A machine learning based predictive maintenance system for industrial robots

Abstract

The application discloses an industrial robot predictive maintenance system based on machine learning, and relates to the technical field of industrial robot maintenance.The application comprises a data acquisition module, a preprocessing module, a feature engineering module, a hybrid prediction model unit, a meta-learning and field adaptation module, a federal learning coordination module, a digital twin sample generation module and a dynamic decision engine unit.The application has the following advantages: through the collaborative mechanism of the meta-learning and field adaptation module, a model-agnostic meta-learning algorithm is used to train a cross-brand general feature extractor;different brand data distributions are confused by combining an adversarial field adaptation network;the "data gap" caused by the definition of sensor parameters and the difference in sampling frequency in traditional modeling is broken through;the model is collaboratively trained under the premise that the data of each brand is not out of the local;the global model containing the commonness of cross-brand faults is generated;and the "data island" problem caused by commercial security requirements in the industrial scene is solved.

Description

A predictive maintenance system for industrial robots based on machine learning Technical Field

[0001] This invention relates to the field of industrial robot maintenance technology, specifically to a predictive maintenance system for industrial robots based on machine learning. Background Technology

[0002] As the core equipment of intelligent manufacturing, industrial robots have been widely used in industrial scenarios such as automobile manufacturing, electronic assembly, and aerospace, undertaking key tasks such as welding, handling, and precision machining. The stability of their operation directly affects the efficiency of the production line, product quality, and safe production. Therefore, an efficient maintenance system is one of the core requirements in the field of industrial automation.

[0003] Existing technologies have certain shortcomings. First, they are difficult to extract unified fault features from heterogeneous data of multi-brand robots, resulting in "data gap" and "data silos" problems and insufficient model generalization ability. Second, existing technologies rely on a large amount of real fault data. When there is a lack of data for new or niche brand robots, it is difficult to achieve early warning and the sample requirement is large. To address these issues, we propose a predictive maintenance system for industrial robots based on machine learning. Summary of the Invention

[0004] The purpose of this invention is to provide a predictive maintenance system for industrial robots based on machine learning.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a predictive maintenance system for industrial robots based on machine learning, the maintenance system comprising:

[0006] Data acquisition module: Real-time acquisition of timing data such as joint torque, motor temperature and vibration frequency from multi-brand robot controllers including ABB and KUKA via Modbus and OPCUA industrial communication protocols, and transmission to the preprocessing module;

[0007] Preprocessing module: Performs digital filtering and outlier detection on the raw data, removes electromagnetic interference and transmission delay noise, and outputs clean data to the feature engineering module;

[0008] Feature engineering module: Extracts time-domain features, frequency-domain features, and time-frequency-domain features from cleaning data to generate equipment health feature vectors, which are then input into the hybrid prediction model unit and the meta-learning and domain adaptation module, respectively.

[0009] Hybrid Prediction Model Unit: A hybrid prediction model is set up, consisting of a base layer and a prediction layer. The base layer identifies multivariate relationships through the random forest algorithm, while the prediction layer learns long-term dependencies of time-series data based on the long short-term memory network and outputs health scores and remaining life prediction values ​​to the dynamic decision engine unit.

[0010] Meta-learning and Domain Adaptation Module: The module uses a model-agnostic meta-learning algorithm to train a cross-brand general fault feature extractor. It uses an adversarial domain adaptive network to obfuscate the data distribution of different brands, generate a cross-domain unified feature vector, and feeds it back to the hybrid prediction model unit to optimize generalization ability.

[0011] Federated learning coordination module: Receives sub-model parameter updates trained locally by robots from various brands, generates a global generalization model and a cross-brand maintenance case library through a differential privacy-preserving weighted aggregation algorithm, and synchronizes it to the hybrid prediction model unit at the edge.

[0012] Digital twin sample generation module: Based on the robot dynamics equations, a digital twin is constructed to simulate fault modes such as gear wear and bearing failure, and synthetic fault samples containing physical meaning are generated and input into the feature engineering module to enhance small sample training;

[0013] Dynamic Decision Engine Unit: Based on health scores, remaining life expectancy predictions, and a cross-brand maintenance case library generated through federated learning, the system sets multi-level early warning thresholds through a rules engine, generates maintenance suggestions, and pushes them to a visual interactive interface.

