A working condition adaptive industrial robot joint fault diagnosis method

Abstract

The application discloses a kind of working condition self-adaptive industrial robot joint fault diagnosis method, belong to industrial robot predictive maintenance and fault diagnosis technical field.It includes: S1, the body running data and working condition data of industrial robot joint are synchronously collected, and the health state baseline parameter is calibrated as residual analysis benchmark;S2, identify current working condition category, extract multidimensional fault feature, output robust feature sequence;S3, construct continuous twin network diagnosis model;S4, according to working condition state and diagnosis performance dynamic optimization diagnosis parameter;S5, predict joint remaining useful life;S6, set multi-condition triggered update mechanism, carry out incremental fine-tuning to diagnosis network, and guarantee model update reliability by verifying rollback mechanism.The above-mentioned application method is adopted, realizes high-precision diagnosis and full life cycle operation and maintenance support to industrial robot joint fault, with strong robustness, working condition self-adaptability and autonomous closed-loop iteration capability.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance and fault diagnosis technology for industrial robots, and in particular to a method for diagnosing joint faults in industrial robots that adapts to different operating conditions. Background Technology

[0002] With the rapid development of the intelligent manufacturing industry, industrial robots have become core equipment on automated production lines. Joints, as the core components that transmit power and enable movement in robots, directly determine the operational reliability and production safety of the equipment. Efficient and accurate joint fault diagnosis is a core component of predictive maintenance systems for industrial robots and a key means to reduce unplanned downtime losses and ensure production continuity. Among these methods, data-driven fault diagnosis based on deep learning has been widely applied in the field of industrial fault diagnosis because it does not rely on precise equipment physical models or extensive prior knowledge.

[0003] However, in actual industrial production scenarios, the operating conditions of industrial robots often change dynamically with production tasks, with parameters such as load, speed, and operating temperature fluctuating frequently. Most existing fault diagnosis models are trained on labeled data under fixed operating conditions, using fixed diagnostic thresholds and feature judgment criteria. When on-site conditions change, the distribution of fault features shifts significantly, directly leading to a substantial decrease in diagnostic accuracy and a significant increase in false alarms and false negatives. Simultaneously, complex noises such as electromagnetic interference and shock vibrations are prevalent in industrial environments, easily generating a large amount of outlier data. Conventional feature engineering solutions are insufficient in suppressing outliers, easily causing feature distortion and further reducing the reliability of diagnostic results.

[0004] On the other hand, existing deep learning diagnostic models generally lack continuous iteration and adaptive update capabilities: when new operating data for new conditions accumulates on-site and incremental training of the model is required, the "catastrophic forgetting" problem is highly likely to occur, where the model's diagnostic knowledge of historical operating conditions is overwritten by new data, leading to a decline in comprehensive diagnostic capabilities across all operating conditions. Simultaneously, most diagnostic solutions lack an online closed-loop update mechanism. As equipment undergoes natural wear and tear and performance aging over long-term operation, the deviation between the initially established health baseline and the actual condition of the equipment continues to widen, and diagnostic performance gradually degrades with the equipment's usage time.

[0005] Furthermore, most current mainstream industrial robot fault diagnosis solutions only focus on fault classification and anomaly warning, lacking methods to quantify the degree of degradation by combining equipment dynamics characteristics, and failing to establish the ability to predict the remaining service life in relation to operating loads and material properties. The diagnostic results cannot directly support the formulation of predictive maintenance plans on site, making it difficult to meet the actual needs of smart factory equipment operation and maintenance throughout its entire life cycle. Summary of the Invention

[0006] The purpose of this invention is to provide a working condition-adaptive method for diagnosing joint faults in industrial robots, thereby solving the aforementioned technical problems.

