Robot fault diagnosis method and device fusing fault mechanism and uncertainty neural network, equipment and medium

Abstract

The invention discloses a robot fault diagnosis method and device fusing a fault mechanism and an uncertainty neural network, equipment and a medium, and relates to the technical field of fault diagnosis, and the method comprises the steps: obtaining the fault type of a flexible bearing of a harmonic reducer; obtaining an effective descending depth according to the fault type; according to the effective descending depth, the contact deformation amount between the rolling body and the raceway is determined; obtaining electromagnetic torque fluctuation of the motor according to the contact deformation; on the basis of the electromagnetic torque fluctuation of the motor, parameter uncertainty interference in a current signal transmission path is introduced, and a simulation current signal is generated; training a probabilistic fault diagnosis network by using the simulation current signal; and obtaining a current signal to be diagnosed, inputting the current signal to be diagnosed into the trained probabilistic fault diagnosis network, and outputting a fault diagnosis result. The method can improve the fault diagnosis accuracy of the harmonic reducer.

Description

Technical Field

[0001] This application relates to the field of fault diagnosis technology, and in particular to a method, apparatus, equipment and medium for robot fault diagnosis that integrates fault mechanism and uncertainty neural network. Background Technology

[0002] Harmonic reducers, with their outstanding advantages of high transmission accuracy, large transmission ratio, and compact structure, occupy a dominant position in high-end equipment fields such as modern industrial robots, precision CNC machine tools, and aerospace. However, under long-term high load, variable working conditions, and complex environments, their main component, the flexible bearing, is prone to localized faults such as radial indentation of the outer raceway, inner race wear, and rolling element cracks. These faults directly affect the positioning accuracy and motion smoothness of the actuator, and in severe cases, can even lead to machine shutdown or safety accidents. Therefore, conducting high-precision fault diagnosis research on harmonic reducers is of great significance for ensuring the reliability and safety of industrial robot systems. Current signals, due to their advantages of non-invasive acquisition and easy sensor installation, have become an important monitoring method in the field of harmonic reducer fault diagnosis.

[0003] However, existing technical solutions have low accuracy in diagnosing faults in harmonic reducers. Summary of the Invention

[0004] This application provides a robot fault diagnosis method, device, equipment, and medium that integrates fault mechanism and uncertainty neural network, which can improve the accuracy of fault diagnosis of harmonic reducers.

[0005] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application provides a robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks, including: Obtain the fault type of the flexible bearing in the harmonic reducer; Based on the fault type, the effective descent depth is obtained; The amount of contact deformation between the rolling element and the raceway is determined based on the effective descent depth. The electromagnetic torque fluctuation of the motor is obtained based on the contact deformation amount; Based on the electromagnetic torque fluctuation of the motor, parameter uncertainty interference in the current signal transmission path is introduced to generate a simulated current signal. The probabilistic fault diagnosis network was trained using the simulated current signal. The system acquires the current signal to be diagnosed and inputs it into a trained probabilistic fault diagnosis network, then outputs the fault diagnosis result.

[0006] Optionally, the step of inputting the current signal to be diagnosed into the trained probabilistic fault diagnosis network and outputting the fault diagnosis result includes: The current signal to be diagnosed is input into the trained probabilistic fault diagnosis network, and multiple random forward propagation samples are performed to obtain multiple predicted probability distributions. Calculate the average prediction entropy value based on the multiple prediction probability distributions; If the average predicted entropy value is greater than or equal to a preset threshold, the fault type corresponding to the current signal to be diagnosed is determined to be an unknown fault and is rejected. If the average prediction entropy value is less than a preset threshold, the fault category corresponding to the highest probability is determined based on the multiple prediction probability distributions output by the trained probabilistic fault diagnosis network, and this fault category is used as the fault diagnosis result.

[0007] Optionally, the fault type includes outer ring fault, inner ring fault, or rolling element fault of the flexible bearing; obtaining the effective descent depth based on the fault type includes: When the fault type is a flexible bearing outer ring fault, the first effective descent depth is obtained; When the fault type is a flexible bearing inner ring fault, a second effective descent depth is obtained; When the fault type is a flexible bearing rolling element fault, a third effective descent depth is obtained.

[0008] Optionally, the preset threshold is determined in the following way: A preset threshold is obtained based on the mean and variance of the predicted entropy of known fault samples.

[0009] Optionally, training the probabilistic fault diagnosis network using the simulated current signal includes: The probabilistic fault diagnosis network is pre-trained in the first stage using simulated current signals and corresponding fault labels to learn the prior distribution of fault features. The pre-trained network is fine-tuned in the second stage using real current signals to update the posterior distribution of the network parameters.

[0010] Optionally, the probabilistic fault diagnosis network is a Transformer network based on Bayesian layer reconstruction, wherein the deterministic linear mapping in at least one of the self-attention mechanism layer, feedforward neural network layer and classification output layer in the Transformer network is replaced with a Bayesian linear layer.

[0011] Optionally, the uncertainty parameter interference includes at least one of the following: uncertainty of flux linkage parameters, rotor angle measurement error, current sensor measurement noise, and statistical distribution characteristics of harmonic amplitude and phase.

[0012] Secondly, this application provides a robot fault diagnosis device that integrates fault mechanisms and uncertainty neural networks, comprising: The acquisition module is used to acquire the fault type of the flexible bearing in the harmonic reducer; The processing module is used to obtain the effective descent depth according to the fault type; determine the contact deformation between the rolling element and the raceway according to the effective descent depth; and obtain the electromagnetic torque fluctuation of the motor according to the contact deformation. The training module is used to introduce parameter uncertainty interference in the current signal transmission path based on the electromagnetic torque fluctuation of the motor to generate a simulated current signal; and to train the probabilistic fault diagnosis network using the simulated current signal. The output module is used to acquire the current signal to be diagnosed, input the current signal to be diagnosed into the trained probabilistic fault diagnosis network, and output the fault diagnosis result.

[0013] Thirdly, this application provides a computing device, including a memory and a processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of the first aspects.

[0014] Fourthly, this application provides a computer-readable storage medium for storing a computer program for performing the method as described in any one of the first aspects.

[0015] As can be seen from the above technical solution, this application has at least the following beneficial effects: In this application, by constructing effective descent depth models under different fault types, the contact deformation of the rolling element when passing through the fault location is accurately quantified, thereby obtaining the electromagnetic torque fluctuation of the motor. Based on this, multi-source parameter uncertainty interference in the current signal transmission path is further introduced to generate a high-fidelity simulated current signal. This method recreates the electromechanical coupling effect caused by the fault at the physical mechanism level, effectively solving the problem of scarce actual fault data and diffuse sample distribution. It provides training samples rich in fault characteristics and close to real-world operating conditions for subsequent diagnostic models, fundamentally improving the quality and diversity of input data for diagnostic models.