[0014] As a further aspect of the present invention: the data acquisition module supports an adjustable sampling frequency from the second level to the minute level, and is compatible with the proprietary communication protocols of ABB Robot Ware and KUKAKRL through a protocol conversion gateway. The acquired time-series data is labeled with device ID, acquisition timestamp and sensor type, and stored in a structured manner in the Influx DB industrial time-series database to form a raw dataset with spatiotemporal identification.

[0015] As a further aspect of the present invention: the preprocessing module uses a 30Hz-50Hz Butterworth low-pass filter to remove electromagnetic interference, and uses the IQR algorithm to remove abnormal data points that exceed 1.5-3 times the interquartile range. The preprocessing module applies a 5min-10min sliding window to divide the data stream into frames, performs Z-Score normalization on each frame of data, eliminates the dimensional differences of current and temperature parameters, and outputs a standardized data sequence.

[0016] As a further aspect of the present invention: the feature engineering module calculates time-domain statistics (mean, root mean square, and kurtosis), frequency-domain features (power spectral density peak frequency, energy percentage), and time-frequency features (wavelet packet energy entropy) for the standardized data sequence to form a healthy feature vector, and uses principal component analysis to reduce the dimensionality of the high-dimensional feature vector, retaining principal components with a cumulative variance contribution rate greater than 85% to reduce the computational complexity of subsequent models.

[0017] As a further aspect of the present invention: the base layer random forest model of the hybrid prediction model contains 50-100 decision trees, and the key parameter combination of torque-temperature coupling degree and vibration frequency-load correlation is selected based on the Gini index. The prediction layer long short-term memory network is set with 2-3 hidden layers, each containing 128-256 neurons. The learning rate (0.001-0.01) and dropout rate (0.1-0.3) are optimized using cross-validation, and the health score (1-100 points) is output with the goal of minimizing the root mean square error.

[0018] As a further aspect of the present invention: the model-agnostic meta-learning algorithm of the meta-learning and domain adaptation module adopts the following parameter update rule:

[0019] ;

[0020] in, To obtain the initial parameters of a meta-model with cross-domain general feature extraction capabilities through pre-training based on historical data from multiple brands, To control the speed of adaptation to new brand data, a meta-learning rate (0.01-0.1) was used. For new brand small sample tasks containing 20-50 cases, The cross-entropy-based fault classification loss function is... To adapt to the first The model parameters were adjusted after each brand, and the prediction error of the model on new brand data was reduced. For meta-model parameters Feature extraction function, For meta-model parameters Gradient calculation, loss function calculated based on backpropagation algorithm. right The gradient vector guides the direction of parameter updates. The adversarial domain adaptive network simulates the distribution of target brand data through the generator and distinguishes the data source through the discriminator, enabling the feature extractor to learn common fault features unrelated to the brand (such as the periodic impact component in vibration signals).

[0021] As a further aspect of the present invention: the federated learning coordination module dynamically adjusts the weights (ranging from 0.1 to 0.5) based on the number of devices from each brand, and adds Laplace noise to the uploaded gradient data. And noise scale satisfy ,in, For privacy budget, a value of 1-10 is acceptable. To ensure data privacy compliance, gradient norm sensitivity is used.

[0022] As a further aspect of the present invention: the digital twin sample generation module constructs a robot dynamics model based on the Lagrange equation, as detailed below:

[0023] ;

[0024] in, It is the Lagrangian function (equal to the kinetic energy minus the potential energy of the robot system). The joint angle vector. The joint angular velocity vector. To control torque, To simulate the disturbance torque caused by a fault (periodic impact torque due to gear wear), based on adjustments The amplitude (0.1N·m-10N·m) and frequency (10Hz-100Hz) are used to generate synthetic data containing more than 10 types of fault modes, including bearing wear and motor eccentricity, to supplement scenarios where real fault samples are insufficient.

[0025] As a further aspect of the present invention: the rule engine of the dynamic decision engine unit establishes a three-level early warning mechanism. A health score of 70-80 points is a blue alert indicating that attention is needed; a score of 60-70 points is a yellow alert requiring investigation within 3 days; and a score below 60 points is a red alert requiring immediate shutdown and maintenance. The case library matches historical fault cases based on the cosine similarity algorithm, and the matching conditions include that the Euclidean distance of the feature vectors is less than 0.2 and the trend change slope is consistent, generating executable maintenance suggestions that include spare parts models, maintenance durations, and historical costs.