[0007] To achieve the above objectives, the present invention provides a working condition adaptive method for diagnosing joint faults in industrial robots, comprising the following steps: S1. Synchronously collect the body operation data and working condition data of the industrial robot joints, establish a single-joint digital twin dynamic baseline model based on Lagrange mechanics, and calibrate the health status baseline parameters as the residual analysis benchmark. S2. Identify the current operating condition category, extract multi-dimensional fault features from the original operating data and dynamic residuals, perform robust Gaussian feature modeling through the Huber weighting mechanism, and output robust feature sequence; S3. Construct a continuous twin network diagnostic model, use metric learning to classify faults, combine the continuous learning mechanism of experience playback to adapt to new working conditions, and integrate multi-structure autoencoders to detect unknown anomalies. S4. Adaptive adjustment of diagnostic parameters is modeled as a Markov decision process, and the SAC deep reinforcement learning algorithm is used to dynamically optimize the diagnostic parameters according to the working conditions and diagnostic performance. S5. Calculate the joint degradation index based on the degree of characteristic deviation, and predict the remaining service life of the joint by combining the material fatigue cumulative damage theory with the current working condition load. S6. Set up a multi-condition triggered update mechanism, use a damped update method to iterate the digital twin baseline parameters, perform incremental fine-tuning of the diagnostic network, and ensure the reliability of model updates through a verification rollback mechanism.

[0008] Preferably, in S1, the body operation data and working condition data are time-stamp aligned via hardware triggering; the body operation data includes joint positions. 、 angular velocity angular acceleration Output torque Motor current Joint shell temperature Triaxial vibration acceleration Operating data includes end-load quality. Ambient temperature Operating speed level.

[0009] Preferably, in S1, the expression for the digital twin dynamics baseline model is: ; In the formula, This represents the equivalent rotational inertia of the joint as the joint angle changes. For the coefficients of Coriolis force and centrifugal force, The gravitational torque term is determined by joint posture and load. Frictional torque is related to rotational speed and joint temperature, and includes static friction, Coulomb friction and viscous friction components.

[0010] Preferably, in S2, real-time operating condition identification uses load, speed, and temperature as inputs, employs a Gaussian mixture model for clustering to obtain typical operating condition categories, and outputs the current operating condition label and the degree of feature distribution offset; the extracted multi-dimensional fault features are as follows: Time-domain statistical characteristics: mean, variance, kurtosis, margin factor, kurtosis factor; Frequency domain characteristics: resonant band peak value, sideband energy ratio, and dominant frequency offset; Dynamic residual characteristics: statistical characteristics of the residual sequence between actual torque and model predicted torque.

[0011] Preferably, in S2, the specific process of Huber weighted robust Gaussian modeling is as follows: Define the Huber loss function. The expression for controlling the outlier suppression strength is: ; In the formula, The residual between the eigenvalues ​​and the initial mean. The Huber threshold is used to balance robustness and feature sensitivity. Based on residuals Calculate the weight of each sample to obtain the robust weighted mean. With robust weighted standard deviation .

[0012] Preferably, in S3, the persistent Siamese network consists of a feature extraction backbone and a metric classification head: the feature extraction backbone is constructed based on a one-dimensional residual convolutional network, with a robust feature sequence as input and an output dimension of [missing information]. The deep feature vector; the metric classification head calculates the cosine distance between the input sample and the features of various fault templates, and determines the fault type by the distance threshold; Continuous Siamese network training uses the SoftTriplet loss function. The expression is: ; In the formula, Cosine similarity for positive sample pairs. Cosine similarity for negative sample pairs. To control the scaling factor of the loss discrimination, The classification margin parameter is used to enforce the distinguishing margin between classes. It is a natural exponential function. It is the natural logarithm function.

[0013] Preferably, in S3, continuous learning adopts a dual-parallel architecture combining a main network and an auxiliary network, while incremental learning is achieved based on an experience replay pool. The main network is responsible for real-time inference and diagnosis, while the auxiliary network is responsible for incremental training on new working condition data. The experience replay pool stores typical samples from historical working conditions, and historical samples are mixed with new samples proportionally during training. For the detection of unknown anomalies, an autoencoder with four structures—LSTM (Long Short-Term Memory), CNN (Convolutional Neural Network), ResNet (Residual Network), and GRU (Gated Recurrent Unit)—is integrated to calculate anomaly scores through weighted reconstruction errors. The expression is: ; In the formula, For the first Reconstruction error of autoencoders The weights for the F1 score of the corresponding autoencoder are determined by offline testing. When an anomaly score exceeds a set threshold, it is judged as an unknown anomaly. This is the sum of the weight coefficients corresponding to the four autoencoders.

