are greater than the preset torque residual amplitude threshold, and detect whether the physical joint continuously maintains a stable joint angular velocity at the measured joint angular velocity data.

[0078] Among them, all target torque residual amplitude values ​​(i.e., absolute values ​​of joint torque residuals) corresponding to the same physical joint correspond to a future timestamp of the future running segment. It can be determined whether the physical joint continuously maintains a stable joint angular velocity at the measured joint angular velocity data by detecting whether the absolute value of the actual difference between the measured maximum joint angular velocity and the measured minimum joint angular velocity of a single physical joint in the future running segment is greater than a preset angular velocity difference amplitude threshold. When the absolute value of the actual difference is not greater than the preset angular velocity difference amplitude threshold, it can be determined that the corresponding physical joint maintains a stable joint angular velocity.

[0079] Step S270: If all target torque residual amplitude values ​​corresponding to the physical joint are detected to be greater than the preset torque residual amplitude threshold, and the physical joint continues to maintain a stable joint angular velocity, it is determined that there is a wear fault or lubrication failure fault inside the physical joint.

[0080] In this embodiment, when all target torque residual amplitude values ​​of a certain physical joint in the future running period are greater than the preset torque residual amplitude threshold, and the physical joint continuously maintains a stable joint angular velocity state in the future running period, it indicates that there is a wear fault or lubrication failure fault inside the physical joint.

[0081] Therefore, by executing the above steps S240 to S270 and combining the results of the robotic arm joint torque prediction, this application can accurately diagnose whether the target robotic arm is involved in joint wear failure or joint lubrication failure in an offline analysis scenario.

[0082] Furthermore, it is understood that for each physical joint of the target robotic arm, when the estimated maximum joint torque of the physical joint in the future running period is close to the rated torque of the joint motor of the physical joint, it indicates that there is a certain defect in the real-time motion trajectory of the physical joint in the future running period (e.g., joint movement exceeding the limit, joint movement jitter, etc.). At this time, the acceleration constraint of the motion trajectory to be implemented can be reduced in the offline trajectory planner of the robotic arm to achieve the offline optimization effect of the robotic arm trajectory.

[0083] In this application, to ensure that the computer device 10 described above can effectively perform... Figure 2 or Figure 5 The offline analysis method for robotic arm joint torque shown in this application implements the aforementioned functions by using a robotic arm joint torque offline analysis device configured in the memory 11 in the form of software functional modules. The specific composition of the robotic arm joint torque offline analysis device provided in this application will be described below.

[0084] Please refer to Figure 6 , Figure 6 This is one of the schematic diagrams of the deep learning-based offline torque analysis device 100 for robotic arms provided in this application embodiment. In this application embodiment, the offline torque analysis device 100 for robotic arms may include a joint data acquisition module 110, a bidirectional dependency learning module 120, and a torque offline prediction module 130.

[0085] The joint data acquisition module 110 is used to acquire offline joint torque data and offline joint kinematics data actually measured by the target robotic arm during the current operating period.

[0086] The bidirectional dependency learning module 120 is used to perform bidirectional temporal inverse dynamic feature fitting and learning based on the joint torque offline data and the joint kinematics offline data to obtain the corresponding bidirectional dependency fused dynamic features.

[0087] The torque offline prediction module 130 is used to use the joint kinematics offline data as an external constraint, call the pre-trained joint torque prediction network to perform joint torque offline prediction based on the bidirectional dependency fused dynamic features, and obtain the joint torque prediction data of the target robotic arm in the future running period.

[0088] It should be noted that the basic principle and technical effects of the robotic arm joint torque offline analysis device 100 provided in this application embodiment are similar to those of the previous one. Figure 2 The offline torque analysis method for the robotic arm joints shown is the same. For brevity, any parts not mentioned in this embodiment can be referred to the above-described methods for... Figure 2 The description of the offline analysis method for joint torque of the robotic arm shown.

[0089] Alternatively, please refer to Figure 7 , Figure 7 This is a second schematic diagram of the composition of the deep learning-based offline torque analysis device 100 for robotic arms provided in this application embodiment. In this application embodiment, it is related to... Figure 6 Compared to the robotic arm joint torque offline analysis device 100 shown, Figure 7 The robotic arm joint torque offline analysis device 100 shown may also include a residual amplitude calculation module 140, a joint state detection module 150, and a joint fault output module 160.

[0090] The joint data acquisition module 110 is also used to acquire measured joint torque data and measured joint angular velocity data of the target robotic arm during the future operating period.

[0091] The residual amplitude calculation module 140 is used to calculate the torque residual amplitude data between the joint torque prediction data and the joint torque measured data.

[0092] The joint state detection module 150 is used to detect, for each physical joint involved in the target robotic arm, whether all target torque residual amplitude values ​​of the physical joint in the torque residual amplitude data are greater than a preset torque residual amplitude threshold, and whether the physical joint continuously maintains a stable joint angular velocity state at the measured joint angular velocity data.

[0093] The joint fault output module 160 is used to determine that there is a wear fault or lubrication failure fault inside the physical joint when it is detected that all target torque residual amplitude values ​​corresponding to the physical joint are greater than the preset torque residual amplitude threshold and the physical joint continues to maintain a stable joint angular velocity.

