Industrial robot electromechanical fault real-time detection and early warning system

Abstract

The invention discloses an industrial robot electromechanical fault real-time detection and early warning system, relates to the technical field of industrial robot fault detection, and is used for solving the problems of single detection dimension, difficulty in early fault capture, easiness in premature convergence of an algorithm, poor variable working condition adaptability and insufficient early warning linkage in the prior art. The system comprises a multi-mode sensing acquisition module, a signal preprocessing module, an improved chaos particle swarm feature extraction module, an LSTMGRU fusion fault identification module, a grading early warning linkage module, a data storage module and a variable working condition adaptive adjustment module. Fiber bragg grating stress-strain sensing, directional capacitive acoustic sensing and WGM optical calibration sensing are used for cooperatively collecting signals, features are extracted through preprocessing and an improved chaos particle swarm algorithm, fault recognition and grade judgment are completed through an LSTMGRU fusion model, and linkage of three-level grading early warning and a robot control system is achieved. The method can accurately capture early weak faults, is high in fault recognition accuracy, is high in variable working condition adaptability, is timely in early warning response, and is suitable for electromechanical fault detection of various industrial robots.

Description

Technical Field

[0001] This invention relates to the field of industrial robot fault detection technology, and in particular to a real-time detection and early warning system for electromechanical faults in industrial robots. Background Technology

[0002] As core equipment in intelligent manufacturing, the stable operation of industrial robots' electromechanical systems (including motors, reducers, servo mechanisms, and transmission components) directly determines production line efficiency and product quality. Currently, industrial robot electromechanical fault detection faces the following major technical bottlenecks, which existing solutions struggle to overcome: I. Limited detection dimensions and weak early fault detection capability: Existing technologies mostly rely on a single vibration sensor or current sensor to collect signals, which makes it difficult to capture weak fault signals such as micro-cracks in joints, slight insulation aging of motor stators, and early wear of gears in reducers. Moreover, fault signals are easily masked by industrial environmental noise, resulting in delayed fault detection. Often, irreversible damage to components has already occurred by the time the fault appears. Second, the algorithm is prone to getting trapped in local optima and has low fault identification accuracy: Traditional fault detection uses a scheme that combines conventional particle swarm optimization with BP neural network, which has the problem of premature convergence. It cannot effectively optimize feature extraction and fault identification parameters, resulting in a fault misjudgment rate of more than 30%, which is difficult to meet the needs of high-precision industrial detection. 3. Poor adaptability to changing working conditions and insufficient generalization ability: The rotation speed and load of industrial robots are constantly changing during actual operation. Existing data-driven detection methods do not consider the impact of changes in working conditions. The accuracy of fault identification in the model drops significantly under changing working conditions and cannot be adapted to complex industrial production scenarios. Fourth, the lack of hierarchical linkage in early warning and the lag in operation and maintenance response: Most existing early warning systems only output a single early warning signal and do not combine the fault level to realize hierarchical early warning and linkage with the robot control system. They cannot perform targeted operations such as load reduction and shutdown according to the severity of the fault, which can easily amplify the fault loss. Summary of the Invention

[0003] In view of this, the purpose of this invention is to overcome the shortcomings of existing technologies in industrial robot electromechanical fault detection, such as single dimension, difficulty in early fault capture, premature convergence of algorithms, poor adaptability to changing working conditions, and insufficient early warning linkage. This invention provides a real-time detection and early warning system for industrial robot electromechanical faults. Through multimodal sensor fusion, improved intelligent algorithms, and hierarchical linkage early warning design, this system achieves early detection, accurate identification, hierarchical early warning, and emergency linkage of electromechanical faults, thereby reducing fault losses and improving the operation and maintenance efficiency of industrial robots.

[0004] The technical solution of this invention is implemented as follows: A real-time detection and early warning system for electromechanical faults in industrial robots includes a multimodal sensing and acquisition module, a signal preprocessing module, an improved chaotic particle swarm feature extraction module, an LSTM-GRU fusion fault identification module, a hierarchical early warning and linkage module, a data storage module, and a variable working condition adaptive adjustment module. The multimodal sensing acquisition module is used to acquire multi-dimensional raw signals from the core electromechanical components of the industrial robot and transmit them synchronously. The signal preprocessing module is used to perform noise suppression, signal repair and standardization on the original signal, and output a clean signal; The improved chaotic particle swarm feature extraction module is used to extract effective features from the purified signal and remove redundant features; The LSTM-GRU fusion fault identification module is used to identify fault types and determine fault levels from the extracted feature vectors. The graded early warning linkage module is used to output corresponding early warning signals according to the fault level and link the industrial robot control system to perform emergency response operations. The data storage module is used to store all process data; The variable operating condition adaptive adjustment module is used to mitigate the impact of operating condition changes on fault identification accuracy and dynamically adjust model parameters. The modules work together in sequence to form a closed loop of signal acquisition, preprocessing, feature extraction, fault identification, early warning linkage, data storage, and model optimization.