[0014] Preferably, in S4, the modeling and optimization logic of the Markov decision process is as follows: The state space includes the current operating condition category, feature distribution offset, diagnostic F1 scores for multiple samples, false positive rate, and false negative rate; the action space is a dynamically adjusted set of diagnostic parameters, including Huber threshold, fault classification distance threshold, anomaly detection score threshold, and weight vectors for each feature dimension, with all parameters having physically defined upper and lower limits; the reward function... Aiming at comprehensive diagnostic performance, while penalizing significant fluctuations in parameters, the expression is: ; In the formula, This is the diagnostic F1 value under the current parameters. To adjust the penalty coefficient, The L2 norm of the parameter adjustment range; The SAC (Soft Actor-Critic) algorithm with entropy regularization is used for policy learning. A dual-critic network is set up to mitigate value overestimation. Soft objectives are used to update the smoothing objective network parameters. S When the F1 value of a diagnostic cycle is lower than the set threshold, parameter optimization is triggered and the optimal parameters are issued.

[0015] Preferably, S5 specifically includes: Degradation index calculated based on the degree of deviation of eigenvectors The range of values ​​is 0 represents healthy, 1 represents unhealthy, and the expression is: ; In the formula, For the current feature vector, For the health baseline feature vector, This is the fault determination threshold vector; Predicting remaining service life based on the theory of cumulative damage from material fatigue The expression is: ; In the formula, This refers to the nominal lifespan of the joint under rated operating conditions. The average operating torque under the current operating conditions. The rated operating torque, The fatigue index is the coefficient of friction for the joint material.

[0016] Preferably, in S6, a model update is triggered when any of the following conditions are met: (1) The cumulative data volume of the new operating condition reaches the set threshold; (2) The F1 value is below the passing line for 10 consecutive cycles; (3) Regular monthly mandatory updates; The dynamic baseline parameters are updated using a damped exponential moving average, expressed as: ; In the formula, For the updated dynamic baseline parameter values, To update the damping coefficient and control the model update speed and anti-interference ability, These are the baseline dynamic parameter values ​​before the update. The measured values ​​of the parameters identified from the current observation data; The diagnostic network is updated using an incremental fine-tuning method, adjusting the top classification layer and the last convolutional layer while retaining the general feature extraction capability of the bottom layer. At the same time, a validation rollback mechanism is set up. The updated model is tested on a reserved validation set. If the F1 score improves, the old model is officially replaced. If the performance degrades, it is rolled back to the previous version of the parameters.

[0017] Therefore, the above-mentioned adaptive industrial robot joint fault diagnosis method has the following beneficial effects: 1. Significantly improved adaptability to changing operating conditions: By using a Gaussian mixture model to identify operating condition categories in real time, and combining the SAC deep reinforcement learning algorithm to dynamically adjust core diagnostic parameters such as Huber threshold, fault classification distance threshold, and feature weights, the average diagnostic F1 value is improved by more than 15% in changing operating condition scenarios compared to static threshold diagnostic solutions. It can effectively adapt to complex scenarios with frequent fluctuations in load, speed, and temperature in industrial sites.

[0018] 2. Excellent diagnostic robustness in high-noise environments: A multi-dimensional feature system integrating time-domain, frequency-domain, and dynamic residual features is constructed, coupled with Huber weighted robust Gaussian modeling mechanism to effectively suppress the interference of impact noise and outlier data; at the same time, four different autoencoders are integrated to realize the detection of unknown anomalies, reduce the risk of missed detection by a single model, and adapt to the harsh operating environment of complex electromagnetic and vibration environments in factories.

[0019] 3. The model can iterate autonomously over a long period of time, avoiding catastrophic forgetting: It adopts a continuous learning architecture of "main network + auxiliary network + experience replay pool". During incremental training, historical typical samples are mixed proportionally and only the top-level parameters of the network are fine-tuned. While adapting to new working conditions, it fully retains historical diagnostic knowledge. With the support of a closed-loop update mechanism triggered by multiple conditions and a verification rollback strategy, the model can iterate autonomously as the equipment wears and ages. After running continuously for 6 months, the diagnostic accuracy can still be maintained above 93%.

