An industrial robot rv reducer diagnosis method based on privileged observation distillation

Abstract

The application discloses an industrial robot RV reducer diagnosis method based on privileged observation distillation, which synchronously collects privileged observation signals such as reducer shell vibration, end inertia signal, motion phase sequence, fault category label and operation condition label in a training stage; a privileged observation teacher network is constructed to extract privileged observation representation and privileged observation phase saliency distribution, and an end inertia diagnosis network containing a motion phase condition weighted aggregation module is constructed; through category probability distillation, feature alignment, phase saliency guidance and condition consistency constraint, the fault sensitive knowledge in the privileged observation is migrated to the end inertia diagnosis network; in an inference stage, the privileged observation teacher network is frozen and discarded, and only the end inertia signal and the motion phase sequence are inputted to output the RV reducer health state or the fault category. The application realizes high-precision identification of the industrial robot RV reducer fault by using the deployable end inertia signal.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment fault diagnosis technology, specifically to a diagnostic method for an industrial robot RV reducer based on privileged observation distillation. Background Technology

[0002] RV reducers are critical transmission components in heavy-duty joints of industrial robots. Their transmission accuracy, torsional stiffness, and dynamic response directly affect the end-effector positioning accuracy, trajectory tracking performance, and operational safety of the industrial robot. When internal gears, cycloidal wheels, pinwheels, or bearings of the RV reducer experience faults such as broken teeth, cracks, pitting, peeling, or wear, the fault disturbances typically first manifest as highly sensitive observation responses at the reducer housing or adjacent transmission components. These disturbances are then transmitted through the joint structure, links, and end effector as attitude fluctuations, angular velocity disturbances, or motion deviations. Therefore, fault diagnosis of RV reducers in industrial robot joints differs from both ordinary rotating machinery bench diagnostics and overall machine condition monitoring based solely on robot controller state variables.

[0003] Existing diagnostic methods for RV reducers mainly include vibration signal analysis, multi-source sensor fusion, deep neural network classification, and health monitoring based on dynamic models. Signals such as vibration, sound, current, torque, and encoder errors near the reducer housing or transmission components typically have strong fault observability, facilitating the capture of gear meshing impacts, modulation sidebands, transmission errors, or load disturbances. However, long-term field deployment may require additional sensors, cables, acquisition modules, and maintenance interfaces, resulting in higher engineering implementation costs. End-effector inertial signals are easy to install and maintain, making them suitable for integration into industrial robot maintenance systems. However, fault information is significantly weakened after transmission through robot joints and links, and is affected by motion phase, attitude changes, load changes, and trajectory conditions, making it difficult for ordinary deep classifiers to reliably distinguish fine-grained gear faults.

[0004] Existing technologies propose abnormal diagnosis and factory evaluation standards for RV reducers based on vibration characteristics, pointing out that vibration characteristics are an important reference indicator reflecting the overall performance of RV reducers, and establishing vibration testing standards for RV40E reducers for manufacturing enterprises. This addresses the lack of clear vibration evaluation standards in production line factory testing. However, its technical approach mainly relies on fixed test benches and near-end vibration testing, failing to address the issue of weak observability of fault characteristics in the end-effector inertial signals after structural transmission in the assembled state of industrial robots. Furthermore, it does not provide a privileged observation distillation diagnostic scheme that utilizes strong fault observation information during the training phase and relies solely on deployable end-effector signals during the inference phase.

[0005] Furthermore, conventional multi-source fusion methods often require multiple sensor signals to be present simultaneously during both the training and inference phases. When it is not possible to collect strong fault observation signals such as shell vibration, sound, current, or torque over a long period in the field, such models are difficult to deploy directly. Conventional knowledge distillation methods typically use overall category probability or global feature alignment, failing to fully consider the differences in fault observability at different motion phases during the cyclic motion of industrial robots. During actual robot motion, certain phase positions are more likely to transmit internal fault disturbances of the RV reducer to the end-effector inertial signal, while other phase positions may be masked by structural flexibility, load changes, or attitude changes. Therefore, there is an urgent need for a method that can use privileged observation signals to guide the end-effector inertial diagnostic model learning during the training phase and complete the diagnosis solely based on the end-effector inertial signal and motion phase sequence during the inference phase. Summary of the Invention

[0006] To address the problems of existing technologies, this invention provides a diagnostic method for industrial robot RV reducers based on privileged observation distillation. During the training phase, privileged observation signals, such as reducer housing vibration, and end-effector inertial signals are simultaneously acquired as heterogeneous observation pairs. Privileged observation signals, which are highly sensitive to faults but inconvenient for long-term deployment during the inference phase, are used only during the training phase to construct a privileged observation teacher network. Through class probability distillation, feature alignment, and phase saliency guidance, the fault-sensitive knowledge and phase saliency structure learned by the privileged observation teacher network are transferred to an end-effector inertial diagnostic network that relies solely on the end-effector inertial signals. During the inference phase, the privileged observation teacher network is frozen and discarded, allowing reducer fault diagnosis to be completed using only the end-effector inertial signals and motion phase sequences.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a diagnostic method for RV reducers of industrial robots based on privileged observation distillation, comprising the following steps:

[0009] Collect privileged observation signals from the RV reducer of the industrial robot joint, robot end effector inertial signals, motion phase sequences, fault category labels, and operating condition labels to construct a training sample set;

[0010] A privileged observation teacher network is constructed, and a privileged observation encoder is used to extract features from the privileged observation signals to obtain a privileged observation phase feature sequence. The privileged observation phase saliency distribution is then generated by a phase saliency calculation head.