[0005] Preferably, the multimodal sensing acquisition module includes a fiber optic stress-strain sensor, a directional capacitive acoustic sensor, and a WGM optical calibration sensor. The fiber optic stress-strain sensor is deployed on the joints and reducer housing of the industrial robot, with a sampling frequency of 1000Hz. The directional capacitive acoustic sensor is deployed near the motor and reducer, with a sampling frequency of 20kHz. The WGM optical calibration sensor is linked with current and temperature sensors, with a sampling frequency of 500Hz, and is used for high-precision calibration of current and temperature signals.

[0006] Preferably, the signal preprocessing module includes the following steps: using bandpass filtering (frequency range 10-1000Hz) and spectral subtraction to suppress noise, using KNN interpolation algorithm to fill in missing signal values, and using a standardization formula. The signal is standardized to map the signal amplitude to the [0,1] interval.

[0007] Preferably, the improvements to the improved chaotic particle swarm feature extraction module include: using Logistic chaotic mapping for particle initialization, the mathematical model of which is as follows: The inertia weight is adjusted using an exponential principle, and the adjustment formula is as follows: ; When the group fitness variance is satisfied, Gaussian expansion mutation is performed on the global extreme value, and the mutation formula is , where τ follows the Gauss(0,1) distribution.

[0008] Preferably, the fitness function of the improved chaotic particle swarm feature extraction module is , where S(x) is the signal-to-noise ratio of the signal, and E(x) is the feature entropy, which is used to optimize feature parameters and remove redundant features.

[0009] Preferably, the LSTM-GRU fusion fault identification module adopts a bidirectional LSTM and GRU fusion structure, including an input layer, a feature fusion layer, and an output layer; the input layer constructs time series samples through a sliding window (window size 60s, step size 30s); the feature fusion layer captures time series features through forward LSTM and backward LSTM, and the GRU layer simplifies and fuses the features; the output layer adopts a dual-task design, the classification head outputs the fault type, and the regression head outputs the fault level. At the same time, the recognition accuracy is calculated through the formula

[0010] Preferably, the hierarchical early warning linkage module adopts a three-level early warning mechanism: when RMSE ≤ 1.5, a yellow early warning is output to monitor the fault evolution in real time; when 1.5 < RMSE ≤ 3.0, an orange early warning is output, and the robot is linked to perform a load reduction operation; when RMSE > 3.0, a red early warning is output, the robot is linked to stop immediately, and the on-site sound and light alarm and operation and maintenance notification are triggered.

[0011] Preferably, the data storage module adopts a dual storage mode of edge server + cloud storage. The edge server stores the real-time data and fault information of the past 30 days, and the cloud server stores the historical data of 1 year, including the original signal, preprocessed signal, feature vector, fault identification result, and early warning record.

[0012] Preferably, the variable working condition adaptive adjustment module adopts domain adaptation technology to alleviate the problem of unstable recognition caused by changes in speed and load through minimizing the perception loss, quantify the difficulty of the migration task, and dynamically adjust the parameters of the LSTM-GRU fusion model to ensure the stability of the fault recognition accuracy under variable working conditions.

[0013] Preferably, the system uses the cloud historical fault data every quarter to iteratively optimize the improved chaotic particle swarm algorithm and the LSTM-GRU fusion model, update the feature extraction parameters and model weights, and improve the fault detection and recognition accuracy.

[0014] Due to the adoption of the above technical solutions in the embodiments of the present invention, it has the following advantages: ​I. Addressing the pain points of "single detection dimension and weak early fault detection capability": An innovative multimodal sensing fusion solution is adopted, combining fiber optic grating sensing, directional capacitive acoustic sensing, and WGM optical calibration sensors. Non-contact and contact sensing work together to accurately capture weak fault signals such as joint micro-cracks, slight insulation aging in motor stators, and early wear of reducer gears. This effectively avoids the masking of fault signals by industrial environmental noise and solves the problem of delayed fault detection. The WGM optical calibration sensor amplifies changes in weak signals, improving the accuracy of the original signal. This sensor is rarely used in industrial robot fault detection. Furthermore, the electromagnetic interference resistance of the fiber optic grating sensor is suitable for strong electromagnetic environments in industry, further improving detection reliability.

[0015] Second, addressing the pain points of "algorithms easily getting trapped in local optima and low fault identification accuracy": Innovative improvements are made to the chaotic particle swarm optimization (PSO) algorithm. Through three improvements—Logistic chaotic initialization, inertia weight exponential adjustment, and Gaussian expansion mutation—the problems of premature convergence and inability to effectively optimize feature extraction and fault identification parameters in the traditional PSO algorithm are completely solved. Combined with the LSTM-GRU fusion model and domain adaptation technology, the fault identification accuracy is significantly improved, increasing the fault identification accuracy to over 99% with RMSE ≤ 1.55. The false fault rate is significantly reduced, far superior to the traditional PSO combined with BP neural network scheme, meeting the high-precision detection needs of industry.