[0020] 4. End-to-end support for predictive maintenance decisions: It establishes a complete technical link from "data acquisition - feature extraction - fault diagnosis - degradation quantification - life prediction". Based on the degradation index in the range of 0 to 1, it combines the theory of material fatigue cumulative damage with the actual operating load to calculate the remaining service life. The output results can directly guide the formulation of maintenance plans, realizing the upgrade from passive fault alarm to proactive predictive operation and maintenance.

[0021] 5. Strong engineering feasibility and wide applicability: The overall solution is built on the standard Lagrange dynamics model of industrial robots and a mature deep learning architecture. All parameters have clear physical meanings and can be quickly deployed through offline calibration experiments and on-site debugging. No large-scale modification of equipment is required. It can be widely adapted to various industrial robot joint fault diagnosis and full life cycle operation and maintenance scenarios.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of an industrial robot joint fault diagnosis method that adapts to working conditions according to the present invention. Figure 2 This is a schematic diagram of the continuous twin diagnostic network and continuous learning mechanism architecture of the present invention; Figure 3 This is a comparison chart of F1 values ​​for different diagnostic methods under varying operating conditions according to the present invention. Figure 4 This is a schematic diagram illustrating the changes in the joint degeneration index and the prediction results of remaining service life according to the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0025] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0026] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] like Figures 1-4 As shown, the present invention provides a working condition adaptive method for diagnosing joint faults in industrial robots, comprising the following steps: S1. Synchronously collect the body operation data and working condition data of the industrial robot joints. The body operation data and working condition data are time-stamp aligned through hardware triggering. Simultaneously, establish a single-joint digital twin dynamic baseline model based on Lagrange mechanics, and calibrate the health state baseline parameters as the residual analysis benchmark. S2. Identify the current working condition category, extract multi-dimensional fault features from the original operation data and dynamic residuals, perform robust Gaussian feature modeling through the Huber weighting mechanism, and output a robust feature sequence. S3. Construct a continuous twin network diagnostic model, use metric learning to achieve fault classification, and combine continuous experience playback. The learning mechanism adapts to new working conditions and integrates multi-structure autoencoders to detect unknown anomalies; S4, the diagnostic parameters are adaptively adjusted and modeled as a Markov decision process, and the SAC deep reinforcement learning algorithm is used to dynamically optimize the diagnostic parameters according to the working condition and diagnostic performance; S5, the joint degradation index is calculated based on the degree of feature deviation, and the remaining service life of the joint is predicted by combining the material fatigue cumulative damage theory and the current working condition load; S6, a multi-condition triggered update mechanism is set, and the digital twin baseline parameters are iterated by a damped update method to incrementally fine-tune the diagnostic network, and the reliability of the model update is ensured by a verification rollback mechanism.

[0028] This embodiment uses the wrist joint (Wrist 3, with an aluminum alloy joint housing) of a six-axis industrial robot in an automotive welding production line as the specific implementation object. The specific implementation steps are as follows: S1. Multi-source data acquisition and digital twin dynamics baseline modeling, specifically including: The industrial robot in this embodiment has a rated load of 5kg and a nominal rated torque of the wrist joint. nominal lifespan Hours. High-precision encoders, torque sensors, temperature sensors, and triaxial accelerometers are installed at the joint ends, while load and speed data from the PLC are collected simultaneously.

[0029] Data synchronization acquisition: Time stamp alignment of all sensors is achieved via hardware trigger signals, with a sampling frequency set to 1kHz; the acquired body operation data includes joint positions. 、 angular velocity angular acceleration Output torque Motor current Joint shell temperature Triaxial vibration acceleration Operating data includes end-load quality. Ambient temperature Operating speed level.

