Industrial robot predictive maintenance system based on machine learning

Abstract

The invention discloses an industrial robot predictive maintenance system based on machine learning, and relates to the technical field of industrial robot maintenance. Comprising a data acquisition module, a preprocessing module, a feature engineering module, a hybrid prediction model unit, a meta-learning and domain adaptive module, a federated learning coordination module, a digital twin sample generation module and a dynamic decision engine unit. A model agnostic meta-learning algorithm is utilized to train a cross-brand universal feature extractor, distribution of different brand data is confused in combination with an adversarial field adaptive network, a'data gap 'caused by sensor parameter definition and sampling frequency difference in traditional modeling is broken through, model cooperative training is achieved on the premise that data of all brands are not out of the local, and the modeling efficiency is improved. A global model containing cross-brand fault generality is generated, and the problem of data islands caused by commercial secrecy requirements in an industrial scene is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial robot maintenance, and in particular to an industrial robot predictive maintenance system based on machine learning. Background Art

[0002] Industrial robots, as core equipment in intelligent manufacturing, have been widely used in industrial scenarios such as automotive manufacturing, electronic assembly, and aerospace. They undertake key tasks such as welding, handling, and precision machining. The stability of their operating status directly affects production line efficiency, product quality, and safe production. Therefore, efficient maintenance systems are one of the core requirements in the field of industrial automation. Existing technologies have certain defects. First, it is difficult to extract unified fault features from heterogeneous data of multi-brand robots. There are problems of "data gap" and "data island" and insufficient model generalization ability. Second, existing technologies rely on a large amount of real fault data. When there is a lack of data on new or niche brand robots, it is difficult to achieve early warning and the sample requirements are large. To this end, we propose a predictive maintenance system for industrial robots based on machine learning. Summary of the Invention

[0003] The object of the present invention is to provide an industrial robot predictive maintenance system based on machine learning.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a machine learning-based predictive maintenance system for industrial robots, the maintenance system comprising: Data acquisition module: This module collects real-time time series data such as joint torque, motor temperature, and vibration frequency from robot controllers from multiple brands, including ABB and KUKA, through Modbus and OPCUA industrial communication protocols, and transmits the data to the preprocessing module. 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; Feature Engineering Module: This module extracts time-domain features, frequency-domain features, and time-frequency-domain features from clean data to generate device health feature vectors, which are then input into the hybrid prediction model unit and the meta-learning and domain adaptation module respectively. Hybrid prediction model unit: A hybrid prediction model is set up, consisting of a base layer and a prediction layer. The base layer uses a random forest algorithm to identify multivariate correlations. The prediction layer uses a long-short-term memory network to learn the long-term dependencies of time series data and outputs health scores and remaining life expectancy prediction values ​​to the dynamic decision engine unit. Meta-learning and domain adaptation module: This module uses a model-agnostic meta-learning algorithm to train a cross-brand universal fault feature extractor. It uses an adversarial domain adaptation network to confuse the data distribution of different brands, generating a unified cross-domain feature vector that is fed back to the hybrid prediction model unit to optimize generalization capabilities. Federated Learning Coordination Module: Receives parameter updates from locally trained sub-models of each brand's robots, generates a global generalization model and a cross-brand maintenance case library using a weighted aggregation algorithm protected by differential privacy, and synchronizes these to the hybrid prediction model unit at the edge. Digital twin sample generation module: This module builds a digital twin based on the robot's dynamic equations, simulates fault modes such as gear wear and bearing failure, and generates synthetic fault samples with physical meaning. These samples are then input into the feature engineering module to enhance small sample training. Dynamic decision engine unit: Based on the health score, remaining life prediction value and the cross-brand maintenance case library generated by federated learning, it sets multi-level warning thresholds through the rule engine, generates maintenance recommendations and pushes them to the visual interactive interface.

[0005] As a further solution of the present invention: the data acquisition module supports an adjustable sampling frequency from seconds to minutes, is compatible with the proprietary communication protocols of ABB Robot Ware and KUKAKRL manufacturers through a protocol conversion gateway, adds device ID, acquisition timestamp and sensor type label to the collected time series data, and stores it in a structured manner in the Influx DB industrial time series database to form an original data set with time and space identification.