[0016] Third, addressing the pain points of "poor adaptability to changing working conditions and insufficient generalization ability": a domain adaptive adjustment mechanism is introduced into the LSTM-GRU fusion fault identification module. By minimizing perception loss, the instability of identification caused by continuous changes in the speed and load of industrial robots is alleviated. The difficulty of the transfer task is quantified, and the model parameters are dynamically adjusted to ensure that the fault identification accuracy remains stable in complex working conditions such as only the load changes, only the speed changes, or both change simultaneously. This effectively adapts to complex industrial production scenarios and solves the deficiency of insufficient generalization ability of existing data-driven detection methods.

[0017] Fourth, addressing the pain points of "lack of hierarchical linkage in early warning and delayed operation and maintenance response": a three-level hierarchical early warning linkage mechanism is designed. The fault level is determined based on the RMSE value of fault identification, and corresponding yellow, orange and red three-level early warning signals are output. It is also linked with the industrial robot control system and operation and maintenance terminal. Targeted emergency response operations such as "real-time monitoring, load reduction operation and immediate shutdown" can be performed according to the severity of the fault to avoid the fault from escalating, reduce fault losses, and solve the problem that the existing early warning system only outputs a single early warning signal and the operation and maintenance response is delayed.

[0018] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a closed-loop diagram of the system workflow of the present invention; Figure 3 This is the core algorithm and detection logic diagram of the present invention. Detailed Implementation

[0021] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0022] It is important to note that terms such as "first," "second," "symmetric," and "array" are used only to distinguish between descriptive and positional descriptions and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, features specified with terms such as "first" or "symmetric" may explicitly or implicitly include one or more of that feature; similarly, when the quantity of certain features is not limited by words such as "two" or "three," it should be noted that such features also explicitly or implicitly include one or more features. In this invention, unless otherwise explicitly specified and limited, terms such as "installation," "connection," and "fixation" should be interpreted broadly; for example, they can refer to a fixed connection, a detachable connection, or an integral molding; they can refer to a mechanical connection, a direct connection, a welding connection, or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the accompanying drawings and specific circumstances.

[0023] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] like Figure 1-3 As shown, this invention provides a real-time detection and early warning system for electromechanical faults in industrial robots, including a multimodal sensing acquisition module, a signal preprocessing module, an improved chaotic particle swarm feature extraction module, an LSTM-GRU fusion fault identification module, a hierarchical early warning linkage module, a data storage module, and a variable operating condition adaptive adjustment module. These modules are linked sequentially to form a closed-loop process of "signal acquisition - preprocessing - feature extraction - fault identification - early warning linkage - data traceability." The specific steps are as follows: Step 1: Multimodal sensing data acquisition Multimodal sensing units are deployed on the core electromechanical components of the industrial robot (joint motor stator, reducer housing, servo driver output, and transmission bearing). These units include fiber optic stress-strain sensors, directional capacitive acoustic sensors, and WGM (whispering-gallery mode) optical calibration sensors, which work together. (1) Fiber grating stress and strain sensor: deployed on the robot joint and reducer housing, with a sampling frequency of 1000Hz, used to collect stress and strain signals during the operation of electromechanical components, capture stress changes caused by joint microcracks and reducer gear wear, and its anti-electromagnetic interference characteristics are suitable for industrial strong electromagnetic environment. (2) Directional capacitive acoustic sensor: Deployed near the motor and reducer, with a sampling frequency of 20kHz, using a non-contact acquisition method to capture weak acoustic emission signals (such as abnormal gear meshing and noise generated by bearing friction) during the operation of electromechanical components, avoiding interference of contact sensing on robot movement; (3) WGM optical calibration sensor: linked with current and temperature sensors, with a sampling frequency of 500Hz, used to perform high-precision calibration on the collected current and temperature signals. It utilizes its high quality factor characteristics to amplify weak signal changes and improve the accuracy of the original signal.

[0025] The signals collected by the three types of sensors (stress-strain signal, acoustic signal, and calibrated current / temperature signal) are synchronously transmitted to the signal preprocessing module to ensure the time synchronization of multi-dimensional data and provide comprehensive and accurate raw data for subsequent feature extraction.

[0026] Step 2: Signal Preprocessing The signal preprocessing module performs layered purification processing on the acquired multimodal raw signals, specifically including three sub-steps, which proceed sequentially: (1) Noise suppression: Bandpass filtering (frequency range 10-1000Hz) is used to filter out low-frequency environmental interference and high-frequency electronic noise. Combined with spectral subtraction, non-stationary background noise is dynamically suppressed. An adaptive filtering algorithm is used for acoustic signals to further eliminate industrial noise interference. (2) Signal repair: The KNN interpolation algorithm is used to fill in the missing values ​​that occur during the sensor acquisition process, so as to avoid the feature extraction deviation caused by missing data and ensure data integrity; (3) Standardization processing: The repaired signal is normalized to map the signal amplitude to the [0,1] interval, eliminating the dimensional differences between signals from different sensors. The formula is as follows: in, The standardized signal value, The original signal value, , These are the maximum and minimum values ​​of this type of signal, respectively.