[0011] An end-effector inertial diagnostic network is constructed. The inertial encoder is used to extract features from the robot's end-effector inertial signals to obtain an inertial phase feature sequence. The inertial phase feature sequence and the motion phase sequence are then input into a motion phase conditional weighted aggregation module to obtain phase aggregation weights and end-effector inertial representation.

[0012] The end-point inertial diagnostic network is trained by distillation using the class probabilities, privileged observation representations, and privileged observation phase saliency distributions of the privileged observation teacher network outputs.

[0013] During the inference phase, the inertial encoder and motion phase condition weighted aggregation module are retained. The end-effector inertial representation of the industrial robot to be diagnosed is output by the end-effector classifier, and the category with the highest probability is taken as the health status or fault category of the RV reducer.

[0014] As a preferred technical solution, a data preprocessing step is also included, specifically:

[0015] For a single robot cycle motion or a diagnostic time window, the privileged observation signal and the end-effector inertial signal are timestamped, synchronized, phase-aligned, and normalized. The synchronization segment is based on the robot motion cycle, the reducer input shaft rotation cycle, the joint trajectory segment, or a fixed time window.

[0016] An alignment index is established based on the timestamp and motion phase relationship, synchronous segmentation is performed, and the signals of each channel are normalized using training set statistics to obtain a training sample set.

[0017] As a preferred technical solution, the privileged observation signal is extracted using a privileged observation encoder to obtain a privileged observation phase feature sequence, which is represented as follows:

[0018] ;

[0019] in, Indicates a privileged observation encoder. Indicates the first Privileged observation phase feature sequence of a sample, Indicates the first The sample at the th Privileged observation features at each phase position, Indicates the phase position index. Indicates the length of the phase sequence.

[0020] As a preferred technical solution, the privileged observation phase saliency distribution is generated by the phase saliency calculation head, specifically including:

[0021] A phase saliency calculation head is set on the privileged observation phase feature sequence, and the phase saliency calculation head calculates the saliency score for each phase position:

[0022] ;

[0023] in, Indicates the first The sample at the th The significance score of each phase position, The learnable weight vector representing the phase saliency calculation head. Representing vectors and The inner product;

[0024] Normalize all phase positions to obtain the privileged observation phase significance weights:

[0025] ;

[0026] in, Indicates the first The sample at the th Privileged observation phase significance weights for each phase position, This represents the natural exponential function. Indicates the summation index;

[0027] The privileged observation characterization is obtained through weighted aggregation:

[0028] ;

[0029] in, Indicates the first Privileged observational characterization of each sample.

[0030] As a preferred technical solution, an inertial encoder is used to extract features from the robot's end effector inertial signal to obtain an inertial phase feature sequence, represented as follows:

[0031] ;

[0032] in, Indicates an inertial encoder. Indicates the first The inertial phase feature sequence of each sample. Indicates the first The sample at the th Inertial characteristics of each phase position.

[0033] As a preferred technical solution, the inertial phase feature sequence and the motion phase sequence are input into the motion phase conditional weighted aggregation module to obtain the phase aggregation weight and the end-effector inertial characterization, specifically including:

[0034] Mapping motion phases to phase embeddings:

[0035] ;

[0036] in, Represents the phase embedding function, Indicates the first The sample at the th Phase embedding vectors at each phase position;

[0037] Calculate the phase observation score:

[0038] ;

[0039] in, Indicates the first The sample at the th Phase observation score at each phase position, , , and All represent learnable parameters; Represents the hyperbolic tangent function;

[0040] The phase aggregation weights are obtained by normalizing the phase observation scores:

[0041] ;

[0042] in, Indicates the first The sample at the th Phase aggregation weights for each phase position;

[0043] The end-effector inertia characterization is obtained by the following formula:

[0044] ;

[0045] in, Indicates the first One sample can be used to represent the end-effector inertia of inference diagnosis. express The inertial characteristics of the t-th phase position. This represents the phase aggregation weight.

[0046] As a preferred technical solution, the distillation training adopts training objectives including class probability distillation, feature alignment, phase saliency guidance, and operating condition consistency constraints.

[0047] As a preferred technical solution, category probability distillation includes: passing the classification outputs of the privileged observation teacher network and the end-point inertial diagnostic network through a normalized exponential function to obtain the teacher category probabilities. and terminal category probability ,

[0048] The class probability distillation loss constructed using relative entropy is:

[0049] ;

[0050] in, Represents the class probability distillation loss. Represents the relative entropy operator, Indicates the first The probability of teacher category for each sample. Indicates the first The terminal class probability of each sample;

[0051] Feature alignment represents the privileged observations of the i-th sample. and end-effector inertial characterization Enter the privileged observation projection head respectively and inertial projection head The privileged observation representation after projection is obtained. and the end-effector inertial characterization after projection Calculate the feature alignment loss:

[0052] ;

[0053] in, This represents the feature alignment loss. Indicates privileged observation of the projection head, Indicates inertial projection head, This represents the privileged observation representation after projection. This represents the end-effector inertia characterization after projection. This indicates that gradient calculation has stopped. Represents the L2 norm operation;