[0026] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0027] 1. This invention utilizes a collaborative mechanism between meta-learning and domain adaptation modules. It trains a cross-brand universal feature extractor using a model-agnostic meta-learning algorithm and combines it with an adversarial domain adaptation network to obfuscate the data distribution of different brands. This enables the system to extract unified fault feature vectors from heterogeneous data of robots from multiple brands such as ABB and KUKA, overcoming the "data gap" caused by differences in sensor parameter definitions and sampling frequencies in traditional modeling. Simultaneously, relying on the privacy-compliant architecture of the federated learning coordination module, it achieves collaborative model training without leaving the local machine with data from each brand. This generates a global model containing common fault characteristics across brands, solving the "data silo" problem caused by commercial confidentiality requirements in industrial scenarios. The combination of these technologies makes the system compatible with controller protocols of different brands of robots. When new brand equipment is connected, only a small number of samples are needed to complete model adaptation, realizing unified status monitoring and fault prediction for multi-brand mixed production lines. The prediction error is reduced compared to traditional solutions, providing a cross-vendor, highly generalized maintenance solution for multi-brand equipment-intensive industries.

[0028] 2. This invention utilizes the physical simulation capabilities of the digital twin sample generation module to construct a digital twin based on robot dynamics equations. It simulates sensor data changes under typical fault modes such as gear wear and bearing failure. By adjusting the amplitude and frequency of the interference torque, it generates synthetic fault samples with physical meaning, effectively filling the gap in real fault data for new robots or niche brand equipment. Combined with the rapid adaptation characteristics of meta-learning, the model-agnostic meta-learning algorithm is trained on small sample tasks, enabling the system to update parameters with only a few samples when accessing new brand data. This avoids the limitations of traditional machine learning, which relies on large amounts of labeled data. This technical approach allows the system to achieve early warning in scenarios without sufficient historical fault data, reducing the sample requirements for maintenance decisions and significantly improving the efficiency of predictive maintenance in emerging industrial robot scenarios. It also provides data-driven flexible support for the dynamic expansion of intelligent manufacturing production lines. Attached Figure Description

[0029] Figure 1 is a schematic diagram of the system flow in an embodiment of the present invention;

[0030] Figure 2 is a schematic diagram of the workflow of the meta-learning and domain adaptation module in an embodiment of the present invention. Detailed Implementation

[0031] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0032] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0033] Please refer to Figures 1 and 2. This invention discloses a predictive maintenance system for industrial robots based on machine learning. The maintenance system includes:

[0034] Data acquisition module: Real-time acquisition of timing data such as joint torque, motor temperature and vibration frequency from multi-brand robot controllers including ABB and KUKA via Modbus and OPCUA industrial communication protocols, and transmission to the preprocessing module;

[0035] Preprocessing module: Performs digital filtering and outlier detection on the raw data, removes electromagnetic interference and transmission delay noise, and outputs clean data to the feature engineering module;

[0036] Feature engineering module: Extracts time-domain features, frequency-domain features, and time-frequency-domain features from cleaning data to generate equipment health feature vectors, which are then input into the hybrid prediction model unit and the meta-learning and domain adaptation module, respectively.

[0037] Hybrid Prediction Model Unit: A hybrid prediction model is set up, consisting of a base layer and a prediction layer. The base layer identifies multivariate relationships through the random forest algorithm, while the prediction layer learns long-term dependencies of time-series data based on the long short-term memory network and outputs health scores and remaining life prediction values ​​to the dynamic decision engine unit.

[0038] Meta-learning and Domain Adaptation Module: The module uses a model-agnostic meta-learning algorithm to train a cross-brand general fault feature extractor. It uses an adversarial domain adaptive network to obfuscate the data distribution of different brands, generate a cross-domain unified feature vector, and feeds it back to the hybrid prediction model unit to optimize generalization ability.

[0039] Federated learning coordination module: Receives sub-model parameter updates trained locally by robots from various brands, generates a global generalization model and a cross-brand maintenance case library through a differential privacy-preserving weighted aggregation algorithm, and synchronizes it to the hybrid prediction model unit at the edge.

