A real-time fault detection method and system for industrial robots

Abstract

The present application relates to the technical field of robot fault detection, and specifically includes an industrial robot real-time fault detection method and system, the method comprising: analyzing the operation state of the industrial robot based on a communication state signal, determining a motion domain and a sensing domain synchronization factor, configuring a cross-domain feature vector, a performance feature vector, and fusing them, using a deep belief network to formulate a fault probability atlas, and identifying feature identifiers to achieve fault reminders, solving the technical problems of fault response strategy solidification, difficulty in adapting to complex and variable industrial scene requirements, and inability to effectively and timely provide fault reminders, achieving the technical effects of constructing cross-domain feature vectors and performance feature vectors, dynamically checking serial control sequences and combining multi-scale decomposition, accurately extracting signal delay fluctuations and retransmission frequencies, using a deep belief network to fuse time and space features to identify fault types, using a fault probability distribution atlas to provide fault reminders, and ensuring the safety of industrial robot operation.

Description

Technical Field

[0001] This invention relates to the field of robot fault detection technology, specifically to a real-time fault detection method and system for industrial robots. Background Technology

[0002] Industrial robots are widely used in various fields such as automobile manufacturing, electronic assembly, and machining. Their operational stability and reliability directly affect production efficiency and product quality. With the increasing complexity of industrial robot application scenarios and the diversification of task requirements, higher demands are placed on the real-time performance and accuracy of their fault detection.

[0003] Current industrial robot fault detection is mostly based on experience threshold judgment, which cannot achieve precise spatial positioning of faults and quantitative assessment of risk levels. Furthermore, fault response strategies are fixed. At the same time, fault detection relies on single sensor data or static parameter analysis, making it difficult to fully capture the correlation between multi-source heterogeneous data such as mechanical joint movement, sensor signals, and communication status. This leads to the failure to detect hidden faults such as mechanical joint overload and servo motor stall.

[0004] In summary, existing technologies suffer from rigid fault response strategies that are difficult to adapt to the needs of complex and ever-changing industrial scenarios and cannot provide timely and effective fault alerts. Summary of the Invention

[0005] This application provides a real-time fault detection method and system for industrial robots, aiming to solve the technical problem that existing fault response strategies are rigid, difficult to adapt to the needs of complex and ever-changing industrial scenarios, and unable to provide timely and effective fault alerts.

[0006] In view of the above problems, the technical solution to achieve the present application is as follows:

[0007] In a first aspect, this application provides a real-time fault detection method for industrial robots. The method includes: performing operational state correlation analysis on the industrial robot based on communication status signals to determine motion domain synchronization factors corresponding to each mechanical joint of the industrial robot and sensing domain synchronization factors corresponding to each monitoring channel of the data monitoring unit; dynamically verifying the check bits of the serial control sequence; configuring a cross-domain feature vector by combining the motion domain synchronization factors and sensing domain synchronization factors; performing multi-scale decomposition on the communication status signals to extract signal delay fluctuations and data packet retransmission frequencies, and configuring a performance feature vector; fusing the cross-domain feature vector and the performance feature vector; using the temporal and spatial features of a deep belief network to formulate a fault probability distribution map; identifying fault feature identifiers; and providing fault alerts.

[0008] Preferably, based on the fault feature identifier, combined with the fault occurrence probability, impact range and duration, a fault risk index is evaluated and compared with a preset risk threshold matrix to set a hierarchical attention mechanism; the hierarchical attention mechanism is used to weight the fault risk index by importance, dynamically focus on high-risk fault areas, and use a control strategy matrix to configure fault alert signals that meet the requirements of fault isolation priority and repair resource scheduling.

[0009] Preferably, an encrypted feature sharing protocol is established to transmit the local defect features of the high-risk fault region through homomorphic encryption; and a contribution incentive mechanism is set according to the local defect features of the high-risk fault region.

[0010] Preferably, the fault feature identifier is associated with mechanical joint overload, servo motor stall, sensor signal drift, and control command packet loss; based on the fault feature identifier, spatial attention branching is used to locate faults in various components of the industrial robot in spatial location.

[0011] Preferably, a lidar and infrared thermal imaging module are deployed to eliminate surface reflection noise through phase unfolding and obtain surface morphology features; based on the surface morphology features, a temporal convolutional network and a spatial attention branch are combined to capture the correlation between mechanical joint displacement and temperature gradient anomalies.

[0012] Preferably, an adversarial generative network is deployed to configure noise-resistant feature samples in an electromagnetic interference scenario and dynamically optimize the spatial attention weight allocation strategy; reinforcement learning is performed based on historical false detection data of metal fatigue cracks in mechanical joints through the spatial attention weight allocation strategy.

[0013] Preferably, the deep belief network of the cloud computing center is transferred to the edge computing node with knowledge; at the edge computing node, the connection weights are structurally pruned.

[0014] Preferably, an integrated digital twin simulation engine is used to pre-simulate multiple repair paths for a mechanical joint jamming scenario; through the multiple repair paths for the mechanical joint jamming scenario, a dynamic reward function based on reinforcement learning is constructed, and the operation log data and industrial robot response data during the execution of the multiple repair paths are used as incremental samples to drive the iterative optimization of the edge computing node.