[0030] Dynamic baseline calibration: Under standard rated operating conditions (load 5kg, ambient temperature 25℃, rated speed), five sets of experiments were conducted at no load, 25% load, 50% load, 75% load, and 100% load. Based on Lagrange mechanics, all parameters of the single-joint dynamic baseline model were identified. ; In the formula, This represents the equivalent rotational inertia of the joint as the joint angle changes. For the coefficients of Coriolis force and centrifugal force, The gravitational torque term is determined by joint posture and load. The frictional torque is related to rotational speed and joint temperature, and includes static friction, Coulomb friction, and viscous friction components. By identifying the equivalent moment of inertia of the joint, the Coriolis force coefficient, the gravitational torque coefficient, and the temperature correlation coefficients of static friction, Coulomb friction, and viscous friction, this set of parameters is determined as the baseline for the healthy state, serving as the benchmark for subsequent residual calculations.

[0031] S2, Condition Awareness and Huber Weighted Robust Gaussian Feature Engineering, specifically includes: Real-time operating condition identification: Using end load, operating speed, and joint temperature as inputs, a Gaussian mixture model is used to cluster historical operating data to obtain four typical operating conditions: light load and low speed, medium load and medium speed, heavy load and high speed, and high temperature and abnormality. During real-time operation, the current operating condition label is output, and the offset between the current feature distribution and the baseline distribution is calculated.

[0032] Multi-dimensional feature extraction: Five-dimensional time-domain statistical features and three-dimensional frequency-domain features are extracted from the original vibration and current signals. The five-dimensional time-domain statistical features include mean, variance, kurtosis, margin factor, and peak factor. The three-dimensional frequency-domain features include resonance band peak value, sideband energy ratio, and main frequency offset. Then, the residual sequence between the actual torque and the baseline model predicted torque is calculated, and eight-dimensional residual statistical features are extracted, for a total of 16-dimensional fault features.

[0033] Huber-weighted robust modeling: Robust Gaussian modeling is performed on each feature dimension under the corresponding operating conditions, with an initial Huber threshold. First, calculate the initial mean of the feature sequence, then calculate the weight of each sample using the Huber loss function, and finally obtain the robust weighted mean. With robust weighted standard deviation Output robust feature sequences, where the Huber loss function is defined. The expression for controlling the intensity of outlier suppression is: ; In the formula, The residuals between the eigenvalues ​​and the initial mean. The Huber threshold is used to balance robustness and feature sensitivity; this mechanism can effectively suppress outliers caused by electromagnetic interference and shock vibration in the welding workshop, and avoid feature distortion.

[0034] S3. Continuous twin network fault classification and anomaly detection, specifically including: Continuous Siamese Network Construction: The network consists of a feature extraction backbone and a metric classification head. The feature extraction backbone is a 3-layer one-dimensional residual convolutional network. The input is a 16-dimensional robust feature sequence, and the output dimension is... The deep feature vector; the metric classification head calculates the input sample and the 5 types of fault templates (wear, clearance, poor lubrication, bearing failure, gear tooth breakage), plus the cosine distance of the healthy template, to determine the fault type by passing the distance threshold; Model Training: Continuous Siamese network training employs the SoftTriplet loss function. Training, the expression is: ; In the formula, Cosine similarity for positive sample pairs. Cosine similarity for negative sample pairs. To control the scaling factor of the loss discrimination, The classification margin parameter is used to enforce the distinguishing margin between classes. It is a natural exponential function. The scaling factor is the natural logarithm function. When optimizing the feature space using the SoftTriplet loss function, a scaling factor is set. Classification interval parameter Training batch size Number of iterations per round The diagnostic accuracy under initial working conditions after training can reach 96.2%.

[0035] Continuous learning mechanism: A dual-parallel architecture of main network and auxiliary network is adopted. The main network is deployed for online real-time inference, and the auxiliary network is responsible for incremental training. An experience replay pool is established to store 200 typical samples of each of the four types of historical working conditions. When new working condition data is introduced, the auxiliary network is trained by mixing historical samples with new samples in a 1:1 ratio to avoid catastrophic forgetting.

[0036] Unknown anomaly detection: An autoencoder integrating four architectures—LSTM, CNN, ResNet, and GRU—calculates anomaly scores through weighted reconstruction errors. The expression is: ; In the formula, For the first Reconstruction error of autoencoders For the corresponding F1 score weights of the autoencoder, The F1 weights of the four autoencoders are obtained through offline testing, summing the weight coefficients of each encoder. The weighted reconstruction error is calculated in real time as an anomaly score. When the score exceeds the set threshold, it is judged as an unknown anomaly and manual review is triggered.