[0054] Phase saliency guidance utilizes privileged observations of phase saliency distribution Phase aggregation weights of constrained end-effector inertial diagnostic networks The phase-guided loss is constructed and calculated as follows:

[0055] ;

[0056] in, Indicates phase guidance loss; Indicates by The distribution of the saliency of the privileged observation phases constituted; Indicates by The phase aggregation weight distribution constituted;

[0057] Operating condition consistency constraints are expressed as follows:

[0058] ;

[0059] in, This indicates a loss of consistency in operating conditions. Indicates the total number of fault categories. Indicates the fault category index, Indicates the first The set of operating conditions included in a fault class, where c represents the index of the operating condition. Indicates the first Class of faults in Mean value of end-effector inertia characterization under various operating conditions Indicates the first Mean value of end-effector inertial characteristics for this type of fault under all operating conditions;

[0060] The total loss function of the final training objective is obtained by weighting and summing the category probability distillation loss, feature alignment loss, phase guidance loss, and working condition consistency loss.

[0061] As a preferred technical solution, the end-effector inertial characterization of the industrial robot to be diagnosed is processed by an end-effector classifier, which outputs probabilities for each category. The category with the highest probability is then used as the health status or fault category of the RV reducer, specifically as follows:

[0062] ;

[0063] in, Indicates the diagnostic results; Represents the end classifier; This indicates the end-effector inertial signal to be diagnosed; This represents the motion phase sequence to be diagnosed. End-effector characterization representing the end-effector inertial signal to be diagnosed; Indicates the end classifier for the first... The predicted probability of a class.

[0064] This invention also provides a diagnostic system for an industrial robot RV reducer based on privileged observation distillation, used to implement the above-mentioned diagnostic method for an industrial robot RV reducer based on privileged observation distillation, including: a data acquisition module, a privileged observation teacher network construction module, an end-effector inertial diagnostic network construction module, a distillation training module, and a diagnostic output module;

[0065] The data acquisition module is used to collect privileged observation signals of the RV reducer of the industrial robot joint, robot end effector inertial signals, motion phase sequences, fault category labels and operating condition labels to construct a training sample set;

[0066] The privileged observation teacher network construction module is used to construct a privileged observation teacher network, extract features from the privileged observation signal using a privileged observation encoder to obtain a privileged observation phase feature sequence, and generate a privileged observation phase saliency distribution using a phase saliency calculation head;

[0067] The end-effector inertial diagnostic network construction module is used to construct an end-effector inertial diagnostic network. It uses an inertial encoder to extract features from the robot's end-effector inertial signals to obtain an inertial phase feature sequence. The inertial phase feature sequence and the motion phase sequence are then input into a motion phase condition weighted aggregation module to obtain phase aggregation weights and end-effector inertial representation.

[0068] The distillation training module is used to perform distillation training on the end-point inertial diagnostic network using the class probabilities, privileged observation representations, and privileged observation phase saliency distributions output by the privileged observation teacher network.

[0069] The diagnostic output module is used to retain the inertial encoder and motion phase condition weighted aggregation module during the inference stage. The end-effector inertial representation of the industrial robot to be diagnosed is output by the end-effector classifier, and the category with the highest probability is taken as the health status or fault category of the RV reducer.

[0070] Compared with the prior art, the advantages and technical effects of the present invention are as follows:

[0071] (1) The training phase of this invention makes full use of the fault sensitivity of privileged observation signals, and the inference phase no longer needs to collect such privileged observation signals, thereby reducing the cost of on-site deployment and maintenance;

[0072] (2) The motion phase condition weighted aggregation module of the present invention can explicitly model the influence of robot motion phase on end fault observability, so that the end inertial diagnostic network can focus on the phase interval of fault information that is more easily propagated to the end.

[0073] (3) This invention uses the combined effects of class probability distillation, feature alignment and phase saliency guidance to transfer fault-sensitive knowledge in privileged observations to the end-point inertial diagnostic network, thereby improving diagnostic stability under weak observable conditions.

[0074] (4) The working condition consistency constraint of the present invention helps to reduce the model’s dependence on a single speed, load or trajectory condition and improve the ability to be applied across working conditions;

[0075] (5) This invention can be used for damage identification of key components of industrial robots, mapping of component and whole machine status, early warning of terminal accuracy degradation and intelligent operation and maintenance decision support, which is compatible with the requirements of component-whole machine level status perception and terminal accuracy early warning of industrial robots. Attached Figure Description

[0076] Figure 1 This is an overall flowchart of the industrial robot reducer diagnostic method based on privileged observation distillation according to the present invention. Detailed Implementation

[0077] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are only for illustrating the present invention and do not constitute a limitation on the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make equivalent substitutions or combinations for sensor type, sampling frequency, window length, number of network layers, feature dimension, loss weight, and fault category, and such substitutions or combinations should all fall within the scope of protection of the present invention.

[0078] like Figure 1 As shown, this embodiment provides a diagnostic method for RV reducers of industrial robots based on privileged observation distillation. Taking the diagnostic task of the joint RV reducer of a six-axis industrial robot as the object, it includes steps S1 to S5 in sequence. Steps S1 to S4 are executed in the training phase, and step S5 is executed in the inference and deployment phase.

[0079] Step S1: Construction of training samples;

[0080] This embodiment establishes an industrial robot reducer fault diagnosis test platform. The test platform is based on a BRTIRUS1510A six-axis industrial robot. An RV reducer is installed on the target diagnostic joint. During the training phase, an acceleration sensor is installed near the RV reducer housing of this joint to collect the vibration signal of the proximal housing as a privileged observation signal. An inertial measurement unit (IMU) is installed at the robot's end effector to collect end-effector inertial signals. .