[0015] Preferably, the basic information of the industrial robot model is obtained; based on the basic information of the industrial robot model, a multi-task transfer learning framework is set up for dynamic adaptation until the cold start detection error meets the allowable error range.

[0016] In a second aspect, this application provides a real-time fault detection system for industrial robots, comprising: a correlation analysis module, used to perform operational status correlation analysis on the industrial robot based on communication status signals, and determine the motion domain synchronization factor corresponding to each mechanical joint of the industrial robot and the sensing domain synchronization factor corresponding to each monitoring channel of the data monitoring unit; a feature vector configuration module, used to dynamically verify the check bits of the serial control sequence, and configure a cross-domain feature vector in combination with the motion domain synchronization factor and the sensing domain synchronization factor; to perform multi-scale decomposition on the communication status signal, extract signal delay fluctuations and data packet retransmission frequency, and configure a performance feature vector; and a fault alert module, used to fuse the cross-domain feature vector and the performance feature vector, use the temporal and spatial features of a deep belief network to formulate a fault probability distribution map, identify fault feature identifiers, and provide fault alerts.

[0017] In summary, one or more technical solutions provided in this application achieve the technical effect of constructing cross-domain feature vectors and performance feature vectors, dynamically verifying serial control sequences and combining multi-scale decomposition to accurately extract signal delay fluctuations and retransmission frequencies, using deep belief networks to fuse temporal and spatial features to identify fault types, and using fault probability distribution maps to provide fault alerts, thereby ensuring the operational safety of industrial robots. Attached Figure Description

[0018] Figure 1 This application provides a flowchart illustrating a real-time fault detection method for industrial robots.

[0019] Figure 2 This application provides a structural schematic diagram of a real-time fault detection system for industrial robots.

[0020] Figure labeling: Correlation analysis module M100, feature vector configuration module M200, fault alert module M300. Detailed Implementation

[0021] Example 1: The present application will be described in detail below with reference to the accompanying drawings, as follows... Figure 1 As shown, this application provides a real-time fault detection method for industrial robots, wherein the method includes:

[0022] S1: Analyze the operational status correlation of the industrial robot based on the communication status signal to determine the motion domain synchronization factor corresponding to each mechanical joint of the industrial robot and the sensing domain synchronization factor corresponding to each monitoring channel of the data monitoring unit; S2: Perform dynamic verification on the check bit of the serial control sequence, and configure a cross-domain feature vector by combining the motion domain synchronization factor and the sensing domain synchronization factor; perform multi-scale decomposition on the communication status signal to extract signal delay fluctuation and data packet retransmission frequency, and configure a performance feature vector.

[0023] Specifically, communication status signals refer to the signals generated by the communication between the internal components of an industrial robot and external devices during operation. These signals include information such as data transmission timing and error checking, reflecting the stability and effectiveness of communication. Operational status correlation analysis refers to analyzing communication status signals to identify the intrinsic relationships between various factors related to the operational status of the industrial robot, in order to gain a more comprehensive understanding of the robot's operational status. Motion domain synchronization factors are related to mechanical joints, reflecting the coordination and consistency of mechanical joints during movement, ensuring that the movements of each joint can be performed according to the predetermined motion trajectory and speed. Sensing domain synchronization factors are related to the monitoring channels of the data monitoring unit, reflecting the temporal synchronization of sensor data acquired by different monitoring channels.

[0024] Dynamic verification refers to the real-time checking and verification of the check bits of the serial control sequence to ensure the accuracy and reliability of control commands; cross-domain feature vector is a feature vector formed by combining relevant factors from the motion domain and the sensing domain, comprehensively reflecting the motion and sensing state of the industrial robot; multi-scale decomposition decomposes the communication state signal into components of different scales in order to extract features such as delay fluctuations and data packet retransmission frequency in the signal; performance feature vector is a vector formed based on these extracted features, used to describe the performance indicators of the communication state signal.

[0025] Execution steps: Based on the communication status signals, analyze the operational status correlation of the industrial robot to determine the motion domain synchronization factor corresponding to each mechanical joint and the sensing domain synchronization factor corresponding to the monitoring channel. Specifically, real-time monitoring and analysis of the communication status signals are required. Through time-series analysis of the communication data, the correlation between the communication status signals and the robot's operational status is determined. Based on the synchronization status of the mechanical joints at different motion stages, the motion domain synchronization factor is obtained. Simultaneously, time synchronization analysis is performed on the data from the monitoring channel to obtain the sensing domain synchronization factor. Dynamic verification of the check bits of the serial control sequence is performed to ensure the correctness of the control commands.

[0026] By combining motion domain synchronization factors and sensing domain synchronization factors to configure cross-domain feature vectors, key information from both motion and sensing domains is integrated to more comprehensively reflect the robot's operational status. Multi-scale decomposition of communication status signals is performed to extract signal delay fluctuations and data packet retransmission frequency. It is important to note that signal delay fluctuations exceeding a certain threshold can lead to robot response delays, impacting production efficiency. Performance feature vectors are then configured, containing key performance indicators of the communication status. Through data acquisition and feature extraction, these steps provide accurate and comprehensive feature information for subsequent fault identification and alerts.

[0027] S3: The cross-domain feature vector is fused with the performance feature vector, and the temporal and spatial features of the deep belief network are used to formulate a fault probability distribution map, identify fault feature identifiers, and provide fault alerts.