[0094] It should be noted that the basic principle and technical effects of the robotic arm joint torque offline analysis device 100 provided in this application embodiment are similar to those of the previous one. Figure 5 The offline torque analysis method for the robotic arm joints shown is the same. For brevity, any parts not mentioned in this embodiment can be referred to the above-described methods for... Figure 5 The description of the offline analysis method for joint torque of the robotic arm shown.

[0095] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of the apparatus, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0096] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. If the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium and includes several instructions to cause a computer device (which may be an intelligent robotic arm, a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes: USB flash drives, mobile hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0097] The above descriptions are merely various embodiments of this application, but 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 technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based offline analysis method for joint torque of a robotic arm, characterized in that, The method includes: Acquire offline joint torque and joint kinematics data of the target robotic arm during the current runtime. Based on the offline joint torque data and the offline joint kinematics data, bidirectional temporal inverse dynamics feature fitting and learning are performed to obtain the corresponding bidirectional dependency fused dynamics features. Specifically, by constructing time series data for the offline joint torque data and the offline joint kinematics data respectively, timestamp-aligned original joint torque sequences and joint motion state sequences are obtained. Then, the original joint torque sequences and joint motion state sequences are fused in time series to obtain the full joint state sequence corresponding to the inverse dynamics equation of the robotic arm. Finally, a bidirectional gated recurrent unit network is called to perform bidirectional temporal dependency capture fusion on the full joint state sequence to obtain the bidirectional dependency fused dynamics features. Using the offline joint kinematics data as an external constraint, a pre-trained joint torque prediction network is invoked to perform offline joint torque prediction based on the bidirectional dependency fused dynamic features, thereby obtaining the joint torque prediction data of the target robotic arm in the future operating period. The joint torque prediction network includes multiple cascaded TimeXer encoders and torque prediction output modules. The step of using the offline joint kinematics data as an external constraint, calling the pre-trained joint torque prediction network to perform offline joint torque prediction based on the bidirectional dependency fused dynamic features, and obtaining the estimated joint torque data for the target robotic arm during future operation includes: The bidirectional dependency fusion dynamic features are subjected to convolutional slice embedding processing to obtain the corresponding dynamic feature slice embedding vector sequence, and a learnable global label embedding vector is injected into the dynamic feature slice embedding vector sequence to obtain the initial dynamic embedding vector sequence. The joint motion state sequence corresponding to the offline joint kinematics data is inverted and embedded to obtain the corresponding kinematic embedding vector sequence. The kinematic embedding vector sequence is used as the exogenous embedding vector input to each of the plurality of TimeXer encoders, and the initial dynamic embedding vector sequence is used as the endogenous embedding vector input to the first TimeXer encoder among the plurality of TimeXer encoders. Each TimeXer encoder performs cross-variable global spatiotemporal fusion prediction based on the input exogenous embedding vector and endogenous embedding vector, and obtains the target cross-variable fusion prediction feature output by the last TimeXer encoder among the plurality of TimeXer encoders. The cross-variable fusion prediction feature output by the first TimeXer encoder in two adjacent TimeXer encoders is used as the endogenous embedding vector input to the second TimeXer encoder. The torque prediction output module is invoked to perform feature flattening mapping based on the target cross-variable fusion prediction features to obtain the joint torque prediction data.

2. The method according to claim 1, characterized in that, The bidirectional gated loop unit network includes a forward gated loop unit module and a reverse gated loop unit module. The step of calling the bidirectional gated loop unit network to perform bidirectional temporal dependency capture fusion on the full joint state sequence to obtain the bidirectional dependency fused dynamic features includes: The positive gating loop unit module is invoked to recursively capture the positive physical causal relationship of the full joint state sequence, thereby obtaining positive causal dynamic features that match the full joint state sequence; The reverse gated loop unit module is invoked to recursively capture the reverse smoothing correction characteristics of the full joint state sequence, thereby obtaining the reverse correction dynamic characteristics that match the full joint state sequence. The positive causal dynamic features and the negative correction dynamic features are spliced ​​and fused to obtain the bidirectional dependency fused dynamic features.

3. The method according to claim 1, characterized in that, For each of the plurality of TimeXer encoders, the step of performing cross-variable global spatiotemporal fusion prediction based on the input exogenous embedding vector and endogenous embedding vector to obtain the cross-variable fusion prediction features output by the TimeXer encoder includes: Multi-head self-attention learning is performed on the input endogenous embedding vector using a self-attention mechanism to obtain the corresponding endogenous self-attention learning result, wherein the endogenous self-attention learning result includes a self-attention local slice feature sequence and a global learned feature vector; The cross-attention mechanism is used to perform cross-attention learning on the global learned feature vector and the input exogenous embedding vector to obtain the corresponding cross-attention learning result. The self-attention local slice feature sequence and the cross-attention learning result are spliced ​​and fused, and the joint torque feedforward prediction is performed based on the corresponding feature splicing and fusion result to obtain the corresponding torque feedforward prediction feature; The feature splicing and fusion result and the torque feedforward prediction feature are subjected to feature overlay processing, and the feature overlay result is subjected to layer normalization processing to obtain the corresponding cross-variable fusion prediction feature.