[0081] In this embodiment, the privileged observation signal is an observation signal that can be collected during the training phase, is highly sensitive to faults, and is not a necessary input during the inference and deployment phase; the privileged observation signal includes one or more of the vibration signals installed on the RV reducer housing, joint housing, or adjacent transmission components, as well as one or more of the motor current signal, joint torque signal, encoder error signal, or sound signal;

[0082] In this embodiment, the end-effector inertial signal Includes one or more of attitude angle, angular velocity, and linear acceleration; motion phase sequence The joint angle, encoder angle, trajectory command, cycle normalized time or a combination thereof output by the robot controller are obtained and mapped to the phase position within the robot's reciprocating motion cycle, joint rotation cycle or reducer input shaft rotation cycle.

[0083] For a single robot cycle motion or a diagnostic time window, the privileged observation signal and the end-effector inertial signal are timestamped, synchronized, phase-aligned, and normalized. The synchronization segment is based on the robot motion cycle, the reducer input shaft rotation cycle, the joint trajectory segment, or a fixed time window.

[0084] Because the sampling frequencies of the privileged observation signal and the end-effector inertial signal are different, this embodiment first establishes an alignment index based on the timestamp and motion phase relationship, then performs synchronous segmentation based on the rotation period of the reducer input shaft, and subsequently normalizes the signals of each channel using training set statistics to obtain the training sample set:

[0085] ;

[0086] in, Represents the training sample set; Indicates the sample sequence number; Indicates the number of training samples; Indicates the first The privileged observation signal of each sample; Indicates the first The end-effector signal of each sample; Indicates the first The motion phase sequence of each sample; Indicates the first Fault category labels for each sample; Indicates the first Operating condition labels for each sample.

[0087] In this embodiment, there are a total of fault categories. The categories are: healthy condition, broken sun gear teeth, cracked planetary gear teeth, and broken planetary gear teeth; the operating condition is determined by the load, with a fixed speed of 2400 rpm and three load levels: 0 kg, 4.8 kg, and 9.6 kg, corresponding to the operating condition labels. It should be noted that the above categories and operating condition settings are only for illustrating this embodiment. The present invention is also applicable to the identification of cycloidal wheels, pin wheels, bearings, or combined fault conditions in RV reducers. The operating condition label can also be determined by speed, load, trajectory speed, trajectory amplitude, or a combination thereof.

[0088] Step S2: Construct the privileged observation teacher network, which consists of a privileged observation encoder, a phase saliency calculation head, and a privileged observation classifier;

[0089] A privileged observation teacher network is used to learn strong fault observation information during the training phase. Given a privileged observation signal... Privileged observation encoder Output phase-aligned privileged observation phase feature sequence:

[0090] ;

[0091] in, Indicates a privileged observation encoder; Indicates the first Privileged observation phase feature sequence of each sample; Indicates the first The sample at the th Privileged observation features at each phase position; Indicates the phase position index; Indicates the length of the phase sequence.

[0092] In this embodiment, the privileged observation signal is the near-end shell vibration signal, with a single window length of 2500 sampling points; the privileged observation encoder... A multi-scale residual one-dimensional convolutional network is employed, comprising a one-dimensional convolutional layer, a normalization layer, a nonlinear activation layer, two residual convolutional blocks, a global phase resampling layer, and a feature output layer. The one-dimensional convolutional layer has a kernel length of 7 and a stride of 2. Each residual convolutional block contains parallel one-dimensional convolutional branches with kernel lengths of 3, 5, and 7, used to extract fault-sensitive features at different time scales from privileged observation signals, such as gear meshing impact, modulation sidebands, broadband vibration response, and transmission errors. The global phase resampling layer aligns the feature sequence to the phase sequence length. The parameters described above are preferred settings for this embodiment and do not limit the scope of protection of this invention.

[0093] To obtain a fault-sensitive phase saliency distribution, a phase saliency calculation head is set up on the privileged observation phase feature sequence. The phase saliency calculation head first calculates the saliency score for each phase position:

[0094] ;

[0095] in, Indicates the first The sample at the th Significance score of each phase position; The learnable weight vector representing the phase saliency calculation head; Representing vectors and The inner product of these values ​​is then normalized across all phase positions to obtain the privileged observation phase saliency weights:

[0096] ;

[0097] in, Indicates the first The sample at the th Privileged observation phase significance weights for each phase position; Represents the natural exponential function; The summation index represents the privileged observation representation obtained through weighted aggregation:

[0098] ;

[0099] in, Indicates the first Privileged observation representation of each sample. In this embodiment, privileged observation representation... privileged observation classifier Output the privileged observation class probability and train the privileged observation teacher network using the cross-entropy loss function; where, This represents the privileged observation classifier. After the privileged observation teacher network is trained, its parameters are fixed and used to guide the end-effector inertial diagnostics network.

[0100] Step S3: Construct an end-effector inertial diagnostic network, which consists of an inertial encoder, a phase embedding module, a motion phase conditional weighted aggregation module, and an end-effector classifier;

[0101] Among them, the motion phase conditional weighted aggregation module takes inertial phase features and phase embedding as input, outputs phase aggregation weights normalized along the phase axis, and obtains the terminal inertial characterization by weighted summation of the inertial phase feature sequence. The phase aggregation weights are supervised by the privileged observation phase saliency distribution through phase saliency guided loss.