[0027] The preprocessed purified signal is transmitted to the improved chaotic particle swarm feature extraction module, which solves the problems of low feature extraction accuracy caused by high noise, inconsistent dimensions, and missing data in the original signal.

[0028] Step 3: Improve the feature extraction of chaotic particle swarm optimization An improved chaotic particle swarm optimization (CPSO) algorithm is used to extract features from the preprocessed multimodal signal. Feature parameters are optimized and redundant features are removed. The specific improvements and extraction steps are as follows: (1) Algorithm Improvement: To address the shortcomings of the traditional PSO algorithm, such as premature convergence and susceptibility to local optima, three improvements are made to ensure the accuracy of feature extraction: Chaotic initialization: Logistic chaotic mapping is used to initialize the particle swarm, improving population diversity and avoiding local optima caused by uneven initial particle distribution. The mathematical model is as follows: Where r is the chaotic attractor, and when r=4, the system is in a completely chaotic state. ; Inertia weight exponential adjustment: The inertia weight ω is adjusted exponentially to achieve a balance between global and local search. The adjustment formula is as follows: in, , These are the initial and final values ​​of the inertia weight (take...). , ), k is the current iteration number, and K is the maximum iteration number (take K=1000); Gaussian expansion mutation: when the population fitness variance (C is the threshold, taken as C=0.01) When the algorithm is found to be trapped in a local optimum, it is determined that the algorithm is not at the global optimum. Gaussian expansion mutation is performed to perturb the direction of particle motion and escape local optima. The mutation formula is as follows: Where τ is a random variable that follows a Gauss(0,1) distribution.

[0029] (2) Feature extraction: The time-domain features (mean, peak value, kurtosis, zero-crossing rate), frequency-domain features (centroid of the spectrum, power spectral density), and time-frequency-domain features (wavelet packet decomposition coefficients) of the signal are used as extraction targets. The feature parameters are optimized by improving the fitness function of the CPSO algorithm and removing redundant features. The fitness function is designed as follows: Where S(x) is the signal-to-noise ratio and E(x) is the feature entropy. , These are the weighting coefficients (0.6 and 0.4 respectively). The higher the fitness value, the stronger the feature effectiveness.

[0030] The extracted effective feature vectors are transmitted to the LSTM-GRU fusion fault identification module, which solves the problems of redundancy and low accuracy in traditional feature extraction and provides core support for accurate fault identification.

[0031] Step 4: LSTM-GRU Fusion Fault Identification A bidirectional LSTM-GRU fusion model is used to identify faults from the extracted feature vectors. Combined with a domain adaptive adjustment mechanism, the identification accuracy under varying operating conditions is improved. The specific steps are as follows: (1) Model structure design: The fusion model is divided into three layers to achieve temporal capture and simplified fusion of features: Input layer: The effective feature vectors are standardized and time series samples are constructed through a sliding window (window size 60s, step size 30s) to preserve the temporal correlation of fault precursors; Feature fusion layer: The forward LSTM layer captures the fault evolution features from history to the present, the backward LSTM layer mines the impact of future trends on the current state, and the GRU layer simplifies and fuses the features extracted by the bidirectional LSTM to reduce redundancy. At the same time, Dropout and LayerNorm are added to prevent overfitting. Output layer: Employing a dual-task design, the classification head outputs the fault type (joint microcracks, motor insulation aging, reducer gear wear, bearing failure, servo driver failure), while the regression head outputs the fault level (minor, moderate, severe). Simultaneously, the root mean square error (RMSE) of fault prediction is calculated to evaluate the recognition accuracy, using the following formula: Where n is the number of samples. This represents the actual fault level value. To predict fault level values.

[0032] (2) Adaptive adjustment under changing working conditions: Introducing domain adaptive technology, the recognition instability caused by changes in working conditions (speed and load changes) is alleviated by minimizing perception loss, quantifying the difficulty of the transfer task, dynamically adjusting the model parameters, and ensuring that the recognition accuracy remains stable in scenarios where only the load changes, only the speed changes, or both change simultaneously.

[0033] The fault identification results (fault type, fault level, identification accuracy) are transmitted to the hierarchical early warning linkage module, which solves the problems of low identification accuracy of a single model and poor adaptability to variable working conditions, and realizes the accurate positioning and level determination of faults.