[0028] Specifically, cross-domain feature vector fusion refers to the comprehensive processing of feature vectors from the motion domain and the sensing domain to fully reflect the operating status of the industrial robot; deep belief networks are deep learning algorithms based on probabilistic generative models that learn complex patterns in data through multi-layer neural network structures, and have the ability to extract both temporal and spatial features; temporal features refer to the changing patterns of data over time, including the characteristics of speed, acceleration, etc., during the movement of the industrial robot; spatial features involve the spatial distribution characteristics of data, such as the spatial distribution of the position and posture of mechanical joints; fault probability distribution maps are a visual representation based on the analysis of the fused feature vectors by deep belief networks, showing the probability of different fault types occurring at a specific time; fault feature identification is achieved by analyzing the fault probability distribution map to identify feature patterns that can represent specific fault types for subsequent fault alerts.

[0029] Execution steps: The cross-domain feature vector and the performance feature vector are fused. Specifically, methods such as weighted averaging and principal component analysis are used to integrate the information of the two feature vectors. The fused feature vector contains information on motion, sensing, and communication performance, which can more comprehensively reflect the operating status of the industrial robot. The temporal and spatial feature extraction capabilities of deep belief networks are used to analyze the fused feature vector. Through its multi-layer neural network structure, deep belief networks abstract and learn features layer by layer from the data to extract temporal and spatial features.

[0030] Using the temporal and spatial features of deep belief networks, further, in terms of temporal feature extraction, patterns of abnormal speed changes are identified by analyzing the motion trajectory data of mechanical joints over a period of time; in terms of spatial feature extraction, spatial distribution characteristics of abnormal coordination between joints are discovered by analyzing the position and posture data of multiple mechanical joints.

[0031] Based on the extracted temporal and spatial features, a fault probability distribution map is constructed, where different regions represent the probability of different fault types occurring. By identifying feature patterns in the fault probability distribution map, fault feature identifiers are determined, and fault alerts are issued accordingly. These alerts can take the form of audible and visual alarms, system prompts, etc. In the above steps, deep learning algorithms are used to perform in-depth analysis on the fused feature vectors, achieving accurate fault identification and early warning, improving the accuracy and timeliness of fault detection, and providing accurate basis for subsequent fault handling.

[0032] Furthermore, the method of this application includes identifying fault characteristics and providing fault alerts:

[0033] Based on the fault feature identifier, combined with the fault occurrence probability, impact range and duration, a fault risk index is evaluated and compared with a preset risk threshold matrix to set a hierarchical attention mechanism; the fault risk index is weighted by importance using the hierarchical attention mechanism to dynamically focus on high-risk fault areas, and a control strategy matrix is ​​used to configure fault alert signals that meet the requirements of fault isolation priority and repair resource scheduling.

[0034] Specifically, the fault risk index is used to quantify the severity of a fault, taking into account factors such as the probability of fault occurrence, the scope of impact, and the duration of the fault; the hierarchical attention mechanism can allocate attention to data according to different levels (such as different components, different fault types, etc.), enabling the model to focus on key parts when processing information; the control strategy matrix is ​​a predefined set of rules used to determine how to configure fault alert signals based on different fault attributes (such as priority, resource requirements, etc.) to ensure the effectiveness and timeliness of alerts; the fault alert signal is used to notify the maintenance system of the fault situation, and can be a visual, auditory, or other form of signal, the configuration of which needs to meet the priority of fault isolation and the scheduling requirements of repair resources.

[0035] Execution steps: Based on the identified fault characteristics, collect data on the probability of occurrence, scope of impact, and duration of the fault. If, through historical data statistics, it is determined that the probability of occurrence of the fault corresponding to mechanical joint overload is 15%, the scope of impact covers 30% of the working area of ​​the industrial robot, and the average duration is 2 hours, for example, the fault risk index is evaluated by weighted summation with weights of 0.4, 0.3, and 0.3 respectively. Then, the risk index corresponding to mechanical joint overload is the standardized value of (0.4×15%+0.3×30%+0.3×2 converted to fractional form).

[0036] The calculated fault risk index is compared with a preset risk threshold matrix, which includes threshold ranges corresponding to different risk levels, such as 0-0.3 for low risk, 0.3-0.6 for medium risk, and 0.6-1.0 for high risk. Based on the comparison results, the current fault risk level is determined, and a corresponding hierarchical attention mechanism is set. If the risk index falls in the high-risk range, the hierarchical attention mechanism will give higher attention to the fault and adjust the model parameters to prioritize the processing of high-risk fault-related data in subsequent analyses. Weighting the fault risk index with the hierarchical attention mechanism means that when dealing with multiple faults, the risk index of high-risk faults will be amplified, thus occupying a more important position in the decision-making process.

[0037] The dynamic focus on high-risk fault areas is achieved by adjusting system resource allocation, such as increasing the monitoring frequency of high-risk areas or allocating more computing resources for in-depth analysis; and by using a control strategy matrix to configure fault alert signals. The control strategy matrix formulates corresponding alert methods and content based on the priority of the fault and the repair resource requirements. In the above steps, by conducting risk assessment and priority ranking of identified faults, and by quantitatively assessing the fault risk and providing strategic alerts, the system can ensure that it can respond to high-risk faults in a timely and effective manner, rationally allocate repair resources, and reduce the impact of faults on production.