[0040] Digital twin sample generation module: Based on the robot dynamics equations, a digital twin is constructed to simulate fault modes such as gear wear and bearing failure, and synthetic fault samples containing physical meaning are generated and input into the feature engineering module to enhance small sample training;

[0041] Dynamic Decision Engine Unit: Based on health scores, remaining life expectancy predictions, and a cross-brand maintenance case library generated through federated learning, the system sets multi-level early warning thresholds through a rules engine, generates maintenance suggestions, and pushes them to a visual interactive interface.

[0042] In one embodiment of the present invention: the data acquisition module supports an adjustable sampling frequency from the second level to the minute level, and is compatible with the proprietary communication protocols of ABB Robot Ware and KUKAKRL through a protocol conversion gateway. The acquired time series data is labeled with device ID, acquisition timestamp and sensor type, and stored in the Influx DB industrial time series database in a structured manner to form a raw dataset with spatiotemporal identification.

[0043] In one embodiment of the present invention: the preprocessing module uses a 30Hz-50Hz Butterworth low-pass filter to remove electromagnetic interference, and uses the IQR algorithm to remove abnormal data points that exceed 1.5-3 times the interquartile range. The preprocessing module applies a 5min-10min sliding window to divide the data stream into frames, performs Z-Score normalization on each frame of data, eliminates the dimensional differences of current and temperature parameters, and outputs a standardized data sequence.

[0044] In one embodiment of the present invention: the feature engineering module calculates time-domain statistics (mean, root mean square and kurtosis), frequency-domain features (power spectral density peak frequency, energy percentage) and time-frequency features (wavelet packet energy entropy) for the standardized data sequence to form a healthy feature vector, and uses principal component analysis to reduce the dimensionality of the high-dimensional feature vector, retaining principal components with a cumulative variance contribution rate greater than 85% to reduce the computational complexity of subsequent models.

[0045] In one embodiment of the present invention: the base layer random forest model of the hybrid prediction model contains 50-100 decision trees, and the combination of key parameters such as torque-temperature coupling degree and vibration frequency-load correlation is selected based on Gini index. The prediction layer long short-term memory network is set with 2-3 hidden layers, each containing 128-256 neurons. The learning rate (0.001-0.01) and dropout rate (0.1-0.3) are optimized using cross-validation, and the health score (1-100 points) is output with the goal of minimizing the root mean square error.

[0046] In one embodiment of the present invention: the model-agnostic meta-learning algorithm of the meta-learning and domain adaptation module adopts the following parameter update rule:

[0047] ;

[0048] in, To obtain the initial parameters of a meta-model with cross-domain general feature extraction capabilities through pre-training based on historical data from multiple brands, To control the speed of adaptation to new brand data, a meta-learning rate (0.01-0.1) was used. For new brand small sample tasks containing 20-50 cases, The cross-entropy-based fault classification loss function is... To adapt to the first The model parameters were adjusted after each brand, and the prediction error of the model on new brand data was reduced. For meta-model parameters Feature extraction function, For meta-model parameters Gradient calculation, loss function calculated based on backpropagation algorithm. right The gradient vector guides the direction of parameter updates. The adversarial domain adaptive network simulates the distribution of target brand data through the generator and distinguishes the data source through the discriminator, enabling the feature extractor to learn common fault features unrelated to the brand (such as the periodic impact component in vibration signals).

[0049] In one embodiment of the invention: the federated learning coordination module dynamically adjusts the weights (ranging from 0.1 to 0.5) based on the number of devices from each brand, and adds Laplace noise to the uploaded gradient data. And noise scale satisfy ,in, For privacy budget, a value of 1-10 is acceptable. To ensure data privacy compliance, gradient norm sensitivity is used.

[0050] In one embodiment of the present invention: the digital twin sample generation module constructs a robot dynamics model based on the Lagrange equation, as detailed below:

[0051] ;

[0052] in, It is the Lagrangian function (equal to the kinetic energy minus the potential energy of the robot system). The joint angle vector. The joint angular velocity vector. To control torque, To simulate the disturbance torque caused by a fault (periodic impact torque due to gear wear), based on adjustments The amplitude (0.1N·m-10N·m) and frequency (10Hz-100Hz) are used to generate synthetic data containing more than 10 types of fault modes, including bearing wear and motor eccentricity, to supplement scenarios where real fault samples are insufficient.