[0034] Step 5: Hierarchical early warning linkage According to the fault identification results, the hierarchical early warning linkage module realizes three-level early warning, and links the industrial robot control system and the operation and maintenance terminal to form a closed-loop disposal. The specific logic is as follows: (1) Minor fault (RMSE ≤ 1.5, the fault does not affect the normal operation of the robot): Output a yellow early warning signal, and the operation and maintenance terminal pushes a fault prompt (including fault type and location). The system does not interfere with the operation of the robot, but monitors the fault evolution trend in real time, and updates the monitoring data every 10 s; (2) Moderate fault (1.5 < RMSE ≤ 3.0, the fault may affect the operation accuracy): Output an orange early warning signal, link the robot control system to perform a load reduction operation (the load is reduced to 60% of the rated load), and the operation and maintenance terminal pushes an emergency disposal prompt, requiring the operation and maintenance personnel to arrive at the scene for investigation within 30 minutes; (3) Severe fault (RMSE > 3.0, the fault may cause component damage or shutdown): Output a red early warning signal, link the robot control system to stop immediately, cut off the power supply of the faulty component, the operation and maintenance terminal pushes an emergency shutdown notice and fault location information, and at the same time trigger on-site sound and light alarms to ensure that the fault does not expand.

[0035] Step 6: Data storage and model optimization The data storage module adopts a dual storage mode of edge server + cloud storage. The edge server stores the real-time monitoring data and fault identification results of the past 30 days for rapid on-site traceability; the cloud stores historical data (storage period: 1 year), including original signals, preprocessed signals, feature vectors, fault identification results, early warning records, etc.

[0036] Every quarter, the system uses the historical fault data in the cloud to iteratively optimize the LSTM-GRU fusion model and the improved CPSO algorithm, update the feature extraction parameters and model weights, further improve the fault detection and identification accuracy, and adapt to the performance degradation during the long-term operation of industrial robots.

[0037] Linkage logic of each module The multimodal sensing acquisition module acquires raw signals → the signal preprocessing module purifies the signals → the improved chaotic particle swarm feature extraction module extracts effective features → the LSTM-GRU fusion fault identification module identifies fault types and levels → the hierarchical early warning and linkage module outputs early warnings and links for handling → the data storage module stores the entire process data → the model is iteratively optimized based on historical data, forming a closed loop of "acquisition-processing-identification-early warning-storage-optimization" to ensure stable and efficient system operation.

[0038] In this embodiment, the present invention operates as follows: Taking the FANUCK RC4 six-DOF industrial robot as the application object, real-time fault detection and early warning are carried out for its core electromechanical components such as joint motors, reducers, and servo drives. The specific workflow is as follows: First, the multimodal sensing acquisition module starts working. FBG-2000 fiber optic stress-strain sensors deployed on the stators of the robot's six joint motors acquire joint stress-strain signals at a frequency of 1000Hz. Two directional capacitive acoustic sensors MIC-3000 deployed on the reducer housing acquire weak acoustic emission signals from electromechanical components non-contactly at a frequency of 20kHz. WGM-100 optical calibration sensors deployed at the output of the servo drive, along with ACS712 current sensors and DS18B20 temperature sensors, perform high-precision calibration of current and temperature signals at a frequency of 500Hz. All multi-dimensional raw signals acquired by the sensors are synchronously transmitted to the signal preprocessing module in the NVIDIA Jetson Xavier NX edge server through a data acquisition card NIcDAQ-9178.

[0039] Subsequently, the signal preprocessing module performs layered purification on the received raw signal. First, it uses 10-1000Hz bandpass filtering and spectral subtraction to filter out low-frequency environmental interference, high-frequency electronic noise, and non-stationary background noise. For acoustic signals, it uses adaptive filtering to further eliminate clutter. Then, it uses the KNN interpolation algorithm to fill in the missing values ​​of the signal. Finally, it uses a standardization formula to map the signal amplitude to the [0,1] interval to eliminate dimensional differences, and obtains the purified multimodal signal, which is then transmitted to the improved chaotic particle swarm feature extraction module.

[0040] Next, the improved chaotic particle swarm feature extraction module starts the improved algorithm. The particle swarm is initialized with the Logistics chaotic mapping r = 4. The initial value of the inertia weight 0.9 and the final value 0.4 are adjusted according to the exponential law, and the maximum number of iterations is 1000. When the variance of the population fitness is less than 0.01, Gaussian expansion mutation is performed on the global extreme value. The feature parameters are optimized through the fitness function f(x)=0.6・S(x)+0.4・E(x), and the effective features such as the mean value, peak value, spectral centroid, and wavelet packet decomposition coefficients of the signal are extracted. After removing the redundant features, the effective feature vector is transmitted to the LSTM-GRU fusion fault identification module.

[0041] After that, the LSTM-GRU fusion fault identification module processes the feature vector. The time series samples are constructed through a sliding window with a 60s window and a 30s step. The bidirectional LSTM layer captures the time series fault features, and the GRU layer simplifies and fuses the features and prevents overfitting through the Dropout coefficient 0.3 and LayerNorm. The output layer's dual tasks respectively identify the fault types such as joint microcracks and motor insulation aging, and the fault levels such as mild, moderate, and severe. At the same time, the RMSE value is calculated to evaluate the identification accuracy; the variable working condition adaptive adjustment module works synchronously. Through domain adaptation technology, it senses the changes in speed and load, and dynamically adjusts the model parameters to ensure the stability of the identification accuracy under variable working conditions.