[0016] Secondly, addressing the limitations of existing deep learning models in quantifying signal uncertainty and exhibiting poor robustness in feature extraction, this solution constructs a Transformer probabilistic fault diagnosis network based on Bayesian layer reconstruction. By replacing the deterministic linear mappings in the self-attention mechanism layer, feedforward layer, and classification layer with Bayesian linear layers, the network gains end-to-end quantization and transmission capabilities for uncertain components in the input signal. When facing complex operating conditions such as load fluctuations and noise interference, this network effectively suppresses the random dispersion of feature distribution, stably extracts fault features, and improves the diagnostic accuracy for known fault types.

[0017] Finally, this scheme introduces an active rejection mechanism for unknown faults based on predicted entropy values, building upon the probabilistic diagnostic network. By performing multiple random forward propagation samplings on the input samples, the average predicted entropy value is calculated, and a dynamic rejection threshold is set. When the entropy value exceeds the threshold, it is determined to be an unknown fault and rejected; only when it falls below the threshold is the specific fault category output. This method endows the diagnostic system with the ability to assess its own cognitive boundaries, effectively avoiding the model's forced classification of non-closed-set data and overconfident output. While ensuring the accuracy of known fault diagnosis, it improves the system's safety and reliability in real, complex industrial environments.

[0018] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0019] Figure 1 A flowchart illustrating a robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks, provided as an embodiment of this application; Figure 2 A schematic diagram of a robot fault diagnosis device that integrates fault mechanism and uncertainty neural network provided in an embodiment of this application; Figure 3 This is a schematic diagram of a computing device provided in an embodiment of this application. Detailed Implementation

[0020] The terms "first," "second," and "third," etc., used in this application specification and accompanying drawings are used to distinguish different objects, not to limit a specific order.

[0021] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0022] To ensure clarity and conciseness in the description of the following embodiments, a brief introduction to the related technologies is given first: Harmonic reducers are precision transmission devices used in high-end equipment such as industrial robots, precision CNC machine tools, and aerospace. With their advantages of high transmission accuracy, large transmission ratio, and compact structure, they have become components of high-end equipment actuators. Their operating status directly determines the positioning accuracy, motion smoothness, and working reliability of the equipment. The core components include wave generators, flexible wheels, rigid wheels, and flexible bearings.

[0023] Flexible bearings are vulnerable components of harmonic reducers. They are thin-walled elastic bearings nested between the wave generator and the flexible wheel. They generate elastic deformation as the wave generator rotates, realizing the meshing transmission between the flexible wheel and the rigid wheel. Under long-term high load, variable working conditions and complex environments, they are prone to local faults such as radial concavity of the outer raceway, wear of the inner raceway, and cracks in the rolling elements. They are the monitoring objects for fault diagnosis of harmonic reducers.

[0024] The main technical problem addressed in this application is that existing robot fault diagnosis schemes that integrate fault mechanisms and uncertain neural networks have low diagnostic accuracy, making it difficult to meet the fault diagnosis needs of high-end equipment such as industrial robots and precision CNC machine tools. Specifically, the accuracy of identifying known faults such as outer ring faults, inner ring faults, and rolling element faults of flexible bearings is insufficient. At the same time, when faced with unknown faults not covered by the training set, forced classification and high-confidence misjudgments are prone to occur, which cannot provide reliable diagnostic support for the safe operation of harmonic reducers, and may even lead to equipment shutdown and safety accidents due to misdiagnosis.

[0025] The main causes of the aforementioned technical problems lie in the combined effects of multiple factors. First, existing technologies do not establish a full-link quantitative correlation between fault types and current signal characteristics based on the physical mechanism of flexible bearing failures. Furthermore, the generated simulated current signals do not incorporate parameter uncertainty interference in the transmission path, resulting in a disconnect from the signal characteristics of real operating conditions and distorting the fault representation from the data source. Second, traditional diagnostic models are deterministic network structures, lacking the ability to quantify and transmit uncertain components in current signals. They also lack the ability to extract interference signal features under complex operating conditions and lack targeted hierarchical training strategies, resulting in poor model generalization ability. Third, the diagnostic system lacks an effective mechanism for identifying unknown faults and has no quantitative indicators to measure the uncertainty of model judgments, making it impossible to accurately reject unknown faults. Ultimately, the overall diagnostic accuracy and reliability fail to meet actual industrial needs.

[0026] In view of this, embodiments of this application provide a robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks, which can be executed by a processing device. This processing device can be a terminal or a server. Terminals include, but are not limited to, smartphones, tablets, laptops, personal digital assistants, or smart wearable devices. Servers can be cloud servers, such as central servers in a central cloud computing cluster or edge servers in an edge cloud computing cluster. Alternatively, servers can be located in a local data center. A local data center refers to a data center directly controlled by the user.

[0027] This application addresses the low accuracy of existing robot fault diagnosis methods that integrate fault mechanisms and uncertain neural networks. First, it establishes an intrinsic correlation between fault physical characteristics and electromagnetic signals by deriving effective descent depth, contact deformation, and motor electromagnetic torque fluctuations based on fault type. Next, it introduces interference from parameter uncertainties in the current signal transmission path to generate simulated current signals, ensuring the simulation data closely reflects real industrial conditions. Then, it employs a Transformer probabilistic fault diagnosis network based on Bayesian layer reconstruction, combined with a hierarchical training strategy using both simulated and real signals, enabling the network to quantify, transmit, and extract features from uncertain current signals. Finally, it introduces a predictive entropy value discrimination mechanism to diagnose known faults and effectively reject unknown faults, forming a complete technical system from fault mechanism modeling to diagnostic decision-making. Ultimately, this improves the accuracy and reliability of harmonic reducer fault diagnosis, meeting the fault diagnosis needs of high-end equipment.

[0028] To make the technical solution of this application clearer and easier to understand, the following description, in conjunction with the accompanying drawings, introduces a robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks, as provided in the embodiments of this application. Figure 1As shown, this figure is a flowchart of a robot fault diagnosis method that integrates fault mechanisms and uncertainty neural networks, according to an embodiment of this application. The method includes: S201, Processing equipment obtains the fault type of the flexible bearing of the harmonic reducer.

[0029] The fault type refers specifically to the typical local fault form of the flexible bearing in the harmonic reducer in this application. The fault type includes outer ring fault, inner ring fault, or rolling element fault of the flexible bearing.

[0030] Flexible bearing outer ring failure refers to various faults occurring in the raceway of the outer ring of the flexible bearing in a harmonic reducer. It is one of the typical fault types of flexible bearings, manifested as radial concavity of the outer ring raceway, and may also include derivative faults such as wear and scratches on the outer ring raceway. The outer ring of the flexible bearing is tightly bonded to the flexible wheel of the harmonic reducer and its tangential position is fixed. Faults in this part will cause the fault width to change periodically as the wave generator rotates, causing the radial contact state of the rolling element to fluctuate regularly when it passes the fault position. This directly affects the radial contact force transmission of the flexible bearing, and in turn, causes periodic fluctuations in the electromagnetic torque of the motor.

[0031] Flexible bearing inner ring failure refers to various failures occurring in the inner ring raceway of the flexible bearing in a harmonic reducer. Typical manifestations include wear and radial depression of the inner ring raceway, and may also include cracks and pitting of the inner ring raceway. The inner ring of the flexible bearing is directly nested with the elliptical wave generator and rotates synchronously with the wave generator. The location of the inner ring failure is fixed relative to the inner ring body, and there is no dynamic change in the failure width. When the rolling element revolves to the location of the inner ring failure, the radial contact deformation will change abruptly due to the physical defects of the raceway, which will break the normal contact force transmission law and cause abnormal fluctuations in the electromagnetic torque of the motor.