[0006] As a further solution 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 eliminate abnormal data points that exceed 1.5-3 times the interquartile range. The preprocessing module uses a 5min-10min sliding window to frame the data stream, performs Z-Score normalization on each frame of data, eliminates the dimensional differences of current and temperature parameters, and outputs a standardized data sequence.

[0007] As a further solution 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 proportion) and time-frequency domain features (wavelet packet energy entropy) of 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 the principal components with a cumulative variance contribution rate greater than 85% to reduce the complexity of subsequent model calculations.

[0008] As a further solution of the present invention: the base layer random forest model of the hybrid prediction model contains 50-100 decision trees, and the key parameter combinations of torque-temperature coupling and vibration frequency-load correlation are screened based on the Gini index. The prediction layer long short-term memory network is set with 2-3 hidden layers and each layer contains 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.

[0009] As a further solution of the present invention, the model-agnostic meta-learning algorithm of the meta-learning and domain adaptation module adopts the following parameter update rules: ; in, To obtain the initial parameters of the meta-model with cross-domain general feature extraction capabilities based on pre-training of multi-brand historical data, To control the meta-learning rate (0.01-0.1) of the adaptation speed to new brand data, For new brand small sample tasks containing 20-50 samples, is the fault classification loss function based on cross entropy, To adapt to the The model parameters after the brand, and the prediction error of the model on the new brand data is reduced, Based on metamodel parameters The feature extraction function, is the meta-model parameter Gradient calculation, loss function is calculated based on the back propagation algorithm right The gradient vector guides the parameter update direction. The adversarial domain adaptation network simulates the target brand data distribution through the generator and the discriminator distinguishes the data source, so that the feature extractor learns the common fault features that are independent of the brand (such as the periodic impact component in the vibration signal).

[0010] As a further solution of the present invention: the federated learning coordination module is dynamically adjusted according to the number of devices of each brand (weight range 0.1-0.5), and Laplace noise is added to the uploaded gradient data. , and the noise scale satisfy ,in, is the privacy budget, ranging from 1 to 10. Gradient norm sensitivity ensures data privacy compliance.

[0011] As a further solution of the present invention: the digital twin sample generation module constructs a robot dynamics model based on the Lagrange equation, as follows: ; in, is the Lagrangian function (equal to the kinetic energy minus the potential energy of the robot system), is the joint angle vector, is the joint angular velocity vector, To control the torque, To simulate the disturbance torque caused by fault (periodic impact torque caused by gear wear), based on the adjustment The amplitude (0.1N.m-10N.m) and frequency (10Hz-100Hz) of the vibration signal are generated to include bearing wear, motor eccentricity 10 and other fault modes, and to supplement the scene of insufficient real fault samples.

[0012] As a further scheme of the application: the rule engine of the dynamic decision engine unit establishes a three-level early warning mechanism, a health score of 70-80 is a blue prompt for attention, 60-70 is a yellow warning for investigation within 3 days, and less than 60 is a red warning for immediate shutdown for maintenance, the case base matches historical fault cases based on the cosine similarity algorithm and the matching conditions include a Euclidean distance of the feature vector less than 0.2 and consistent trend change slope, and generates executable maintenance suggestions including spare parts model, maintenance time and historical cost.