[0042] After the identification is completed, the hierarchical warning linkage module performs corresponding warning and linkage operations according to the RMSE value: when RMSE≤1.5 for mild faults, such as slight wear of the reducer gear, a yellow warning is output, and the operation and maintenance terminal pushes a fault prompt. The system monitors the fault evolution in real time and updates the monitoring data every 10s; when 1.5<RMSE≤3.0 for moderate faults, such as motor insulation aging, an orange warning is output, and the linkage robot control system reduces the load to 60% of the rated load, pushes an emergency handling prompt, and requires the operation and maintenance personnel to arrive at the scene for investigation within 30 minutes; when RMSE>3.0 for severe faults, such as bearing jamming, a red warning is output, and the linkage robot control system immediately shuts down and cuts off the power supply of the faulty component, triggers the on-site sound and light alarm, and pushes the emergency shutdown and fault location information to the operation and maintenance terminal.

[0043] Meanwhile, the data storage module starts a dual storage mode. The edge server with a 1TB storage capacity stores the real-time monitoring data, fault identification results, and warning records for nearly 30 days for quick on-site traceability; the cloud server Alibaba Cloud ECS stores the historical data for 1 year, including the full-process data such as the original signal, preprocessed signal, and feature vector. Every quarter, the system uses the historical fault data in the cloud to iteratively optimize the improved chaotic particle swarm algorithm and the LSTM-GRU fusion model, updates the feature extraction parameters and model weights, continuously improves the fault detection and identification accuracy, adapts to the performance attenuation of the robot during long-term operation, and forms a complete working loop.

[0044] The following are several other specific embodiments of the application of this invention: Example 1: SCARA Robot Application Scenarios This embodiment uses a SCARA robot as the application object to carry out real-time detection and early warning of electromechanical faults in its horizontal joint motor, harmonic reducer, and servo driver. The core technical solution is consistent with the present invention, and the specific implementation details are as follows: System Deployment (1) Deployment of multimodal sensing acquisition module: One fiber optic stress-strain sensor (FBG-2000) is installed on each of the two horizontal joint motor stators of the SCARA robot, one directional capacitive acoustic sensor (MIC-3000) is installed on the housing of the harmonic reducer, and one WGM optical calibration sensor (WGM-100) is installed at the output of the servo driver. At the same time, a current sensor (ACS712) and a temperature sensor (DS18B20) are linked. All sensors are connected to the edge server (NVIDIA Jetson Nano) through a data acquisition card (NIcDAQ-9178) to ensure synchronous signal acquisition. (2) Module hardware configuration: The signal preprocessing module, the improved chaotic particle swarm feature extraction module, and the LSTM-GRU fusion fault identification module are all deployed on the edge server. The hierarchical early warning linkage module is connected to the SCARA robot control system via Ethernet. The data storage module adopts a dual storage mode of 500GB edge server storage capacity plus Alibaba Cloud ECS cloud server. The operation and maintenance terminal adopts an industrial computer that supports wireless communication.

[0045] System operation process (1) Signal acquisition: The fiber optic grating sensor acquires the horizontal joint stress and strain signal at a frequency of 1000Hz, the acoustic sensor acquires the weak acoustic emission signal at a frequency of 20kHz, and the WGM optical calibration sensor calibrates the current and temperature signal at a frequency of 500Hz, which are then synchronously transmitted to the signal preprocessing module.

[0046] (2) Signal preprocessing: bandpass filtering of 10-1000Hz is used to filter out noise, spectral subtraction is used to suppress background clutter, KNN interpolation is used to fill missing values, and the signal amplitude is mapped to the [0,1] interval through the standardization formula to obtain the purified multimodal signal.

[0047] (3) Feature extraction: Start the improved chaotic particle swarm algorithm, set the Logistic chaotic mapping r=4, the initial value of the inertial weight is 0.9 and the final value is 0.4, the maximum number of iterations is 1000, optimize the feature parameters through the fitness function f(x)=0.6・S(x)+0.4・E(x), extract the effective features such as the mean, peak value, spectral centroid, and wavelet packet decomposition coefficients of the signal, and remove redundant features to obtain the feature vector.

[0048] (4) Fault identification: Input the feature vector into the bidirectional LSTM-GRU fusion model, with a sliding window size of 60s and a step size of 30s, and a Dropout coefficient of 0.3. The model outputs the fault type and level, and calculates the RMSE value. If a minor wear fault of the harmonic reducer is detected, RMSE=1.3, and a yellow warning is output. If a moderate insulation aging fault of the horizontal joint motor is detected, RMSE=2.5, an orange warning is output and the load is reduced. If a severe fault of the servo drive is detected, RMSE=3.2, a red warning is output and the machine is stopped.