[0053] In one embodiment of the present invention: the rule engine of the dynamic decision engine unit establishes a three-level early warning mechanism. A health score of 70-80 points is a blue alert indicating that attention is needed; a score of 60-70 points is a yellow alert requiring investigation within 3 days; and a score below 60 points is a red alert requiring immediate shutdown and maintenance. The case library matches historical fault cases based on the cosine similarity algorithm, and the matching conditions include that the Euclidean distance of the feature vectors is less than 0.2 and the trend change slope is consistent, generating executable maintenance suggestions that include spare parts models, maintenance durations, and historical costs.

[0054] In one embodiment of the present invention: the physical constraint data synthesized by digital twins (such as...) Amplitude / frequency significantly improves the generalization efficiency of meta-learning feature extractors:

[0055] During the aggregation phase of federated learning, the feature vectors derived from physical simulation improved the accuracy of cross-brand fault commonality identification by 22.5% (compared to a baseline without digital twins).

[0056] The number of training iterations for few-shot tasks in meta-learning was reduced from 50 to 15 (see Implementation Case 2, F1 score of 0.85 achievement condition).

[0057] In one embodiment of the present invention: the meta-learning and domain adaptation module achieves cross-brand feature unification through a three-level mechanism of "pre-training-adversarial alignment-rapid adaptation", and the specific interaction process is as follows:

[0058] (1) Meta-learning pre-training stage (corresponding to Figure 2 "Meta-model pre-training"):

[0059] (2) Use historical data from multiple brands to initialize and train the meta-model. Optimize the feature extractor parameters through a model-agnostic meta-learning algorithm so that it can extract cross-domain common fault features. At this time, the feature extractor can identify common fault modes of robots from different brands, such as vibration and shock spectrum features corresponding to bearing wear, but the data distribution differences between brands have not been completely eliminated.

[0060] (3) Adversarial domain adaptive training (corresponding to Figure 2 "Adversarial domain adaptive training"):

[0061] Generator: Receives feature vectors from the source domain brand, simulates the data distribution of the target domain brand through a neural network, and generates pseudo-target domain features;

[0062] Discriminator: It takes the real target domain features and the pseudo features output by the generator as input, and calculates the loss function LD through a binary classification task (determining the source of the features), forcing the generator to improve the simulation accuracy;

[0063] Feature extractor collaborative optimization: During the backpropagation of the discriminator, the feature extractor adjusts its parameters synchronously. Its optimization goal is to minimize the classification accuracy of the discriminator, thereby forcing the output features to ignore brand-specific data distribution differences (such as sampling frequency and units of measurement) and retain only essential features related to the fault (such as the periodic peak frequency of the vibration signal).

[0064] (3) Small sample rapid adaptation stage (corresponding to Figure 2 "Small sample task adaptation"):

[0065] When integrating data from a new brand robot, a rapid parameter update rule is implemented using meta-learning: Using small sample tasks (Includes 20-50 new brand feature vectors) for pre-trained parameters Perform gradient updates and generate adaptation parameters. At this point, the feature extractor combines the cross-brand common features learned from adversarial training with the features of a small number of samples from the new brand. It can quickly adapt to the unique noise patterns of the new brand (such as the electromagnetic interference features of the new sensor) while maintaining generalization ability.

[0066] Example 1: Predictive maintenance of multi-brand automotive welding lines (see Figures 1-2).

[0067] Application scenarios

[0068] A car factory has deployed 200 industrial robots, including 120 ABBIRB6700s (using the RobotWare protocol) and 80 KUKAKR210R2700s (using the KRL protocol). It needs to perform cross-brand unified prediction of faults such as joint bearing wear and motor overheating in the welding robots.

[0069] Key parameters and processes

[0070] Data acquisition module

[0071] Sampling frequency: Set to 100Hz (second level) to account for the high-frequency vibration characteristics of the welding process.

[0072] Protocol conversion: Interacts with the ABB controller via the OPCUA gateway and with the KUKA controller via the Modbus TCP protocol, real-time acquisition of joint torque (N·m), motor temperature (°C), and vibration acceleration (m / s²). 2 )data;

[0073] Data tags: Add a unique ID (e.g., “ABB-001”, “KUKA-002”), data collection timestamp (accurate to milliseconds), and sensor type tag (e.g., “Joint 1 vibration sensor”) to each device and store them in the InfluxDB database;

[0074] Preprocessing module

[0075] Digital filtering: A 40Hz Butterworth low-pass filter is used to filter high-frequency electromagnetic interference generated by the welding arc;

[0076] Outlier detection: Outliers exceeding twice the interquartile range (such as instantaneous torque jumps caused by welding spatter) are removed using the IQR algorithm.