[0037] S4 and SAC deep reinforcement learning diagnostic parameter dynamic optimization specifically include: The problem of adaptive diagnostic parameters is modeled as a Markov decision process, and a SAC reinforcement learning agent is deployed: State space: Input the current working condition category, feature distribution offset, diagnostic F1 score of the last 100 samples, false positive rate, and false negative rate. The F1 score is a core evaluation metric in machine learning classification tasks, comprehensively measuring the diagnostic accuracy of the model; it is the harmonic mean of precision and recall. The F1 score ranges from [value missing in original text]. The closer the F1 value is to 1, the better the overall diagnostic performance of the model, with both false positives and false negatives at a low level. In this embodiment, the F1 value specifically refers to the F1 index calculated from the fault diagnosis results of the most recent 100 running samples, which is used to quantify the overall performance of the current diagnostic model in real time.

[0038] Motion space: Adjustable parameters include Huber threshold (Range 0.5~3), fault classification distance threshold (range 0.2~0.8), anomaly detection score threshold, 16-dimensional feature weight vector, all parameters are set with physical upper and lower limits to avoid exceeding the limits.

[0039] Reward function: set to In the formula, This is the diagnostic F1 value under the current parameters. To adjust the penalty coefficient, The L2 norm of the parameter adjustment range; this embodiment sets a penalty coefficient. This improves diagnostic accuracy while avoiding frequent parameter fluctuations.

[0040] Algorithm configuration: SAC algorithm with entropy regularization is used, discount factor Soft target update smoothing coefficient The configuration of a dual-critic network mitigates the overvaluation.

[0041] Triggering logic: When consecutive When the F1 value of a diagnostic cycle is lower than 0.9, SAC agent optimization is triggered. In this embodiment, when the robot switches to heavy-load high-speed working condition and the diagnostic F1 drops to 87%, the agent automatically adjusts the Huber threshold to 1.2 and the fault classification distance threshold to 0.35. After optimization, the diagnostic F1 rises back to 94.8%.

[0042] S5. Failure degradation assessment and remaining useful life prediction, specifically including: Degradation Index Quantification: The degradation index is calculated based on the Euclidean distance between the current 16-dimensional feature vector and the feature vector of the healthy baseline. , The range of values ​​is 0 represents complete health, and 1 represents complete failure. The expression is: ; In the formula, For the current feature vector, For the health baseline feature vector, This is the fault determination threshold vector; in this embodiment, the fault determination threshold vector corresponds to... The warning line and The alarm line; when a bearing wear-related fault is detected, the actual measurement... It is in a stage of mild degradation.

[0043] Predicting remaining service life based on the theory of cumulative damage from material fatigue Material fatigue index of aluminum alloy joints The average operating torque under current operating conditions is 1.2 times the rated value. Substituting this into the formula, the following calculation is performed: ; In the formula, This refers to the nominal lifespan of the joint under rated operating conditions. The average operating torque under the current operating conditions. The rated operating torque, The fatigue index is the value of the joint material. This embodiment... The system triggered a level 2 maintenance reminder and recommended scheduling a lubrication and wear check within 3 months.

[0044] S6. Online iteration and diagnostic closed-loop update of digital twin models, specifically including: Update trigger: A model update is triggered when any of the following conditions are met: (1) The cumulative data volume of the new operating condition reaches the set threshold; (2) The F1 value is below the passing line for 10 consecutive cycles; (3) Regular monthly mandatory updates; The triple triggering conditions set in this embodiment are: the cumulative number of new working condition data exceeds 500 sets, the F1 value is lower than 0.9 for 10 consecutive periods, and a forced update is required on the 1st of each month. A mechanism combining monthly forced updates and abnormal triggering is adopted.

[0045] Dynamic baseline damping update: Baseline parameters are updated using a damped exponential moving average, with the damping coefficient... This means that new observations contribute only 10% of the weight in each update, avoiding baseline drift caused by a single outlier. Substituting this into the formula, we can calculate: ; In the formula, For the updated dynamic baseline parameter values, To update the damping coefficient and control the model update speed and anti-interference ability, These are the baseline dynamic parameter values ​​before the update. The measured values ​​of the parameters identified from the current observation data; Incremental fine-tuning of the diagnostic network: Using newly added labeled data in the current month, only the top classification layer and the last convolutional layer of the network are fine-tuned, while the weights of the bottom convolutional layers remain unchanged, balancing update efficiency and retention of historical knowledge.