[0102] In this embodiment, the end-effector inertial diagnostic network is used for inference-stage diagnostics. Given an end-effector inertial signal... Inertial encoder Output inertial phase feature sequence:

[0103] ;

[0104] in, Indicates an inertial encoder; Indicates the first The inertial phase feature sequence of each sample; Indicates the first The sample at the th Inertial characteristics of each phase position.

[0105] In this embodiment, the end effector inertial signal is resampled for each motion cycle. Each phase point contains multiple channels for attitude angle and angular velocity; inertial encoder. To extract weak fault features transmitted through robot joints and links from end-effector inertial signals, a channel-gated temporal convolutional network is employed, comprising a channel-gated layer, a first dilated temporal convolutional layer, a second dilated temporal convolutional layer, a residual connection layer, and a normalized activation layer. The channel-gated layer suppresses noise channels and enhances fault-sensitive channels; the dilated temporal convolutional layer captures slow torsional responses and end-effector attitude perturbations across phases.

[0106] Since the fault information in the end-effector inertial signal is not uniformly visible throughout the entire motion cycle, this embodiment sets up a motion phase condition weighted aggregation module, which includes three sub-steps in sequence: phase embedding, phase observation score calculation, and weighted aggregation.

[0107] First, the motion phase is mapped to a phase embedding:

[0108] ;

[0109] in, Represents the phase embedding function; Indicates the first The sample at the th Phase embedding vectors at each phase position;

[0110] Then calculate the phase observation score:

[0111] ;

[0112] in, Indicates the first The sample at the th Phase observation scores for each phase position; , , and All represent learnable parameters; This represents the hyperbolic tangent function.

[0113] The phase aggregation weights are obtained by normalizing the phase observation scores:

[0114] ;

[0115] in, Indicates the first The sample at the th Phase aggregation weights for each phase position;

[0116] The end-effector inertia characterization is obtained by the following formula:

[0117] ;

[0118] in, Indicates the first One sample can be used to represent the end-effector inertia of inference diagnosis. express The inertial characteristics of the t-th phase position. This represents the phase aggregation weight.

[0119] The above aggregation process differs from ordinary global average pooling; its key lies in phase aggregation weights. Phase saliency distribution generated by privileged observation teacher network The guidance enables the end-effector inertial diagnostic network to learn which motion phases are more observable for faults.

[0120] Step S4: Privileged observation distillation training;

[0121] In order to transfer fault-sensitive knowledge from the privileged observation teacher network to the end-point inertial diagnostic network, this embodiment adopts training objectives including category probability distillation, feature alignment, phase saliency guidance, and condition consistency constraints.

[0122] Category probabilistic distillation is used to transfer category discrimination knowledge from privileged observation teacher networks to end-point inertial diagnostic networks.

[0123] Category probability distillation involves passing the classification outputs of the privileged observation teacher network and the terminal inertial diagnostic network through a normalized exponential function to obtain the teacher category probabilities. and terminal category probability The relative entropy is used to measure the difference between the teacher class probability and the terminal class probability. The class probability distillation loss is constructed using the relative entropy as follows:

[0124] ;

[0125] in, Represents the class probability distillation loss; Represents the relative entropy operator; Indicates the first The probability of teacher category for each sample; Indicates the first The terminal class probability of each sample.

[0126] Feature alignment is used to constrain the privileged observation representation and end-point inertia representation of the same sample to be close in the projection space, thereby bringing the privileged observation representation of the i-th sample closer together. and end-effector inertial characterization Enter the privileged observation projection head respectively and inertial projection head The privileged observation representation after projection is obtained. and the end-effector inertial characterization after projection :

[0127] ;

[0128] The feature alignment loss is calculated using the L2 distance with stopping gradient operator, constraining the projected end-point inertia representation to approximate the projected privileged observation representation. The feature alignment loss is specifically expressed as follows:

[0129] ;

[0130] in, Indicates feature alignment loss; Indicates privileged observation of the projection head; Indicates an inertial projection head; This represents the privileged observation representation after projection; This represents the end-effector inertial characterization after projection. This indicates that gradient calculation has stopped; This represents the L2 norm operation.

[0131] Phase saliency guidance is used to constrain the phase aggregation weights of the end-point inertial diagnostic network to make them close to the fault-sensitive phases identified by the privileged observation teacher network. The phase saliency guidance loss uses relative entropy to measure the difference between the phase aggregation weight distribution and the privileged observation phase saliency distribution.

[0132] Specifically, using privileged observation phase saliency distribution Phase aggregation weights of constrained end-effector inertial diagnostic networks The phase-guided loss is constructed and calculated as follows:

[0133] ;

[0134] in, Indicates phase guidance loss; Indicates by The distribution of the significance of the privileged observation phases; Indicates by The phase aggregation weight distribution constituted;

[0135] Operating condition consistency constraints are used to reduce the model's dependence on a single operating condition. For the same fault category... Under different operating conditions The mean value of the end-effector inertia characterization is calculated under various operating conditions (such as different rotational speeds, loads, trajectory speeds, or trajectory amplitudes), and is kept close to the overall mean value of the fault category under all operating conditions. The operating condition consistency constraint uses the L2 distance between the mean value of the end-effector inertia characterization of the same fault category under each operating condition and the overall mean value of the fault category. The operating condition consistency constraint is expressed as follows:

[0136] ;

[0137] in, This indicates a loss of consistency in operating conditions; Indicates the total number of fault categories; Indicates the fault category index; Indicates the first The set of operating conditions included in the fault class; c represents the operating condition index; Indicates the first Class of faults in Mean value of end-effector inertia characterization under each operating condition; Indicates the first The average end-effector inertia characteristic of this type of fault under all operating conditions. In this embodiment, and Each by satisfying and The samples and satisfying End inertial characterization of the sample The average is obtained.