[0077] Frame processing: The continuous data stream is divided into 3000 samples / frames (100Hz×300 seconds) using a 5-minute sliding window. Z-Score normalization is performed on each frame of data to eliminate dimensional differences.

[0078] Feature Engineering Module

[0079] Temporal characteristics: Calculate the mean, root mean square, and kurtosis of each frame of data (e.g., the mean joint torque reflects the load trend, and kurtosis characterizes the impact characteristics).

[0080] Frequency domain characteristics: The power spectral density of the vibration signal is calculated by Fast Fourier Transform (FFT), and the peak frequency (e.g., the characteristic frequency of bearing wear is 120Hz) and energy proportion are extracted.

[0081] Time-frequency domain characteristics: Wavelet packet transform is used to calculate energy entropy and capture the impact of dynamic load changes on equipment health during welding;

[0082] Dimensionality reduction: The first 20 principal components (cumulative variance contribution rate of 88%) are retained by PCA, and the 100-dimensional feature vector is compressed into 20 dimensions.

[0083] Meta-learning and Domain Adaptive Module

[0084] Meta-model training: The meta-model was pre-trained using historical fault data from ABB and KUKA (500 cases each), with initial parameters... Learn common characteristics across brands (such as the frequency domain patterns of periodic impacts in vibration signals);

[0085] Small sample adaptation: For new brand equipment (such as newly added UR collaborative robots), input the feature vectors of 20 cases from the 30 days prior to the failure as the task. Through meta-learning rate =0.05 Update Parameters This reduces the prediction error of the model on new brand data;

[0086] Adversarial training: The generator simulates the vibration data distribution of KUKA robots, the discriminator distinguishes the data sources, and forces the feature extractor to ignore brand differences and focus on the common features of bearing wear (such as the kurtosis value of vibration acceleration > 3.5).

[0087] Federal Learning Coordination Module

[0088] Parameter aggregation: ABB robots account for 60%, KUKA robots account for 40%, so the aggregation weights are 0.6 and 0.4 respectively;

[0089] Privacy Protection: Set a Privacy Budget =5, calculate gradient norm sensitivity =0.1, then the noise scale ≥ =50, add Laplace noise to the gradient data (50);

[0090] Case library generation: After aggregation, cross-brand maintenance cases are generated, such as "When the average joint torque is >80 N·m and the peak vibration frequency is 120 Hz, 85% of the cases require bearing replacement within 7 days".

[0091] Hybrid prediction model

[0092] Base layer: The random forest contains 80 decision trees, which are selected using the Gini index to determine torque-temperature coupling (correlation coefficient > 0.7) and vibration frequency-load correlation (…). Key parameter combinations such as values ​​<0.01;

[0093] Prediction layer: The LSTM has 2 hidden layers, each with 256 neurons, a learning rate of 0.002, a dropout rate of 0.2, and outputs a health score (e.g., a KUKA robot with a health score of 58 points triggers a yellow warning).

[0094] Dynamic Decision Engine Unit

[0095] Warning threshold: A health score of 60 points triggers a yellow warning, indicating the maintenance recommendation to "check joint bearing 1 within 3 days";

[0096] Case matching: Historical cases are matched using the cosine similarity algorithm, such as "An ABB robot once exhibited similar vibration characteristics, and the score recovered to 85 points after the bearing was replaced", generating a work order that includes the spare part model (SKF6205) and maintenance duration (4 hours).

[0097] Example 2: Please refer to Figures 1-2 for small-sample fault prediction of a novel collaborative robot.

[0098] Application scenarios

[0099] An electronics factory introduced 50 new domestically produced collaborative robots (brand X, no historical fault data). It needs to provide early warning of wear failure in the reducer gears. There are only 10 actual fault samples.