[0038] Furthermore, to dynamically focus on high-risk fault areas, the method in this application also includes:

[0039] Establish an encrypted feature sharing protocol to transmit local defect features of the high-risk fault region through homomorphic encryption; set up a contribution incentive mechanism based on the local defect features of the high-risk fault region.

[0040] Specifically, the encrypted feature sharing protocol refers to a communication protocol used to protect data privacy and security, specifying how to securely share fault data containing local defect features between different systems or devices; homomorphic encryption allows specific computational operations to be performed on encrypted data without first decrypting the data, thereby protecting the confidentiality of data during transmission and processing; local defect features refer to specific physical or performance characteristics related to faults in high-risk fault areas, such as tiny cracks in mechanical joints or abnormal temperature distributions, which can provide key information for accurate fault identification and analysis; the contribution incentive mechanism is a mechanism used to incentivize participants to actively contribute their data or resources, by quantifying the contribution of each participant's local defect features to the overall fault detection and providing corresponding rewards or resource allocations.

[0041] Execution steps: Establish an encrypted feature sharing protocol, defining the data sharing format, encryption standard, and permissions of participating parties. For example, the protocol stipulates that the AES-256 encryption algorithm is used to encrypt the data, and only authorized recipients can decrypt it using a specific key. Use homomorphic encryption technology to encrypt the local defect features of high-risk fault areas. For example, use the Paillier homomorphic encryption algorithm to encrypt the temperature anomaly data and surface crack size data of mechanical joints before transmission to ensure data security.

[0042] The receiving party sets up a contribution incentive mechanism based on the decrypted local defect features, evaluating the contribution of each data provider based on the completeness, accuracy, and timeliness of the local defect features provided. According to preset reward rules, corresponding points or resource rewards are given to ensure the secure sharing and efficient utilization of high-risk fault data, while also incentivizing all parties to actively participate in data sharing, thereby improving the performance and reliability of the entire fault detection system. In the above steps, the use of encrypted feature sharing protocols and homomorphic encryption technology not only protects data privacy and security but also expands the data sources for fault detection, enabling the system to gain a more comprehensive understanding of fault conditions. The contribution incentive mechanism further improves data quality and the timeliness of sharing, promoting continuous system optimization.

[0043] Furthermore, the application methods also include:

[0044] The fault feature identifiers are associated with mechanical joint overload, servo motor stall, sensor signal drift, and control command packet loss; based on the fault feature identifiers, spatial attention branching is used to locate faults in various components of the industrial robot in spatial space.

[0045] Specifically, fault feature identification refers to the characteristic patterns identified through fault data analysis that represent specific fault types, playing a crucial role in fault detection; mechanical joint overload refers to the load on a robot joint exceeding its design or safety threshold during operation, potentially leading to joint damage or performance degradation; servo motor stall refers to the inability of a servo motor to rotate normally due to mechanical obstacles or other reasons; sensor signal drift refers to the shift or change in the sensor output signal over time, usually caused by sensor aging, environmental changes, etc., affecting measurement accuracy; control command packet loss refers to the loss of control command data packets during robot control due to communication interference or other problems, affecting the robot's normal action execution; spatial attention branch is a deep learning-based mechanism that focuses on processing and analyzing spatial information in input data, capturing spatial distribution features in the data to achieve more accurate fault location.

[0046] Execution steps: Based on the identified fault feature identifiers, they are correlated and matched with common fault types such as mechanical joint overload, servo motor stall, sensor signal drift, and control command packet loss. For example, by analyzing the abnormal current fluctuation characteristics in the fault feature identifiers and comparing them with the typical current characteristics when a mechanical joint is overloaded, combined with historical data statistics, it is shown that the standard deviation of current fluctuations increases when a mechanical joint is overloaded, thus determining the correlation between the fault feature identifier and mechanical joint overload.

[0047] Spatial attention branching is used to perform spatial fault localization on various components of industrial robots. Specifically, spatial attention branching extracts spatial features from sensor data and operational status data from different components. Commonly, by analyzing the spatial distribution of sensor data such as temperature, pressure, and displacement from various joints of the robot, the specific location of the fault can be determined. If the spatial attention branching finds that the temperature sensor signal of the robot's right arm joint is abnormally drifting through the data analysis, combined with fault feature identification, the existence of a fault can be determined, and spatial positioning technology can be used to precisely locate the fault at the right arm joint, thus achieving accurate identification of the fault type and precise location of the fault.

[0048] Furthermore, this application employs spatial attention branching for spatial fault localization of various components of an industrial robot. The method includes:

[0049] Deploy a lidar and infrared thermal imaging module to eliminate surface reflection noise on metal surfaces through phase unrolling and obtain surface morphology features; based on the surface morphology features, combine a temporal convolutional network and a spatial attention branch to capture the correlation between mechanical joint displacement and temperature gradient anomalies.

[0050] Specifically, LiDAR refers to a sensor that generates three-dimensional environmental information by emitting a laser beam and measuring the return time of the reflected light, and is often used for distance measurement and spatial modeling; an infrared thermal imaging module is a device that can detect infrared radiation from an object's surface and convert it into a thermal image, used to detect the temperature distribution and changes on the object's surface; phase unwrapping is used to eliminate measurement errors caused by phase wrapping, especially effective when processing periodic signals (such as phase changes of reflected light); surface morphology features refer to the microscopic or macroscopic geometric features of an object's surface, including roughness, waviness, and contour; temporal convolutional networks are used to process time-series data, extracting features in the time dimension through convolution operations; temperature gradient anomalies refer to temperature changes in space or time that exceed the normal range, usually related to overheating, overload, or other faults of mechanical components.