[0032] Rolling element failure in flexible bearings refers to various failures occurring in the rolling element components, manifested as cracks and radial indentations in the rolling elements, as well as wear and chipping. The rolling elements are the moving parts of the flexible bearing, simultaneously revolving around the bearing center and rotating around their own center. There is no fixed fault contact location for this part; the timing of contact between the fault point and the inner / outer raceway must be determined based on the revolution and rotation angles of the rolling elements. When the fault point contacts the raceway, a sudden change in radial contact deformation occurs, which in turn causes irregular fluctuations in the electromagnetic torque of the motor. This is the most complex type of fault in terms of contact state among the three types of faults.

[0033] The processing equipment uses pre-collected operating data of the harmonic reducer, fault sample annotations, or typical fault categories of flexible bearings determined through fault mechanism analysis to obtain the specific fault type (outer ring fault, inner ring fault, or rolling element fault) of the flexible bearing that needs to be quantitatively analyzed and diagnosed. After obtaining the fault type, the processing equipment uses it as the input basis for subsequent derivation of fault quantification parameters such as effective drop depth and contact deformation, laying the foundation for establishing a full-link correlation between fault physical characteristics, motor electromagnetic torque fluctuations, and current signal characteristics. This is the starting point for realizing the transformation from abstract fault types to specific quantitative characteristics.

[0034] S202. The processing equipment obtains the effective descent depth based on the fault type.

[0035] The effective drop depth characterizes the actual indentation displacement in the radial direction when the rolling elements come into contact with the faulty parts (outer ring raceway, inner ring raceway, and rolling element body) after a failure in a flexible bearing. It is a parameter that quantifies the physical characteristics of flexible bearing failures. Different failure types correspond to different effective drop depth calculation methods. It is a quantitative indicator that connects the mechanical failure of flexible bearings with subsequent electrical signal fluctuations.

[0036] Specifically, when the fault type is a fault in the outer ring of the flexible bearing, a first effective descent depth is obtained; when the fault type is a fault in the inner ring of the flexible bearing, a second effective descent depth is obtained; and when the fault type is a fault in the rolling element of the flexible bearing, a third effective descent depth is obtained.

[0037] After obtaining the specific fault type, the processing device will trigger a preset classification calculation logic to match the fault type, the specific calculation model, and the effective descent depth. When a fault type of flexible bearing outer ring fault is identified, the processing equipment calls the effective descent depth calculation model specific to outer ring faults. Combining parameters such as the initial position angle of the outer ring fault, the fault width, and the change in the radius of curvature caused by the rotation of the wave generator, the effective descent depth for this type of fault is obtained through numerical calculation and defined as the first effective descent depth. The calculation expression for the first effective descent depth is:

[0038]

[0039]

[0040]

[0041] in, This represents the first effective descent depth of the outer ring fault in the flexible bearing at time t, expressed in millimeters (mm), and indicates the radial indentation displacement caused by the rolling element passing through the fault location. Indicates the radius of the rolling element of a flexible bearing. This indicates the outer ring pitch circle radius of the flexible bearing. This indicates the revolution angle of the rolling element, expressed in radians (rad). This indicates the dynamic fault width after curvature radius correction. This represents the real-time angle of the fault location relative to the initial position after the wave generator rotates. Indicates fault width The corresponding angle range, This means that taking the modulus of the revolution angle of the rolling element by 2π yields its equivalent angle in the range of 0 to 2π. This indicates that the wave generator has rotated to an angle. The radius of curvature correction factor at that time. This indicates the original width of the outer ring fault in the flexible bearing. and The major and minor axes of the elliptical profile of the wave generator are represented. This indicates the initial position angle of a fault in the outer ring of a flexible bearing. Indicates the angular velocity of the wave generator. Represents a time variable.

[0042] When a fault type of flexible bearing inner ring fault is identified, the processing equipment calls the inner ring fault-specific calculation model, and calculates the effective descent depth by combining parameters such as the inner ring fault location and the revolution angle of the rolling element. The calculated effective descent depth is defined as the second effective descent depth; the calculation expression for the second effective descent depth is:

[0043] in, This indicates the second effective descent depth, i.e., the effective descent depth for inner circle faults. This indicates the radius of the inner pitch circle.

[0044] When a fault type of flexible bearing rolling element failure is identified, the processing equipment calls upon a rolling element-specific calculation model. Combining parameters such as the rolling element's rotation angle, revolution angle, and the contact position relationship between the fault point and the inner and outer rings, the resulting effective descent depth is defined as the third effective descent depth. The expression for calculating the third effective descent depth is:

[0045] in, This indicates the third effective descent depth, i.e., the effective descent depth of a rolling element failure. Indicates the rotation angle of the rolling element at the point of failure. Indicates the revolution angle of the rolling element. Represents an integer variable, used to represent integers that are periodically contacted. Indicates the range of fault contact angles.

[0046]

[0047] in, Indicates the effective descent depth.

[0048] This classification rule is crucial for quantifying the physical characteristics of faults. It avoids distortion of the effective descent depth caused by a one-size-fits-all approach and ensures that the calculation process of the processing equipment is targeted and standardized. Different fault types correspond to fixed calculation models and result definitions, guaranteeing the accuracy and consistency of input parameters for subsequent derivation steps such as contact deformation and electromagnetic torque fluctuation. Simultaneously, this step builds upon the preceding fault type acquisition and provides standardized quantitative parameters for the subsequent unified derivation of contact deformation. This creates a standardized and accurate closed loop in the derivation chain of fault type, effective descent depth, and contact deformation, serving as a vital intermediate link in establishing the correlation between mechanical faults and electrical signals.

[0049] S203. The processing equipment determines the amount of contact deformation between the rolling element and the raceway based on the effective descent depth.

[0050] Rolling elements are the moving parts of flexible bearings. Driven by a wave generator, they simultaneously revolve and rotate, serving as the carriers for transmitting radial contact forces. Their contact state with the raceway directly determines the operating characteristics of the flexible bearing.

[0051] The raceway is the annular contact surface between the inner and outer rings of a flexible bearing, providing a motion track for the rolling elements. After a failure occurs, the depression or wear on the raceway surface will change the contact position and stress state of the rolling elements.

[0052] Contact deformation is the radial elastic deformation generated when the rolling element contacts the raceway under the influence of the effective descent depth. It is an intermediate parameter connecting the physical characteristics of the fault and the subsequent electromagnetic torque fluctuation, and directly reflects the degree of influence of the fault on the transmission of contact force.

[0053] After obtaining the effective descent depth corresponding to the fault type, the processing equipment uses it as input, and combines it with the contact stiffness, contact damping and geometric parameters (such as rolling element radius, raceway curvature radius, etc.) of the flexible bearing, calls the preset contact deformation calculation model, and determines the contact deformation between the rolling element and the raceway through numerical calculation.