[0013] Compared with the prior art, the beneficial effects of the application are as follows: 1、The application uses the collaborative mechanism of meta-learning and domain adaptation module, trains a cross-brand general feature extractor using a model-agnostic meta-learning algorithm, and combines an adversarial domain adaptation network to confuse different brand data distributions, so that the system can extract unified fault feature vectors from heterogeneous data of ABB, KUKA and other multi-brand robots, break through the "data gap" caused by differences in sensor parameter definition and sampling frequency in traditional modeling, and at the same time, relying on the privacy compliance architecture of the federal learning coordination module, through the weighted aggregation algorithm of differential privacy protection, the model collaborative training is realized under the premise that the brand data does not leave the local, the global model containing cross-brand fault commonality is generated, the "data island" problem caused by commercial confidentiality requirements in industrial scenarios is solved, the above technical combination makes the system compatible with the controller protocols of different brand robots, and when a new brand device is connected, only a small amount of samples are needed to complete model adaptation, realizing unified state monitoring and fault prediction of multi-brand mixed production lines, the prediction error is reduced compared with the traditional scheme, providing a cross-vendor, highly generalized maintenance solution for multi-brand device intensive industries; 2. The present invention uses the physical simulation capability of the digital twin sample generation module to construct a digital twin based on the robot dynamics equation, simulates the changes in sensor data under typical fault modes such as gear wear and bearing failure, and generates synthetic fault samples with physical significance by adjusting the amplitude and frequency of the interference torque, effectively filling the gap in real fault data of new robots or niche brand equipment. Combined with the rapid adaptation characteristics of meta-learning, the model-agnostic meta-learning algorithm is trained through small sample tasks, so that when the system accesses new brand data, only a few samples are needed to complete parameter updates, avoiding the limitations of traditional machine learning that relies on a large amount of labeled data. This technical path enables the system to achieve early warning in scenarios without sufficient historical fault data, reduces the sample requirements for maintenance decisions, and significantly improves the implementation efficiency of predictive maintenance in emerging industrial robot scenarios, providing data-driven flexible support for the dynamic expansion of intelligent manufacturing production lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A schematic diagram of a system flow in an embodiment of the present invention; Figure 2 Schematic diagram of the workflow of the meta-learning and domain adaptation module in an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0016] In addition, 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.

[0017] Please see the attached Figure 1 -Attached Figure 2 The present invention provides a predictive maintenance system for industrial robots based on machine learning, the maintenance system comprising: Data acquisition module: This module collects real-time time series data such as joint torque, motor temperature, and vibration frequency from robot controllers from multiple brands, including ABB and KUKA, through Modbus and OPCUA industrial communication protocols, and transmits the data to the preprocessing module. 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; Feature Engineering Module: This module extracts time-domain features, frequency-domain features, and time-frequency-domain features from clean data to generate device health feature vectors, which are then input into the hybrid prediction model unit and the meta-learning and domain adaptation module respectively. Hybrid prediction model unit: A hybrid prediction model is set up, consisting of a base layer and a prediction layer. The base layer uses a random forest algorithm to identify multivariate correlations. The prediction layer uses a long-short-term memory network to learn the long-term dependencies of time series data and outputs health scores and remaining life expectancy prediction values ​​to the dynamic decision engine unit. Meta-learning and domain adaptation module: This module uses a model-agnostic meta-learning algorithm to train a cross-brand universal fault feature extractor. It uses an adversarial domain adaptation network to confuse the data distribution of different brands, generating a unified cross-domain feature vector that is fed back to the hybrid prediction model unit to optimize generalization capabilities. Federated Learning Coordination Module: Receives parameter updates from locally trained sub-models of each brand's robots, generates a global generalization model and a cross-brand maintenance case library using a weighted aggregation algorithm protected by differential privacy, and synchronizes these to the hybrid prediction model unit at the edge. Digital twin sample generation module: This module builds a digital twin based on the robot's dynamic equations, simulates fault modes such as gear wear and bearing failure, and generates synthetic fault samples with physical meaning. These samples are then input into the feature engineering module to enhance small sample training. Dynamic decision engine unit: Based on the health score, remaining life prediction value and the cross-brand maintenance case library generated by federated learning, it sets multi-level warning thresholds through the rule engine, generates maintenance recommendations and pushes them to the visual interactive interface.

[0018] In one embodiment of the present invention: the data acquisition module supports an adjustable sampling frequency from seconds to minutes, is compatible with the proprietary communication protocols of ABB Robot Ware and KUKAKRL manufacturers through a protocol conversion gateway, adds device ID, acquisition timestamp and sensor type label to the collected time series data, and stores it in a structured manner in the Influx DB industrial time series database to form an original data set with time and space identification.

[0019] 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 eliminate abnormal data points that exceed 1.5-3 times the interquartile range. The preprocessing module uses a 5min-10min sliding window to frame the data stream, performs Z-Score normalization on each frame of data, eliminates the dimensional differences of current and temperature parameters, and outputs a standardized data sequence.

[0020] 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 proportion) and time-frequency domain features (wavelet packet energy entropy) of 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 the principal components with a cumulative variance contribution rate greater than 85% to reduce the complexity of subsequent model calculations.

[0021] 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 key parameter combinations of torque-temperature coupling and vibration frequency-load correlation are screened based on the Gini index. The prediction layer long short-term memory network is set with 2-3 hidden layers and each layer contains 128-256 neurons. The learning rate (0.001-0.01) and dropout rate (0.1-0.3) are optimized by cross-validation, and the health score (1-100 points) is output with the goal of minimizing the root mean square error.