[0138] The ultimate training objective is:

[0139] ;

[0140] in, Represents the total loss function; This represents the classification loss of the end-effector inertial diagnostics network; , , , These represent the non-negative weight coefficients of the category probability distillation loss, feature alignment loss, phase guidance loss, and condition consistency loss, respectively. In this embodiment, , , , The values ​​were set to 1.0, 0.5, 1.0, and 0.1 respectively. The model training adopted the adaptive moment estimation optimization algorithm with an initial learning rate of 0.001 that decayed with the number of training rounds. The batch size was 64, the number of training rounds was 100, and an early stopping strategy and validation set monitoring were set to improve training stability.

[0141] Step S5: Inference diagnosis process. The input of the inference stage only includes the end inertial signal and motion phase sequence, and no more privileged observation signals are required. It supports both K-type fault classification output and health-abnormal binary classification threshold discrimination based on the predicted probability of health category.

[0142] During the inference phase, privileged observation signals are no longer collected; the privileged observation teacher network is frozen and discarded, and only the inertial encoder is retained. Motion phase conditional weighted aggregation module and end classifier For the industrial robot to be diagnosed, only the end effector inertial signal is collected. And obtain the corresponding motion phase sequence ;Will Input inertial encoder Obtain the inertial phase feature sequence, and then compare the inertial phase feature sequence with... The end-effector inertial characterization is obtained by inputting motion phase condition weighted aggregation module. Finally, the end classifier Output the probability of each category, and take the category with the highest probability as the health status or fault category of the RV reducer:

[0143] ;

[0144] in, Indicates the diagnostic results; Represents the end classifier; This indicates the end-effector inertial signal to be diagnosed; This represents the motion phase sequence to be diagnosed. End-effector characterization representing the end-effector inertial signal to be diagnosed; Indicates the end classifier for the first... The predicted probability of a class. When this invention is used for binary classification of health and abnormality, the predicted probability of the health status class can be taken and compared with a judgment threshold. If the probability is lower than the threshold, it is judged as abnormal; otherwise, it is judged as healthy.

[0145] To verify the diagnostic effectiveness of this embodiment, data was collected and experiments were conducted on the BRTIRUS1510A industrial robot RV reducer test platform under three load conditions: healthy state, broken sun gear tooth, cracked planetary gear tooth, and broken planetary gear tooth, for four health states: healthy state, broken sun gear tooth, cracked planetary gear tooth, and broken planetary gear tooth. The samples were divided into training set, validation set, and test set in a 6:2:2 ratio. All comparison methods used the same data partitioning and training protocol. The diagnostic performance of the method in this embodiment and the comparison methods are shown in Table 1.

[0146] Table 1. Comparison of diagnostic performance between this embodiment and the comparative method.

[0147]

[0148] As shown in Table 1, the accuracy of a single diagnostic network trained solely on end-effector inertial signals is only 82.32%, while the conventional knowledge distillation method, which only incorporates overall category probability distillation, can only improve it to 83.94%. In contrast, the method in this embodiment achieves a diagnostic accuracy of 99.90% using only end-effector inertial signals and motion phase sequences for inference, representing an improvement of 17.58 percentage points compared to the single end-effector inertial diagnostic network and 15.96 percentage points compared to the conventional knowledge distillation method. This indicates that this embodiment, through the synergistic effect of the motion phase condition weighted aggregation module and category probability distillation, feature alignment, phase saliency guidance, and operating condition consistency constraints, can effectively transfer fault-sensitive knowledge from near-effector shell vibration signals to end-effector inertial representation.

[0149] To further verify the contribution of each module, an ablation experiment was conducted, and the results are shown in Table 2. The values ​​in the table are the mean and standard deviation of multiple independent repeated experiments.

[0150] Table 2 Ablation Experiment Results of This Example

[0151]

[0152] Table 2 shows that the performance degradation was most significant when the motion phase condition weighted aggregation module was removed, with the accuracy dropping from 98.65% to 86.81%. This indicates that the motion phase condition weighted aggregation module and the guidance of the phase aggregation weights by the privileged observation phase saliency distribution are the main sources of diagnostic performance in this embodiment. Removing category probability distillation and feature alignment, as well as removing the operating condition consistency constraint, also led to significant performance degradation, indicating that the above modules work synergistically. Further analysis shows that the end-effector single diagnostic network significantly confuses planetary gear faults with similar morphologies, namely planetary gear cracks and planetary gear tooth breakage. However, the method in this embodiment can significantly reduce this confusion when using only the end-effector inertial signal, indicating that this embodiment has indeed successfully transferred the fine-grained discrimination structure contained in the near-end shell vibration signal to the end-effector inertial characterization.