[0046] Validation rollback: The updated model is tested on a reserved validation set of 200 sets. If the F1 score is higher than the old model, it is officially replaced; if the F1 score drops by more than 2%, it is rolled back to the previous version of parameters, and data continues to accumulate until the next update. After running continuously for 6 months, the diagnostic accuracy of this embodiment remains above 93%, with no performance degradation.

[0047] In summary, the method of this invention constructs a complete technical closed loop of "data acquisition - feature engineering - fault diagnosis - parameter optimization - life prediction - closed-loop update", which effectively improves the diagnostic robustness in industrial scenarios with changing operating conditions and high noise. It covers the entire operation and maintenance process from anomaly warning and fault classification to degradation assessment and life prediction, and can provide a complete technical solution for predictive maintenance of industrial robots throughout their entire life cycle.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for diagnosing joint faults in an industrial robot that adapts to different operating conditions, characterized in that, Includes the following steps: S1. Synchronously collect the body operation data and working condition data of the industrial robot joints, establish a single-joint digital twin dynamic baseline model based on Lagrange mechanics, and calibrate the health status baseline parameters as the residual analysis benchmark. S2. Identify the current operating condition category, extract multi-dimensional fault features from the original operating data and dynamic residuals, perform robust Gaussian feature modeling through the Huber weighting mechanism, and output robust feature sequence; S3. Construct a continuous twin network diagnostic model, use metric learning to classify faults, combine the continuous learning mechanism of experience playback to adapt to new working conditions, and integrate multi-structure autoencoders to detect unknown anomalies. S4. Adaptive adjustment of diagnostic parameters is modeled as a Markov decision process, and the SAC deep reinforcement learning algorithm is used to dynamically optimize the diagnostic parameters according to the working conditions and diagnostic performance. S5. Calculate the joint degradation index based on the degree of characteristic deviation, and predict the remaining service life of the joint by combining the material fatigue cumulative damage theory with the current working condition load. S6. Set up a multi-condition triggered update mechanism, use a damped update method to iterate the digital twin baseline parameters, perform incremental fine-tuning of the diagnostic network, and ensure the reliability of model updates through a verification rollback mechanism.

2. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 1, characterized in that: In S1, the body's operational data and working condition data are time-stamp aligned via hardware triggering; the body's operational data includes joint positions. 、 angular velocity angular acceleration Output torque Motor current Joint shell temperature Triaxial vibration acceleration ; Operating data includes end-load quality Ambient temperature Operating speed level.

3. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 2, characterized in that: In S1, the expression for the digital twin dynamics baseline model is: ； In the formula, This represents the equivalent rotational inertia of the joint as the joint angle changes. For the coefficients of Coriolis force and centrifugal force, The gravitational torque term is determined by joint posture and load. Frictional torque is related to rotational speed and joint temperature, and includes static friction, Coulomb friction and viscous friction components.

4. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 1, characterized in that: In S2, real-time operating condition identification uses load, speed, and temperature as inputs, and employs a Gaussian mixture model for clustering to obtain typical operating condition categories. It outputs the current operating condition label and the degree of feature distribution offset. The extracted multi-dimensional fault features are as follows: Time-domain statistical characteristics: mean, variance, kurtosis, margin factor, kurtosis factor; Frequency domain characteristics: resonant band peak value, sideband energy ratio, and dominant frequency offset; Dynamic residual characteristics: statistical characteristics of the residual sequence between actual torque and model predicted torque.

5. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 1, characterized in that, In S2, the specific process of Huber weighted robust Gaussian modeling is as follows: Define the Huber loss function. The expression for controlling the outlier suppression strength is: ； In the formula, The residuals between the eigenvalues ​​and the initial mean. The Huber threshold is used to balance robustness and feature sensitivity. Based on residuals Calculate the weight of each sample to obtain the robust weighted mean. With robust weighted standard deviation .

6. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 1, characterized in that: In S3, the Continuous Siamese Network consists of a feature extraction backbone and a metric classification head: the feature extraction backbone is built based on a one-dimensional residual convolutional network, with robust feature sequences as input and an output dimension of [missing information]. The deep feature vector; the metric classification head calculates the cosine distance between the input sample and the features of various fault templates, and determines the fault type by the distance threshold; Continuous Siamese network training uses the SoftTriplet loss function. The expression is: ； In the formula, Cosine similarity for positive sample pairs. Cosine similarity for negative sample pairs. To control the scaling factor of the loss discrimination, The classification margin parameter is used to enforce the distinguishing margin between classes. It is a natural exponential function. It is the natural logarithm function.

7. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 1, characterized in that: In S3, continuous learning employs a dual-parallel architecture combining a main network and an auxiliary network, while incremental learning is achieved based on an experience replay pool. The main network handles real-time inference and diagnosis, while the auxiliary network handles incremental training on new working condition data. The experience replay pool stores typical samples from historical working conditions, and historical samples are mixed with new samples proportionally during training. For the detection of unknown anomalies, an autoencoder with four structures—LSTM, CNN, ResNet, and GRU—is integrated, and anomaly scores are calculated through weighted reconstruction errors. The expression is: ； In the formula, For the first Reconstruction error of autoencoders The weights for the F1 score of the corresponding autoencoder are determined by offline testing. When an anomaly score exceeds a set threshold, it is judged as an unknown anomaly. This is the sum of the weight coefficients corresponding to the four autoencoders.

8. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 1, characterized in that, In S4, the modeling and optimization logic of the Markov decision process is as follows: The state space includes the current operating condition category, feature distribution offset, diagnostic F1 scores for multiple samples, false positive rate, and false negative rate; the action space is a dynamically adjusted set of diagnostic parameters, including Huber threshold, fault classification distance threshold, anomaly detection score threshold, and weight vectors for each feature dimension, with all parameters having physically defined upper and lower limits; the reward function... Aiming at comprehensive diagnostic performance, while penalizing significant fluctuations in parameters, the expression is: ； In the formula, This is the diagnostic F1 value under the current parameters. To adjust the penalty coefficient, The L2 norm of the parameter adjustment range; The SAC algorithm with entropy regularization is used for policy learning. A dual-commentator network is set up to alleviate value overestimation. Soft targets are used to update the smooth target network parameters. S When the F1 value of a diagnostic cycle is lower than the set threshold, parameter optimization is triggered and the optimal parameters are issued.

9. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 1, characterized in that, S5 specifically includes: Degradation index calculated based on the degree of deviation of eigenvectors The range of values ​​is 0 represents healthy, 1 represents unhealthy, and the expression is: ； In the formula, For the current feature vector, For the health baseline feature vector, This is the fault determination threshold vector; Predicting remaining service life based on the theory of cumulative damage from material fatigue The expression is: ； In the formula, This refers to the nominal lifespan of the joint under rated operating conditions. The average operating torque under the current operating conditions. The rated operating torque, The fatigue index is the coefficient of friction for the joint material.

10. The method for diagnosing joint faults in an industrial robot with adaptive operating conditions according to claim 1, characterized in that: In S6, a model update is triggered when any of the following conditions are met: (1) The cumulative data volume of the new operating condition reaches the set threshold; (2) The F1 value is below the passing line for 10 consecutive cycles; (3) Regular monthly mandatory updates; The dynamic baseline parameters are updated using a damped exponential moving average, expressed as: ； In the formula, For the updated dynamic baseline parameter values, To update the damping coefficient and control the model update speed and anti-interference ability, These are the baseline dynamic parameter values ​​before the update. The measured values ​​of the parameters identified from the current observation data; The diagnostic network is updated using an incremental fine-tuning method, adjusting the top classification layer and the last convolutional layer while retaining the general feature extraction capability of the bottom layer. At the same time, a validation rollback mechanism is set up. The updated model is tested on a reserved validation set. If the F1 score improves, the old model is officially replaced. If the performance degrades, it is rolled back to the previous version of the parameters.

Images