[0054] The formula for calculating contact deformation is:

[0055] in, Indicates the amount of contact deformation. Indicates the radial displacement of the inner ring. Indicates the radial displacement of the outer ring. Indicates the initial clearance of the bearing. Indicates the effective descent depth.

[0056] The logic behind this step is that the effective descent depth changes the initial contact position between the rolling element and the raceway, which in turn causes a change in the amount of contact deformation. The processing equipment quantifies this change, transforming the abstract fault indentation into a contact deformation response that can participate in dynamic modeling.

[0057] S204. The processing equipment obtains the electromagnetic torque fluctuation of the motor based on the amount of contact deformation.

[0058] The electromagnetic torque of a motor is the torque generated by the interaction between the stator magnetic field and the rotor magnetic field of the motor. It is the power source for driving the wave generator of the harmonic reducer to rotate, and its fluctuation directly reflects the impact of flexible bearing failure on the transmission system.

[0059] Electromagnetic torque fluctuation is the periodic or non-periodic change in the electromagnetic torque of a motor that deviates from its rated value under the influence of contact force changes caused by a failure of a flexible bearing. It is a manifestation of the fault characteristics being transmitted from the mechanical domain to the electrical domain.

[0060] After acquiring the contact deformation, the processing equipment uses it as input, and combines it with the contact stiffness and contact damping of the flexible bearing, as well as the transmission parameters such as the average radial distance and equivalent pressure angle of the harmonic reducer. It then calls upon the preset electromechanical coupling dynamics model and sequentially solves the contact mechanics equation, torque transmission equation, and motor torque balance equation to obtain the curve of the motor electromagnetic torque changing with time. Based on this, by comparing it with the fault-free reference torque, the fluctuation of the motor electromagnetic torque is extracted.

[0061] When a flexible bearing fails, the amount of contact deformation It will directly cause contact force The change in is calculated using the following formula:

[0062] in, This represents the contact force between the rolling elements and the raceway of a flexible bearing. Indicates contact stiffness, Indicates contact damping, This represents the rate of change of contact deformation.

[0063] The cam of the flexible bearing is embedded in the inner ring of the flexible bearing, and outputs a radial contact force outward during rotation. For the drive motor, this contact force is converted into input torque through the transmission mechanism, driving the wave generator of the harmonic reducer to rotate. The calculation formula is as follows:

[0064] in, This represents the torque input to the wave generator. This represents the contact force between the rolling elements and the raceway of a flexible bearing. Indicates the average radial distance. This represents the equivalent pressure angle, expressed in radians.

[0065] The wave generator of the harmonic reducer is directly driven by the drive motor, therefore the electromagnetic torque of the motor provides the input torque for the wave generator. The torque balance formula constructed using the lumped parameter method is as follows:

[0066] in, This represents the electromagnetic torque of the motor. This represents the moment of inertia of the motor shaft. This represents the angular acceleration of the motor rotor. This represents the damping coefficient of the motor shaft. This represents the angular velocity of the motor rotor. This represents the torque input to the wave generator.

[0067] Obtaining the electromagnetic torque of the motor Then, the rated electromagnetic torque of the motor under fault-free operating conditions can be obtained first. Using the reference torque as a reference, the difference between the electromagnetic torque and the reference torque at each moment under fault conditions is calculated to obtain the torque deviation. Finally, the torque deviation is analyzed in the time domain or frequency domain to extract its amplitude, frequency and phase characteristics, so as to obtain the electromagnetic torque fluctuation law of the motor. This fluctuation law directly reflects the impact of flexible bearing failure on the transmission system.

[0068] S205. The processing equipment introduces parameter uncertainty interference in the current signal transmission path based on the electromagnetic torque fluctuation of the motor to generate a simulated current signal.

[0069] Parameter uncertainty interference refers to various non-ideal factors existing in the current signal transmission path in industrial scenarios. Uncertain parameter interference includes at least one of the following: uncertainty of flux linkage parameters, rotor angle measurement error, current sensor measurement noise, and statistical distribution characteristics of harmonic amplitude and phase.

[0070] Uncertainty in flux linkage parameters refers to the fact that the flux linkage of permanent magnets is not a constant value under actual operating conditions, but rather fluctuates randomly within a certain range around the rated value due to factors such as temperature, demagnetization, and manufacturing tolerances. This fluctuation alters the mapping relationship between the motor's electromagnetic torque and stator current, causing a deviation between the ideal current model output and the actual current, and is the main source of uncertainty at the motor body level.

[0071] Rotor angle measurement error refers to the deviation between the measured value and the true value when position sensors such as encoders and rotary transformers measure the rotor angle of a motor, caused by factors such as installation eccentricity, noise, and quantization errors. This error directly affects the decoupling accuracy of the d / q axis current, leading to deviations in current control and electromagnetic torque calculation.

[0072] Current sensor measurement noise refers to the random noise superimposed on the actual current when the current sensor acquires the stator current signal. It can usually be modeled as zero-mean Gaussian white noise. Noise sources include thermal noise from the sensor's internal circuitry, electromagnetic interference (EMI), and interference during signal transmission. These noises can cause high-frequency jitter in the acquired current signal, affecting the accuracy of fault feature extraction.

[0073] The statistical distribution characteristics of harmonic amplitude and phase refer to the presence of various harmonics in the stator current of a motor, in addition to the fundamental wave, caused by factors such as inverter nonlinearity, cogging effect, and load imbalance. The amplitude and phase of these harmonics are not fixed values, but follow a certain statistical distribution (such as normal distribution or uniform distribution), and will change randomly under different operating conditions, resulting in distortion of the current waveform and increasing the complexity of the signal.

[0074] Simulated current signals refer to current signal samples generated by processing equipment based on electromagnetic torque fluctuations and superimposed with parameter uncertainty interference, which closely match the real industrial operating conditions and are used to train probabilistic fault diagnosis networks.

[0075] After acquiring the electromagnetic torque fluctuation of the motor, the processing equipment uses it as input and combines it with the motor's electromagnetic parameters (such as the number of pole pairs, permanent magnet flux linkage, stator resistance, inductance, etc.). It then calls a preset motor electromagnetic model to convert the electromagnetic torque fluctuation into an ideal three-phase stator current signal. Subsequently, based on this ideal current signal, various parameter uncertainties in the current signal transmission path (such as flux linkage fluctuation, measurement noise, harmonic distortion, etc.) are introduced. A simulated current signal is generated through numerical superposition or statistical sampling to ensure that the distribution characteristics of the simulated data are consistent with the current signals collected in real industrial scenarios.

[0076] The logic behind this step is as follows: the ideal current signal is derived from electromagnetic torque fluctuations through an electromagnetic model. However, there are many interference factors in real industrial scenarios. Diagnostic models trained directly using ideal signals will have insufficient generalization ability in practical applications. The processing equipment introduces parameter uncertainty interference to make the simulated current signal closer to the real working conditions, thereby improving the robustness and accuracy of the subsequent diagnostic network.