[0051] Execution steps: Deploy LiDAR and infrared thermal imaging modules. These modules are installed at key locations on the industrial robot to monitor critical components such as mechanical joints in real time. The LiDAR acquires precise position and motion trajectory data of the mechanical joints, while the infrared thermal imaging module monitors the temperature distribution within the joints. It's important to note that the LiDAR's accuracy can reach millimeter-level, precisely capturing minute displacement changes in the mechanical joints. The infrared thermal imaging module has a temperature resolution of 0.1℃, sensitively detecting even minute temperature fluctuations.

[0052] Reflection noise from metal surfaces can cause errors in lidar measurement signals. Phase unrolling can effectively correct these errors and improve the accuracy of measurement data. Surface morphology features are acquired by analyzing lidar and infrared thermal imaging data to extract morphological features of the mechanical joint surface, such as roughness, displacement changes, and temperature distribution patterns. By combining a temporal convolutional network with a spatial attention branch, the extracted surface morphology features are analyzed in depth. The temporal convolutional network is used to process the time-series changes in displacement and temperature data of the mechanical joint, capturing the dynamic changes of these data over time.

[0053] By analyzing displacement and temperature data of mechanical joints in different operating cycles, temporal convolutional networks can identify abnormal vibration patterns or temperature rise trends. The spatial attention branch focuses on analyzing the spatial distribution of these features to pinpoint the specific location of the fault. When analyzing the temperature gradient of the mechanical joint, the spatial attention branch can detect abnormal temperature increases in a particular area, indicating localized wear or overload. Through these steps, advanced sensor technology and deep learning models enable precise monitoring and correlation analysis of abnormal displacement and temperature gradients in mechanical joints, providing strong support for early fault detection and accurate diagnosis.

[0054] Furthermore, the method of this application includes:

[0055] A generative adversarial network is deployed to configure noise-resistant feature samples in an electromagnetic interference scenario and dynamically optimize the spatial attention weight allocation strategy. Through the spatial attention weight allocation strategy, reinforcement learning is performed based on historical false detection data of metal fatigue cracks in mechanical joints.

[0056] Specifically, Generative Adversarial Networks (GANs) consist of a generator and a discriminator. The generator produces realistic data samples, while the discriminator distinguishes between real and generated data. They train against each other, and the generator can produce high-quality, noise-resistant feature samples. Electromagnetic interference scenarios refer to interference signals generated by various electromagnetic devices in industrial environments, which can affect the normal operation of sensors and control systems. Noise-resistant feature samples refer to data samples that accurately reflect fault characteristics such as metal fatigue cracks in mechanical joints even in the presence of electromagnetic interference, and are used to train and optimize fault detection models. Spatial attention weight allocation strategies are used to adjust and optimize the parameters of the spatial attention mechanism in deep learning models, dynamically adjusting weights to better focus on key fault feature regions. Reinforcement learning involves an agent interacting with the environment and learning from reward signals received from the environment to optimize the agent's performance.

[0057] Execution steps: Deploy a generative adversarial network (GAN) to generate sample data containing noise-resistant features by simulating electromagnetic interference scenarios. For example, in a simulated electromagnetic interference environment, the noise-resistant feature samples generated by the generator are continuously optimized with real fault samples under the supervision of the discriminator, so that the similarity between the features of the generated samples and the real fault features reaches more than 95%, thereby improving the model's robustness to noise. Use these noise-resistant feature samples to dynamically optimize the spatial attention weight allocation strategy. For example, during training, the spatial attention weight is dynamically adjusted according to the distribution of fault features in the samples, so that the model increases the attention weight of key fault areas such as mechanical joint metal fatigue cracks, thereby capturing fault features more effectively.

[0058] By employing a spatial attention weight allocation strategy, reinforcement learning is performed based on historical false detection data of metal fatigue cracks in mechanical joints. Specifically, historical false detection data is used as training samples, and the accuracy of fault detection is used as a reward signal to continuously adjust the model's parameters and weight allocation strategy. In these steps, generative adversarial networks enhance the model's resistance to noise such as electromagnetic interference, and reinforcement learning optimizes the spatial attention mechanism to improve the detection accuracy of key fault features. This enhances the reliability and accuracy of the entire fault detection system, reducing false negatives or missed positives caused by noise interference and false detections.

[0059] Furthermore, the method of this application includes:

[0060] The deep belief network in the cloud computing center is migrated to the edge computing nodes with knowledge; at the edge computing nodes, the connection weights are structurally pruned.

[0061] Specifically, a cloud computing center refers to a centralized pool of computing resources that provides powerful computing capabilities and data storage; deep belief networks are deep learning algorithms based on probabilistic generative models that learn complex patterns in data through multi-layered neural network structures; knowledge transfer refers to the process of applying knowledge learned in one environment (such as model parameters, feature representations, etc.) to another environment; edge computing nodes are computing units located close to data sources or terminal devices that can perform data processing and analysis locally, reducing dependence on cloud computing centers; and structured pruning reduces the number of model parameters and computational complexity by removing unimportant connections or neurons from a neural network, while maintaining model performance as much as possible.