[0022] 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 rules: ; in, To obtain the initial parameters of the meta-model with cross-domain general feature extraction capabilities based on pre-training of multi-brand historical data, To control the meta-learning rate (0.01-0.1) of the adaptation speed to new brand data, For new brand small sample tasks containing 20-50 samples, is the fault classification loss function based on cross entropy, To adapt to the The model parameters after the brand, and the prediction error of the model on the new brand data is reduced, Based on metamodel parameters The feature extraction function, is the meta-model parameter Gradient calculation, loss function is calculated based on the back propagation algorithm right The gradient vector guides the parameter update direction. The adversarial domain adaptation network simulates the target brand data distribution through the generator and the discriminator distinguishes the data source, so that the feature extractor learns the common fault features that are independent of the brand (such as the periodic impact component in the vibration signal).

[0023] In one embodiment of the present invention, the federated learning coordination module dynamically adjusts the number of devices of each brand (weight range 0.1-0.5) and adds Laplace noise to the uploaded gradient data. , and the noise scale satisfy ,in, is the privacy budget, ranging from 1 to 10. Gradient norm sensitivity ensures data privacy compliance.

[0024] In one embodiment of the present invention, the digital twin sample generation module constructs a robot dynamics model based on the Lagrange equation, as follows: ; in, is the Lagrangian function (equal to the kinetic energy minus the potential energy of the robot system), is the joint angle vector, is the joint angular velocity vector, To control the torque, To simulate the disturbance torque caused by fault (periodic impact torque caused by gear wear), based on the adjustment The amplitude (0.1N·m-10N·m) and frequency (10Hz-100Hz) of the generator are used to generate synthetic data including more than 10 types of fault modes such as bearing wear and motor eccentricity, supplementing scenarios where real fault samples are insufficient.

[0025] In one embodiment of the present invention: the rule engine of the dynamic decision engine unit establishes a three-level warning mechanism, a health score of 70-80 points is a blue prompt that requires attention, a score of 60-70 points is a yellow warning that it should be checked within 3 days, and a score below 60 points is a red warning that the machine should be shut down for maintenance immediately. The case library matches historical fault cases based on the cosine similarity algorithm and the matching conditions include the Euclidean distance of the feature vector being less than 0.2 and the trend change slope being consistent, to generate executable maintenance suggestions including spare part model, maintenance time and historical cost.

[0026] In one embodiment of the present invention: the physical constraint data synthesized by the digital twin (such as Amplitude / frequency) significantly improves the generalization efficiency of the meta-learning feature extractor: In the aggregation phase of federated learning, the physically simulated feature vectors improve the accuracy of cross-brand fault commonality identification by 22.5% (compared to a baseline without digital twins); The number of training iterations for the small sample task of meta-learning was reduced from 50 to 15 (see the conditions for achieving an F1 score of 0.85 in Implementation Case 2).

[0027] 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 - fast adaptation". The specific interaction process is as follows: (1) Meta-learning pre-training stage (corresponding to Figure 2 "meta-model pre-training"): (2) The meta-model is initialized and trained using historical data from multiple brands. The parameters of the feature extractor are optimized through a model-agnostic meta-learning algorithm, enabling it to extract cross-domain universal fault features. At this time, the feature extractor can identify common fault modes of robots from different brands, such as the vibration shock spectrum characteristics corresponding to bearing wear, but the data distribution differences between brands have not been completely eliminated. (3) Adaptive training in adversarial domains (corresponding to Figure 2"Adversarial Domain Adaptation Training"): Generator: Receives the feature vector of the source domain brand, simulates the data distribution of the target domain brand through a neural network, and generates pseudo target domain features; Discriminator: Inputs the real target domain features and the pseudo features output by the generator, calculates the loss function LD through a binary classification task (determining the source of the features), and forces the generator to improve the simulation accuracy; Feature extractor collaborative optimization: During the discriminator's backpropagation process, the feature extractor simultaneously adjusts its parameters to minimize the discriminator's classification accuracy. This forces the output features to ignore brand-specific data distribution differences (such as sampling frequency and dimensional units) and retain only essential fault-related features (such as the periodic peak frequency of the vibration signal). (3) Small sample rapid adaptation stage (corresponding to Figure 2 "Small sample task adaptation"): When new brand robot data is introduced, the fast parameter update rules are updated through meta-learning: , using small sample tasks (Contains 20-50 new brand feature vectors) for pre-training parameters Perform gradient updates to generate adaptation parameters At this time, the feature extractor combines the common cross-brand features learned through adversarial training with the features of a small number of samples of the new brand. It can quickly adapt to the unique noise patterns of the new brand (such as the electromagnetic interference characteristics of the new sensor) while maintaining the generalization ability.