[0153] In summary, this embodiment, through privileged observation distillation and motion phase condition weighted aggregation mechanism, significantly improves the fault diagnosis accuracy and cross-condition robustness of RV reducers while retaining the convenience of deploying end-point inertial signals, and has strong engineering application value.

[0154] Example 2

[0155] This embodiment provides a diagnostic system for an industrial robot RV reducer based on privileged observation distillation, used to implement the diagnostic method for an industrial robot RV reducer based on privileged observation distillation in Embodiment 1 above. The system includes: a data acquisition module, a privileged observation teacher network construction module, an end-effector inertial diagnostic network construction module, a distillation training module, and a diagnostic output module.

[0156] In this embodiment, the data acquisition module is used to collect privileged observation signals of the RV reducer of the industrial robot joint, robot end effector inertial signals, motion phase sequences, fault category labels, and operating condition labels to construct a training sample set;

[0157] In this embodiment, the privileged observation teacher network construction module is used to construct the privileged observation teacher network, extract features from the privileged observation signal using the privileged observation encoder to obtain the privileged observation phase feature sequence, and generate the privileged observation phase saliency distribution using the phase saliency calculation head;

[0158] In this embodiment, the end-effector inertial diagnostic network construction module is used to construct the end-effector inertial diagnostic network. It uses an inertial encoder to extract features from the robot's end-effector inertial signals to obtain an inertial phase feature sequence. The inertial phase feature sequence and the motion phase sequence are then input into the motion phase condition weighted aggregation module to obtain the phase aggregation weight and the end-effector inertial representation.

[0159] In this embodiment, the distillation training module is used to perform distillation training on the end-point inertial diagnostic network using the class probabilities, privileged observation representations, and privileged observation phase saliency distributions output by the privileged observation teacher network.

[0160] In this embodiment, the diagnostic output module is used to retain the inertial encoder and motion phase condition weighted aggregation module during the inference stage. The end-effector inertial representation of the industrial robot to be diagnosed is output by the end-effector classifier, and the category with the highest probability is taken as the health status or fault category of the RV reducer.

Claims

1. A diagnostic method for an industrial robot RV reducer based on privileged observation distillation, characterized in that, Includes the following steps: Collect privileged observation signals from the RV reducer of the industrial robot joint, robot end effector inertial signals, motion phase sequences, fault category labels, and operating condition labels to construct a training sample set; A privileged observation teacher network is constructed, and a privileged observation encoder is used to extract features from the privileged observation signals to obtain a privileged observation phase feature sequence. The privileged observation phase saliency distribution is then generated by a phase saliency calculation head. An end-effector inertial diagnostic network is constructed. The inertial encoder is used to extract features from the robot's end-effector inertial signals to obtain an inertial phase feature sequence. The inertial phase feature sequence and the motion phase sequence are then input into a motion phase conditional weighted aggregation module to obtain phase aggregation weights and end-effector inertial representation. The end-point inertial diagnostic network is trained by distillation using the class probabilities, privileged observation representations, and privileged observation phase saliency distributions of the privileged observation teacher network outputs. During the inference phase, the inertial encoder and motion phase condition weighted aggregation module are retained. The end-effector inertial representation of the industrial robot to be diagnosed is output by the end-effector classifier, and the category with the highest probability is taken as the health status or fault category of the RV reducer.

2. The diagnostic method for industrial robot RV reducers based on privileged observation distillation according to claim 1, characterized in that, It also includes data preprocessing steps, specifically: For a single robot cycle motion or a diagnostic time window, the privileged observation signal and the end-effector inertial signal are timestamped, synchronized, phase-aligned, and normalized. The synchronization segment is based on the robot motion cycle, the reducer input shaft rotation cycle, the joint trajectory segment, or a fixed time window. An alignment index is established based on the timestamp and motion phase relationship, synchronous segmentation is performed, and the signals of each channel are normalized using training set statistics to obtain a training sample set.

3. The diagnostic method for industrial robot RV reducers based on privileged observation distillation according to claim 1, characterized in that, The privileged observation signal is used to extract features from the privileged observation signal using a privileged observation encoder, resulting in a privileged observation phase feature sequence, which is represented as follows: ； in, Indicates a privileged observation encoder. Indicates the first Privileged observation phase feature sequence of a sample, Indicates the first The sample at the th Privileged observation features at each phase position, Indicates the phase position index. Indicates the length of the phase sequence.

4. The diagnostic method for industrial robot RV reducers based on privileged observation distillation according to claim 1, characterized in that, The privileged observation phase significance distribution is generated from the phase significance calculation head, specifically including: A phase saliency calculation head is set on the privileged observation phase feature sequence, and the phase saliency calculation head calculates the saliency score for each phase position: ； in, Indicates the first The sample at the th The significance score of each phase position, The learnable weight vector representing the phase saliency calculation head. Representing vectors and The inner product; Normalize all phase positions to obtain the privileged observation phase significance weights: ； in, Indicates the first The sample at the th Privileged observation phase significance weights for each phase position, This represents the natural exponential function. Indicates the summation index; The privileged observation characterization is obtained through weighted aggregation: ； in, Indicates the first Privileged observational characterization of each sample.

5. The diagnostic method for industrial robot RV reducers based on privileged observation distillation according to claim 1, characterized in that, The inertial phase feature sequence is obtained by extracting features from the robot's end effector inertial signal using an inertial encoder, as shown below: ； in, Indicates an inertial encoder. Indicates the first The inertial phase feature sequence of each sample. Indicates the first The sample at the th Inertial characteristics of each phase position.