[0077] In the simulation of permanent magnet synchronous motors (PMSMs), the uncertainties of the following parameters are considered and are all modeled as normally distributed: Uncertainty in flux linkage parameters:

[0078] in, Let represent the flux linkage of a permanent magnet, be a random variable following a normal distribution, and characterize the uncertainty of the flux linkage parameters. This indicates the nominal value of the magnetic flux linkage, which is the rated reference value for the magnetic flux linkage. It represents the standard deviation of magnetic flux, reflecting fluctuations caused by factors such as temperature and aging.

[0079] Rotor angle measurement error:

[0080] in, The rotor angle measurement value is represented by a random variable that follows a normal distribution, characterizing the uncertainty in the angle measurement. This indicates the nominal value of the rotor angle, which is the rated reference value for the angle. It represents the standard deviation of angle measurement error, reflecting the degree of angle deviation caused by encoder calibration error, quantization error, etc.

[0081] Current sensor measurement noise:

[0082] in, The noise in the current sensor measurement is a random process that follows a zero-mean normal distribution, characterizing the uncertainty in the sensor measurement. It represents the standard deviation of measurement noise, reflecting the degree of measurement error caused by sensor bias, circuit noise, electromagnetic interference, etc.

[0083] Statistical distribution of harmonic amplitude and phase:

[0084] in, Let represent the amplitude of the nth harmonic, and be a random variable following a normal distribution. This represents the mean amplitude of the nth harmonic. The standard deviation of the amplitude of the nth harmonic reflects the degree of amplitude fluctuation. Let represent the phase of the nth harmonic, and be a random variable following a normal distribution. This represents the mean value of the phase of the nth harmonic. It represents the standard deviation of the phase of the nth harmonic, reflecting the degree of phase fluctuation.

[0085] For permanent magnet synchronous motors, the following is adopted: The vector control strategy, considering inverter PWM modulation, dead-time effect and harmonic components, constructs the phase current model as follows:

[0086] in, This represents the phase a current signal, i.e., the final generated simulated current. This represents the electromagnetic torque of the motor at time t. Indicates the number of pole pairs of the motor. Indicates from the normal distribution The magnetic flux sample value obtained from the sampling process, Indicates the angular velocity of the motor rotor. Indicates from the normal distribution The initial angle sample value obtained from the sampling process, Indicates from the normal distribution The sample value of the nth harmonic amplitude obtained from the sampling process. Indicates from the normal distribution The nth harmonic phase sample value obtained from the sampling process. Indicates from the normal distribution The measurement noise sample value obtained from the sampling process, Represents a time variable.

[0087] S206. The processing equipment uses simulated current signals to train the probabilistic fault diagnosis network.

[0088] Probabilistic fault diagnosis networks are intelligent diagnostic models based on deep learning or probabilistic models. They can extract features and make probabilistic inferences from input current signals and output probability distributions of different fault types and severity, rather than a single deterministic diagnostic result. They have stronger uncertainty handling capabilities and robustness.

[0089] The probabilistic fault diagnosis network is a Transformer network reconstructed from Bayesian layers, in which the deterministic linear mapping in at least one of the self-attention mechanism layer, feedforward neural network layer and classification output layer in the Transformer network is replaced with a Bayesian linear layer.

[0090] Bayesian layer reconstruction refers to replacing the deterministic weight layers (such as linear layers) in traditional deep learning networks with Bayesian linear layers, so that the weights are no longer fixed values, but random variables that follow a specific prior distribution (such as Gaussian distribution), thereby quantifying the uncertainty of the model.

[0091] Transformer networks are a deep learning architecture based on self-attention mechanisms. By capturing long-range dependencies in input sequences (such as current signals), they enable efficient feature extraction from complex time-series data and perform well in fields such as fault diagnosis and natural language processing.

[0092] The self-attention mechanism layer is a component of Transformer. By calculating the attention weights between different positions in the sequence, it dynamically focuses on the main features and effectively captures fault modes in current signals.

[0093] The feedforward neural network layer is a fully connected network in the Transformer used to perform nonlinear transformations on the output of the self-attention layer, further enhancing the feature representation capability.

[0094] The classification output layer is the final layer of the network, which maps the extracted features to predictions of fault type or severity.

[0095] Deterministic linear mapping is a linear transformation with fixed weights in traditional neural networks, which cannot quantify the uncertainty of the model itself.

[0096] The Bayesian linear layer models the weights of the linear layer as random variables that follow a prior distribution. The posterior distribution of the weights is obtained through Bayesian inference, thereby outputting prediction results with probability confidence.

[0097] In the probabilistic fault diagnosis network of this application, a Transformer network architecture based on Bayesian layer reconstruction is adopted. Its innovation lies in replacing at least one layer in the standard Transformer network—the self-attention mechanism layer, the feedforward neural network layer, and the classification output layer—with a Bayesian linear layer instead of a traditional deterministic linear mapping. The logic of this reconstruction is as follows: The weights of traditional Transformer networks are fixed deterministic parameters, and their prediction results lack quantification of the uncertainty of the model itself. When faced with complex noise and interference in industrial scenarios, they are prone to giving overconfident but unreliable diagnostic conclusions.

[0098] By introducing a Bayesian linear layer, the network weights are modeled as random variables, the distribution of which is determined by both prior assumptions and training data. During the inference phase, the network not only outputs the fault diagnosis result but also provides the probability confidence level of that result, thereby enabling a quantitative assessment of diagnostic uncertainty.

[0099] This reconstruction approach makes the diagnostic network more robust and interpretable, enabling it to accurately identify flexible bearing faults under complex operating conditions and provide reliable probabilistic basis for operation and maintenance decisions.

[0100] This application proposes a Transformer fault diagnosis architecture based on improved Bayesian inference, which uses a full-link probabilistic strategy to transform the input sequence (in For sequence length, The model dimensions are sequentially fed into the Transformer architecture containing Bayesian layers to achieve the propagation of uncertainty in features.

[0101] The self-attention layer is used to compute queries. ,key ,value The linear transformations are all implemented using Bayesian linear layers:

[0102] in, This represents the query vector matrix, used to calculate attention weights with the key vector. The input sequence feature matrix has dimensions L×d and represents the feature sequence of the simulated current signal after preprocessing. This represents a key vector matrix used to calculate attention weights with the query vector. This represents a value vector matrix. Attention weights are applied to this matrix to obtain the final attention output. These are the Bayesian linear layer weight matrices corresponding to the query, key, and value, respectively. They are random variables following a Gaussian distribution, not fixed values. These are the Bayesian linear layer bias vectors corresponding to the query, key, and value, respectively, and are also random variables that follow a Gaussian distribution.

[0103] The output of the attention layer then enters the feedforward network for nonlinear transformation. This layer also uses a Bayesian linear layer and incorporates the ReLU activation function to enhance feature representation capabilities.

[0104] in, This represents the output feature matrix of the feedforward network. This represents the input feature matrix of the feedforward network, i.e., the output of the self-attention layer. Let represent the weight matrices of the two Bayesian linear layers in the feedforward network, and let be random variables following a Gaussian distribution. Let represent the bias vectors of two Bayesian linear layers in the feedforward network, which are also random variables that follow a Gaussian distribution.