[0100] Key parameters and processes

[0101] Digital twin sample generation module

[0102] Dynamics Modeling: Constructing a Dynamics Model of the Robot Arm Based on the Lagrange Equations

[0103]

[0104] Among them, kinetic energy Potential energy , Indicates the total mass of the robot. Represents gravitational acceleration. Indicates the robot's center of gravity height and controls torque. Driven by real operational data;

[0105] Fault simulation: Setting interference torque Simulated gear wear:

[0106] Amplitude: Gradually increase from 0.5 N·m (initial wear) to 5 N·m (severe wear);

[0107] Frequency: Corresponds to the gear meshing frequency, set to three levels: 50Hz, 80Hz and 120Hz;

[0108] Synthetic data: 1,000 simulated fault samples were generated, including normal state (500 cases), light wear (300 cases), and heavy wear (200 cases), with features including torque fluctuation amplitude, temperature rise rate, etc.

[0109] Meta-learning and few-shot training

[0110] Data augmentation: Mix 10 real-world fault data points with 1000 synthetic data points, and divide the training set / test set in a 7:3 ratio;

[0111] Meta-model adaptation: The meta-model is pre-trained using synthetic data, and the initial parameter θ is used to learn the characteristic patterns of gear wear (such as periodic spikes in the torque signal).

[0112] Quick update: Input 10 real samples as the task Meta-learning rate =0.1, parameter after 5 iterations This enabled the model to achieve an F1 score of 0.85 on the test set.

[0113] Feature Engineering and Prediction

[0114] Feature extraction: Targeting the low-load characteristics of collaborative robots, the focus is on extracting time-domain features (root mean square increment > 15%) and time-frequency domain features (wavelet energy entropy abrupt change).

[0115] Model Deployment: The trained LSTM model is deployed to the edge server to monitor gearbox vibration data in real time. When the root mean square increment of three consecutive frames is greater than 20% and the energy entropy is greater than 0.7, a red warning is triggered to "stop the machine immediately and check the reducer".

[0116] Maintenance strategy optimization

[0117] Sample feedback: After maintenance, the case of "the score was restored to 90 points after replacing the gearbox" will be added to the federated learning case library;

[0118] Model update: The model is automatically retrained weekly, and the threshold is optimized by combining newly collected normal operation data (about 5,000 cases), reducing the false positive rate from 15% to 5%.

[0119] Based on the above embodiments, it can be concluded that by combining the technical combination of meta-learning and domain-adaptive module collaboration mechanism, federated learning coordination module privacy compliance architecture, and the synergistic effect of digital twin sample generation module and meta-learning rapid adaptation characteristics, the heterogeneity and privacy barriers of multi-brand robot data can be effectively overcome, filling the gap in fault data of new robots. With the synergistic cooperation of various algorithms and models in each module, it is possible to achieve good results in compatibility with multi-brand robots, reduce the need for new brand model adaptation samples and prediction errors, and improve the efficiency of predictive maintenance in multi-brand mixed production lines and emerging industrial robot scenarios.