[0062] Execution steps: First, perform knowledge transfer on the deep belief network model trained in the cloud computing center. This involves migrating the model's parameters, structure, and trained feature representations from the cloud computing center to edge computing nodes. For example, by designing a lightweight model conversion algorithm, the multi-layer parameters of the deep belief network can be optimized and compressed to enable efficient operation under the limited resources of edge computing nodes. Second, perform structured pruning of the model's connection weights on the edge computing nodes. Specifically, based on the actual operation of the model on the edge nodes and resource constraints, determine unimportant connection weights and prune them.

[0063] By analyzing the model's performance on edge nodes for detecting fatigue cracks in mechanical joint metals, it was found that certain connection weights had a relatively small impact on the detection results. Pruning these weights optimized the model's deployment and operational efficiency, enabling the deep belief network to run efficiently on resource-constrained edge computing nodes. In these steps, knowledge transfer and structured pruning reduced reliance on cloud computing centers, lowered data transmission latency, and improved the system's real-time performance and reliability. This allows the fault detection system to respond to and handle fault situations more quickly, adapting to complex industrial environments.

[0064] Furthermore, based on reinforcement learning using historical false detection data of metal fatigue cracks in mechanical joints, the method of this application also includes:

[0065] An integrated digital twin simulation engine is used to pre-simulate multiple repair paths for a mechanical joint jamming scenario. Based on these multiple repair paths, a dynamic reward function based on reinforcement learning is constructed. The operation log data and industrial robot response data during the execution of these multiple repair paths are used as incremental samples to drive the iterative optimization of the edge computing nodes.

[0066] Specifically, a digital twin simulation engine is a technology that uses digital models to accurately simulate and analyze physical entities. Through real-time data synchronization and advanced simulation algorithms, it can reproduce the operating state and behavior of physical entities in a virtual environment. Mechanical joint jamming scenarios refer to situations in industrial robots where mechanical joints cannot move normally due to mechanical failures, control problems, etc., which usually leads to production interruptions or equipment damage. The dynamic reward function in reinforcement learning refers to a function that adjusts the reward value based on dynamic information from environmental feedback during reinforcement learning, used to guide the agent in learning the optimal behavioral strategy. Iterative optimization of edge computing nodes refers to continuously adjusting and optimizing the parameters of the model on edge computing nodes to improve the model's performance and adaptability.

[0067] Execution steps: Integrate a digital twin simulation engine to establish an accurate digital model of the industrial robot and its mechanical joints, and simulate multiple possible repair paths for mechanical joint jamming scenarios. Specifically, in the simulation environment, the robot's operation after different repair measures such as recalibration, joint lubrication, and component replacement can be pre-simulated, generating multiple repair paths for analysis. Based on the pre-simulation results of these repair paths, a dynamic reward function of reinforcement learning is constructed, and reward values ​​are set according to indicators such as the success rate, time consumption, and cost of the repair path. The path that is successfully repaired and has the shortest time receives the highest reward.

[0068] The operation log data and industrial robot response data during the execution of the multiple repair paths are used as incremental samples and fed back to the edge computing nodes. This records data such as operator steps, robot response time, and post-repair performance metrics. These incremental samples drive iterative optimization of the edge computing nodes, for example, by adjusting model parameters through online learning algorithms, enabling the model to better adapt to new fault conditions and repair strategies. In the above steps, the combination of digital twin simulation and reinforcement learning optimizes fault repair strategies, improving the success rate and efficiency of repairs. Simultaneously, the iterative optimization of the edge computing nodes allows for rapid adaptation to new fault modes and repair scenarios, giving the entire fault detection and repair system stronger adaptability and continuous improvement capabilities.

[0069] Furthermore, the method of this application also includes:

[0070] Obtain basic model information of the industrial robot; based on the basic model information of the industrial robot, set up a multi-task transfer learning framework for dynamic adaptation until the cold start detection error meets the allowable error range.

[0071] Specifically, the basic information of an industrial robot model refers to fundamental data such as the robot's specific type, specifications, manufacturer, design parameters, and control system configuration. This information forms the basis for targeted fault detection and optimization. Cold start detection error refers to the deviation between the detection results and the actual situation when the system is first started or the model is first deployed to new equipment, due to a lack of historical data or the model not yet being adapted to the new environment. A multi-task transfer learning framework is a machine learning method that allows the model to share knowledge and feature representations across multiple related tasks, while simultaneously making specific adjustments for each task to improve the model's learning efficiency and performance on new tasks. Dynamic adaptation refers to automatically adjusting model parameters or configurations based on real-time data and environmental changes to ensure the model's accuracy and effectiveness.

[0072] Execution steps: Obtain basic information about the industrial robot model. Collect key data such as the robot's mechanical structure, range of motion, load capacity, sensor type and distribution by consulting equipment manuals, reading parameter configurations in the robot control system, or connecting to the manufacturer's database. This includes the maximum speed and rated load of the industrial robot's mechanical joints. Based on this basic model information, construct a multi-task transfer learning framework. In the framework, preset multiple fault detection tasks related to this robot model, such as mechanical joint overload detection and servo motor temperature anomaly detection. Extract common features in the shared layer of the model and make targeted adjustments in the task-specific layer.