[0105] The final feature vector output by the encoder is fed into the Bayesian classification layer. Through a probabilistic fully connected mapping, the feature vector is mapped to the fault category space. Due to the random sampling characteristics of the weights, the model not only outputs the predicted fault category but also the uncertainty of each prediction. This uncertainty provides the theoretical basis for subsequent identification of unknown faults.

[0106] Specifically, the probabilistic fault diagnosis network is pre-trained in the first stage using simulated current signals and corresponding fault labels to learn the prior distribution of fault features. The pre-trained network is fine-tuned in the second stage using real current signals to update the posterior distribution of the network parameters.

[0107] Fault labels are annotation information that corresponds one-to-one with simulated current signals, clearly indicating the fault type (such as outer race fault, inner race fault, rolling element fault) and severity of the signal, and serve as supervision signals for network training.

[0108] The first stage of pre-training refers to the process of initially training the probabilistic fault diagnosis network using large-scale simulated current signals and corresponding fault labels. The goal is to enable the network to learn the prior distribution of fault features based on mechanism and establish a basic mapping relationship between fault modes and current signals.

[0109] The prior distribution of fault characteristics refers to the probabilistic cognition of fault characteristics learned by the network through simulation data before it comes into contact with real data. It reflects the statistical regularity of fault characteristics under ideal operating conditions and serves as the knowledge basis for subsequent fine-tuning.

[0110] The real current signal refers to the current signal collected by the sensor during the actual operation of the harmonic reducer. It includes noise, interference and unknown fault characteristics under real working conditions and is used for network fine-tuning.

[0111] The second stage of fine-tuning refers to the process of updating the parameters of the pre-trained network using real current signals. The core objective is to adapt the network to the uncertainty of real working conditions, update the posterior distribution of fault features, and improve the diagnostic robustness in real-world scenarios.

[0112] The posterior distribution of network parameters refers to the probability distribution of parameters updated by Bayesian inference after the network comes into contact with real data. It integrates the prior knowledge of simulation data and the observation information of real data, and is closer to the fault characteristics and laws under real working conditions.

[0113] In the robot fault diagnosis method integrating fault mechanism and uncertainty neural network in this application, the network training adopts a two-stage learning strategy to balance fault mechanism cognition and adaptability to real working conditions: In the first stage of pre-training, the processing equipment uses large-scale simulated current signals and corresponding fault labels to initially train the probabilistic fault diagnosis network. Although simulated signals cannot completely reproduce real-world scenarios, they contain core fault evolution mechanism information and have the advantages of absolutely accurate labels and a sufficient sample size. Through pre-training, the network learns the prior distribution of fault features based on mechanism understanding, establishes a basic mapping relationship between fault modes and current signals, and forms an initial understanding of faults.

[0114] The second stage, fine-tuning, involves the processing device fine-tuning the network using current signals collected under real-world operating conditions after pre-training. This stage uses the prior distribution obtained from pre-training as initial cognition and updates the posterior distribution of network parameters with observational information from real data. This allows the network to accurately capture the unique uncertainties in real-world operating conditions while retaining fault knowledge from the simulation domain, avoiding overfitting and ultimately improving the model's diagnostic robustness in real-world scenarios.

[0115] First-stage pre-training loss function: In this stage, the model is pre-trained using a simulation dataset. The optimization objective is to maximize the variational lower bound, and the corresponding loss function is:

[0116] in, This represents the total loss value of the first stage of pre-training. This represents the expectation operation based on the variational posterior distribution. This represents the variational posterior distribution of the model parameters. This represents the set of Bayesian parameters of the model. This represents the log-likelihood loss (reconstruction error). This represents the simulated current signal input sample. This indicates the fault category label corresponding to the simulation sample. This represents the loss balance hyperparameter. This represents the Körbek-Leibler divergence (KL divergence). This represents the initial prior distribution of the model parameters.

[0117] Second stage: Fine-tuning the loss function: In this stage, fine-tuning is performed using a real experimental dataset. The variational posterior distribution learned in the simulation domain in the first stage is used as the new prior distribution for training in the real domain, and the loss function is adjusted as follows:

[0118] in, This represents the total loss value for the second-stage fine-tuning. This represents the actual experimental current signal input sample. This represents the loss balance hyperparameter. This represents the variational posterior distribution obtained from training in the simulation domain.

[0119] S207. The processing device acquires the current signal to be diagnosed, inputs the current signal to be diagnosed into the trained probabilistic fault diagnosis network, and outputs the fault diagnosis result.

[0120] Specifically, the current signal to be diagnosed is input into a trained probabilistic fault diagnosis network, and multiple random forward propagation samples are performed to obtain multiple predicted probability distributions.

[0121] The purpose of this process is to take advantage of the fact that the weights in the probabilistic fault diagnosis network follow a probability distribution, to perform multiple random samplings of the network weights during the inference phase, and to perform a complete forward propagation calculation after each sampling, so as to obtain a set of independent fault category prediction probability distributions.

[0122] Specifically, since the weights of the Bayesian linear layers in the network are not fixed values, but random variables that follow a specific posterior distribution, the network randomly samples a set of parameters from the posterior distribution of the weights during each forward propagation. Based on this set of parameters, the network performs feature extraction and classification prediction on the input current signal to be diagnosed, and obtains a set of probability assignments for various types of faults. By repeating this random sampling and forward propagation process N times, N independent prediction probability distributions can be obtained.

[0123] The differences between these distributions directly reflect the model's uncertainty about the current diagnostic results: when multiple predicted probability distributions are highly concentrated in a certain fault category, it indicates that the model has a high degree of confidence in the diagnostic conclusion, the fault characteristics are clear, and the interference is small; when multiple distributions are relatively dispersed and fluctuate between multiple fault categories, it indicates that the model has a large degree of uncertainty about the current diagnostic results, which may be due to unclear fault characteristics, strong operating condition interference, or the existence of unknown fault modes.

[0124] In this way, not only can fault diagnosis conclusions be obtained, but the reliability of the diagnosis results can also be quantified, providing maintenance personnel with a more comprehensive basis for decision-making, effectively avoiding the risk of misjudgment that may be caused by a single deterministic diagnosis, and improving the robustness and interpretability of the fault diagnosis system in complex industrial scenarios.

[0125] Secondly, the processing device calculates the average prediction entropy value based on multiple prediction probability distributions.

[0126] During the inference phase, the processing device first obtains multiple predicted probability distributions through multiple Monte Carlo samplings. Then, it calculates the average predicted entropy value based on these distributions to quantify the uncertainty of the diagnosis. Specifically, the processing device first averages the predicted probability vectors obtained from multiple samplings for each category, obtaining an average predicted probability vector. This vector integrates the results of multiple predictions by the model, reflecting the model's overall confidence tendency towards the current input. Next, it calculates the average predicted entropy value based on this average probability vector. The magnitude of the entropy value directly reflects the uncertainty of the diagnosis: when the input is a known fault type, the multiple prediction results are highly consistent, and the average probability vector will be significantly concentrated in a certain category. In this case, the entropy value is extremely low, indicating high diagnostic confidence. When the input is an unknown fault or the features are not obvious, the multiple prediction results diverge drastically, and the average probability vector tends to be uniformly distributed. In this case, the entropy value increases significantly, indicating high diagnostic uncertainty.