[0049] (5) Early warning linkage: The operation and maintenance terminal receives early warning information and fault location, monitors minor faults in real time, investigates moderate faults within 30 minutes, and immediately shuts down and handles serious faults.

[0050] (6) Data storage and optimization: The edge server stores data for the past 30 days, and the cloud stores historical data. The model parameters are iteratively optimized every quarter using historical fault data to improve recognition accuracy.

[0051] Implementation effect In this embodiment, the system operates stably, with a fault detection response time of ≤25ms, a fault identification accuracy of 98.8%, and an RMSE of 1.58. It can detect early minor faults 65 hours in advance. Under varying operating conditions (speed 80-400r / min, load 40%-100%), the identification accuracy remains above 97.5%. The hierarchical early warning and linkage mechanism reduces downtime by 75% and maintenance costs by 60%. It is adapted to the high-speed and high-precision operation requirements of SCARA robots, verifying the versatility and feasibility of the technical solution of this invention.

[0052] Example 2: Parallel Robot Application Scenarios This embodiment uses a parallel robot as the application object to carry out real-time detection and early warning of electromechanical faults in its moving platform joint motors, planetary reducers, and servo drives. The core technical solution is consistent with the present invention, and the specific implementation details are as follows: 4.1 System Deployment (1) Deployment of multimodal sensing acquisition module: One fiber optic stress-strain sensor (FBG-2000) is installed on the stator of each of the three moving platform joint motors of the parallel robot. Two directional capacitive acoustic sensors (MIC-3000) are installed on the planetary reducer housing. One WGM optical calibration sensor (WGM-100) is installed at the output of the servo driver. At the same time, a current sensor (ACS712) and a temperature sensor (DS18B20) are linked. All sensors are connected to the edge server (NVIDIA Jetson AGX Orin) through a data acquisition card (NIcDAQ-9178) to ensure synchronous signal acquisition. (2) Module hardware configuration: The signal preprocessing module, the improved chaotic particle swarm feature extraction module, and the LSTM-GRU fusion fault identification module are all deployed on the edge server. The hierarchical early warning linkage module is connected to the parallel robot control system via Ethernet. The data storage module adopts a dual storage mode of 1TB edge server storage capacity plus Alibaba Cloud ECS cloud server. The operation and maintenance terminal adopts an industrial tablet that supports wireless communication.

[0053] 4.2 System Operation Process (1) Signal acquisition: The fiber optic grating sensor acquires the joint stress and strain signal of the moving platform at a frequency of 1000Hz, the acoustic sensor acquires the weak acoustic emission signal at a frequency of 20kHz, and the WGM optical calibration sensor calibrates the current and temperature signal at a frequency of 500Hz, which are then synchronously transmitted to the signal preprocessing module.

[0054] (2) Signal preprocessing: bandpass filtering of 10-1000Hz is used to filter out noise, spectral subtraction is used to suppress background clutter, KNN interpolation is used to fill missing values, and the signal amplitude is mapped to the [0,1] interval through the standardization formula to obtain the purified multimodal signal.

[0055] (3) Feature extraction: Start the improved chaotic particle swarm algorithm, set the Logistic chaotic mapping r=4, the initial value of the inertial weight is 0.9 and the final value is 0.4, the maximum number of iterations is 1000, optimize the feature parameters through the fitness function f(x)=0.6・S(x)+0.4・E(x), extract the effective features such as the mean, peak value, spectral centroid, and wavelet packet decomposition coefficients of the signal, and remove redundant features to obtain the feature vector.

[0056] (4) Fault identification: Input the feature vector into the bidirectional LSTM-GRU fusion model, with a sliding window size of 60s and a step size of 30s, and a Dropout coefficient of 0.3. The model outputs the fault type and level, and calculates the RMSE value. If a minor fault of slight wear of planetary reducer gear is detected, RMSE=1.4, and a yellow warning is output. If a moderate fault of wear of joint motor bearing of moving platform is detected, RMSE=2.7, and an orange warning is output and the load is reduced. If a serious fault of short circuit of servo drive is detected, RMSE=3.6, and a red warning is output and the machine is stopped.

[0057] (5) Early warning linkage: The operation and maintenance terminal receives early warning information and fault location, monitors minor faults in real time, investigates moderate faults within 30 minutes, and immediately shuts down and handles serious faults.

[0058] (6) Data storage and optimization: The edge server stores data for the past 30 days, and the cloud stores historical data. The model parameters are iteratively optimized every quarter using historical fault data to improve recognition accuracy.

[0059] 4.3 Implementation Results In this embodiment, the system operates stably, with a fault detection response time of ≤22ms, a fault identification accuracy of 99.0%, and an RMSE of 1.56. It can detect early minor faults 68 hours in advance. Under varying operating conditions (speed 100-500r / min, load 50%-100%), the identification accuracy remains above 98%. The graded early warning and linkage mechanism reduces downtime by 78% and maintenance costs by 62%. It is suitable for the heavy-load and high-rigidity operation requirements of parallel robots, further verifying the versatility and feasibility of the technical solution of this invention.