[0073] Through a dynamic adaptation mechanism, the model is deployed to edge computing nodes, and cold start detection errors are continuously monitored. Using an initial set of test data, the initial cold start detection errors of the model on different fault types are calculated. By dynamically adjusting model parameters, such as the learning rate and regularization coefficient, the detection accuracy of the newly deployed fault detection system gradually improves from the initial level during the cold start phase, reducing the error to an acceptable range (e.g., within 5%). This ensures that the fault detection system can quickly adapt to different models of industrial robots and provide accurate fault detection results in the early stages of system startup. In the above steps, the multi-task transfer learning framework effectively utilizes knowledge from relevant tasks, reducing reliance on large amounts of specific model data. Simultaneously, the dynamic adaptation mechanism ensures the accuracy and reliability of the model in practical applications.

[0074] In summary, the beneficial effects of the embodiments of this application are:

[0075] This application provides a real-time fault detection method and system for industrial robots. By analyzing the operational status correlation of the industrial robot based on communication status signals, it determines the motion domain synchronization factor corresponding to each mechanical joint of the industrial robot and the sensing domain synchronization factor corresponding to each monitoring channel of the data monitoring unit. It dynamically verifies the check bits of the serial control sequence and configures cross-domain feature vectors by combining the motion domain synchronization factors and sensing domain synchronization factors. It also performs multi-scale decomposition of the communication status signals to extract signal delay fluctuations and data packet retransmission frequencies, configuring performance feature vectors. Finally, it fuses the cross-domain feature vectors and performance feature vectors, uses the temporal and spatial features of a deep belief network to construct a fault probability distribution map, identifies fault characteristics, and provides fault alerts. This achieves the technical effect of ensuring the operational safety of industrial robots by constructing cross-domain feature vectors and performance feature vectors, dynamically verifying the serial control sequence and combining multi-scale decomposition, accurately extracting signal delay fluctuations and retransmission frequencies, using a deep belief network to fuse temporal and spatial features to identify fault types, and providing fault alerts using a fault probability distribution map.

[0076] Example 2, based on the same inventive concept as the real-time fault detection method for industrial robots in the foregoing examples, such as... Figure 2 As shown in the figure, this application embodiment provides a real-time fault detection system for industrial robots, wherein the system includes:

[0077] The correlation analysis module M100 is used to perform operational status correlation analysis on the industrial robot based on the communication status signal, and to determine the motion domain synchronization factor corresponding to each mechanical joint of the industrial robot and the sensing domain synchronization factor corresponding to each monitoring channel of the data monitoring unit.

[0078] The feature vector configuration module M200 is used to dynamically verify the check bits of the serial control sequence, and configure cross-domain feature vectors in combination with the motion domain synchronization factor and the sensing domain synchronization factor; it also performs multi-scale decomposition on the communication state signal, extracts signal delay fluctuations and data packet retransmission frequency, and configures performance feature vectors.

[0079] The fault alert module M300 is used to fuse the cross-domain feature vector with the performance feature vector, use the temporal and spatial features of the deep belief network to formulate a fault probability distribution map, identify fault feature identifiers, and provide fault alerts.

[0080] Furthermore, the fault alert module M300 is used to perform the following method:

[0081] Based on the fault feature identifier, combined with the fault occurrence probability, impact range and duration, a fault risk index is evaluated and compared with a preset risk threshold matrix to set a hierarchical attention mechanism; the fault risk index is weighted by importance using the hierarchical attention mechanism to dynamically focus on high-risk fault areas, and a control strategy matrix is ​​used to configure fault alert signals that meet the requirements of fault isolation priority and repair resource scheduling.

[0082] Furthermore, the fault alert module M300 is also used to perform the following method:

[0083] Establish an encrypted feature sharing protocol to transmit local defect features of the high-risk fault region through homomorphic encryption; set up a contribution incentive mechanism based on the local defect features of the high-risk fault region.

[0084] Furthermore, the fault alert module M300 is also used to perform the following method:

[0085] The fault feature identifiers are associated with mechanical joint overload, servo motor stall, sensor signal drift, and control command packet loss; based on the fault feature identifiers, spatial attention branching is used to locate faults in various components of the industrial robot in spatial space.

[0086] Furthermore, the fault alert module M300 is also used to perform the following method:

[0087] Deploy a lidar and infrared thermal imaging module to eliminate surface reflection noise on metal surfaces through phase unrolling and obtain surface morphology features; based on the surface morphology features, combine a temporal convolutional network and a spatial attention branch to capture the correlation between mechanical joint displacement and temperature gradient anomalies.

[0088] Furthermore, the fault alert module M300 is also used to perform the following method:

[0089] A generative adversarial network is deployed to configure noise-resistant feature samples in an electromagnetic interference scenario and dynamically optimize the spatial attention weight allocation strategy. Through the spatial attention weight allocation strategy, reinforcement learning is performed based on historical false detection data of metal fatigue cracks in mechanical joints.

[0090] Furthermore, the fault alert module M300 is also used to perform the following method:

[0091] The deep belief network in the cloud computing center is migrated to the edge computing nodes with knowledge; at the edge computing nodes, the connection weights are structurally pruned.