[0028] Example 1, please refer to the attached Figure 1 -Attached Figure 2 , predictive maintenance of multi-brand automobile welding lines: Application Scenario An automobile factory deployed 200 industrial robots, including 120 ABB IRB6700s (using the RobotWare protocol) and 80 KUKAKR210R2700s (using the KRL protocol). They needed to perform unified cross-brand predictions for welding robot joint bearing wear, motor overheating, and other faults. Key parameters and processes Data acquisition module Sampling frequency: Based on the high-frequency vibration characteristics of the welding process, it is set to 100 Hz (second level); Protocol conversion: connect to ABB controller via OPCUA gateway and connect to KUKA controller via ModbusTCP protocol, collect joint torque (N・m), motor temperature (℃) and vibration acceleration (m / s) in real time 2 )data; Data labeling: Add a unique ID (such as "ABB-001" or "KUKA-002"), a collection timestamp (accurate to milliseconds), and a sensor type label (such as "Joint 1 Vibration Sensor") to each device and store them in the InfluxDB database. Preprocessing module Digital filtering: 40Hz Butterworth low-pass filter is used to filter high-frequency electromagnetic interference generated by the welding arc; Outlier detection: The IQR algorithm is used to eliminate outliers exceeding 2 times the interquartile range (such as instantaneous torque jumps caused by welding spatter). Frame processing: Use a 5-minute sliding window to split the continuous data stream into 3000 samples / frame (100 Hz × 300 seconds), and perform Z-Score normalization on each frame of data to eliminate dimensional differences.

[0029] Feature Engineering Module Time domain features: Calculate the mean, root mean square, and kurtosis of each frame of data (for example, the mean joint torque reflects the load trend, and the kurtosis represents the impact characteristics); Frequency domain characteristics: Calculate the power spectrum density of the vibration signal through fast Fourier transform (FFT), extract the peak frequency (for example, the characteristic frequency of bearing wear is 120Hz) and energy proportion; Time-frequency domain features: Wavelet packet transform is used to calculate energy entropy to capture the impact of dynamic load changes on equipment health during welding; Dimensionality reduction: PCA is used to retain the first 20 principal components (cumulative variance contribution rate 88%) and compress the 100-dimensional feature vector to 20 dimensions.

[0030] Meta-learning and domain adaptation modules Metamodel training: Use ABB and KUKA historical fault data (500 cases each) to pre-train the metamodel, with initial parameters Learning common features across brands (e.g., frequency domain patterns of periodic shocks in vibration signals); Small sample adaptation: For new brand equipment (such as the newly added UR collaborative robot), input the feature vectors of 20 failures 30 days before the failure as the task , through the meta-learning rate =0.05 Update parameters , which reduces the prediction error of the model on new brand data; Adversarial training: The generator simulates the vibration data distribution of KUKA robots, and the discriminator distinguishes the data sources, forcing the feature extractor to ignore brand differences and focus on the common characteristics of bearing wear (such as the kurtosis value of vibration acceleration > 3.5).

[0031] Federated Learning Coordination Module Parameter aggregation: ABB robots account for 60% and KUKA accounts for 40%, so the aggregation weights are 0.6 and 0.4 respectively; Privacy Protection: Setting a Privacy Budget =5, calculate the gradient norm sensitivity =0.1, then the noise scale ≥ =50, add Laplace noise to the gradient data (50); Case library generation: After aggregation, cross-brand maintenance cases are generated, such as "When the average joint torque is greater than 80N·m and the peak vibration frequency is 120Hz, 85% of cases require bearing replacement within 7 days."