[0127] The formula for calculating the average prediction entropy is:

[0128] in, Represents the average predicted probability vector. This indicates the total number of samples taken in Monte Carlo. Let represent the predicted probability vector for the m-th sample.

[0129] Based on the average prediction probability vector Calculate the average prediction entropy value as a quantitative indicator of the model's uncertainty regarding the current input:

[0130] in, This represents the average predicted entropy value. This represents the total number of known fault categories. This represents the probability value of class c in the average predicted probability vector.

[0131] If the average prediction entropy value is greater than or equal to the preset threshold, the fault type corresponding to the current signal to be diagnosed is determined to be an unknown fault and rejected; if the average prediction entropy value is less than the preset threshold, the fault category corresponding to the highest probability is determined according to the multiple prediction probability distributions output by the trained probabilistic fault diagnosis network, and the fault category is taken as the fault diagnosis result.

[0132] The preset threshold is determined in the following way: A preset threshold is obtained based on the mean and variance of the predicted entropy of known fault samples.

[0133] After the probabilistic fault diagnosis network completes two-stage training, a batch of known fault samples with clearly defined fault types are selected. For each known fault sample, the same inference process as the current signal to be diagnosed is executed: multiple random forward propagation samples are input into the trained probabilistic fault diagnosis network to calculate the average prediction entropy value of each known fault sample. After obtaining the average prediction entropy values ​​of all known fault samples, these entropy values ​​are statistically analyzed to calculate the mean and variance of the average prediction entropy values ​​of all known fault samples. The mean reflects the average uncertainty level of the model when diagnosing known faults, and the variance reflects the degree of uncertainty fluctuation when the model diagnoses different known fault samples. The calculated mean and variance are then substituted into the preset threshold calculation expression, combined with an adjustable coefficient, to finally calculate the preset threshold. This preset threshold is not arbitrarily set but is based on the actual diagnostic uncertainty statistical characteristics of known fault samples. It matches the uncertainty distribution of the model under known fault modes, providing a scientific and reliable critical judgment standard for subsequently distinguishing between known and unknown faults and achieving the rejection of unknown faults.

[0134] The formula for calculating the preset threshold is:

[0135] in, Indicates the preset threshold. This represents the mean predicted entropy of known faulty samples. This represents the predicted entropy variance of known fault samples. This represents the adjustable coefficient.

[0136] If the average predicted entropy of the current signal to be diagnosed is greater than or equal to a preset threshold, it indicates that the model's diagnostic result for the signal is highly uncertain, the probability distribution of multiple predictions is relatively scattered, and a reliable judgment cannot be formed. In this case, the model determines that the fault corresponding to the signal is an unknown fault and rejects it, avoiding misclassifying unknown faults into known fault modes, thereby improving the robustness of the diagnostic system.

[0137] If the average prediction entropy is less than a preset threshold, it indicates that the model's diagnosis of the signal is relatively certain, and the probability distribution of multiple predictions is highly concentrated. In this case, the model finds the fault category with the highest average probability from multiple prediction probability distributions and uses it as the final fault diagnosis result, ensuring the accuracy of diagnosis under known fault modes.

[0138] The advantage of this decision-making logic is that it can not only accurately identify known faults, but also effectively identify and reject unknown faults, avoiding the problem of blind confidence and misclassification in the face of unknown operating conditions in traditional deterministic diagnostic models, and providing maintenance personnel with a more reliable and comprehensive basis.

[0139] Based on the above description, this application has the following beneficial effects: In this application, by constructing effective descent depth models under different fault types, the contact deformation of the rolling element when passing through the fault location is accurately quantified, thereby obtaining the electromagnetic torque fluctuation of the motor. Based on this, multi-source parameter uncertainty interference in the current signal transmission path is further introduced to generate a high-fidelity simulated current signal. This method recreates the electromechanical coupling effect caused by the fault at the physical mechanism level, effectively solving the problem of scarce actual fault data and diffuse sample distribution. It provides training samples rich in fault characteristics and close to real-world operating conditions for subsequent diagnostic models, fundamentally improving the quality and diversity of input data for diagnostic models.

[0140] Secondly, addressing the limitations of existing deep learning models in quantifying signal uncertainty and exhibiting poor robustness in feature extraction, this solution constructs a Transformer probabilistic fault diagnosis network based on Bayesian layer reconstruction. By replacing the deterministic linear mappings in the self-attention mechanism layer, feedforward layer, and classification layer with Bayesian linear layers, the network gains end-to-end quantization and transmission capabilities for uncertain components in the input signal. When facing complex operating conditions such as load fluctuations and noise interference, this network effectively suppresses the random dispersion of feature distribution, stably extracts fault features, and improves the diagnostic accuracy for known fault types.

[0141] Finally, this scheme introduces an active rejection mechanism for unknown faults based on predicted entropy values, building upon the probabilistic diagnostic network. By performing multiple random forward propagation samplings on the input samples, the average predicted entropy value is calculated, and a dynamic rejection threshold is set. When the entropy value exceeds the threshold, it is determined to be an unknown fault and rejected; only when it falls below the threshold is the specific fault category output. This method endows the diagnostic system with the ability to assess its own cognitive boundaries, effectively avoiding the model's forced classification of non-closed-set data and overconfident output. While ensuring the accuracy of known fault diagnosis, it improves the system's safety and reliability in real, complex industrial environments.

[0142] The above text combined Figure 1 The robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks provided in the embodiments of this application has been described in detail. The apparatus and equipment provided in the embodiments of this application will be described below with reference to the accompanying drawings.

[0143] like Figure 2 As shown in the figure, this is a schematic diagram of a robot fault diagnosis device that integrates fault mechanisms and uncertainty neural networks according to an embodiment of this application. The device includes: The acquisition module 301 is used to acquire the fault type of the flexible bearing of the harmonic reducer; The processing module 302 is used to obtain the effective descent depth according to the fault type; determine the contact deformation amount between the rolling element and the raceway according to the effective descent depth; and obtain the electromagnetic torque fluctuation of the motor according to the contact deformation amount. Training module 303 is used to introduce parameter uncertainty interference in the current signal transmission path based on the electromagnetic torque fluctuation of the motor to generate a simulated current signal; and to train the probabilistic fault diagnosis network using the simulated current signal. The output module 304 is used to acquire the current signal to be diagnosed, input the current signal to be diagnosed into the trained probabilistic fault diagnosis network, and output the fault diagnosis result.

[0144] Optionally, the output module 304 is specifically used to input the current signal to be diagnosed into the trained probabilistic fault diagnosis network, perform multiple random forward propagation sampling, and obtain multiple predicted probability distributions. Calculate the average prediction entropy value based on the multiple prediction probability distributions; If the average predicted entropy value is greater than or equal to a preset threshold, the fault type corresponding to the current signal to be diagnosed is determined to be an unknown fault and is rejected. If the average prediction entropy value is less than a preset threshold, the fault category corresponding to the highest probability is determined based on the multiple prediction probability distributions output by the trained probabilistic fault diagnosis network, and this fault category is used as the fault diagnosis result.