4. The method according to any one of claims 1-3, characterized in that, The method further includes: Acquire measured joint torque and joint angular velocity data of the target robotic arm during the future operating period; Calculate the torque residual amplitude data between the estimated joint torque data and the measured joint torque data; For each physical joint involved in the target robotic arm, it is detected whether all target torque residual amplitude values ​​of the physical joint in the torque residual amplitude data are greater than a preset torque residual amplitude threshold, and whether the physical joint continuously maintains a stable joint angular velocity at the measured joint angular velocity data. If all target torque residual amplitude values ​​corresponding to the physical joint are detected to be greater than the preset torque residual amplitude threshold, and the physical joint continues to maintain a stable joint angular velocity, it is determined that there is a wear fault or lubrication failure fault inside the physical joint.

5. A deep learning-based offline analysis device for robotic arm joint torque, characterized in that, The device includes: The joint data acquisition module is used to acquire offline joint torque data and offline joint kinematics data that are actually measured by the target robotic arm during the current running period. A bidirectional dependency learning module is used to perform bidirectional temporal inverse dynamics feature fitting and learning based on the offline joint torque data and the offline joint kinematics data to obtain the corresponding bidirectional dependency fused dynamics features. Specifically, the bidirectional dependency learning module constructs time series data from the offline joint torque data and the offline joint kinematics data respectively to obtain a timestamp-aligned original joint torque sequence and a joint motion state sequence. Then, it fuses the original joint torque sequence and the joint motion state sequence in time series to obtain a full joint state sequence corresponding to the inverse dynamics equation of the robotic arm. Finally, it calls a bidirectional gated recurrent unit network to perform bidirectional temporal dependency capture and fusion on the full joint state sequence to obtain the bidirectional dependency fused dynamics features. The torque offline prediction module is used to use the joint kinematics offline data as an external constraint, call the pre-trained joint torque prediction network to perform joint torque offline prediction based on the bidirectional dependency fused dynamic features, and obtain the joint torque prediction data of the target robot arm in the future running period. The joint torque prediction network includes multiple cascaded TimeXer encoders and torque prediction output modules. The offline torque prediction module uses the offline joint kinematics data as an external constraint and calls the pre-trained joint torque prediction network to perform offline joint torque prediction based on the bidirectional dependency fused dynamic features. The method for obtaining the estimated joint torque data of the target robotic arm during future operation includes: The bidirectional dependency fusion dynamic features are subjected to convolutional slice embedding processing to obtain the corresponding dynamic feature slice embedding vector sequence, and a learnable global label embedding vector is injected into the dynamic feature slice embedding vector sequence to obtain the initial dynamic embedding vector sequence. The joint motion state sequence corresponding to the offline joint kinematics data is inverted and embedded to obtain the corresponding kinematic embedding vector sequence. The kinematic embedding vector sequence is used as the exogenous embedding vector input to each of the plurality of TimeXer encoders, and the initial dynamic embedding vector sequence is used as the endogenous embedding vector input to the first TimeXer encoder among the plurality of TimeXer encoders. Each TimeXer encoder performs cross-variable global spatiotemporal fusion prediction based on the input exogenous embedding vector and endogenous embedding vector, and obtains the target cross-variable fusion prediction feature output by the last TimeXer encoder among the plurality of TimeXer encoders. The cross-variable fusion prediction feature output by the first TimeXer encoder in two adjacent TimeXer encoders is used as the endogenous embedding vector input to the second TimeXer encoder. The torque prediction output module is invoked to perform feature flattening mapping based on the target cross-variable fusion prediction features to obtain the joint torque prediction data.

6. The apparatus according to claim 5, characterized in that, The device also includes a residual amplitude calculation module, a joint status detection module, and a joint fault output module; The joint data acquisition module is also used to acquire measured joint torque data and measured joint angular velocity data of the target robotic arm during the future operating period; The residual amplitude calculation module is used to calculate the torque residual amplitude data between the estimated joint torque data and the measured joint torque data; The joint state detection module is used to detect, for each physical joint involved in the target robotic arm, whether all target torque residual amplitude values ​​of the physical joint in the torque residual amplitude data are greater than a preset torque residual amplitude threshold, and whether the physical joint continuously maintains a stable joint angular velocity state at the measured joint angular velocity data. The joint fault output module is used to determine that there is a wear fault or lubrication failure fault inside the physical joint when it is detected that all target torque residual amplitude values ​​corresponding to the physical joint are greater than the preset torque residual amplitude threshold and the physical joint continues to maintain a stable joint angular velocity.

7. A computer device, characterized in that, It includes a processor and a memory, the memory storing a computer program executable by the processor, the processor being able to execute the computer program to implement the deep learning-based offline analysis method for joint torque of a robotic arm as described in any one of claims 1-4 for a target robotic arm.

8. A readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a computer device, it implements the deep learning-based offline analysis method for robotic arm joint torque as described in any one of claims 1-4.

Images