6. The diagnostic method for industrial robot RV reducers based on privileged observation distillation according to claim 1, characterized in that, The inertial phase feature sequence and the motion phase sequence are input into the motion phase conditional weighted aggregation module to obtain the phase aggregation weight and the end-effector inertial characterization, specifically including: Mapping motion phases to phase embeddings: ； in, Represents the phase embedding function, Indicates the first The sample at the th Phase embedding vectors at each phase position; Calculate the phase observation score: ； in, Indicates the first The sample at the th Phase observation score at each phase position, , , and All represent learnable parameters; Represents the hyperbolic tangent function; The phase aggregation weights are obtained by normalizing the phase observation scores: ； in, Indicates the first The sample at the th Phase aggregation weights for each phase position; The end-effector inertia characterization is obtained by the following formula: ； in, Indicates the first One sample can be used to represent the end-effector inertia of inference diagnosis. express The inertial characteristics of the t-th phase position. This represents the phase aggregation weight.

7. The diagnostic method for industrial robot RV reducers based on privileged observation distillation according to claim 1, characterized in that, The distillation training employs training objectives including class probability distillation, feature alignment, phase saliency guidance, and operating condition consistency constraints.

8. The diagnostic method for industrial robot RV reducers based on privileged observation distillation according to claim 7, characterized in that, Category probability distillation involves passing the classification outputs of the privileged observation teacher network and the terminal inertial diagnostic network through a normalized exponential function to obtain the teacher category probabilities. and terminal category probability , The class probability distillation loss constructed using relative entropy is: ； in, Represents the class probability distillation loss. Represents the relative entropy operator, Indicates the first The probability of teacher category for each sample. Indicates the first The terminal class probability of each sample; Feature alignment represents the privileged observations of the i-th sample. and end-effector inertial characterization Enter the privileged observation projection head respectively and inertial projection head The privileged observation representation after projection is obtained. and the end-effector inertial characterization after projection Calculate the feature alignment loss: ； in, This represents the feature alignment loss. Indicates privileged observation of the projection head, Indicates inertial projection head, This represents the privileged observation representation after projection. This represents the end-effector inertia characterization after projection. This indicates that gradient calculation has stopped. Represents the L2 norm operation; Phase saliency guidance utilizes privileged observations of phase saliency distribution Phase aggregation weights of constrained end-effector inertial diagnostic networks The phase-guided loss is constructed and calculated as follows: ； in, Indicates phase guidance loss; Indicates by The distribution of the significance of the privileged observation phases; Indicates by The phase aggregation weight distribution constituted; Operating condition consistency constraints are expressed as follows: ； in, This indicates a loss of consistency in operating conditions. Indicates the total number of fault categories. Indicates the fault category index, Indicates the first The set of operating conditions included in a fault class, where c represents the index of the operating condition. Indicates the first Class of faults in Mean value of end-effector inertia characterization under various operating conditions Indicates the first Mean value of end-effector inertial characteristics for this type of fault under all operating conditions; The total loss function of the final training objective is obtained by weighting and summing the category probability distillation loss, feature alignment loss, phase guidance loss, and working condition consistency loss.

9. The diagnostic method for industrial robot RV reducers based on privileged observation distillation according to claim 1, characterized in that, The end-effector inertial representation of the industrial robot to be diagnosed is processed by the end-effector classifier, which outputs probabilities for each category. The category with the highest probability is taken as the health status or fault category of the RV reducer, specifically as follows: ； in, Indicates the diagnostic results; Represents the end classifier; This indicates the end-effector inertial signal to be diagnosed; This represents the motion phase sequence to be diagnosed. End-effector characterization representing the end-effector inertial signal to be diagnosed; Indicates the end classifier for the first... The predicted probability of a class.

10. A diagnostic system for an industrial robot RV reducer based on privileged observation distillation, characterized in that, The method for diagnosing an industrial robot RV reducer based on privileged observation distillation as described in any one of claims 1-9 includes: a data acquisition module, a privileged observation teacher network construction module, an end-effector inertial diagnosis network construction module, a distillation training module, and a diagnostic output module. The data acquisition module is used to collect privileged observation signals of the RV reducer of the industrial robot joint, robot end effector inertial signals, motion phase sequences, fault category labels and operating condition labels to construct a training sample set; The privileged observation teacher network construction module is used to construct a privileged observation teacher network, extract features from the privileged observation signal using a privileged observation encoder to obtain a privileged observation phase feature sequence, and generate a privileged observation phase saliency distribution using a phase saliency calculation head; The end-effector inertial diagnostic network construction module is used to construct an end-effector inertial diagnostic network. It uses an inertial encoder to extract features from the robot's end-effector inertial signals to obtain an inertial phase feature sequence. The inertial phase feature sequence and the motion phase sequence are then input into a motion phase condition weighted aggregation module to obtain phase aggregation weights and end-effector inertial representation. The distillation training module is used to perform distillation training on the end-point inertial diagnostic network using the class probabilities, privileged observation representations, and privileged observation phase saliency distributions output by the privileged observation teacher network. The diagnostic output module is used to retain the inertial encoder and motion phase condition weighted aggregation module during the inference stage. The end-effector inertial representation of the industrial robot to be diagnosed is output by the end-effector classifier, and the category with the highest probability is taken as the health status or fault category of the RV reducer.

Images