[0145] Optionally, the processing module 302 is specifically used to obtain a first effective descent depth when the fault type is a flexible bearing outer ring fault; When the fault type is a flexible bearing inner ring fault, a second effective descent depth is obtained; When the fault type is a flexible bearing rolling element fault, a third effective descent depth is obtained.

[0146] Optionally, the processing module 302 is specifically used to obtain a preset threshold based on the mean and variance of the predicted entropy of known fault samples.

[0147] Optionally, the training module 303 is specifically used to perform a first-stage pre-training of the probabilistic fault diagnosis network using simulated current signals and corresponding fault labels to learn the prior distribution of fault features. The pre-trained network is fine-tuned in the second stage using real current signals to update the posterior distribution of the network parameters.

[0148] The robot fault diagnosis device that integrates fault mechanisms and uncertain neural networks according to the embodiments of this application can correspondingly execute the method described in the embodiments of this application, and the other operations and / or functions of each module / unit of the robot fault diagnosis device that integrates fault mechanisms and uncertain neural networks are respectively for implementing Figure 1 For the sake of brevity, the corresponding processes of each method in the illustrated embodiments will not be described in detail here.

[0149] This application also provides a computing device. For example... Figure 3 As shown in the figure, this is a schematic diagram of a computing device provided in an embodiment of this application. The computing device 700 includes a bus 701, a processor 702, a communication interface 703, and a memory 704. The processor 702, the memory 704, and the communication interface 703 communicate with each other via the bus 701.

[0150] The 701 bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0151] The processor 702 can be any one or more of the following processors: central processing unit (CPU), graphics processing unit (GPU), microprocessor (MP), or digital signal processor (DSP).

[0152] The communication interface 703 is used for communication with external devices.

[0153] Memory 704 may include volatile memory, such as random access memory (RAM). Memory 704 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0154] The memory 704 stores executable code, and the processor 702 executes the executable code to perform the aforementioned robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks.

[0155] Specifically, in achieving Figure 2 In the case of the illustrated embodiment, and Figure 2 When the modules or units of the robot fault diagnosis device that integrates fault mechanisms and uncertain neural networks described in the embodiments are implemented in software, the following steps are performed: Figure 2 The software or program code required for the functions of each module / unit can be partially or entirely stored in memory 704. Processor 702 executes the program code corresponding to each unit stored in memory 704, and executes the aforementioned robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks.

[0156] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a computing device can store, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the computing device to execute the aforementioned robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks.

[0157] This application also provides a computer program product comprising one or more computer instructions. When the computer instructions are loaded and executed on a computing device, all or part of the processes or functions described in this application are generated.

[0158] The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, or data center to another website, computer, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0159] When the computer program product is executed by a computer, the computer performs any of the aforementioned methods of the robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks. The computer program product can be a software installation package; when any of the aforementioned methods of the robot fault diagnosis method that integrates fault mechanisms and uncertain neural networks needs to be used, the computer program product can be downloaded and executed on the computer.

[0160] The descriptions of the processes or structures corresponding to the above figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0161] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be covered within the scope of protection of this application.

Claims

1. A robot fault diagnosis method integrating fault mechanism and uncertainty neural network, characterized in that, The method includes: Obtain the fault type of the flexible bearing in the harmonic reducer; Based on the fault type, the effective descent depth is obtained; The amount of contact deformation between the rolling element and the raceway is determined based on the effective descent depth. The electromagnetic torque fluctuation of the motor is obtained based on the contact deformation amount; Based on the electromagnetic torque fluctuation of the motor, parameter uncertainty interference in the current signal transmission path is introduced to generate a simulated current signal. The probabilistic fault diagnosis network is trained using the simulated current signal. The system acquires the current signal to be diagnosed and inputs it into a trained probabilistic fault diagnosis network, then outputs the fault diagnosis result.

2. The method according to claim 1, characterized in that, The process of inputting the current signal to be diagnosed into the trained probabilistic fault diagnosis network and outputting the fault diagnosis result includes: The current signal to be diagnosed is input into the trained probabilistic fault diagnosis network, and multiple random forward propagation samples are performed to obtain multiple predicted probability distributions. Calculate the average prediction entropy value based on the multiple prediction probability distributions; If the average predicted entropy value is greater than or equal to a preset threshold, the fault type corresponding to the current signal to be diagnosed is determined to be an unknown fault and is rejected. If the average prediction entropy value is less than a preset threshold, the fault category corresponding to the highest probability is determined based on the multiple prediction probability distributions output by the trained probabilistic fault diagnosis network, and this fault category is used as the fault diagnosis result.

3. The method according to claim 1, characterized in that, The fault types include outer ring fault, inner ring fault, or rolling element fault of the flexible bearing; obtaining the effective descent depth based on the fault type includes: When the fault type is a flexible bearing outer ring fault, the first effective descent depth is obtained; When the fault type is a flexible bearing inner ring fault, a second effective descent depth is obtained; When the fault type is a flexible bearing rolling element fault, a third effective descent depth is obtained.

4. The method according to claim 2, characterized in that, The preset threshold is determined in the following way: The preset threshold is obtained based on the mean and variance of the predicted entropy of known fault samples.

5. The method according to claim 1, characterized in that, The step of training the probabilistic fault diagnosis network using the simulated current signal includes: The probabilistic fault diagnosis network is pre-trained in the first stage using simulated current signals and corresponding fault labels to learn the prior distribution of fault features. The pre-trained network is fine-tuned in the second stage using real current signals to update the posterior distribution of the network parameters.

6. The method according to claim 1, characterized in that, The probabilistic fault diagnosis network is a Transformer network based on Bayesian layer reconstruction, wherein the deterministic linear mapping in at least one of the self-attention mechanism layer, feedforward neural network layer and classification output layer in the Transformer network is replaced with a Bayesian linear layer.

7. The method according to claim 1, characterized in that, The uncertain parameter interference includes at least one of the following: uncertainty of flux linkage parameters, rotor angle measurement error, current sensor measurement noise, and statistical distribution characteristics of harmonic amplitude and phase.

8. A robot fault diagnosis device integrating fault mechanism and uncertainty neural network, characterized in that, The device includes: The acquisition module is used to acquire the fault type of the flexible bearing in the harmonic reducer; The processing module is used to obtain the effective descent depth according to the fault type; determine the contact deformation between the rolling element and the raceway according to the effective descent depth; and obtain the electromagnetic torque fluctuation of the motor according to the contact deformation. The training module is used to introduce parameter uncertainty interference in the current signal transmission path based on the electromagnetic torque fluctuation of the motor to generate a simulated current signal; and to train the probabilistic fault diagnosis network using the simulated current signal. The output module is used to acquire the current signal to be diagnosed, input the current signal to be diagnosed into the trained probabilistic fault diagnosis network, and output the fault diagnosis result.

9. A computing device, characterized in that, Including memory and processor; The memory stores one or more computer programs, the one or more computer programs including instructions; when the instructions are executed by the processor, the computing device performs the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store a computer program for performing the method as described in any one of claims 1 to 7.

Images