[0120] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. A predictive maintenance system for industrial robots based on machine learning, characterized in that, The maintenance system includes: a feature engineering module: extracting time-domain features, frequency-domain features, and time-frequency-domain features from cleaning data to generate equipment health feature vectors; a hybrid prediction model unit: equipped with a hybrid prediction model consisting of a base layer and a prediction layer. The base layer uses a random forest algorithm to identify multivariate relationships, while the prediction layer learns long-term dependencies in time-series data based on a long short-term memory network, outputting health scores and remaining life prediction values; and a meta-learning and domain adaptation module: training a cross-brand universal fault feature extractor using a model-agnostic meta-learning algorithm, and using an adversarial domain adaptation network to obfuscate the data distribution of different brands, generating a unified feature vector across domains. The model-agnostic meta-learning algorithm of the meta-learning and domain adaptation module uses the following parameter update rules: ;in, To obtain the initial parameters of a meta-model with cross-domain general feature extraction capabilities through pre-training based on historical data from multiple brands, To control the meta-learning rate of the speed at which new brand data is adapted, For new brand small sample tasks containing 20-50 cases, The cross-entropy-based fault classification loss function is... To adapt to the first Model parameters after each brand For meta-model parameters Feature extraction function, For meta-model parameters Gradient calculation, loss function calculated based on backpropagation algorithm. right The gradient vector guides the direction of parameter updates. The adversarial domain adaptive network simulates the target brand data distribution based on the generator and distinguishes data sources based on the discriminator, enabling the feature extractor to learn common fault features unrelated to the brand. The federated learning coordination module receives parameter updates from the sub-models trained locally by each brand robot and uses a differential privacy-preserving weighted aggregation algorithm to generate a global generalization model and a cross-brand maintenance case library. The digital twin sample generation module constructs a digital twin based on the robot dynamics equations, simulates fault modes, generates synthetic fault samples containing physical meaning, and inputs them into the feature engineering module to enhance small-sample training. The digital twin sample generation module constructs a robot dynamics model based on the Lagrange equations, as detailed below: ;in, For Lagrange functions, The joint angle vector. The joint angular velocity vector. To control torque, To simulate the disturbance torque caused by a fault, based on adjustment The system generates synthetic data including bearing wear and motor eccentricity fault modes based on amplitude and frequency. The dynamic decision engine unit uses a rule engine to set multi-level warning thresholds based on health scores, remaining life predictions, and a cross-brand maintenance case library generated through federated learning. It sets three levels of warnings based on health score thresholds, including a red warning triggered when the health score is ≤60, generating a shutdown and maintenance suggestion. It also generates a spare parts list based on historical fault characteristics matched to the cross-brand maintenance case library, generates maintenance suggestions, and pushes them to a visual interactive interface. The rule engine of the dynamic decision engine unit establishes a three-level warning mechanism: a blue warning for health scores of 70-80 (requiring attention), a yellow warning for 60-70 (requiring investigation within 3 days), and a red warning for below 60 (requiring immediate shutdown and maintenance). The case library matches historical fault cases based on a cosine similarity algorithm, with matching conditions including a feature vector Euclidean distance less than 0.2 and consistent trend slopes, generating executable maintenance suggestions including spare parts models, maintenance duration, and historical costs.

2. The predictive maintenance system for industrial robots based on machine learning according to claim 1, characterized in that: The maintenance system also includes a data acquisition module and a preprocessing module. The data acquisition module collects real-time time-series data of joint torque, motor temperature, and vibration frequency from multi-brand robot controllers, including ABB and KUKA, based on Modbus and OPCUA industrial communication protocols. The data acquisition module supports adjustable sampling frequencies from seconds to minutes and is compatible with ABB Robot Ware and KUKA AKRL's proprietary communication protocols based on a protocol conversion gateway. The acquired time-series data is labeled with device ID, acquisition timestamp, and sensor type, and stored in a structured manner in the Influx DB industrial time-series database to form a raw dataset with spatiotemporal identifiers.

3. The predictive maintenance system for industrial robots based on machine learning according to claim 2, characterized in that: The preprocessing module performs digital filtering and outlier detection on the raw data to remove electromagnetic interference and transmission delay noise, outputting clean data. The preprocessing module uses a 30Hz-50Hz Butterworth low-pass filter to remove electromagnetic interference and uses the IQR algorithm to remove outlier data points that exceed 1.5-3 times the interquartile range. The preprocessing module applies a 5-10 minute sliding window to divide the data stream into frames, performs Z-Score normalization on each frame of data, eliminates the dimensional differences of current and temperature parameters, and outputs a standardized data sequence.

4. The predictive maintenance system for industrial robots based on machine learning according to claim 3, characterized in that: The feature engineering module calculates time-domain statistics, frequency-domain features, and time-frequency features for standardized data sequences to form a healthy feature vector. Principal component analysis is then used to reduce the dimensionality of the high-dimensional feature vector, retaining principal components with a cumulative variance contribution rate greater than 85%.

5. The predictive maintenance system for industrial robots based on machine learning according to claim 4, characterized in that: The base layer of the hybrid prediction model is a random forest model containing 50-100 decision trees. It uses the Gini index to select key parameter combinations such as torque-temperature coupling degree and vibration frequency-load correlation. The prediction layer is a long short-term memory network with 2-3 hidden layers, each containing 128-256 neurons. Cross-validation is used to optimize the learning rate and dropout rate, and a health score is output.

6. The predictive maintenance system for industrial robots based on machine learning according to claim 5, characterized in that: The federated learning coordination module dynamically adjusts based on the number of devices from each brand, adding Laplace noise to the uploaded gradient data. And noise scale satisfy ,in, For privacy budget, a value of 1-10 is acceptable. For gradient norm sensitivity.

Images