[0032] Hybrid prediction model Basic layer: Random forest contains 80 decision trees, and the Gini index is used to screen out torque-temperature coupling (correlation coefficient>0.7), vibration frequency-load correlation ( value <0.01) and other key parameter combinations; Prediction layer: LSTM is configured with two hidden layers, 256 neurons in each layer, a learning rate of 0.002, a dropout rate of 0.2, and outputs a health score (for example, a KUKA robot's health score is 58, triggering a yellow alert).

[0033] Dynamic decision engine unit Warning threshold: A health score of 60 points is a yellow warning, triggering the maintenance recommendation "Inspect joint 1 bearing within 3 days"; Case matching: The cosine similarity algorithm is used to match historical cases. For example, "An ABB robot once had similar vibration characteristics, and the score returned to 85 points after the bearing was replaced." This generates a work order containing the spare part model (SKF6205) and maintenance time (4 hours).

[0034] Example 2, please refer to the attached Figure 1 -Attached Figure 2 , Small sample fault prediction of new collaborative robots: Application Scenario An electronics factory introduced 50 new domestic collaborative robots (brand X, with no historical failure data) and needed early warning of reducer gear wear failures. However, there were only 10 actual failure samples. Key parameters and processes Digital twin sample generation module Dynamic modeling: Construct a dynamic model of the robot arm based on the Lagrange equation:

[0035] Among them, kinetic energy , potential energy , represents the total mass of the robot, represents the acceleration due to gravity, Indicates the height of the robot's center of gravity and the control torque Driven by real operating data; Fault simulation: Setting the disturbance torque Simulating gear wear: Amplitude: gradually increases from 0.5N·m (initial wear) to 5N·m (severe wear); Frequency: corresponds to the gear meshing frequency and can be set to 50Hz, 80Hz and 120Hz; Synthetic data: Generate 1,000 simulated fault samples, including normal state (500 cases), mild wear (300 cases), and severe wear (200 cases). Characteristics include torque fluctuation amplitude, temperature rise rate, etc.

[0036] Meta-learning and few-shot training Data augmentation: 10 real fault data sets were mixed with 1,000 synthetic data sets, and the training set / test set was divided into a ratio of 7:3; Metamodel adaptation: Using synthetic data to pre-train the metamodel, the initial parameters θ learn the characteristic patterns of gear wear (such as periodic spikes in the torque signal); Quick update: Input 10 real samples as tasks , meta-learning rate =0.1, after 5 iterations, the parameter The model achieves an F1 score of 0.85 on the test set.

[0037] Feature Engineering and Prediction Feature extraction: Focusing on the low-load characteristics of collaborative robots, we extract time domain features (RMS increment > 15%) and time-frequency domain features (wavelet energy entropy mutation). Model deployment: The trained LSTM model is deployed to the edge server to monitor the gearbox vibration data in real time. When the root mean square increment of three consecutive frames is detected to be greater than 20% and the energy entropy is greater than 0.7, a red alert is triggered: "Immediately stop and check the reducer."

[0038] Maintenance strategy optimization Sample reflux: After maintenance, the case of "the score was restored to 90 points after the gearbox was replaced" was added to the federated learning case library; Model update: The model is automatically retrained weekly, and the threshold is optimized based on newly collected normal operation data (approximately 5,000 cases). The false alarm rate has been reduced from 15% to 5%.

[0039] According to the above embodiment, through the technical combination of the meta-learning and field adaptation module cooperative mechanism, the federal learning coordination module privacy compliance architecture, and the synergistic effect of the digital twin sample generation module and the meta-learning rapid adaptation characteristics, the data heterogeneity and privacy barriers of multi-brand robots are effectively broken through, the new robot fault data blank is filled, and with the synergistic cooperation of various algorithms and models in each module, the good effect of compatible multi-brand robots, reducing the sample demand and prediction error of new brand model adaptation, and improving the landing efficiency of predictive maintenance in multi-brand mixed production lines and emerging industrial robot scenarios can be achieved.

[0040] Although the present application is disclosed in the preferred embodiments, it is not intended to limit the present application, and any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, any modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, which does not deviate from the technical solutions of the present application, falls within the protection scope defined by the claims of the present application.