[0060] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in the present invention, and these should all be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A real-time detection and early warning system for electromechanical faults in industrial robots, characterized in that, It includes a multi-modal sensing and acquisition module, a signal preprocessing module, an improved chaotic particle swarm feature extraction module, an LSTM-GRU fusion fault identification module, a hierarchical early warning and linkage module, a data storage module, and a variable working condition adaptive adjustment module; The multi-modal sensing and acquisition module is used to collect multi-dimensional original signals of the core electromechanical components of the industrial robot and transmit them synchronously; The signal preprocessing module is used to perform noise suppression, signal repair, and normalization processing on the original signals, and output purified signals; The improved chaotic particle swarm feature extraction module is used to effectively extract features from the purified signals and remove redundant features; The LSTM-GRU fusion fault identification module is used to identify the fault type and determine the fault level of the extracted feature vectors; The hierarchical early warning and linkage module is used to output corresponding early warning signals according to the fault level, and link the industrial robot control system to perform emergency disposal operations; The data storage module is used to store the whole-process data; The variable working condition adaptive adjustment module is used to alleviate the influence of working condition changes on the fault identification accuracy and dynamically adjust the model parameters; Each module is linked in sequence to form a closed loop of signal acquisition, preprocessing, feature extraction, fault identification, early warning linkage, data storage, and model optimization.

2. The system according to claim 1, characterized in that, The multi-modal sensing and acquisition module includes a fiber Bragg grating stress and strain sensor, a directional capacitive acoustic sensor, and a WGM optical calibration sensor. The fiber Bragg grating stress and strain sensor is deployed on the joints and reducer casings of the industrial robot, and the sampling frequency is 1000Hz; the directional capacitive acoustic sensor is deployed near the motor and reducer, and the sampling frequency is 20kHz; the WGM optical calibration sensor is linked with current and temperature sensors, and the sampling frequency is 500Hz, which is used to perform high-precision calibration on current and temperature signals.

3. The system according to claim 1, characterized in that, The signal preprocessing module includes the following steps: noise suppression using bandpass filtering (frequency range 10-1000Hz) and spectral subtraction; filling in missing signal values ​​using the KNN interpolation algorithm; and applying a standardization formula. The signal is standardized to map the signal amplitude to the [0,1] interval.

4. The system according to claim 1, characterized in that, The improvements to the improved chaotic particle swarm feature extraction module include: using Logistic chaotic mapping for particle initialization, the mathematical model of which is as follows: The inertia weight is adjusted using an exponential principle, and the adjustment formula is as follows: When the group fitness variance When the global extremum is reached, Gaussian expansion and mutation are performed, and the mutation formula is: , where τ follows a Gauss(0,1) distribution.

5. The system according to claim 1, characterized in that, The fitness function of the improved chaotic particle swarm feature extraction module is: Where S(x) is the signal-to-noise ratio and E(x) is the feature entropy, used to optimize feature parameters and remove redundant features.

6. The system according to claim 1, characterized in that, The LSTM-GRU fusion fault identification module adopts a bidirectional LSTM and GRU fusion structure, including an input layer, a feature fusion layer, and an output layer; the input layer constructs time series samples through a sliding window (window size 60s, step size 30s); The feature fusion layer captures time series features through forward LSTM and reverse LSTM, and the GRU layer simplifies and fuses the features; the output layer adopts a dual-task design, the classification head outputs the fault type, the regression head outputs the fault level, and at the same time through the formula Calculate the recognition accuracy.

7. The system according to claim 1, characterized in that, The hierarchical early warning and linkage module adopts a three-level early warning mechanism: when RMSE≤1.5, it outputs a yellow early warning and monitors the fault evolution in real time; when 1.5<RMSE≤3.0, it outputs an orange early warning and links the robot to perform a load reduction operation; when RMSE>3.0, it outputs a red early warning, links the robot to stop immediately, and triggers on-site sound and light alarms and operation and maintenance notifications.

8. The system according to claim 1, characterized in that, The data storage module adopts a dual storage mode of edge server + cloud storage. The edge server stores real-time data and fault information for nearly 30 days, and the cloud server stores historical data for 1 year, including original signals, preprocessed signals, feature vectors, fault identification results, and early warning records.

9. The system according to claim 1, characterized in that, The variable operating condition adaptive adjustment module adopts domain adaptive technology to alleviate the recognition instability caused by changes in speed and load by minimizing perception loss, quantifies the difficulty of the transfer task, and dynamically adjusts the parameters of the LSTM-GRU fusion model to ensure stable fault recognition accuracy under variable operating conditions.

10. The system according to any one of claims 1-9, characterized in that, The system uses historical fault data from the cloud every quarter to iteratively optimize the improved chaotic particle swarm optimization algorithm and the LSTM-GRU fusion model, update feature extraction parameters and model weights, and improve the accuracy of fault detection and identification.

Images