[0092] Furthermore, the fault alert module M300 is also used to perform the following method:

[0093] An integrated digital twin simulation engine is used to pre-simulate multiple repair paths for a mechanical joint jamming scenario. Based on these multiple repair paths, a dynamic reward function based on reinforcement learning is constructed. The operation log data and industrial robot response data during the execution of these multiple repair paths are used as incremental samples to drive the iterative optimization of the edge computing nodes.

[0094] Furthermore, the fault alert module M300 is also used to perform the following method:

[0095] Obtain basic model information of the industrial robot; based on the basic model information of the industrial robot, set up a multi-task transfer learning framework for dynamic adaptation until the cold start detection error meets the allowable error range.

[0096] In summary, any step can be stored as a computer instruction or program in an unrestricted computer memory and can be called and recognized by an unrestricted computer processor; no further restrictions are imposed here.

[0097] Furthermore, the above technical solutions only embody the preferred technical solutions of the embodiments of this application. Any changes that those skilled in the art may make to certain parts of these solutions embody the novel principles of the embodiments of this application. Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope of this application.

Claims

1. A real-time fault detection method for industrial robots, characterized in that, The method includes: Based on the communication status signals, the operation status correlation analysis of the industrial robot is performed to determine the motion domain synchronization factor corresponding to each mechanical joint of the industrial robot and the sensing domain synchronization factor corresponding to each monitoring channel of the data monitoring unit. The check bits of the serial control sequence are dynamically checked, and a cross-domain feature vector is configured by combining the motion domain synchronization factor and the sensing domain synchronization factor; the communication state signal is decomposed into multiple scales to extract signal delay fluctuations and data packet retransmission frequency, and a performance feature vector is configured. The cross-domain feature vector is fused with the performance feature vector, and the temporal and spatial features of the deep belief network are used to formulate a fault probability distribution map, identify fault feature identifiers, and provide fault alerts. This includes identifying fault characteristic markers and providing fault alerts, including: Based on the fault feature identifier, combined with the fault occurrence probability, impact range and duration, the fault risk index is evaluated, compared with the preset risk threshold matrix, and a hierarchical attention mechanism is set. Using the aforementioned hierarchical attention mechanism, the fault risk index is weighted by importance, and high-risk fault areas are dynamically focused. A control strategy matrix is ​​used to configure fault alert signals that meet the requirements of fault isolation priority and repair resource scheduling. Among these, dynamically focusing on high-risk fault areas also includes: Establish an encrypted feature sharing protocol to transmit local defect features of the high-risk fault region using homomorphic encryption; Based on the local defect characteristics of the high-risk fault area, a contribution incentive mechanism is set up; The fault feature identifiers are associated with mechanical joint overload, servo motor stall, sensor signal drift, and control command packet loss. Based on the fault feature identifiers, spatial attention branching is used to locate faults in various components of the industrial robot in spatial location. Among them, the spatial attention branch is used for spatial localization of faults in various components of industrial robots, including: Deploy lidar and infrared thermal imaging modules to eliminate surface reflection noise and obtain surface morphology features through phase unfolding; Based on the surface morphology features, a temporal convolutional network and a spatial attention branch are combined to capture the correlation between mechanical joint displacement and temperature gradient anomalies. This includes: Deploy a generative adversarial network, configure noise-resistant feature samples for electromagnetic interference scenarios, and dynamically optimize the spatial attention weight allocation strategy. The spatial attention weight allocation strategy is used to perform reinforcement learning based on historical false detection data of metal fatigue cracks in mechanical joints.

2. The real-time fault detection method for industrial robots as described in claim 1, characterized in that, The method includes: Migrate the knowledge of deep belief networks in cloud computing centers to edge computing nodes; At the edge computing node, the connection weights are structurally pruned.

3. The real-time fault detection method for industrial robots as described in claim 2, characterized in that, The method further includes reinforcement learning based on historical false positive data of metal fatigue cracks in mechanical joints: An integrated digital twin simulation engine is used to simulate multiple repair paths for mechanical joint jamming scenarios; By constructing multiple repair paths for the mechanical joint jamming scenario, a dynamic reward function based on reinforcement learning is developed. The operation log data and industrial robot response data during the execution of these multiple repair paths are used as incremental samples to drive the iterative optimization of the edge computing nodes.

4. The real-time fault detection method for industrial robots as described in claim 2, characterized in that, The method further includes: Obtain basic information about the industrial robot model; Based on the basic information of the industrial robot model, a multi-task transfer learning framework is set up for dynamic adaptation until the cold start detection error meets the allowable error range.

5. A real-time fault detection system for industrial robots, characterized in that, The system, used for implementing the real-time fault detection method for an industrial robot according to any one of claims 1-4, comprises: The correlation analysis module is used to perform operational status correlation analysis on the industrial robot based on the communication status signals, and to determine the motion domain synchronization factor corresponding to each mechanical joint of the industrial robot and the sensing domain synchronization factor corresponding to each monitoring channel of the data monitoring unit. The feature vector configuration module is used to dynamically verify the check bits of the serial control sequence, and configure cross-domain feature vectors in combination with the motion domain synchronization factor and the sensing domain synchronization factor; it also performs multi-scale decomposition on the communication state signal, extracts signal delay fluctuations and data packet retransmission frequency, and configures performance feature vectors. The fault alert module is used to fuse the cross-domain feature vector with the performance feature vector, use the temporal and spatial features of the deep belief network to formulate a fault probability distribution map, identify fault feature identifiers, and provide fault alerts.

Images