Claims

1. A machine learning-based predictive maintenance system for industrial robots, characterized in that: The maintenance system includes: Feature Engineering Module: Extracts time domain features, frequency domain features, and time-frequency domain features from clean data to generate device health feature vectors; Hybrid prediction model unit: A hybrid prediction model is set up, which consists of a base layer and a prediction layer. The base layer uses the random forest algorithm to identify multivariate correlations. The prediction layer uses the long-short-term memory network to learn the long-term dependency of time series data and outputs the health score and remaining life expectancy prediction value; Meta-learning and domain adaptation module: This module uses a model-agnostic meta-learning algorithm to train a cross-brand universal fault feature extractor and an adversarial domain adaptation network to confuse the data distribution of different brands and generate a unified cross-domain feature vector. Federated Learning Coordination Module: Receives parameter updates from sub-models trained locally by robots from various brands and uses a weighted aggregation algorithm protected by differential privacy to generate a global generalization model and a cross-brand maintenance case library. Digital twin sample generation module: This module builds a digital twin based on the robot's dynamic equations, simulates failure modes, and generates synthetic failure samples with physical meaning. This sample is then input into the feature engineering module to enhance small sample training. Dynamic decision engine unit: Based on the health score, remaining life prediction value and the cross-brand maintenance case library generated by federated learning, the rule engine is used to set multi-level warning thresholds. A three-level warning is set based on the health score threshold, including triggering a red warning when the health score is ≤ 60 points and generating a shutdown and maintenance recommendation. A spare parts list is generated based on the cross-brand maintenance case library by matching historical fault characteristics, generating maintenance recommendations and pushing them to the visual interactive interface.

2. The machine learning-based predictive maintenance system for industrial robots 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 joint torque, motor temperature and vibration frequency time series data in real time from multi-brand robot controllers including ABB and KUKA based on Modbus and OPCUA industrial communication protocols. The data acquisition module supports adjustable sampling frequency from seconds to minutes, and is compatible with the proprietary communication protocols of ABB Robot Ware and KUKAKRL manufacturers based on a protocol conversion gateway. The collected time series data is added with device ID, acquisition timestamp and sensor type label, and is structured and stored in the Influx DB industrial time series database to form an original data set with time and space identification.

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

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

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

6. The machine learning-based predictive maintenance system for industrial robots according to claim 5, characterized in that: The model-agnostic meta-learning algorithm of the meta-learning and domain adaptation module adopts the following parameter update rules: ； in, To obtain the initial parameters of the meta-model with cross-domain general feature extraction capabilities based on pre-training of multi-brand historical data, To control the meta-learning rate of adapting to new brand data, For new brand small sample tasks containing 20-50 samples, is the fault classification loss function based on cross entropy, To adapt to the Model parameters after branding, Based on metamodel parameters The feature extraction function, is the meta-model parameter Gradient calculation, loss function is calculated based on the back propagation algorithm right The gradient vector guides the parameter update direction. The adversarial domain adaptation network simulates the target brand data distribution based on the generator and the discriminator distinguishes the data source, so that the feature extractor learns the common fault features that are independent of the brand.

7. The machine learning-based predictive maintenance system for industrial robots according to claim 6, characterized in that: The federated learning coordination module dynamically adjusts according to the number of devices of each brand and adds Laplace noise to the uploaded gradient data. , and the noise scale satisfy ,in, is the privacy budget, ranging from 1 to 10. is the gradient norm sensitivity.

8. The machine learning-based predictive maintenance system for industrial robots according to claim 7, characterized in that: The digital twin sample generation module constructs a robot dynamics model based on the Lagrange equation, as follows: ； in, is the Lagrangian function, is the joint angle vector, is the joint angular velocity vector, To control the torque, To simulate the disturbance torque generated by the fault, based on the adjustment The amplitude and frequency of the φ are used to generate synthetic data containing bearing wear and motor eccentricity failure modes.

9. The machine learning-based predictive maintenance system for industrial robots according to claim 8, characterized in that: The rule engine of the dynamic decision engine unit establishes a three-level early warning mechanism. A health score of 70-80 is blue, prompting attention, 60-70 is yellow, and it is checked within 3 days. A score below 60 is red, and it is immediately shut down for maintenance. The case library matches historical fault cases based on the cosine similarity algorithm, and the matching conditions include the Euclidean distance of the feature vector being less than 0.2 and the trend change slope being consistent, to generate executable maintenance suggestions including spare part model, maintenance time and historical cost.

Images